The Elevate Blog banner: Real Classrooms. Practical Ideas. Higher Expectations. Three students climb a rocky ridge at sunrise, each pulling the next one up, toward a teacher reaching down to help.

Author: Clay Shumate

  • Cognitive Load Theory: What It Claims, and What the Evidence Supports in Grades 6–12

    Cognitive Load Theory: What It Claims, and What the Evidence Supports in Grades 6–12

    By Clay Shumate

    Cognitive load theory says working memory handles only a few new elements at once, so instruction should strip out everything that is not the learning itself. Its named instructional effects have real evidence behind them — worked examples pool at g = 0.43 — but most of that evidence comes from adults and undergraduates rather than teenagers.

    Key Takeaways

    • The capacity figure teachers are usually given is out of date. Cowan (2001) puts the central working-memory limit at about four chunks, with individual variation from roughly two to six, and says Miller’s famous seven was “meant more as a rough estimate and a rhetorical device than as a real capacity limit.”
    • The worked-example effect is medium, not huge. A meta-analysis of 49 studies and 185 effect sizes found g = 0.43 (95 percent confidence interval 0.31 to 0.55). Forty-one of those 185 effects were negative.
    • Where the age breakdown exists, the effect is not flat across grades. In a split-attention meta-analysis, grades 9–12 came in at g = 0.81 and grades 6–8 at g = 0.34, with K–5 not significant.
    • Cognitive load is mostly measured by asking people how hard it felt. The standard instrument is a single self-report rating scale, and the theory’s own authors say it cannot easily separate one kind of load from another.
    • Classroom transfer is the open question, not the mechanism. The Education Endowment Foundation, which funds large school trials, says positive results “often come from small or tightly controlled studies” and that we do not yet know how well these approaches work in everyday classroom conditions.

    What is cognitive load theory?

    Cognitive load theory is an instructional theory built on one claim: working memory processes only a small number of new elements at a time, while long-term memory has no known practical limit. Teaching works best when it respects that bottleneck.

    John Sweller developed the theory in the 1980s out of research on problem solving. The instructional advice that falls out of it is mostly subtraction — take out whatever is competing for the bottleneck and is not the thing you want learned.

    The original argument was about how students solve unfamiliar problems. Conventional problem solving pushes students into means-ends analysis, working backward from the goal, and the authors’ own retrospective calls that process “exceptionally expensive of working memory capacity”. A student can grind through a problem that way, get the right answer, and learn almost nothing about the pattern, because every spare slot went into the search rather than into building the mental model.

    That is the piece worth keeping even if you never read another word about the theory. Finishing the task and learning the material are two different outcomes, and a hard enough task can produce the first without the second.

    How much can working memory actually hold?

    About four chunks, not seven. Cowan’s 2001 review argues for “a single, central capacity limit averaging about four chunks,” with a range of “two to six chunks in individuals.” Miller’s 1956 figure of seven, the one that made it onto every in-service slide, was never presented as a hard limit. Working inside that limit is also why pairing a picture with spoken words rather than stacking text on text tends to help.

    The number matters less than what a chunk is. A chunk is whatever the student already holds as one unit, which means the same slide is a different load for different students in the same room. For a student who already knows what a primary source is, a document analysis prompt is one chunk. For a student who does not, it is four: what the document is, who made it, why that matters, and what the question is asking. The theory’s own 2019 restatement makes this explicit — complexity depends on the material and on prior knowledge together, not on the material alone. That has an equity edge worth saying out loud: the student with the least background knowledge pays the highest price for a badly laid-out handout, because they have the fewest slots to spare in the first place.

    The New South Wales education department’s widely circulated teacher summary uses the figure of about four chunks as well. If your school’s professional development is still handing out seven plus or minus two, it is quoting a seventy-year-old rhetorical device.

    Three stacked panels defining intrinsic, extraneous and germane cognitive load, with a note that germane load was largely abandoned as a separate category after 2010.

    What are the three types of cognitive load?

    Cognitive load theory splits load into three. Intrinsic load comes from how many elements the material forces a student to juggle at once. Extraneous load is added by how the material is presented. Germane load is the effort that builds the mental model.

    Extraneous load is the one a teacher fully controls, which is why almost every practical recommendation in this article is about it.

    The third category has had a rough time. Critics argued it was doing too much explanatory work, and by the early 2010s the theory’s own camp had largely stopped treating germane load as a separate source of load at all — a review of the debate records that Sweller (2010) and Kalyuga (2011) appear to have given up on the notion. When you see a planning template with three boxes to fill in, one of them is a category the field quietly retired.

    Intrinsic load has its own problem, which is that it is supposed to be fixed by the material and plainly is not. de Jong (2010) points out that instructional approaches can change intrinsic load, which makes the clean separation between the three types harder to maintain than the diagrams suggest. In practice, two of the three boxes are contested and the useful one is extraneous load.

    Is cognitive load theory evidence based?

    Parts of it are well evidenced and the whole of it is not. The named instructional effects — worked examples, split attention, redundancy — have decades of experiments behind them and respectable pooled effect sizes.

    The underlying construct, cognitive load itself, is measured almost entirely by self-report, and that is the part critics go after.

    The worked-example effect is the headline. A meta-analysis of 49 studies and 185 effect sizes in mathematics found an average of g = 0.43, confidence interval 0.31 to 0.55. That is a medium benefit, and it is not uniform: effects ranged from −1.80 to 4.66, 41 of the 185 were negative, and heterogeneity was extremely high. Awkwardly for the theory, no moderator reached significance — including prior knowledge, which is the mechanism cognitive load theory invokes to explain the effect. The authors’ own plain-language summary of the published version says the benefit did not consistently depend on what learners already knew.

    Split attention is the effect with the age data. A meta-analysis of 58 comparisons and 2,426 learners, read here through the Technical University of Munich’s review summary rather than the original article, reports g = 0.63 overall — and a grade breakdown worth more to a secondary teacher than the headline: g = 0.81 for grades 9–12, g = 0.34 for grades 6–8, and a non-significant g = 0.24 for K–5.

    Table of four cognitive load theory effects - worked example, split attention, redundancy and expertise reversal - listing the best verified effect size and the population tested for each.

    The redundancy effect is the one to be careful quoting. The experiment usually cited found that audio alone beat audio plus matching on-screen text, and it ran on 34 first-year trade apprentices in Sydney. The finding is real. The sample is not a high school class.

    Does cognitive load theory hold up in middle and high school?

    Nobody has shown that it does not, and almost nobody has tested it on teenagers. The literature is dominated by undergraduates and adult trainees, which is a limitation the theory’s critics name directly rather than something this site is inventing.

    de Jong’s review puts it bluntly: “it is common for the research subjects to be psychology students”, often studying material they have no interest in for a very short time, and results from those conditions “raise problems when these results are translated into practical recommendations.” Look at the primary studies and he is right. The expertise reversal papers ran on electrical trainees, machinery trainees, mechanical apprentices and adults. The redundancy experiment ran on apprentices. The worked-example meta-analysis does not report grade levels at all.

    The clearest secondary-aged classroom test I could find is Bentley and Yates (2017), two small experiments with Year 7 students in Adelaide learning proportional reasoning. Worked examples beat conventional problems in both: 44 students, g = 0.90 at post-test and 0.68 delayed in the first; 38 students, g = 0.87 and 0.77 in the second. Encouraging, and two studies at one grade level in one country is not an evidence base.

    So the honest statement for a grades 6–12 teacher is narrower than the slide deck: the mechanism is plausible, the classroom effects that have been measured on students your age are few, and the one meta-analysis that breaks results out by grade says middle school gets about half of what high school gets.

    How is cognitive load measured?

    Usually by asking learners, after a task, how much mental effort it took, on a single rating scale introduced by Paas in 1992. That is the field’s standard instrument, and it is the weakest joint in the whole structure.

    This is not an outside attack. Sweller, van Merriënboer and Paas say in their own retrospective that the scale gives an overall measure and cannot easily distinguish intrinsic from extraneous load. Ouwehand and colleagues (2021), with Paas as a co-author, go further: the reliability and validity of subjective measures are “debatable,” the scales have been used “in an inconsistent way,” and changing the item’s wording changes which kind of load you capture.

    de Jong’s version is harder still. The values from the questionnaire “have no agreed upon meaning,” the critical level that would indicate overload is unknown, and the measure therefore cannot detect the thing the theory is named after. A teacher reading a research summary that says an approach “reduced cognitive load” is usually reading that a group of adults ticked a slightly lower number on a nine-point scale.

    What are the main criticisms of cognitive load theory?

    Three, and they are worth knowing before you defend the theory in a faculty meeting: it is hard to falsify, its measures are contested, and its classroom transfer is unproven.

    • Falsifiability. de Jong argues the theory “is constructed in such a way that it is hard or even impossible to falsify”, because every outcome fits after the fact — if learners do better, the higher load is reclassified as germane. The partial retirement of germane load by the theory’s own authors is a real concession to that charge.
    • Measurement. Covered above, and conceded in part by the theory’s founders. Separate measurement of the three load types is widely held not to be feasible.
    • Transfer to real classrooms. The Education Endowment Foundation’s own research agenda says “there are limitations to the evidence base,” that positive results “often come from small or tightly controlled studies,” and that “we don’t yet know how well these approaches work in everyday classroom conditions.” Its completed working-memory and arithmetic trial found no additional impact.

    What none of that supports is the claim that cognitive load theory has been debunked. Contested measurement, narrow samples and unproven classroom transfer are ordinary problems for an applied theory, and the subtraction advice it produces is cheap and mostly sensible. It is the confident teacher-facing summaries that oversell it. The New South Wales guide calls the evidence base “robust” and lists no limitations at all.

    Why do worked examples stop working?

    Because the support that replaces a novice’s missing knowledge gets in the way once the student has that knowledge. Cognitive load theory calls this the expertise reversal effect: an instructional technique that is “highly effective with inexperienced learners” can lose its effectiveness, and eventually reverse, as expertise grows.

    This is the most practically useful thing the theory says and the part teachers skip. It means a scaffold has a shelf life. The worked example that carried a class in September is, by November, a page of text a student has to reconcile with what they already know — which is itself extra load. In one study of mechanical trade apprentices, novices learned more from worked examples while more experienced learners did better solving problems outright.

    Two honest caveats. The 2003 review that named the effect reports no effect sizes and no sample sizes — it is a qualitative review plus mental-effort ratings, so do not go quoting a number for it. And the effects shrink before they flip, which means there is no clean day to stop. That is the same call as deciding when to pull a temporary support, made the same way: by watching what students can do without it.

    Doesn’t this contradict desirable difficulty?

    It looks like it does and it mostly does not. Cognitive load theory says to cut load; research on mixing problem types on purpose says to add difficulty. The difference is which load you are adding.

    Interleaving adds difficulty on purpose to the decision a student has to make about which method applies, and that decision is the learning. A cluttered worksheet adds difficulty to finding the question.

    This site has published both arguments, so it is worth stating the boundary plainly. Cutting extraneous load means cutting the clip art, the duplicated on-screen text and the diagram whose labels live on another page. It does not mean cutting the thinking. The same line separates genuinely letting a student wrestle with a problem from leaving them stuck because the instructions were unreadable — one of those is a design choice and the other is a mistake.

    Numbered list of five ways to cut extraneous cognitive load: integrate labels into diagrams, say it or show it rather than both, work one example fully, fade support deliberately, and cut decoration.

    How do you reduce cognitive load in a grades 6–12 lesson?

    Start with presentation, not content. Extraneous load is the only one of the three you control outright, and most of the gain is in five unglamorous moves.

    None of them needs a purchase, a platform or a day of release time, which is more than can be said for most things that arrive with a research citation attached.

    1. Put the words on the picture. If a diagram needs a key, a caption or a set of labels, integrate them into the image rather than setting them underneath it. This is the split-attention effect, and it has the strongest high school numbers behind it. You are not redrawing a textbook diagram to do this — projecting it and writing the labels onto it, or handing out a copy with the key already on the image, is the same move.
    2. Say it or show it, not both. Reading your slide aloud while students read the same words is the redundancy effect in action. Pick one.
    3. Work the first example all the way through. Then hand out one that is half finished, then a bare problem. Three stages inside one period, with the examples on the same page as the problems.
    4. Fade the support deliberately, and say that you are doing it. Decide in advance what students must be able to do before the worked example comes off, check it, and then tell them the example is going away and why. From a desk, support that quietly disappears reads as a teacher who has stopped helping, and the first attempt without it is supposed to feel worse than the one before.
    5. Name the chunk before you use it. If a task assumes students hold “thesis statement” or “primary source” as one unit, confirm that before the task, not during it. Everything you can make familiar in advance is a slot you get back.

    Notice what is not on that list: shorter texts, easier questions, fewer sources. Reducing extraneous load is a presentation decision, and nothing in cognitive load theory licenses lowering how demanding the thinking is.

    Where does cognitive load theory go wrong in schools?

    It gets used as an argument for doing less of everything. A theory about presentation becomes a mandate to cut group work, discussion, projects and open questions, usually on the grounds that those things raise cognitive load. They do. So does reading a difficult book, and nobody suggests stopping that.

    The second failure is treating load as a property of the task rather than of the student. The same activity is low load for half the room and high load for the other half. The answer to that is not a simpler task; it is knowing who is carrying which chunks, which looks more like checking what students actually do when they study than like redesigning a slide.

    The third is the one the measurement problem creates. Because load is read off a rating scale, any lesson students call easy can be reported as a lesson that managed load well. Easy and well designed are not the same thing, and a class that reports low effort may simply have not done much.

    Clay’s take: it is a theory about presentation, not about ambition

    I find the subtraction advice useful and I would not hand a colleague the slide deck that usually comes with it.

    Integrating labels into a diagram, cutting duplicated text, working one example all the way down — those cost nothing, they are easy to check, and the high school numbers on split attention are better than most of what we get told to do.

    What I will not do is let it become a reason to ask teenagers for less. It was built on adults learning unfamiliar procedures under lab conditions, and its own authors concede they cannot separate the three types of load you are being asked to plan around. Treat it as presentation rules with decent evidence and it earns its place. Treat it as a complete theory of learning and you end up with a quiet room where nothing hard gets asked.

    What to do next week

    Take one lesson you already teach and change only the presentation. Move the labels onto the diagram. Delete the slide text you are about to say out loud anyway.

    Put the worked example on the same page as the problems rather than on the board behind you. Then teach it and listen to whether the questions are about the content instead of about where something is. That is a twenty-minute edit, and it is the part of this theory with evidence behind it.

    Frequently Asked Questions

    What is cognitive load theory in simple terms?

    It is the idea that working memory can only handle a few new pieces of information at once, so instruction should remove anything competing for that space. The theory splits load into intrinsic load from the material, extraneous load from how it is presented, and germane load spent building understanding. Extraneous load is the part a teacher controls.

    How many things can working memory hold?

    About four chunks, according to Cowan’s 2001 review, with individual variation from roughly two to six. Miller’s better-known seven plus or minus two was described by Cowan as a rough estimate and a rhetorical device rather than a real capacity limit. A chunk is whatever a particular student already holds as a single unit, so the same material is a different load for different students.

    Is cognitive load theory evidence based?

    Its instructional effects are better evidenced than its central construct. Worked examples pool at g = 0.43 across 49 studies, and split attention at g = 0.63 across 58 comparisons. Cognitive load itself is usually measured by a single self-report rating scale that the theory’s own authors say cannot cleanly separate load types.

    What are the criticisms of cognitive load theory?

    The three serious ones are that the theory is hard to falsify because results can be reclassified after the fact, that its measures are subjective and contested, and that classroom transfer is unproven. The Education Endowment Foundation notes that positive results often come from small or tightly controlled studies rather than everyday classrooms.

    Does cognitive load theory apply to high school students?

    The mechanism is assumed to, but the evidence base is mostly adults and undergraduates. The one meta-analysis that reports a grade breakdown found split attention at g = 0.81 in grades 9 to 12 and g = 0.34 in grades 6 to 8. A pair of Australian Year 7 experiments found worked examples beat conventional problems with effect sizes between 0.68 and 0.90.

    What is the expertise reversal effect?

    It is the finding that instructional support which helps inexperienced learners can lose its benefit, and eventually hurt, as learners gain expertise. A worked example that carries a novice becomes extra material an experienced student has to reconcile with what they already know. The review that named the effect publishes no effect size for it.

    Does reducing cognitive load mean making work easier?

    No. It means removing load that comes from presentation, not load that comes from thinking. Integrating labels into a diagram, cutting duplicated on-screen text and placing worked examples beside the problems all reduce extraneous load. Shortening the text, simplifying the question or lowering the demand reduces something else entirely.

    How do teachers measure cognitive load in the classroom?

    Realistically, they do not. Research measures it with a post-task self-report scale whose values, as de Jong notes, have no agreed meaning and no known threshold for overload. In a classroom the usable proxy is what students ask about: questions about where to find something or what the directions mean are extraneous load you can design away.

    Sources

    • Sweller, J., van Merriënboer, J. J. G., & Paas, F. (2019). Cognitive Architecture and Instructional Design: 20 Years Later. Educational Psychology Review, 31(2), 261–292. Open-access copy. Source of the theory’s own restatement of means-ends analysis, element interactivity and prior knowledge, and of the authors’ concession about the Paas rating scale.
    • Cowan, N. (2001). The magical number 4 in short-term memory: A reconsideration of mental storage capacity. Behavioral and Brain Sciences, 24. Publisher PDF. Source of the four-chunk figure, the two-to-six range, and the statement about Miller’s seven.
    • Clerjuste, S., Chawla, K., Barbieri, C. A., & Miller-Cotto, D. (2022). A Meta-Analysis of the Worked Examples Effect on Mathematics Performance. University of Delaware. Conference poster. Source of g = 0.43, the confidence interval, the 49 studies and 185 effect sizes, and the 41 negative effects. The poster does not report grade levels.
    • Adler, R. (2024). Worked examples. The Learning Scientists. Guest post. Plain-language summary of the published version of the meta-analysis above; source of the statement that the benefit did not consistently depend on prior knowledge. Used as a summary because the published article could not be opened.
    • Schroeder, N. L., & Cenkci, A. T. (2018). Spatial Contiguity and Spatial Split-Attention Effects in Multimedia Learning Environments. Educational Psychology Review, 30(3). Read through the Clearing House Unterricht review at the Technical University of Munich, not the original article. Source of g = 0.63 and the grade-band breakdown.
    • Kalyuga, S., Ayres, P., Chandler, P., & Sweller, J. (2003). The Expertise Reversal Effect. Educational Psychologist, 38(1), 23–31. Full text. Source of the definition and of the apprentice and trainee samples. It reports no effect sizes, which is why none is quoted here.
    • Kalyuga, S., Chandler, P., & Sweller, J. (1999). Managing Split-attention and Redundancy in Multimedia Instruction. Applied Cognitive Psychology, 13, 351–371. Open-access copy. Source of the redundancy result and of the sample of 34 first-year trade apprentices.
    • de Jong, T. (2010). Cognitive load theory, educational research, and instructional design: some food for thought. Instructional Science, 38, 105–134. University of Twente copy. Source of the falsifiability argument, the critique of the rating scale, and the statement about psychology-student samples.
    • Ouwehand, K., van der Kroef, A., Wong, J., & Paas, F. (2021). Measuring Cognitive Load: Are There More Valid Alternatives to Likert Rating Scales? Frontiers in Education, 6:702616. Full text. Source of the reliability and validity criticism, from a paper co-authored by the scale’s originator.
    • Bentley, B., & Yates, G. C. R. (2017). Facilitating proportional reasoning through worked examples: Two classroom-based experiments. Cogent Education, 4. University of Adelaide copy. Source of the Year 7 effect sizes and sample sizes.
    • Centre for Education Statistics and Evaluation, NSW Department of Education (2017). Cognitive load theory: Research that teachers really need to understand. Report. Cited as practitioner guidance, not as evidence appraisal; it uses the four-chunk figure and lists no limitations.
    • Education Endowment Foundation. Research agenda theme: cognitive science. Programme page. Source of the statement that positive results often come from small or tightly controlled studies and that classroom transfer is not established.
    • Claes, J. Yet another reconsideration of cognitive load theory. Ghent University. Conference paper. Source for the record that Sweller (2010) and Kalyuga (2011) stepped back from germane load as a separate source, and for Schnotz and Kürschner’s position on separate measurement.

    About Clay Shumate

    Clay Shumate is a certified secondary Social Studies teacher in the public schools of West Alabama, with seven years of classroom experience, a B.A. in History, and an M.Ed. in Secondary Education. He writes about project-based learning, student responsibility, respect, and practical ways to hold young people to a higher standard while giving them room to learn from mistakes. He is a member of the Society of Professional Journalists and writes to its Code of Ethics; this site’s editorial standards and corrections policy are published in full. More about Clay.

  • The Flipped Classroom in Grades 6–12: What the Evidence Shows, and How to Flip One Lesson

    The Flipped Classroom in Grades 6–12: What the Evidence Shows, and How to Flip One Lesson

    By Clay Shumate

    A flipped classroom moves direct instruction out of class time, usually to a short video, so class is spent on practice the teacher can watch. Meta-analyses find a small positive effect on achievement, around 0.35 of a standard deviation, but roughly nine out of ten of those studies were run on college students rather than teenagers.

    Key Takeaways

    • The effect is real and small. van Alten, Phielix, Janssen and Kester (2019) pooled 114 studies and found g = 0.36 on assessed learning, with a 95 percent confidence interval of 0.28 to 0.44. Student satisfaction did not move at all.
    • The evidence is about undergraduates. In Låg and Sæle (2019), 239 of 272 learning comparisons came from higher education and 20 from secondary schools. Kapur, Hattie, Grossman and Sinha (2022) reports that 90 percent of the students in its 173 studies were university-based.
    • Two moderators are worth more than the headline number. Flipped classes that cut face-to-face time did worse. Flipped classes that added a quiz on the preparation did better. Both come from the same meta-analysis.
    • Do not cite one number. A review of 46 flipped-learning meta-analyses found overall effects ranging from 0.19 to 2.29. Whichever figure is on the in-service slide was chosen from a very wide menu.
    • The homework-access problem is smaller than it was and still real. Federal data put 5 percent of students in homes with no internet, 7 percent in rural areas, and 12 percent among students below the poverty threshold against 2 percent in high-income families.

    What is a flipped classroom?

    A flipped classroom is an arrangement where first exposure to new material happens before class, usually through a short video or reading, so class time can be used for the work students would otherwise have done alone. The lecture becomes the homework and the homework becomes the lesson.

    The formal definition comes from The Flipped Learning Network, which published its four-pillar version in March 2014: flipped learning is an approach “in which direct instruction moves from the group learning space to the individual learning space.” The same organization draws a line that most people skip. “These terms are not interchangeable,” it says. “Flipping a class can, but does not necessarily, lead to Flipped Learning.”

    That distinction is the whole argument in one sentence. Recording your lecture and assigning it is a change in scheduling. Whether anything improves depends entirely on what you do with the fifty minutes you just freed up, and the research below says most teachers do not change that part at all. One concrete answer to that question is splitting the period into stations students move through.

    Diagram contrasting a traditional lesson, where class time is spent on the lecture and practice happens alone at home, with a flipped lesson, where first exposure happens at home and class time is spent on practice the teacher can watch and correct.

    Who invented the flipped classroom?

    Nobody cleanly. The term is usually credited to Jonathan Bergmann and Aaron Sams, two high school chemistry teachers who began recording lectures around 2007, but the practice predates the name by at least a decade.

    a 2020 review of the research traces an earlier line: Lage, Platt and Treglia (2000), three economics professors, described an “inverted classroom” in a peer-reviewed journal in 2000, and Eric Mazur’s peer instruction at Harvard in the 1990s is described as “one of the earliest instances of flipped learning.” That same review calls Bergmann and Sams only “believed to be first,” which is the honest level of confidence. Both of those earlier efforts were with university students.

    It matters less than it sounds, but it explains something. A practice named by classroom teachers and evidenced by university professors is going to have a research base that does not match the people using it, and that is exactly what happened.

    Does the flipped classroom actually work?

    Modestly, on average, in the settings where it has been tested most. The best-powered estimates land around a fifth to a third of a standard deviation on achievement, and shrink further when you correct for the studies that never got published.

    Table of four flipped classroom meta-analyses showing the effect size each reported, how much of its evidence came from higher education rather than schools, and what happened to the number after correcting for publication bias.

    van Alten, Phielix, Janssen and Kester (2019) is the most quoted. Across 114 studies it found g = 0.36 on assessed learning outcomes, 95 percent confidence interval 0.28 to 0.44. Two things in it get left off the slide. Perceived learning was not significant, and student satisfaction came in at g = 0.05 with a confidence interval spanning zero — students did not enjoy it more. The authors call flipping “a promising pedagogical approach when appropriately designed,” which is a careful sentence doing careful work.

    Låg and Sæle (2019) is the one to read if you want the number after the air comes out. The headline is g = 0.35 across 272 comparisons. Restrict it to adequately powered studies and it falls to 0.24. Adjust for publication bias and the estimate lands at g = 0.17. Satisfaction ran the same way: a small positive 0.16 that became −0.03 after the same correction. The authors are careful to note that the correction gives “approximate indications,” not a precise number, but the direction is unambiguous.

    Two findings in van Alten, Phielix, Janssen and Kester (2019) are worth more to a working teacher than any of those effect sizes. Flipped classes where face-to-face time was reduced performed significantly worse, at −0.26. Flipped classes that added quizzes on the pre-class material performed better, at +0.19. In plain terms: if the plan is to flip so the class can meet less often, the evidence says do not. If you flip, check the preparation.

    How much of the flipped classroom research is about teenagers?

    Very little, and this is the single most important thing to know before quoting any of it in a secondary school. Three independent syntheses put the higher-education share at roughly 85 to 90 percent.

    The numbers are specific. In Låg and Sæle (2019), the learning outcomes break down as 239 higher education, 20 secondary, 12 primary and one further education — and the pass-rate analysis, with 45 comparisons, contains zero school studies at all. In van Alten, Phielix, Janssen and Kester (2019), 11 of 114 studies were secondary. Kapur, Hattie, Grossman and Sinha (2022) states that “the majority of students were university-based (90%),” while reporting its effect sizes as 3 percent primary, 42 percent secondary and 55 percent higher education. Those two figures count different things, students versus effect sizes, and the paper gives both.

    There is one meta-analysis built only on schools. Zhu (2021) pooled 53 effect sizes from 27 studies of K-12 students and reported a weighted average of d = .54 — larger than any of the higher-education estimates. Before anyone puts that on a slide, the author’s own analysis flags publication bias by funnel plot, trim-and-fill and Egger’s test, and heterogeneity sits at I² = 81.2 percent. Twenty-seven studies with an acknowledged bias problem is better than nothing and is not a settled answer.

    One scale distinction belongs with those numbers. Nearly all of this research tests one teacher flipping one course, and almost none of it tests a school flipping everything. A building-wide flip is a different proposition: it multiplies the home-preparation load across every subject a student takes on the same evening, and no student’s evening was studied that way. A district deciding this at scale is making a decision the evidence base does not cover.

    Lo and Hew’s review of K-12 flipped classrooms is the other school-specific piece, covering 15 published K-12 studies. Achievement results were neutral or positive against traditional classrooms, and student attitudes were mixed. That is a fair summary of where this sits for grades 6 through 12: probably not harmful, possibly helpful, not proven at the size the enthusiasm implies.

    Why do the effect sizes disagree so much?

    Because “flipped classroom” describes a schedule rather than a method, and because the syntheses pooling those studies have real methodological problems. The range across the literature is roughly twelvefold.

    Kapur, Hattie, Grossman and Sinha (2022) reviewed 46 flipped-learning meta-analyses and reports: “We found the overall effects from 0.19 to 2.29.” Their own meta-analysis of 173 studies landed at g = 0.37, confidence interval 0.32 to 0.41, with I² = 91.27 percent and a statistically significant test for publication bias. An I² above 90 means the studies are not estimating one common effect, and pooling them produces an average of things that are not alike.

    Hew, Bai, Dawson and Lo (2021) examined 19 of those meta-analyses, covering more than 495 primary studies, and found six elements “inadequately conducted or reported” across them: searching grey literature, assessing risk of bias, handling non-independent effect sizes, controlling for students’ initial differences, controlling for instructor equivalence, and assessing publication bias. Their conclusion is that positive evidence exists but there is “reason to question how this evidence has been synthesized.” Controlling for prior achievement and for who taught the class are not fine points. They are the two most obvious ways a flipped section could look better without the flipping doing anything.

    There is also a content problem underneath the statistics. Kapur, Hattie, Grossman and Sinha (2022) observes that “most in-class sessions are not modified based on the flipped implementation.” If the video moves home and the class period stays the same, the study is measuring a scheduling change. That is roughly what one chemistry teacher told Education Week asked teachers who had tried it more than a decade ago, calling flipping “a time-shifting tool that is grounded in the same didactic, lecture-based philosophy.”

    What happens when students don’t watch the video?

    You teach the lesson twice, badly, to a split room. This is the failure mode every teacher who has tried it names first, and the evidence points at one specific fix: make the preparation produce something you collect.

    The quiz moderator in van Alten, Phielix, Janssen and Kester (2019) is that fix with a number attached: +0.19 for flipped classes that added quizzes on the pre-class material, across 36 of 114 studies. It is the difference between assigning a video and assigning a task. Three or four questions that take two minutes and that you can see before the bell is enough, which is also the difference between a check and a grade in practice.

    Checklist of six findings from the flipped classroom research: check the preparation with a quiz, keep the whole class period, keep videos under six minutes, plan for students without home internet, do not expect higher student satisfaction, and do not leave the class period unchanged.

    The honest reports from practitioners are blunter. In Education Week asked teachers who had tried it, an instructional coach described students who “may have been in class, but they weren’t engaged,” and said that for some, “self-paced became no pace,” so teachers had to put deadlines back. Another teacher put it plainly: “Just because you flipped your classroom doesn’t mean your students will watch the videos.”

    Tell families what you are doing before they work it out from their kitchen table. From outside, “my child watches videos for homework and does worksheets in class” sounds exactly like a teacher who stopped teaching. Two sentences in a syllabus or an email — the explanation moved home so the practice could happen where I can help — prevents a conversation that is otherwise very hard to have in March.

    Two practical consequences follow. Plan the first ten minutes so the room works whether or not everyone prepared — a short recap station, a partner summary, something like earning the first five minutes of the period — rather than opening with a task that assumes compliance. And decide in advance what happens to the student who did not watch it, because that is a policy question, not a mood. If the answer is a zero, you are back to a late work policy you can actually defend and all its problems.

    What about students without reliable internet at home?

    It is a smaller problem than it was and it is not solved, and the answer depends on which students you teach. Federal data put about 5 percent of students in homes with no internet access at all, rising to 12 percent for students below the poverty threshold.

    the federal figures on rural students and the internet give the most useful breakdown for school-age students: roughly 2.4 million students, about 5 percent, lived in homes without internet access. Seven percent of students in rural areas and 11 percent in remote rural areas were without it. Fixed broadband at home reached 82 percent of students overall, 76 percent in rural areas. The poverty gap is the number that should decide your planning: 12 percent of students below the poverty threshold had no home internet against 2 percent of students in high-income families, and among rural students fixed broadband ran 57 percent for low-income families against 82 percent for high-income ones.

    A caution about any single figure here. the federal data on children’s home internet access reports 97 percent of 3- to 18-year-olds with home internet in 2021 and 3 percent without — but notes on the same page that a different federal survey put the no-access figure at 12 percent, because the two surveys define internet access differently. Report both or neither. And note that 4 percent had access only through a smartphone, which is access in the survey and not access for watching a ten-minute video and taking notes, something worth holding alongside what the research says about phones in class.

    a Pew Research Center survey of teenagers found that 17 percent of teens were at least sometimes unable to finish homework because they lacked reliable computer or internet access, and 35 percent sometimes did homework on a cellphone, rising to 45 percent in households under $30,000. Those figures are from 2018 and pre-date the pandemic push on connectivity, so treat them as the shape of the problem rather than its current size. The practical move is cheap either way: ask the class, on paper, how they would watch something tonight, and build the fallback before you need it.

    How long should a flipped classroom video be?

    Short. Under six minutes if you can, and never longer than about nine. The most-cited evidence for this comes from online adult learners rather than teenagers, which is worth knowing before you treat it as a law.

    Guo, Kim and Rubin (2014) analyzed 862 videos and nearly 6.9 million viewing sessions and found that “video length was by far the most significant indicator of engagement.” Median engagement time was “at most 6 minutes, regardless of total video length,” and students “often make it less than halfway through videos longer than 9 minutes.” The rate at which viewers went on to attempt the associated problems fell steadily with length: 56 percent for the shortest videos, then 48, 43, 41 and 31 percent for the longest. The recommendation is explicit: segment into chunks “ideally less than 6 minutes.”

    Prep time is the honest cost, and the one figure available is a 2013 survey of 309 college faculty, in which 75 percent said preparing a flipped class takes more time than a traditional one. That sample was self-selected, from 2013, and made up of college instructors rather than school teachers, so treat it as a direction rather than a measurement. The thing that makes the cost bearable is the one nobody mentions: a six-minute explanation of how to balance an equation does not expire. You pay the recording cost once and spend it every year, which is why flipping your most-repeated lesson is better economics than flipping your most interesting one.

    The population caveat is not a technicality. Those were edX learners — self-selected adults with no attendance obligation and nothing at stake if they close the tab — and the measure was voluntary watch time, not learning. Your students have different incentives in both directions. Take the six-minute figure as a sensible default, not a finding about teenagers.

    Two other findings from the same study survive translation. Videos made for the purpose beat chopped-up recordings of live lectures, 55 percent against 41 percent on problem attempts, so recording your actual class period is the weakest version of this. And informal, close-up delivery beat studio production, which is good news for anyone recording at a kitchen table. On design, Richard Mayer’s own summary of the multimedia research reports segmenting a lesson into shorter learner-paced parts at a median d of 0.67 across its tests, while two of the moves teachers reach for first come out near zero: putting your on-screen text alongside your own narration sits at 0.10, and showing your face on screen at 0.20. Mayer notes himself that “most of our research has not targeted special populations” and calls for more work in real classrooms.

    How do you flip one lesson?

    Pick the lesson where you spend the most time talking and the least time watching students work. Flip that one, keep everything else, and judge it on what you could see during the period rather than on whether it felt modern.

    1. Choose a lesson that is mostly explanation. A procedure, a vocabulary set, a background-knowledge lecture. Do not flip the lesson that depends on discussion — you would be moving the good part home.
    2. Make the video under six minutes, or find one. Khan Academy says its library is “always free,” and your own three-minute recording beats a polished twelve-minute one. Record for the purpose rather than filming your class period.
    3. Attach a task that produces something you collect — and subtract the old homework. Three questions, a one-sentence summary, one thing they did not follow. This is the moderator with evidence behind it. But say the swap out loud, because from a desk this reads as an addition: students who used to watch nothing and do problems at home now watch a video, answer questions, and do the problems. If the old assignment does not actually come off the list, you have not flipped anything, you have added a second homework. Edpuzzle’s free plan caps the free plan at 20 saved activities and will embed questions in a video; a paper half-sheet does the same job for nothing.
    4. Decide the no-access and no-watch plan before you assign it. A transcript on paper, a lunch viewing, or a five-minute recap station at the start. Not every student has watchable internet at home and some are on a phone. Keep the consequence off the gradebook: a student whose wifi failed is not a student who refused, and the two are indistinguishable from the front of the room. Grading that difference means grading household income.
    5. Read the collected task before the bell and group the room with it. The quiz is only worth the moderator it earned if it changes the next fifty minutes. Three piles — got it, shaky, did not prepare — take two minutes to sort and tell you exactly who to start with. Collecting data you do not act on is the most common way this step gets hollowed out.
    6. Plan the class period as practice you can see. This is the step that decides whether anything improves. Work in pairs you circulate through, problems you watch them attempt, and real time spent letting students wrestle with something before you rescue them. If the period is still a lecture, you have moved a video and changed nothing.
    7. Do not shorten the class. The one moderator that came out negative was cutting face-to-face time. The freed minutes belong to practice, not to the calendar.
    8. Judge it on one question. Did you see more of their thinking than you usually do? If yes, flip the next lesson of that type. If no, the problem is the class period, not the video.

    Can you flip a classroom without videos or technology?

    Yes, and for many classes it is the better version. Nothing in the definition requires video. It requires that first exposure happen before class, which a reading, a one-page primer or a short problem set does equally well.

    The Flipped Learning Network’s own wording is about where direct instruction happens, not what format carries it. A two-page excerpt with four margin questions is a flipped lesson. So is a diagram students annotate before they arrive. Paper also removes the access problem in one move, which is why it is often the right call in a building where home connectivity is uneven.

    One thing has changed since most of this research was done, and it argues for paper rather than against preparation. students are already using AI on work you send home, and a take-home task whose output is a paragraph is now a task a chatbot can finish in seconds. A preparation task that asks for a question the student still has, or a prediction they will test in class, is both harder to outsource and more useful to you when the period starts.

    Clay’s take

    The flipped classroom is a decent idea carrying a reputation it did not earn, and the part worth keeping is not the video.

    The genuinely good argument for flipping has nothing to do with technology. It is that watching students practice is the most useful thing a teacher does with a class period, and most schedules spend that period on the thing a recording could do instead. If flipping one lesson gets you from talking for thirty minutes to circulating for thirty, that is a real improvement whether the effect size is 0.17 or 0.54.

    What does not hold up is the version sold in an in-service, where a single large number appears on a slide with no mention that nine of ten studies behind it were run on college students who chose to be there. Teenagers at eight in the morning are a different population with different incentives, and a model that depends on unsupervised preparation is a model that depends hardest on the students with the least support at home. That is worth saying out loud before a department commits to it.

    So flip selectively and keep the class time. The evidence’s clearest instruction is the one nobody repeats: do not use this to meet less. Everything good about the approach lives in the period you freed up, and if you spend that period the way you spent it before, you have given your students homework and yourself a recording project. The same restraint belongs in the homework policy itself, which holds up better when it sets how often you assign rather than how many minutes it should take.

    What to do next

    Flip one lesson this month. Not a unit and not a semester — one lesson, the one where you talk the most.

    Keep the video under six minutes, attach three questions you actually collect, and write down beforehand what happens for the student with no internet and the student who simply didn’t watch it. Then spend the whole period watching them work, and judge the experiment on whether you saw more of their thinking than usual. That is the only measure that matters here, and it is one you can read the same day.

    Frequently Asked Questions

    What is the flipped classroom model?

    It is an arrangement where first exposure to new material happens before class, usually through a short video or reading, so class time can be spent on practice the teacher can observe. The Flipped Learning Network defines it as direct instruction moving from the group learning space to the individual learning space. Flipping a class does not automatically produce flipped learning.

    Does the flipped classroom improve test scores?

    On average, a little. The largest meta-analyses report effects around 0.35 of a standard deviation, falling to roughly 0.17 to 0.24 once underpowered studies and publication bias are accounted for. One meta-analysis built only on K-12 students reported a larger average of 0.54, but its own author flags publication bias in that set.

    Is the flipped classroom research based on high school students?

    Mostly not. Across the major syntheses, roughly 85 to 90 percent of the evidence comes from higher education. One review found 239 of 272 learning comparisons were university-based against 20 from secondary schools, and its pass-rate analysis contained no school studies at all. Quote those numbers with the population attached.

    What do you do when students don’t watch the video?

    Attach a short task to the preparation that you collect, and plan an opening that works for a split room. Adding quizzes on the pre-class material was one of only two moderators that significantly improved outcomes in a 114-study meta-analysis. Decide in advance what happens for the student who did not prepare, rather than improvising it.

    How long should a flipped lesson video be?

    Under six minutes where possible, and under nine at most. In an analysis of 862 videos and almost 6.9 million viewing sessions, median engagement was at most six minutes regardless of video length, and viewers often stopped less than halfway through videos over nine minutes. That study measured adult online learners, not teenagers.

    Can you flip a classroom without technology?

    Yes. The definition concerns where direct instruction happens, not what carries it. A two-page reading with margin questions, an annotated diagram or a short primer all move first exposure before class. Paper also removes the home-internet problem entirely, which matters most in schools where connectivity is uneven.

    What are the main drawbacks of a flipped classroom?

    Students not preparing, unequal access to devices and internet at home, and the preparation time it costs the teacher. In a survey of 309 college faculty, 75 percent said preparing a flipped class takes more time than a traditional one. The deeper risk is flipping the video but leaving the class period unchanged.

    Does flipping let a school reduce class time?

    The evidence says the opposite. In a meta-analysis of 114 studies, flipped classes that reduced face-to-face time performed significantly worse than their comparison groups, at an estimate of negative 0.26. The benefit of the model lives in what the freed class time is used for, so removing that time removes the mechanism.

    Sources

    • Flipped Learning Network (2014). Definition of Flipped Learning. Page. Source of the March 2014 release, the four pillars, and the statement that flipping a class does not necessarily lead to flipped learning. The full definition sentence was read on a separately hosted copy of the network’s handout.
    • Hutton, A. (2020). The Efficacy of the Flipped Classroom Technique in Undergraduate Math Education: A Review of the Research. arXiv:2010.11393. Full text. Source of the origin history, including Mazur’s peer instruction and the hedged attribution to Bergmann and Sams. Used as a review describing those earlier works, not as a primary account of them.
    • Lage, M. J., Platt, G. J., and Treglia, M. (2000). Inverting the Classroom: A Gateway to Creating an Inclusive Learning Environment. The Journal of Economic Education, 31(1), 30–43. Bibliographic record. Citation verified on a bibliographic record only; the paper itself was not reachable, so no claim about its contents is made here beyond what the review above reports.
    • van Alten, D. C. D., Phielix, C., Janssen, J., and Kester, L. (2019). Effects of flipping the classroom on learning outcomes and satisfaction: A meta-analysis. Educational Research Review, 28, 100281. Full text. Source of g = 0.36, the null satisfaction result, the negative class-time moderator and the positive quiz moderator. 114 studies, of which 11 were secondary education.
    • Låg, T., and Sæle, R. G. (2019). Does the Flipped Classroom Improve Student Learning and Satisfaction? A Systematic Review and Meta-Analysis. AERA Open, 5(3). Full text. Source of the 0.35, 0.24 and 0.17 progression, the satisfaction reversal, and the education-level table showing 239 of 272 learning comparisons in higher education. The authors describe their bias correction as approximate.
    • Zhu, G. (2021). Is flipping effective? A meta-analysis of the effect of flipped instruction on K-12 students’ academic achievement. Educational Technology Research and Development, 69(2), 733–761. Institutional record. The only K-12-only meta-analysis found. Source of d = .54 from 53 effect sizes across 27 studies, the publication-bias flags and I² = 81.2 percent. Read on the author’s university record because the publisher’s page was not reachable; no confidence interval is quoted because none appears there.
    • Kapur, M., Hattie, J., Grossman, I., and Sinha, T. (2022). Fail, flip, fix, and feed: Rethinking flipped learning. Frontiers in Education, 7, 956416. Full text. Source of the 0.19 to 2.29 range across 46 meta-analyses, the authors’ own g = 0.37, I² = 91.27 percent, the publication-bias test, the 90 percent university figure and the observation that most in-class sessions are not modified. The paper’s separate comparison of in-class active learning contains an internal inconsistency and is not used here.
    • Hew, K. F., Bai, S., Dawson, P., and Lo, C. K. (2021). Meta-analyses of flipped classroom studies: A review of methodology. Educational Research Review, 33, 100393. Institutional record. Source of the 19 meta-analyses, the 495-plus primary studies, the 0.19 to 1.13 cognitive range and the six inadequately handled methodological elements. Read on institutional records; the full text was not reachable.
    • Lo, C. K., and Hew, K. F. (2017). A critical review of flipped classroom challenges in K-12 education. Research and Practice in Technology Enhanced Learning, 12. Record. Source of the 15 K-12 publications, the neutral-or-positive achievement finding and the mixed attitudes. Abstract only; the full text was not reachable, so no specific challenge figures are attributed to it.
    • National Center for Education Statistics. Rural Students’ Access to the Internet. Indicator. 2019 American Community Survey data on 5- to 17-year-old students. Source of the 5 percent, 7 percent rural, 11 percent remote rural, 82 and 76 percent broadband, and the 12-versus-2 percent poverty gap.
    • National Center for Education Statistics. Children’s Internet Access at Home. Indicator. 2021 data on 3- to 18-year-olds. Source of the 97 percent, the 4 percent smartphone-only figure, and the page’s own note that a second federal survey put no-access at 12 percent rather than 3 percent because the surveys define access differently.
    • Pew Research Center (2018). Nearly one-in-five teens can’t always finish their homework because of the digital divide. Report. Survey of 743 U.S. teens aged 13 to 17, fielded spring 2018. Source of the 17 percent, 35 percent and 45 percent figures. Pre-pandemic data; no more recent Pew survey on this question was found.
    • Ash, K. (2012). Educators Evaluate ‘Flipped Classrooms’. Education Week. Article. Source of the practitioner quotations on students not watching, “self-paced became no pace,” and flipping as a time-shifting tool. Journalism rather than research, and quoted as such.
    • Reich, J. (2013). Flipping the New York Times. Education Week opinion. Article. Read for its caveats on the widely circulated schoolwide-flip turnaround story, including that the reported gains were internal markers and that enrollment shifted substantially over the same period. No failure-rate figure from that story is repeated here, because no peer-reviewed evaluation of it was found.
    • Faculty Focus (2013). Survey Confirms Growth of the Flipped Classroom. Report. Source of the 75 percent preparation-time figure. 309 higher education faculty, a self-selected 2013 sample whose publisher states the results cannot be projected to the whole population. It is the only quantified preparation-time figure found, and it is not about school teachers.
    • Guo, P. J., Kim, J., and Rubin, R. (2014). How Video Production Affects Student Engagement: An Empirical Study of MOOC Videos. Proceedings of Learning at Scale 2014. Full text. Source of the six-minute median, the nine-minute drop-off, the problem-attempt percentages and the purpose-made-versus-recorded comparison. Population: 127,839 edX learners across 862 videos — self-selected adult online learners, and the outcome is voluntary watch time, not learning.
    • Mayer, R. E. The Future of Multimedia Learning. Journal of Applied Instructional Design, 11(4). Full text. Source of the median effect sizes by principle, including segmenting at 0.67, redundancy at 0.10 and image at 0.20, and of Mayer’s own statement that most of the research has not targeted special populations. The page gives no publication year and reports one principle’s median differently in two places.
    • Edpuzzle. What does Edpuzzle cost? Help center. Source of the free plan’s 20-activity limit. No price is quoted for any paid tier because none appears on the vendor pages that could be read.
    • Khan Academy. Does it cost money to use Khan Academy? Help center. Source of the statement that learning at Khan Academy is always free.
    • Google. Education editions FAQ. Admin knowledge site. Source of the statement that Education Fundamentals is available at no cost for qualifying institutions. The page does not define qualifying institutions and lists no prices for the paid editions.

    About Clay Shumate

    Clay Shumate is a certified secondary Social Studies teacher in the public schools of West Alabama, with seven years of classroom experience, a B.A. in History, and an M.Ed. in Secondary Education. He writes about project-based learning, student responsibility, respect, and practical ways to hold young people to a higher standard while giving them room to learn from mistakes. He is a member of the Society of Professional Journalists and writes to its Code of Ethics; this site’s editorial standards and corrections policy are published in full. More about Clay.

  • Universal Design for Learning: What UDL Is, and What the Evidence Supports in Grades 6–12

    Universal Design for Learning: What UDL Is, and What the Evidence Supports in Grades 6–12

    By Clay Shumate

    Universal Design for Learning is a framework for planning a lesson up front so fewer students need a retrofit later, built on three principles: multiple means of engagement, representation, and action and expression. Federal law calls it a scientifically valid framework. The research is thinner than that sounds, and thinnest of all in grades 6 through 12.

    Key Takeaways

    • UDL is a planning stance, not a set of activities. CAST’s UDL Guidelines 3.0 asks you to design the lesson expecting variability, rather than build one version and adapt it when it fails.
    • Congress called it “scientifically valid” in 2008, and that was a statutory label rather than a research finding. 20 U.S.C. § 1003(24) wrote the phrase into law, and four sections of the Every Student Succeeds Act reference UDL.
    • The most-cited meta-analysis says the outcomes were not demonstrated. Capp’s 2017 meta-analysis concludes UDL improves the learning process and then states plainly that “the impact on educational outcomes has not been demonstrated.”
    • At secondary level the base is 22 studies, two of them randomized. the systematic review of UDL at second level found all 22 were North American, none were longitudinal, and only five implemented the whole framework.
    • The parts have better evidence than the bundle. Explicit vocabulary instruction, comprehension-strategy instruction and intensive intervention are all rated STRONG for these grades by the Institute of Education Sciences. The framework that packages them is not.

    What is Universal Design for Learning?

    Universal Design for Learning is a framework for designing instruction so that predictable differences between students are planned for in advance instead of handled as exceptions. The term comes from architecture, where a curb cut is built in rather than added after a complaint.

    The education version is maintained by CAST, a nonprofit that has published the UDL Guidelines since the late 1990s. The current version, CAST’s UDL Guidelines 3.0, was released in July 2024. CAST states the goal as “learner agency that is purposeful & reflective, resourceful & authentic, strategic & action-oriented” — a fair example of why the framework needs translating before it reaches a department meeting.

    The plain-English version is shorter. Before you teach the lesson, ask three questions. Why would a student in this room care? How will they take the information in? How will they show me they have it? If you only have one answer to any of those, you have built a lesson that works for the students who happen to match it.

    Chart of the three Universal Design for Learning principles with the planning question each one asks: engagement asks why a student would care, representation asks how they take information in, and action and expression asks how they show what they know.

    What are the three UDL principles?

    The three principles are engagement, representation, and action and expression. CAST frames each as “design multiple means of” that thing. Under them sit nine guidelines, and under those sit the items that Guidelines 3.0 renamed from “checkpoints” to “considerations.”

    • Engagement — guidelines 7 through 9: welcoming interests and identities, sustaining effort and persistence, and emotional capacity. This is the one teachers skip, and it is also the one with the weakest evidence behind it.
    • Representation — guidelines 1 through 3: perception, language and symbols, and building knowledge. Captions, read-aloud, glossaries, graphic organizers, background knowledge.
    • Action and expression — guidelines 4 through 6: interaction, expression and communication, and strategy development. Letting a student record an answer instead of writing it, or build the argument as a chart before writing a paragraph.

    What changed in 3.0 is mostly framing. About Guidelines 3.0 describes the theme as “emphasizing identity as part of variability,” and says the revision “emphasizes addressing barriers rooted in biases and systems of exclusion.” CAST does not publish a side-by-side comparison with the previous version on that page. No new evidence claim arrived with the update.

    Is UDL an evidence-based practice?

    Not in the sense most teachers mean by that phrase. UDL is written into federal law as a “scientifically valid framework,” and a large body of research is attached to it. What it lacks is a reliable estimate of how much it moves achievement.

    That gap is widest in middle and high school, which is the band this article is about. The two facts sit oddly together, and the people who noticed first were not critics. A peer-reviewed article in Educational Policy by researchers friendly to the framework is titled Universal Design for Learning: When Policy Changes Before Evidence. That is the whole problem in a title.

    The legal side first. 20 U.S.C. § 1003(24), from the Higher Education Opportunity Act of 2008, defines UDL as “a scientifically valid framework for guiding educational practice” that “reduces barriers in instruction, provides appropriate accommodations, supports, and challenges.” The Every Student Succeeds Act adopted that definition and references UDL in four places, per CAST’s summary of UDL in the ESSA. Worth knowing: the language is permissive. Section 1005 asks states to apply UDL to assessment design “to the extent practicable.” It is an endorsement, not a mandate.

    A phrase in a statute is not a finding. Congress wrote “scientifically valid” into the definition itself, which means every citation of the law reads as a citation of evidence. It is not one. CAST’s own research-evidence page lists foundational thinkers including Piaget, Vygotsky and Bloom, states that “neuroscience provides a solid foundation” for the framework, and acknowledges that “updating and expanding the research base is an ongoing process.” There is no strength-of-evidence rating per guideline.

    What did the big UDL meta-analyses actually find?

    They found a positive direction and almost no precision. The three most-quoted syntheses disagree by an order of magnitude, and the largest number in the literature comes from the weakest set of studies. Read the sentence after the headline in each one.

    Table comparing four UDL research syntheses by number of studies, how many used randomized designs, and what each one actually concluded, showing that the largest pooled effect size comes from the set with the fewest control groups.

    Capp’s 2017 meta-analysis reviewed 18 peer-reviewed studies from 2013 to 2016 and is the one most often cited as proof. It reports no single pooled effect size, and its participants run from preschoolers through university students. Its conclusion is that “UDL is an effective teaching methodology for improving the learning process for all students.” The next sentence is the one nobody quotes: “The impact on educational outcomes has not been demonstrated.” Capp also notes that his own positive findings “may be due to the lack of empirical evidence involving a pre- and post-test methodology.” Several of the largest effect sizes in his set measured teacher lesson planning or teacher motivation, not student learning.

    One caution worth passing on: the copy of Capp’s paper that comes up first in a search carries an auto-generated summary claiming UDL “improves accessibility for 100% of students.” The paper says nothing of the kind. That is roughly how the UDL figures in a slide deck get there.

    Almeqdad and colleagues (2023) is the source of the largest figure, a pooled Hedges’ g of 3.56. That is not a usable estimate and the authors say so: the 95 percent confidence interval runs from 1.03 to 6.08, individual study effects range from 0.33 to 13.18, and heterogeneity is I² = 99.8 percent. Of the 13 studies pooled, ten had no control group, and eight measured teachers rather than students. An I² that high means the studies are not measuring the same thing.

    One distinction has to travel with all of this, because it is the one a district leader will reasonably press on. Thin evidence is not evidence that something does not work. Most of what is wrong with the UDL research is that the studies are short, small, rarely randomized and rarely implement the whole framework — which means the honest verdict is unproven at this grade level, not disproven. That is a real difference, and it cuts both ways: it is not grounds for dismissing the framework, and it is not grounds for a mandate either.

    The most useful contrary work is Bissonnette’s reanalysis of a 2023 meta-analysis of 20 experimental UDL studies. He applied three ordinary filters — at least twelve weeks long, a regular student population, taught in person — and reports that “of the 20 studies in this meta-analysis, only two meet these criteria.” Both of those two were null on their primary outcomes. One, in fifth grade, “had no significant effect on overall reading comprehension.” The other, a full-year fourth-grade science study, “showed no statistically significant impact on academic performance or motivation” overall, though it found large positive effects of 0.82 to 1.01 for students with disabilities on curriculum-aligned measures that did not transfer to the national science test. Both studies are elementary, which matters for a grades 6 through 12 reader.

    What does the evidence say about UDL in middle and high school?

    At secondary level it is thin. the systematic review of UDL at second level is the only synthesis built specifically for this band, and it found 22 empirical studies. Two used randomized pre- and post-test designs. None were longitudinal, none were national in scale, and all 22 were North American.

    The detail matters more than the count. Only five of the 22 were comprehensive UDL implementations, and three made no reference to specific principles at all. Coverage was lopsided toward the principle teachers control most easily: representation appeared in 14 studies, engagement in 13, action and expression in 11, self-regulation in four. Exactly one study measured student engagement as an outcome — a strange gap in a literature built on a principle of that name. The reviewers ran a quality-appraisal tool and did not report the results. Sixteen of 22 studies were positive or somewhat positive, four mixed, two found no effect.

    The single hardest finding in the secondary set is King-Sears and colleagues (2015). Students in four co-taught high school chemistry classes were randomly assigned to a UDL treatment or a comparison condition, with a pre-test, a post-test and a four-week delayed post-test. The result: “there were no significant differences between conditions; however, there was an interaction effect between students with and without disabilities for post-tests.” When Capp recalculated that study, the effect was 0.97 for students with disabilities and −0.49 for students without them on the delayed test. The original authors call for “disaggregated learning outcomes” in future work, which is a polite way of saying a UDL package can help one group while not helping, or slightly costing, everyone else.

    Widen it to all of Pre-K through 12 and it does not improve. Ok, Rao, Bryant and McDougall (2017) reviewed 13 studies covering 3,550 students, reported that “none of the 13 studies utilized a true-experimental research design,” and found six of the 13 never connected their intervention to UDL principles at all.

    What is the difference between UDL, differentiation, and accommodations?

    Timing and target. UDL designs the lesson before you meet the students. Differentiation adjusts it once you know them. An accommodation changes conditions for one named student because a legal document says so. The three are layers, not competitors.

     When it happensWho it is forWhat it changesIs it optional?
    Universal Design for LearningBefore the lesson is taughtThe whole class, by designThe materials, the task options and the routes into the contentYes — a planning choice
    Differentiated instructionDuring the unit, in response to what you seeGroups or individuals you have identifiedPacing, grouping, questioning, the complexity of the taskYes — a teaching choice
    AccommodationsWhenever the plan saysA named student with an IEP or 504 planConditions: time, setting, format, toolsNo — legally required

    CAST’s own brief on UDL and differentiated instruction puts the first distinction this way: “UDL emphasizes proactive design of the environment and curricula while DI emphasizes responding the individual needs,” and calls it “a conceptual shift from thinking about ‘fixing the student’ to ‘fixing the curricula.’” In practice the line is softer than the brief makes it sound. Most of what the site already covers about adjusting a lesson once you already know who needs what overlaps heavily with what a UDL advocate would call good design, and deciding when to take a support away is a question both frameworks have to answer.

    Does UDL replace an IEP or a 504 plan?

    No. Nothing in the statute, the guidelines or the research supports dropping a required accommodation because a lesson was designed well. Federal law puts accommodations inside the definition of UDL rather than opposite it.

    That is worth being precise about, because the claim drifts. 20 U.S.C. § 1003(24) says UDL “reduces barriers in instruction, provides appropriate accommodations, supports, and challenges.” Read it again: provides accommodations. The most defensible version of the stronger claim appears inside Lowrey, Classen and Sylvest (2019), quoting other authors describing UDL as “lessening the number of accommodations and modifications necessary for students with disabilities.” Fewer needed is not none required, and that sentence is a description of a hope rather than a measured result.

    Good design makes an accommodation unremarkable, because everyone in the room has the same tool. It does not retire it. If anyone frames UDL as a way to reduce paperwork, that is the moment to ask which specific plan is being changed and who agreed to it.

    The fairness objection is worth answering plainly, because families raise it and students feel it. If one student records an answer and another writes one, has the second been treated worse? No — both were asked the same question and graded against the same criteria. What changed is the delivery, not the standard. The version that would be unfair is a second format with a softer rubric attached, which is exactly what the step above is written to prevent.

    Which parts of UDL have the strongest evidence?

    The components, not the bundle. Several practices UDL packages have been evaluated on their own in these grade bands, and some carry federal STRONG ratings. The framework that assembles them has nothing comparable.

    Chart of practices that Universal Design for Learning bundles together, each paired with the strength of its own separate evidence base, from federally rated STRONG for explicit vocabulary and comprehension-strategy instruction down to small effects for offering students choice.

    Reading and language come out best. the What Works Clearinghouse guide for grades 4–9 rates three recommendations STRONG for students in those grades: build decoding skills for multisyllabic words, provide purposeful fluency-building activities, and routinely use comprehension-building practices. the IES practice guide on adolescent literacy rates explicit vocabulary instruction, explicit comprehension-strategy instruction and intensive individualized intervention for struggling readers as STRONG, and makes an honest caveat on the first one: only some of the underlying studies showed effects on standardized reading comprehension rather than on vocabulary learning alone. On this site, teaching the words directly is the same recommendation written as a classroom routine.

    Two specific UDL tools have their own numbers. a meta-analysis of text-to-speech tools for students with reading disabilities found an average weighted effect of d = 0.35 with a 95 percent confidence interval of 0.14 to 0.56 across 22 studies, and the authors still write that “it is not clear how effective text-to-speech is at improving reading comprehension.” The grade levels in that meta-analysis are not reported, so do not read it as a secondary finding. For strategy development, the review that rated practice testing and distributed practice highest out of ten study techniques points at retrieval practice and spacing, which is also what which study techniques hold up and which do not comes down to.

    Choice is the interesting one, because it is the UDL move teachers reach for first and the evidence is more modest than the enthusiasm. Patall, Cooper and Robinson (2008) pooled 41 studies and found choice raised intrinsic motivation by d = 0.30 under fixed effects, shrinking by roughly a third after correcting for publication bias but staying positive. Two details should change what you build: the effect was strongest for instructionally irrelevant choices and weaker for the academically relevant ones teachers actually offer, and it was strongest with two to four successive choices rather than many. The samples were North American and excluded students with learning or behavioral disorders. So a real choice instead of a decorative one means three good options, not twelve — and three genuinely different ones. Students notice immediately when the three choices are three versions of the same essay, and a choice that is not really a choice buys none of the effect and costs you the credibility of offering it.

    What does UDL look like in a grades 6 through 12 lesson?

    It looks like one planning pass over a lesson you already have. You are not writing three versions. You are checking that each of the three principles has at least two answers, and adding the cheapest second answer where there is only one.

    1. Write the goal without the method in it. “Explain why the compromise failed” is a goal. “Write a five-paragraph essay explaining why the compromise failed” is a goal welded to one format. Separate them and the rest of the pass gets easier.
    2. Count the ways in. If the content arrives only as a lecture, add the text or the chart. If only as a text, add the three-minute explanation. Two routes is the bar, not five.
    3. Count the ways out. Pick one task where the format is not the skill being graded and allow a second format there. One task, not every task — two formats means two grading mental models, and across five sections that cost is real. Keep the rubric identical by writing its criteria about the thinking rather than the format: “uses evidence from at least two sources” works on a paragraph and on a three-minute recording. A criterion like “well-organized paragraphs” does not, and that is how a second format quietly becomes the easier option.
    4. Make one support available to everyone. Captions on, directions posted as well as spoken, word bank on the board. Nobody gets handed the visibly different sheet.
    5. Name the hard part out loud — then ask them. Say what the task is actually testing and which support is there for what; a support students do not understand is a support they will not use. Then take ninety seconds and ask the class which of your two formats they would pick and why. Every step above has you guessing from the front of the room at something the people at the desks could just tell you, and the framework’s own engagement principle is about their agency, not your planning.
    6. Check what you added. Options are not free: more choices mean more decisions for students who lose the assignment between hearing it and starting it. If a student spent the period choosing, the design failed.

    Why does UDL implementation fail in real schools?

    Because it is a whole-practice change delivered as a one-day workshop, and because fidelity collapses quietly. The implementation data are more consistent than the outcome data, and they point one way: most UDL that gets counted as UDL is a fraction of the framework.

    The fidelity numbers from Bissonnette’s reanalysis are the clearest. In the full-year fourth-grade study, teachers logged into the platform an average of 9.63 days across the entire school year, and students completed 6.53 investigations. In the fifth-grade reading study, only 54 percent of the treatment group completed the program, averaging 6.96 of eight planned texts. Those are not tests of UDL. They are tests of a partial dose.

    The practical reading of that is not that UDL is untrainable. It is that a one-day workshop is the wrong delivery for a change in how someone plans. The fidelity numbers describe people who were introduced to a framework and then left with it, and the studies that look best are the ones that ran twelve weeks or longer. If a building is going to do this, the honest version is one principle a semester with planning time attached, not three principles in an August in-service.

    Training transfers badly too. Lowrey, Classen and Sylvest (2019) ran a UDL seminar with eight preservice teachers and found it “did not create a substantial overall change in the practice of planning and instruction.” Three of the eight increased their use of UDL guidelines in lesson plans and five decreased. Their plans still sorted students into high, average and low groups afterward. That is eight people and the authors call it preliminary, so take it as an illustration rather than a finding.

    Two deeper criticisms are worth knowing, because they explain the thin evidence rather than just complain about it. Baglieri’s critical review calls UDL’s central claim “a loose and ambiguous claim” and observes that “in most cases, UDL seems to be practiced as an organizational solution for inclusion or as an intervention for disability” rather than a redesign. A framework defined loosely is hard to test, which is one reason 22 secondary studies produced no comparable outcome measure. And an annotated bibliography of UDL work from 2023 to 2025 records a finding that cuts against the enthusiasm directly: in one study “too much interactivity, particularly in group discussions, was sometimes overwhelming” for some neurodivergent students. More options is not automatically better for the students the framework is most often invoked for.

    Clay’s take

    UDL is a good habit with an oversold research file, and both halves of that sentence matter if you work in a building where it is being rolled out.

    The habit is genuinely useful. Asking whether a lesson has more than one way in and more than one way out is a better question than most of what gets handed down in professional development, it costs a planning period rather than a budget line, and it tends to produce lessons that need fewer emergency adjustments on the day. Designing for the range you know is coming is just competent planning, and it is worth doing whether or not a meta-analysis ever lands on a number.

    What is not defensible is how it gets sold. When a district quotes the law’s “scientifically valid framework” language as though it settles the evidence question, a statutory phrase is doing work research has not done. Teachers told a framework is proven, who try it faithfully and see nothing move on their assessments, reasonably conclude they were oversold — and then discount the next thing too, including the things that really do have strong evidence. That is how you spend credibility you will need later.

    So: use the three questions, and spend the scarce hours on what has been tested on its own in these grades — explicit vocabulary and comprehension instruction, retrieval practice, and real intervention for students behind in reading. Treat the full framework as a sensible organizing idea, not an intervention with a known effect size.

    What to do next

    Take one lesson you are teaching this week and run the three questions over it. Not a unit, one lesson, and one you already consider good.

    Strip the format out of the goal. Count the routes in and the routes out, and where either count is one, add the cheapest second option with an identical rubric. Then make a single support available to the whole room rather than only to the students who have it in writing. That is the pass, and it takes about twenty minutes once you have done it twice. If you would rather spend the hour on the practice with the strongest evidence behind it, start with the words themselves and leave the framework for later.

    Frequently Asked Questions

    What does UDL stand for in education?

    UDL stands for Universal Design for Learning, a framework for planning instruction so that predictable differences between students are built into the lesson rather than handled as exceptions. It is maintained by the nonprofit CAST, which has published the UDL Guidelines since the late 1990s and released version 3.0 in July 2024. The term is borrowed from universal design in architecture.

    Is UDL evidence based?

    It depends what you mean. UDL is defined in federal law as a scientifically valid framework and has a large attached literature, but no reliable estimate of its effect on achievement, particularly in secondary grades. The most-cited meta-analysis states that the impact on educational outcomes has not been demonstrated. Several practices UDL bundles do have strong evidence on their own.

    Is UDL the same as differentiation?

    No. UDL is proactive and designed before you meet the students, aiming at the whole class by default. Differentiated instruction is responsive, adjusting pacing, grouping or task complexity once you know who needs what. CAST describes the difference as a shift from fixing the student to fixing the curriculum. In daily practice the two overlap heavily.

    Does UDL work in high school?

    The evidence does not say clearly. The only systematic review built for second-level education found 22 studies, only two of which used randomized designs, none longitudinal and all North American. One randomized study in high school chemistry classes found no significant difference between the UDL condition and the comparison condition, with different results for students with and without disabilities.

    Can UDL replace an IEP or 504 accommodation?

    No. Required accommodations stay required regardless of how well a lesson is designed. The statutory definition of UDL explicitly includes providing appropriate accommodations rather than replacing them. Good universal design can make an accommodation feel unremarkable because everyone has access to the same tool, but the legal obligation in a student’s plan is unchanged.

    What are some UDL examples for middle and high school?

    Posting directions in writing as well as saying them. Leaving captions on every video for the whole class. Allowing a recorded explanation instead of a written one where writing is not the skill being graded. Offering three task options rather than twelve. Providing a word bank everyone can use. Each is a second route in or out, not a separate lesson.

    How long does it take to plan a UDL lesson?

    Roughly twenty minutes per lesson once the habit is established, because it is a review pass over a lesson you already have rather than a rewrite. The failure mode is treating it as building several versions of everything. Implementation research suggests the real cost is sustained practice over a year, not the first planning session.

    Why is the research on UDL so weak if it is in federal law?

    Because the policy arrived first. Congress wrote the scientifically valid framework language into statute in 2008, and researchers friendly to UDL later published a paper titled Universal Design for Learning: When Policy Changes Before Evidence. Critics add that the framework is defined loosely enough to be hard to test, which is part of why studies rarely share an outcome measure.

    Sources

    • CAST. UDL Guidelines 3.0. Full guidelines and About Guidelines 3.0. Source of the three principles, the nine guidelines, the July 2024 release, the learner-agency goal statement and the stated themes of the revision. The framework’s publisher, not an independent evaluation.
    • CAST. Research Evidence. Evidence page. Source of CAST’s neuroscience claim and of its own caveat that expanding the research base is ongoing. No per-guideline strength rating is published.
    • United States Code. 20 U.S.C. § 1003(24), definition of universal design for learning, from the Higher Education Opportunity Act of 2008. Text. Source of the “scientifically valid framework” language and of the clause including accommodations. Read on a commercial reproduction of the U.S. Code rather than a government site.
    • CAST. UDL in the ESSA, 2016. Summary. Source of the four Every Student Succeeds Act sections referencing UDL and of the “to the extent practicable” qualifier. No reference to UDL was found in IDEA.
    • Capp, M. J. (2017). The effectiveness of universal design for learning: a meta-analysis of literature between 2013 and 2016. International Journal of Inclusive Education. doi:10.1080/13603116.2017.1325074. Full text. Source of the 18 studies, the learning-process conclusion, the “impact on educational outcomes has not been demonstrated” statement, and the recalculated King-Sears effects. The paper’s own limitations section gives a different study count than its methods section.
    • Devitt, A., Banks, J., Bray, A., and colleagues (2025). A Systematic Literature Review on the Effectiveness of Universal Design for Learning in Second-level Education. International Journal for Universal Design and Universal Design for Learning, vol. 1, no. 1. Full text. Source of every secondary-specific count used here. Published in the first issue of a new journal, and its abstract and methods report different screening totals.
    • King-Sears, M. E., Johnson, T. M., Berkeley, S., and colleagues (2015). An Exploratory Study of Universal Design for Teaching Chemistry to Students with and without Disabilities. Learning Disability Quarterly, 38(2), 84–96. Record. Source of the randomized high school chemistry null result. Sample size is not stated on the record page; only “four co-taught classes.”
    • Ok, M. W., Rao, K., Bryant, B. R., and McDougall, D. (2017). Universal design for learning in Pre-K to Grade 12 classrooms: A systematic review of research. Exceptionality, 25(2), 116–138. Full text. Source of the 13 studies, the 3,550 students, and the statement that none used a true-experimental design.
    • Almeqdad, Q. I., Alodat, A. M., Alquraan, M. F., and colleagues (2023). The effectiveness of universal design for learning: A systematic review of the literature and meta-analysis. Cogent Education, article 2218191. Abstract. Source of the pooled g = 3.56, its confidence interval, the 99.8 percent heterogeneity, and the study composition. Read as an abstract record rather than the full paper.
    • Bissonnette, S. UDL and Academic Achievement: Critical Analysis of the Meta-analysis by King-Sears and Colleagues (2023). TÉLUQ University. English translation. Source of the two-of-twenty filter result, both null primary outcomes, and the fidelity figures. The document carries internal citation inconsistencies and no stated year. The underlying King-Sears 2023 meta-analysis could not be opened, so no pooled effect size is quoted for it here.
    • Hollingshead, A., Lowrey, K. A., and Howery, K. (2020). Universal Design for Learning: When Policy Changes Before Evidence. Educational Policy. doi:10.1177/0895904820951120. Record. Interviews with 19 experts; read as an abstract record. Cited for its framing and for the point that definitional variability is acknowledged inside the field.
    • Baglieri, S. Toward Inclusive Education? Focusing a Critical Lens on Universal Design for Learning. Canadian Journal of Disability Studies. Full text. Source of the ambiguity critique and the observation that UDL is often practised as an organizational solution. Year, volume and pages are not printed on the PDF.
    • CAST (2013). UDL and Differentiated Instruction brief. Copy hosted by the Oregon Department of Education. Source of the proactive-versus-responsive distinction and the fixing-the-curricula quotation. The original CAST URL now redirects.
    • Institute of Education Sciences. Providing Reading Interventions for Students in Grades 4–9, WWC 2022007. Overview. Source of the three STRONG-rated reading recommendations for a grade band overlapping 6 through 12.
    • Kamil, M. L., Borman, G. D., Dole, J., and colleagues (2008). Improving adolescent literacy: Effective classroom and intervention practices, NCEE 2008-4027. Full guide. Source of the STRONG ratings for explicit vocabulary instruction, comprehension-strategy instruction and intensive intervention, the MODERATE rating for motivation and engagement, and the caveat about standardized comprehension outcomes.
    • Wood, S. G., Moxley, J. H., Tighe, E. L., and Wagner, R. K. (2018). Does use of text-to-speech and related read-aloud tools improve reading comprehension for students with reading disabilities? A meta-analysis. Journal of Learning Disabilities, 51(1), 73–84. Summary record. Source of d = 0.35 and its confidence interval. Grade levels are not reported on the records read, so this is not presented as a secondary-specific result.
    • Patall, E. A., Cooper, H., and Robinson, J. C. (2008). The effects of choice on intrinsic motivation and related outcomes: A meta-analysis of research findings. Psychological Bulletin, 134(2), 270–300. Full text. Source of the choice effect sizes, the publication-bias correction, the two-to-four-choices finding, and the result that instructionally relevant choices produced weaker effects. North American samples excluding students with learning or behavioral disorders.
    • Association for Psychological Science. Which Study Strategies Make the Grade? Publisher release for Dunlosky, J., Rawson, K. A., Marsh, E. J., Nathan, M. J., and Willingham, D. T. (2013), Psychological Science in the Public Interest. Release. Source of the finding that practice testing and distributed practice received the highest utility ratings of ten techniques. A publisher release, not the paper; no effect sizes are quoted from it.
    • Lowrey, K. A., Classen, A., and Sylvest, A. (2019). Exploring Ways to Support Preservice Teachers’ Use of UDL in Planning and Instruction. Journal of Educational Research and Practice, 9(1), 261–281. Full text. Source of the three-of-eight and five-of-eight seminar result and of the quoted description of UDL lessening the number of accommodations needed. Eight participants; the authors call it preliminary.
    • Western Norway University of Applied Sciences. What’s new in UDL? Selected articles 2023–2025, annotated bibliography. PDF. Used as a description of other papers rather than as a primary source, including the finding that excessive interactivity was overwhelming for some neurodivergent students. No named author or date on the document.

    About Clay Shumate

    Clay Shumate is a certified secondary Social Studies teacher in the public schools of West Alabama, with seven years of classroom experience, a B.A. in History, and an M.Ed. in Secondary Education. He writes about project-based learning, student responsibility, respect, and practical ways to hold young people to a higher standard while giving them room to learn from mistakes. He is a member of the Society of Professional Journalists and writes to its Code of Ethics; this site’s editorial standards and corrections policy are published in full. More about Clay.

  • Digital Citizenship in Grades 6–12: What to Teach, and What the Evidence Shows

    Digital Citizenship in Grades 6–12: What to Teach, and What the Evidence Shows

    By Clay Shumate

    Digital citizenship is instruction in how students manage their identity, privacy, conduct and sources online. The ISTE standard breaks it into four jobs: digital identity, how you treat people, well-being, and privacy. Any school taking federal E-rate money already has to teach part of it. The strongest classroom evidence is for source evaluation.

    Key Takeaways

    • Students are worse at this than adults assume. A Stanford assessment of 3,446 high school students, working with a live internet connection, found 96 percent never discovered that a climate website was tied to the fossil fuel industry and two-thirds could not tell news stories from ads.
    • Six lessons measurably changed what students could do. In a randomized study across one district’s six high schools, students who got the curriculum did twice as well at spotting dubious websites — and still earned only half the available points.
    • It is already required, in part, almost everywhere. Taking E-rate discounts obliges a school to teach students about appropriate online behavior and cyberbullying. Most districts satisfy that with a slideshow nobody remembers.
    • Cyberbullying programs work, but less well than you would hope. A review of 24 evaluations found reductions of roughly 10 to 15 percent in perpetration and about 14 percent in victimization, below what face-to-face bullying programs achieve.
    • Do not measure this with a survey. In the one study that measured both, students’ self-reported source-checking barely moved while their actual behavior changed enormously. They overestimate themselves.

    What is digital citizenship?

    Digital citizenship is teaching students to act responsibly in the online spaces they already live in — managing their identity and footprint, treating people decently, protecting their privacy and data, and judging what they read. It is a conduct-and-judgment subject, not a technology-skills subject.

    Chart of the four jobs in the ISTE Digital Citizen standard: identity and footprint, how you treat people online, digital well-being and time spent online, and privacy and personal data, with a note that intellectual property no longer appears among the sub-points.

    The reference definition most districts point at is the ISTE Standards for Students. Standard 1.2, Digital Citizen, reads: “Students recognize the responsibilities and opportunities for contributing to their digital communities.” Its four sub-points cover managing digital identity and the lasting impact of online behavior, demonstrating empathetic and inclusive interactions, safeguarding well-being by being intentional about time spent online, and protecting privacy and personal data.

    One thing worth knowing if your district’s slide deck is a few years old: intellectual property used to sit under this standard and no longer appears among its sub-points, while digital well-being and time online were added. Plenty of curriculum maps still quote the retired version.

    Two things this is not. It is not the “nine elements of digital citizenship” you will find all over search results — that is a framework from a book ISTE published, not from the standards. And it is not the set of media literacy activities covered separately on this site, which is one component of the job rather than the whole of it.

    Is digital citizenship required to be taught?

    Partly, and the requirement is older and broader than most teachers realize. It arrives through federal E-rate funding rather than through a curriculum mandate, which is why nobody announced it.

    Where it comes fromWho it bindsWhat it actually requires
    Protecting Children in the 21st Century Act, through the E-rate programAny school or district taking E-rate discountsEducate minors about appropriate online behavior, including interacting on social networking sites and in chat rooms, plus cyberbullying awareness and response
    State media literacy and information literacy statutesVaries by state; some high-school only, some K–12Varies widely — from a required unit of instruction to a resolution encouraging one
    ISTE Standards for StudentsNobody, legallyA reference framework districts adopt voluntarily

    The federal piece is specific. Guidance published by state education agencies summarizing the E-rate rules states that districts “must educate minors about appropriate online behavior,” including interactions on social networking sites and in chat rooms, “as well as cyberbullying awareness and response.” That has been a condition of the funding since the 2012 funding year.

    The state picture is messier than the headline. Media Literacy Now reports that 25 states have media literacy laws on their books. Read the entries and that number covers everything from binding statutes to resolutions that merely encourage instruction, so the count of states that actually require it is considerably smaller. Illinois is one of the real ones: every public high school must deliver a unit of instruction covering accessing information, analyzing and evaluating media messages, creating media, reflecting on media consumption, and social responsibility and civics.

    The tracker that counts those laws also declines to evaluate whether states implemented them, and notes that passing a bill is where the work starts rather than ends. That is the implementation risk for anyone at district level: the mandate usually arrives without the training or the curriculum money, and lands on whichever teacher has a spare period.

    How good are students at this already?

    Worse than nearly every adult expects. The largest assessment of American high school students’ online reasoning gave 3,446 of them a live internet connection and six open-response tasks, and the results were bad across every one.

    Ninety-six percent, shown a climate website, never discovered the organization’s ties to the fossil fuel industry. Two-thirds could not distinguish news stories from ads on a popular site’s front page. More than half believed an anonymously posted video, shot in Russia, was strong evidence of voter fraud in the United States. The sample reflected the demographic profile of American high school students, so this is not a story about one struggling school.

    It is worth pairing that with how much of their lives this covers. Pew Research Center’s survey of 1,458 teens found 97 percent use the internet daily, 40 percent describe themselves as online almost constantly, and 64 percent have used an AI chatbot, with about three in ten using one every day. Whatever a school teaches about online judgment is competing with a great deal of practice.

    And do not let anyone in a meeting tell you the kids have this handled because they grew up with it. The review that put the phrase under a microscope concluded there is no such thing as a digital native, and that people are not capable of multitasking when real thinking is involved.

    Does teaching digital citizenship actually work?

    For source evaluation, yes, with a measurable change from a small number of lessons. For everything else, the evidence is thinner than the curriculum catalogs suggest, and most of it measures what students say rather than what they do.

    Chart of what digital citizenship instruction moved and what it did not: spotting a dubious website, evaluating credibility in content classes and lateral source-checking improved, while students' self-reported behavior and the results of a single ninety-minute session did not.

    The strongest school result comes from a study of about 500 high school students in an urban Midwestern district, in a required government class at each of the district’s six high schools. After six 50-minute lessons, the curriculum group did twice as well as before at spotting dubious websites. The researchers were straight about the ceiling: those students still earned only half the total available points. The study was funded by Google.org, which is worth saying out loud given that Google also publishes a competing free curriculum.

    A follow-up put the same lessons inside biology and world geography classes with 574 ninth graders, taught by content teachers rather than a specialist, and reported statistically significant growth in evaluating credibility. No control group is described, and no effect size is published on the accessible version, so read it as encouraging rather than conclusive.

    The sharpest finding for anyone planning to evaluate a program is from a college study of 230 students. The share earning full marks on a source-checking task went from about 6 percent to 61 percent — while their self-reported source-checking barely moved and the difference was not statistically significant. If you survey students about their habits, you will measure their confidence.

    Do cyberbullying lessons reduce cyberbullying?

    Somewhat. A review of 24 evaluations of school-based programs found reductions of roughly 10 to 15 percent in cyberbullying perpetration and about 14 percent in victimization. Real, worth having, and smaller than the same team’s figures for face-to-face bullying programs.

    Those face-to-face programs, across more than 100 evaluations, reduced bullying perpetration by about 19 to 20 percent and victimization by roughly 15 to 16 percent. The researchers themselves describe the effects as modest, and note that almost all of the outcomes rest on students reporting on themselves.

    The scale of the problem is not in doubt. Pew found that 46 percent of teens have experienced at least one of six online abusive behaviors, with offensive name-calling at 32 percent and false rumors at 22 percent. In the same survey, roughly four in ten teens said teachers are doing a good or excellent job addressing online harassment — which means six in ten said otherwise, and they are the ones who would know.

    That is the number I would put in front of a faculty before adopting anything. A ten to fifteen percent reduction is worth the time. Telling students the problem is handled, when most of them can see it is not, costs credibility that a lesson cannot buy back. This is the same reason repair after harm has to be followed through rather than announced.

    Which free digital citizenship curriculum should I use?

    Pick by what you are trying to change. For source evaluation, the Stanford-derived civic online reasoning lessons have the strongest evidence. For broad coverage across a whole school, Common Sense has the most material. All of the options below are free; two of them want an account.

    SourceWhat you getGradesAccount needed?Best for
    Civic Online Reasoning (Digital Inquiry Group)30 free lessons plus assessments; the middle school module is 5 lessons over about 5 daysMiddle school to collegeNot stated on the pageThe one with the classroom evidence behind it
    Common Sense EducationSix lessons per grade by the page’s own listing, 45–60 minutes each in the 9–12 collectionK–12Not stated; sign-in offeredBreadth across every grade in a building
    Be Internet Awesome (Google)Five topics with lesson plans and a gameNot stated for the core curriculumNo — the page says no personal information or loginA no-friction starting point
    NetSmartz (NCMEC)One presentation for middle school, one for high schoolK–12Yes for downloads — log in or registerA single assembly-style session

    One caution on the largest of them. Common Sense has announced that its long-running digital citizenship curriculum is being archived and is steering teachers to a replacement called Digital Literacy & Well-Being. If your building’s scope and sequence names specific lessons from the old set, check the links before you plan a semester around them.

    Common Sense is also the publisher of the screen-time figures most often quoted to justify buying a digital citizenship program, and its own page carrying those figures states no sample size or methodology. Good material, and worth knowing who is counting.

    Who should teach it, and how long does it take?

    Content teachers, inside real subjects, in blocks of about five lessons. Both studies that moved the needle ran that way: six 50-minute lessons in a government class, and the same lessons taught by biology and world geography teachers.

    What has not been shown to work is the format most schools choose. A single 90-minute middle school session improved two narrow behaviors and left students relying on novice strategies for everything else. The advisory-period version is worse still: in a student newspaper editorial from a California high school, the editor’s verdict on her district’s mandated 20-minute lessons was that they are not making an impact and are wasting students’ and teachers’ time. She is describing exactly the delivery model a compliance requirement produces.

    There is a more uncomfortable reason to put this inside content classes. A government teacher asking students to check a source is doing history; an advisory teacher asking the same question in a twenty-minute block is doing a module. Students can tell the difference and they grade us on it. The practical version is to attach the lessons to assignments that already exist — the research paper, the current-events discussion, the document analysis you already run — so that checking a source is part of doing the work rather than a separate unit about being a good person.

    What should you skip?

    Gamified misinformation apps, anything built on the digital natives idea, and self-report as your measure of success. Each is popular, and each has evidence against it rather than merely an absence of evidence for it.

    Chart of what to skip in digital citizenship instruction: gamified misinformation games, materials built on the digital natives idea, treating the nine elements framework as a standard, and using student self-report as evidence of skill.

    The inoculation games are the clearest case. A reanalysis of 13 experiments covering two widely promoted games found they did not improve students’ ability to tell true news from false — players simply began calling more things false, including the true ones. The authors warn the games may be counterproductive. A class that leaves believing nothing is not more informed than one that leaves believing everything.

    It is also worth knowing how thin the foundation under internet safety education is. A systematic review had to examine 31 meta-analyses from other prevention fields, because the internet safety programs themselves had not been evaluated rigorously enough to review — and the authors noted there had not even been efforts to define what behavior the programs were trying to change.

    Define the behavior first, then pick the lessons. “Students will be responsible online” is not a measurable outcome. “Students will open a second tab and search the name of an unfamiliar site before quoting it” is. The civic online reasoning materials ship with their own assessment tasks, including baseline and matched post versions, which is more than most free curricula give you.

    One warning from the student side of the desk, because it is the fastest way to lose a class. Do not grade the ritual. If the points are for opening a second tab, students will open a second tab, and the one who already knew the source was solid gets marked down for not performing a check they did not need. Score the sources that end up in the work and ask how they know. And do not open the unit with the statistic that 96 percent of students failed — teenagers hear that as an adult telling them they are gullible, from a generation that forwards things on Facebook. Hand them the climate website first and let them find it themselves.

    Clay’s take: teach the part you can actually check

    Most digital citizenship curricula are a values unit with a technology coat of paint, and students read them that way. The part with real evidence behind it is narrower: can a student find out who is behind a website before believing it?

    That one is less inspiring than the brochure, and it is teachable in about five lessons and testable with a laptop, which the rest of the brochure is not.

    Keep it procedural and it survives a school board meeting. The tasks in the research use climate and election material, and a unit that teaches students which conclusions are correct will draw a complaint it deserves. A unit that teaches who published this, what the publisher wants, and what other sources say is a method, and a method applies to every side of every argument. That distinction is worth putting in writing before a parent asks.

    It is also not an incident-response plan, and compliance paperwork tends to blur the two. Teaching students about cyberbullying in September does not tell an administrator what to do about a group chat in February, and a district that has checked the instruction box has not answered the harder question.

    I would spend the instructional time there and handle the rest the way we handle other conduct — in the moment, with the actual incident, in the room where it happened. A kindness lesson in October does not compete with what a kid sees in a group chat in February. A teacher who responds to the group chat does.

    The honest limit on everything above: even the best study here left students at half marks, the cyberbullying effects are in the ten to fifteen percent range, and the whole field runs on students reporting on themselves. Anyone selling certainty about this is selling something. The same applies to the parts of this that live elsewhere on the site — what the evidence says about phones in a classroom, how to make student AI use count for something, and what happens after an integrity case — are separate problems with separate evidence, and digital citizenship is not a substitute for any of them.

    Where to start, concretely: take one assignment this semester that requires outside sources, add two lessons on checking who published a site, and then grade the next assignment on the quality of the sources that show up in it. That is four lessons fewer than the research used, and it puts the points on the outcome rather than on the ritual. Add the rest when you have proof the first part stuck.

    Frequently Asked Questions

    What are the main elements of digital citizenship?

    The ISTE Digital Citizen standard groups them into four: managing your digital identity and the lasting impact of what you post, treating people online with empathy and inclusion, safeguarding your own well-being and time spent online, and protecting your privacy and personal data. Source evaluation sits alongside these and has the strongest classroom evidence of any of them.

    Is digital citizenship required to be taught in schools?

    Partly. Any school taking federal E-rate discounts must educate minors about appropriate online behavior, including social networking and chat rooms, plus cyberbullying awareness and response. Beyond that it varies by state: 25 states have media literacy laws on the books, but those entries range from binding statutes requiring a unit of instruction down to resolutions that merely encourage one.

    Does teaching digital citizenship actually work?

    For source evaluation, yes. High school students who received six 50-minute lessons scored twice as well afterward at identifying dubious websites, though they still earned only half the available points. Evidence for the broader conduct and well-being topics is much thinner, and most of it rests on students reporting on their own behavior rather than on anything measured.

    How long does it take to teach digital citizenship?

    The studies that changed what students could do used about five or six lessons of 45 to 60 minutes. A single 90-minute session improved two narrow behaviors and left students relying on novice strategies otherwise. Short advisory-period modules are the most common format in schools and have no evidence behind them at all.

    Who should teach digital citizenship?

    Content teachers, inside their own subjects. The strongest study ran in a required government class, and a follow-up put the same lessons in biology and world geography taught by those teachers rather than a specialist. Attaching the lessons to assignments that already need sources makes source-checking part of the work instead of a separate unit about being a good person.

    What is the best free digital citizenship curriculum?

    It depends on the goal. The Stanford-derived civic online reasoning lessons have the research behind them and ship with their own assessments. Common Sense Education has the broadest coverage across grades, though it is archiving its long-running digital citizenship curriculum in favor of a replacement. Google’s Be Internet Awesome requires no login at all; NetSmartz asks you to register for downloads.

    Do cyberbullying prevention programs reduce cyberbullying?

    By some amount. A review of 24 school-based program evaluations found reductions of roughly 10 to 15 percent in perpetration and about 14 percent in victimization, which is below what the same researchers found for face-to-face bullying programs. The authors describe the effects as modest and note that nearly all of the outcomes depend on students reporting about themselves.

    How do I measure whether digital citizenship instruction worked?

    Watch what students do, not what they say. In the study that measured both, the share of students checking sources properly went from about 6 percent to 61 percent while their self-reported checking barely moved at all. Give students a live browser and an unfamiliar website, and score the sources that end up in their actual work.

    Sources

    • ISTE. ISTE Standards for Students, 2024 edition. Full text. Source of the wording of Standard 1.2, Digital Citizen, and its four sub-points. Intellectual property appeared under this standard in the previous edition and does not appear among the 2024 sub-points, which is why older curriculum maps quote it differently.
    • Breakstone, J., Smith, M., Wineburg, S., Rapaport, A., Carle, J., Garland, M., & Saavedra, A. (2021). Students’ Civic Online Reasoning: A National Portrait. Educational Researcher. Stanford Digital Repository record. Source of the 3,446-student sample, the six tasks completed with a live internet connection, and the 96 percent, two-thirds and over-half figures. The full text is paywalled; the figures above are the ones stated on the open repository record, which reports results for three of the six tasks.
    • Wineburg, S., Breakstone, J., Smith, M., Ortega, T., & McGrew, S. (2022). Reported by Stanford Graduate School of Education, It Doesn’t Take Long to Learn How to Spot Misinformation Online. Article. Source of the roughly 500 students, six district high schools, six 50-minute lessons, the doubled score, the half-of-available-points ceiling, and the Google.org funding. The journal article itself is paywalled; no effect size is reported on the accessible page, so none is quoted here.
    • McGrew, S., & Breakstone, J. (2023). Civic Online Reasoning Across the Curriculum. AERA Open, vol. 9. Repository record. Source of the 574-student sample, drawn from grade 9 biology and world geography classes, and the statistically significant growth. No control group is described and no effect sizes appear on the accessible record.
    • Brodsky, J. E., Brooks, P. J., Scimeca, D., Todorova, R., Galati, P., Batson, M., Grosso, R., Matthews, M., Miller, V., & Caulfield, M. (2021). Improving College Students’ Fact-Checking Strategies Through Lateral Reading Instruction in a General Education Civics Course. Cognitive Research: Principles and Implications, 6, 23. Full text. Source of the 230-student sample, the rise from 5.9 to 61.0 percent on the performance measure, and the non-significant self-report result. First-year university students, not secondary students.
    • University of Cambridge Violence Research Centre, summarizing Gaffney, H., Farrington, D. P., Espelage, D. L., & Ttofi, M. M. (2019), Are Cyberbullying Intervention and Prevention Programs Effective? Aggression and Violent Behavior, 45. Summary page. Source of the 24 evaluations, the 10 to 15 percent and 14 percent cyberbullying reductions, and the 19 to 20 percent and 15 to 16 percent figures for face-to-face programs from over 100 evaluations. The journal article was not reachable; these figures come from the research centre’s own summary of it. The team’s Campbell systematic review describes the effects as modest and notes that self-reported measures may over- or under-estimate them.
    • Faverio, M., & Sidoti, O. Teens, Social Media and AI Chatbots 2025. Pew Research Center. Report. Source of the 97 percent daily internet use, 40 percent almost constantly online, and 64 percent AI chatbot figures. Survey of 1,458 U.S. teens ages 13 to 17 recruited through their parents; all figures are self-reported.
    • Vogels, E. A. Teens and Cyberbullying 2022. Pew Research Center. Report. Source of the 46 percent figure, the 32 percent name-calling and 22 percent false-rumor breakdown, and the finding that roughly four in ten teens think teachers are doing a good or excellent job addressing online harassment. Survey of 1,316 U.S. teens ages 13 to 17.
    • Delaware Department of Education. E-Rate and CIPA guidance for internet safety policies. Guidance document. Source of the quoted federal requirement to educate minors about appropriate online behavior, social networking and chat room interactions, and cyberbullying awareness and response, under the Protecting Children in the 21st Century Act. A state agency’s summary of the federal rule, not the statute itself; the requirement has applied since the 2012 E-rate funding year.
    • Media Literacy Now. U.S. Media Literacy Policy & Impact Report. Report page. Source of the statement that 25 states have media literacy laws on their books, covering legislation through October 30, 2025. The report’s own entries mix binding statutes with resolutions and encouragements and it states no single total, and it does not evaluate whether states implemented or funded what they passed.
    • Illinois General Assembly. Public Act 102-0055. Text. Source of the five required components of the Illinois high school media literacy unit and the requirement that it apply to every public high school.
    • Modirrousta-Galian, A., & Higham, P. A. (2023). Gamified Inoculation Interventions Do Not Improve Discrimination Between True and Fake News. Journal of Experimental Psychology: General. Accepted manuscript. Source of the reanalysis of 13 experiments and the finding that the games elicited more “false” responses to all news items rather than improving discrimination. The samples are adults, not school students.
    • Kirschner, P. A., & De Bruyckere, P. (2017). The Myths of the Digital Native and the Multitasker. Teaching and Teacher Education, 67. Full text. Source of the conclusion that there is no such thing as a digital native and that people cannot multitask when thinking is involved. A narrative review rather than new data; most of the studies it draws on used university students.
    • Jones, L. M., Mitchell, K. J., & Walsh, W. A. (2014). A Systematic Review of Effective Youth Prevention Education: Implications for Internet Safety Education. Crimes Against Children Research Center, University of New Hampshire. Full text. Source of the finding that internet safety education lacked rigorous evaluation and of the review’s method of borrowing from 31 meta-analyses in other prevention fields. It did not itself evaluate internet safety programs.
    • Curriculum pages read for the comparison table: Civic Online Reasoning (30 free lessons) and its middle school module; Common Sense Education, which states the curriculum will be archived in December 2026 and points to its Digital Literacy & Well-Being replacement; Be Internet Awesome, which states its materials require no personal information or login; and NetSmartz, whose downloads prompt users to log in or register. Lesson counts per grade for Common Sense are counted from the page listing, which states no total.
    • Midlam, S. Mandatory digital citizenship lessons fail to engage. The Iliad, Poway High School student news. Editorial. Source of the student verdict on 20-minute mandated advisory lessons. One student’s opinion piece, not evidence — included because it is a primary account of how the compliance version lands with the people sitting through it.

    About Clay Shumate

    Clay Shumate is a certified secondary Social Studies teacher in the public schools of West Alabama, with seven years of classroom experience, a B.A. in History, and an M.Ed. in Secondary Education. He writes about project-based learning, student responsibility, respect, and practical ways to hold young people to a higher standard while giving them room to learn from mistakes. He is a member of the Society of Professional Journalists and writes to its Code of Ethics; this site’s editorial standards and corrections policy are published in full. More about Clay.

  • Interleaving: Why Mixing Up Practice Beats Doing Twenty of the Same Problem

    Interleaving: Why Mixing Up Practice Beats Doing Twenty of the Same Problem

    By Clay Shumate

    Interleaving means mixing different problem types inside one practice set instead of grouping them. In a randomized trial of 787 seventh graders, interleaved math practice produced 61 percent on a later test against 38 percent for blocked practice. It feels worse while students are doing it, which is why most sets go back to blocked.

    Key Takeaways

    • The strongest evidence is in secondary math, and it is strong. Rohrer, Dedrick, Hartwig and Cheung (2020) randomized 54 seventh-grade math classes across five schools and reported 61 percent against 38 percent on an unannounced test, d = 0.83, with a positive effect for every one of the 15 teachers.
    • Interleaving does not work on everything. The meta-analysis of 238 effect sizes found a small positive effect for math tasks, nothing reliable for expository text, and an effect pointing the other way for vocabulary words.
    • The practice-day score is supposed to drop. In a grades 9–12 science study the interleaved quizzes ran 55 percent against 59 percent for blocked quizzes, and the interleaved group still came out ahead a month later.
    • Mix similar things, not unrelated things. Mixing two methods students keep confusing is the intervention. Mixing a science unit with a foreign-language unit is not, and has not been found to help.
    • Nobody has tested it on students who are already behind. The flagship trial ran in a course the district labeled Honors. That is the single biggest gap between the research and most classrooms.

    What is interleaving?

    Interleaving is practicing several different kinds of problem in one mixed set, instead of doing all of one kind and then all of the next. The opposite is called blocked practice, and it is what almost every textbook section and almost every worksheet does by default.

    Comparison of blocked and interleaved practice: blocked groups all ten slope problems together then all ten graph problems and produces high worksheet scores, interleaved shuffles the same twenty problems so students must identify the problem type before choosing a method, nothing about the problems changes except the order, and practice-day accuracy drops.

    The distinction matters because of what each one asks a student to do. On a blocked page, the student solves the first problem, works out which method applies, and then reuses that method nineteen more times without ever asking the question again. On a mixed page, every problem starts with the same question a test asks: what kind of problem is this? That is the skill that goes missing when a kid can do every problem in the chapter and then freezes on the unit test.

    Worth saying early, because the two get swapped constantly: interleaving is about order within a set. Spacing is about gaps between sessions. They often travel together and they are not the same move.

    Does interleaving actually work in a real classroom?

    Yes, in secondary math, in real classrooms, with large effects. Four classroom studies have tested interleaving with students in grades 7 through 12, and all four found the mixed sets beat the blocked ones on a delayed test.

    That is unusually good evidence for a practice that costs nothing to adopt.

    StudyWhoDesignDelayed test resultEffect
    Rohrer, Dedrick, Hartwig & Cheung (2020)787 seventh graders, 54 classes, 5 schoolsRandomized by class, 15 teachers, test 33 days after review61% vs 38%d = 0.83
    Rohrer, Dedrick & Burgess (2014)140 seventh graders, 8 classesCrossover, test two weeks after last assignment72% vs 38%d = 1.05
    Rohrer, Dedrick & Stershic (2015)126 seventh graders, 9 classesWithin-subject, delay randomized74% vs 42% at 30 daysd = 0.79
    Sana & Yan (2022)155 students, grades 9–12, 8 science classesWithin-subject quizzing, test one month later63% vs 54% (no-quiz control 47%)d = 0.35

    The 2020 trial is the one to know. Fifty-four classes, fifteen teachers, five middle schools, assigned at the class level, with eight worksheets spread across the semester and an unannounced test a month after the last one. The authors report a positive effect for every single teacher in the study, which is rare and is the part that should interest a department.

    Sana and Yan (2022) is the only one outside math: biology, chemistry, physics and general science classes in one Canadian public high school, using weekly ten-minute quizzes. Smaller effect, one school, one study. It is the best evidence that this travels beyond a math worksheet, and it is one building.

    One limitation nobody puts on the slide: three of those four classroom studies come from the same research group. That is not a reason to dismiss them — the 2020 trial is a proper randomized design across five schools and fifteen teachers — but a district deciding policy should know that the secondary-math evidence base is one research program and one high school science study, not a broad literature.

    Why does interleaving feel like it is not working?

    Because the mixed set is harder to do, and the payoff does not arrive until weeks later. Students score lower on the practice itself, feel less confident, and tell you they liked the old worksheets better. Every one of those signals points the wrong way.

    Sana and Yan measured it: on the practice quizzes themselves, the interleaved version ran 55 percent against 59 percent for the blocked version. Same students, same content, four points worse in the moment — and nine points better a month out.

    The classic demonstration of the illusion is Kornell and Bjork (2008), where 78 percent of participants performed better after interleaved study and 78 percent said the blocked version was as good or better. Be careful quoting that figure at a faculty meeting, though: those were UCLA undergraduates learning to recognize paintings by twelve artists, not teenagers doing math. The illusion is real; that specific number is not a classroom finding.

    This is the same shape as the argument for letting students wrestle with something before you rescue them. Difficulty that produces learning and difficulty that produces nothing feel identical from the front of the room on Tuesday. The only way to tell them apart is to look at a delayed measure, which is exactly the thing a weekly grading cycle never shows you.

    What does interleaving not work for?

    Vocabulary words, long expository text, and anything mixed with something unrelated. The meta-analysis that pooled 238 effect sizes from 59 papers found the overall benefit was g = 0.42, but that the number moves enormously depending on what is being learned.

    Chart of what the evidence says to mix: confusable problem types and visual categories show benefits, vocabulary words and expository text do not, and mixing unrelated subjects such as science terms with foreign-language vocabulary has not been found to help.

    Paintings came out best at g = 0.67. Mathematical tasks came out at g = 0.34, which the authors call small. Expository texts and tastes produced nonsignificant effects. And for words, the effect ran the other way: g = −0.39, an advantage for blocked study. If a reading coach tells your department to interleave the vocabulary list, that is the number to bring.

    Mixing unrelated material does not help either. The 2021 systematic review by Firth, Rivers and Boyle reports that studies interleaving unrelated items have not found a benefit — one of them mixed science terms with foreign-language vocabulary. The point is not variety for its own sake. The point is forcing a choice between options a student confuses.

    Nobody has tested interleaving whole lessons or whole units, either. Firth and colleagues say so plainly. A department that decides to interleave the curriculum itself is extrapolating from twelve-item worksheets.

    Why do the classroom trials and the meta-analyses disagree?

    They measure different things. The classroom math trials of interleaving report effects between d = 0.35 and d = 1.05, while meta-analyses pooling mostly laboratory studies land near 0.42. Both numbers are accurate. Neither one describes the other.

    This site has published the lower number itself: the rundown of which study techniques actually work lists interleaved practice at 0.47, inside the low band of the meta-analysis it comes from. That article is not wrong and this one is not either, which is worth explaining rather than leaving a reader to find the gap.

    Three reasons for the gap, and they all run the same direction. First, the pooled literature is mostly not school. Firth and colleagues found that across their 26 included studies, none used high school participants — three used adult online samples and the rest were undergraduates. Second, it is mostly not math: the meta-analysis authors say visual materials dominate the field, and the math estimate rests on a thin slice. Third, the meta-analysis of learning techniques behind that 0.47 figure included nothing published after 2014, which means the two biggest classroom trials in this article are not in it.

    So the honest summary is narrower than either headline. Averaged across paintings, word lists and laboratory tasks, interleaving is a moderate technique with real exceptions. Applied to confusable problem types on a secondary math worksheet, it is one of the largest effects anybody has measured in a real classroom. Those are not the same claim and you should not let a slide deck collapse them.

    How do I build an interleaved practice set?

    Keep the problems you already assign and change the order they appear in. Agarwal and Agostinelli, writing in American Educator, put it in one line: there is no need to change the problems, simply mix up what you assign. The work is in choosing what to mix.

    Five steps for building an interleaved practice set: block first until students can do one problem type, mix the confusable pairs rather than unrelated units, keep the problems and change the order with two current and three earlier, expect practice accuracy to fall, and delay the test.
    1. Block first, then mix. Deans for Impact recommends blocked practice at the start and a move to interleaved, spaced practice once initial fluency is achieved. Mixing on the day you introduce a method is not the intervention these studies ran.
    2. Mix the pair students keep swapping. Agarwal and Agostinelli are explicit that the key is mixing up similar ideas. In practice that means the two formulas that look alike, the two graph types with the same axes, the two document types that call for different moves.
    3. Keep the set small. The American Educator piece describes a five-problem homework set from a middle school math classroom: two problems on this week’s content and three pulled from last week, last month and last year. That is not extra work, it is the same work in a different order. The cheapest source of the three old problems is the set you already assigned three weeks ago.
    4. Strip the labels off the page. If the heading says Section 4.3, the student knows which method to use before reading the problem and the mixing buys you nothing. That is also the one-second look-for if somebody is coaching this: can a student tell which method to use from the layout, before reading a word of the problem?
    5. Tell the students what you are doing and why. They are the ones whose scores dip, and if nobody explains the design they will reasonably conclude they are getting worse at math. Say it plainly before the first mixed set: this page is built to be harder, the first-attempt score is not the measure, and the comparison that counts is the unit test.
    6. Give answers back fast. Rohrer and colleagues list feedback as one of the conditions their trial cannot rule in or out, and their students got worked solutions. Until somebody tests the version without feedback, assume mixed practice without it is guessing practice.
    7. Judge it on a delayed, unannounced check. The 2020 trial’s gap appeared 33 days after the review assignment. A Friday quiz on Friday’s content will show you nothing, which is also the difference between a check used to steer teaching and a check used to record a grade.

    Does interleaving confuse students who are already behind?

    Nobody knows, and anyone who tells you otherwise is going past the evidence. This is the first question every teacher asks and it is the question the research has not answered.

    Start with the flagship trial: the students were in a course the district called Honors Advanced Grade 7 Mathematics. The authors also note that interleaving might be less effective, or simply too hard, without some prior blocked practice. In a German study of 236 third graders — younger than this site’s grade band, so read it as a signal and not a finding — the authors raise the same concern directly, that the cognitive-load demands of interleaving may matter for weaker students.

    What there is evidence for is that the benefit is not universal. Yan and Sana (2021) found about 30 percent of participants showed no interleaving benefit on average — but also that who benefited on one set of materials barely predicted who benefited on another. There is no evidence of a stable type of student who needs the blocked version, which cuts against sorting kids into two worksheets.

    The practical answer is the sequencing one. Block until a student can actually do the thing, then mix. The trials interleaved review of material students had already been taught, not first exposure, and that distinction is doing a lot of work.

    Is interleaving the same as spaced practice?

    No. Spacing is about putting time between practice sessions. Interleaving is about mixing problem types inside a session. A mixed set naturally revisits older material, so the two overlap in a real worksheet, but they are separate decisions with separate evidence.

    They also do not simply stack. Firth and colleagues report that adding spacing on top of interleaving was not additive, and in places reduced the benefit. For a classroom that is a relief rather than a problem: you do not have to engineer both at once. Pull three old problems into this week’s set and you have already got most of both, which is the same logic behind a do-now that asks about last week instead of last night.

    Doesn’t the federal practice guide already recommend interleaving?

    No. The federal practice guide recommends interleaving worked example solutions with problem-solving exercises, which means alternating a worked-out solution with a problem the student tries alone. That is a different intervention from mixing problem types inside one set.

    The Institute of Education Sciences guide on organizing instruction and study rates that recommendation at a moderate evidence level, and its first recommendation, to space learning over time, is also moderate. Neither one is about problem sets that mix topic types, and the confusion between the two meanings of the word is everywhere.

    This matters because “the federal government recommends interleaving” is the kind of sentence that ends an argument in a PD session, and in this case it ends it about the wrong practice. The real federal-grade support for mixing problem types is the randomized trial, not the guide.

    Clay’s take: this is the cheapest thing on the list, and the easiest to quit

    Most research-backed practices cost money, training, or a schedule change. This one costs the five minutes it takes to reorder a problem set, which makes it the rare recommendation a department can adopt on Monday without asking anybody for anything.

    Keep the transition out of the gradebook. If mixed sets are scored for accuracy while students are still learning to read the page, a practice designed to raise learning has been converted into a grade penalty for the kids who are slowest to adjust — and the email from a parent asking why their child’s homework average fell in October is both inevitable and fair. Score completion, or score nothing, until the first unit test tells you something.

    The catch is that everything about the first two weeks tells you to stop. Scores dip. Kids complain, and their complaint is accurate — the work is harder. Pull a formative check on Thursday and the numbers back the complaint. That is the design, and it is why I would not hand this to a team without also telling them what the first fortnight looks like. A practice that gets abandoned on week two is worse than one nobody tried, because now everyone has evidence it does not work.

    Two honest limits worth saying out loud to students. Rohrer and colleagues state that interleaved practice probably takes more time, and that they did not measure time on task — so part of the gain may simply be that a mixed set is slower work. And this is practice design, not a difficulty setting: a mixed worksheet of recall questions is still a worksheet of recall questions, and how hard the thinking is is a separate decision you still have to make.

    Where to start: pick the two topics your students most reliably mix up on the unit test. Build one five-problem set a week that includes both, strip the section headings, hand back answers the next day, and do not look at the results until the unit test. If it works, the place it will show up is the question your students have always gotten wrong for the same reason — not the practice page. For review days, the same logic applies to the formats you use to review: mixing the question types is what makes the game a review instead of a round of trivia.

    Frequently Asked Questions

    What is interleaved practice?

    Interleaved practice means mixing different problem types inside one practice set instead of doing all of one kind and then all of the next. The opposite, blocked practice, is what most textbook sections and worksheets do by default. The point of mixing is that every problem forces the student to decide which method applies before using it.

    Is interleaving better than blocked practice?

    On a delayed test, yes, in secondary math. A randomized trial across 54 seventh-grade classes found 61 percent for interleaved practice against 38 percent for blocked, with a positive effect for every teacher in the study. On the practice itself blocked usually looks better, which is why the comparison has to be made weeks later rather than the same day.

    Does interleaving work for subjects other than math?

    Partly. One study in grades 9 through 12 science classrooms found interleaved quizzes beat blocked quizzes a month later, 63 percent against 54. Outside that, the meta-analytic picture is mixed: visual categories benefit, expository text shows no reliable effect, and vocabulary words show an advantage for blocked study instead.

    Why do my students score worse on mixed practice?

    Because that is how it works. Choosing a method is harder than reusing one, so accuracy during practice falls. In one high school science study the interleaved quizzes ran 55 percent against 59 percent for the blocked ones, and the same students were nine points ahead on a test a month later. The practice-day score is not the measure.

    Does interleaving confuse students who are already behind?

    Nobody has tested it properly. The flagship trial ran in a course the district labeled Honors, and the authors note interleaving may be too difficult without some prior blocked practice. The safe sequence is to block until a student can do the method, then mix for review. There is no evidence supporting a separate blocked worksheet for struggling students.

    Is interleaving the same as spaced practice?

    No. Spacing puts time between practice sessions; interleaving mixes problem types inside one session. They overlap in a real worksheet, because a mixed set naturally revisits older material. One systematic review found that adding spacing on top of interleaving was not additive and sometimes reduced the benefit, so there is no need to engineer both at once.

    How many old problems should go in a practice set?

    A five-problem set with two problems from this week and three from earlier weeks is the pattern described in American Educator by a middle school math teacher. That keeps the workload identical to a standard homework set while making most of the page cumulative. The cheapest source for the older three is the assignment you gave three weeks ago.

    What should I tell a parent whose child’s practice scores dropped?

    That the practice set changed, not the student. Mixed practice is designed to be harder in the moment and the research measures its payoff weeks later, on a delayed test. The honest version also includes what you are doing about it: scoring the transition for completion rather than accuracy, so a deliberate difficulty does not turn into a grade penalty.

    Sources

    • Rohrer, D., Dedrick, R. F., Hartwig, M. K., & Cheung, C.-N. (2020). A Randomized Controlled Trial of Interleaved Mathematics Practice. Journal of Educational Psychology, 112(1), 40–52. Accepted manuscript. Source of the 61 against 38 percent result, d = 0.83, the 787-student analyzed sample across 54 classes and five schools, the 33-day delay, the positive effect for each of 15 teachers, and the authors’ own caveats. The course was described by the district as Honors Advanced Grade 7 Mathematics, and the authors state that interleaved practice probably takes more time and that time on task was not measured. Read via the author manuscript because the journal version is paywalled.
    • Rohrer, D., Dedrick, R. F., & Burgess, K. (2014). The Benefit of Interleaved Mathematics Practice Is Not Limited to Superficially Similar Kinds of Problems. Psychonomic Bulletin & Review. Full text. Source of the 72 against 38 percent result and d = 1.05 with 140 seventh graders, tested two weeks after the final assignment.
    • Rohrer, D., Dedrick, R. F., & Stershic, S. (2015). Interleaved Practice Improves Mathematics Learning. Journal of Educational Psychology. Full text. Source of the 74 against 42 percent figure at a 30-day delay with 126 seventh graders. The copy read is the advance-online version, which is headed with a different year and volume than the final citation; the figures above are from the text itself.
    • Sana, F., & Yan, V. X. (2022). Interleaving Retrieval Practice Promotes Science Learning. Psychological Science, 33(5), 782–788. Full text. Source of the 63, 54 and 47 percent final-test figures, d = 0.35, and the 55 against 59 percent practice-quiz figures. 155 students in grades 9–12 at one Canadian public high school; the authors note the study was not preregistered, the outcome measure was researcher-designed, and generalization to standardized exams is untested.
    • Brunmair, M., & Richter, T. (2019). Similarity Matters: A Meta-Analysis of Interleaved Learning and Its Moderators. Psychological Bulletin. DOI 10.1037/bul0000209. Accepted manuscript. Source of g = 0.42 overall across 59 papers, 158 samples and 238 effect sizes; paintings 0.67; mathematical tasks 0.34; words −0.39 favoring blocking; nonsignificant results for expository texts and tastes. The authors state that visual materials dominate the field and that most studies examined college students.
    • Firth, J., Rivers, I., & Boyle, J. (2021). A Systematic Review of Interleaving as a Concentration and Learning Strategy. Review of Education. Full text. Source of the statement that none of the 26 included studies used high school participants, the finding that interleaving unrelated items has not been beneficial, the absence of research on interleaving whole lessons, and the finding that adding spacing to interleaving was not additive.
    • Kornell, N., & Bjork, R. A. (2008). Learning Concepts and Categories: Is Spacing the “Enemy of Induction”? Psychological Science, 19(6). Full text. Source of the 78 percent figures. Participants were UCLA undergraduates learning to recognize paintings by twelve artists, not secondary students, and the paper’s “spaced” condition is what this article calls interleaved.
    • Yan, V. X., & Sana, F. (2021). The Robustness of the Interleaving Benefit. Journal of Applied Research in Memory and Cognition, 10. Full text. Source of the finding that about 30 percent of participants showed no interleaving benefit and that benefit on one material set barely predicted benefit on another. Undergraduate samples, visual categories only.
    • Pashler, H., Bain, P., Bottge, B., Graesser, A., Koedinger, K., McDaniel, M., & Metcalfe, J. (2007). Organizing Instruction and Study to Improve Student Learning. NCER 2007-2004, Institute of Education Sciences. Full text. Source of Recommendation 2, “interleave worked example solutions with problem-solving exercises,” at a moderate evidence level, and Recommendation 1, space learning over time. This guide does not evaluate mixing problem types, which is the practice this article is about.
    • Nemeth, L., Werker, K., Arend, J., Vogel, S., & Lipowsky, F. (2019). Interleaved Learning in Elementary School Mathematics. Frontiers in Psychology, 10:86. Full text. Source of the authors’ concern that the cognitive-load demands of interleaving may matter for weaker students. 236 German third graders, outside this site’s grade band, and the study combined interleaving with comparison prompts so the two cannot be separated. The paper carries a published corrigendum correcting its description of the interleaving meta-analysis.
    • Agarwal, P. K., & Agostinelli, A. (2020). Interleaving in Math: A Research-Based Strategy to Boost Learning. American Educator, Spring 2020. Article. Source of the instruction to mix similar ideas without changing the problems and of the five-problem homework pattern, which is described there as one middle school math teacher’s own classroom routine. A practitioner article, not a study.
    • Deans for Impact. The Science of Learning, second edition. Full text. Source of the recommendation that practice move to interleaved, spaced practice once initial fluency is achieved. A summary document for teacher preparation, not primary research.

    About Clay Shumate

    Clay Shumate is a certified secondary Social Studies teacher in the public schools of West Alabama, with seven years of classroom experience, a B.A. in History, and an M.Ed. in Secondary Education. He writes about project-based learning, student responsibility, respect, and practical ways to hold young people to a higher standard while giving them room to learn from mistakes. He is a member of the Society of Professional Journalists and writes to its Code of Ethics; this site’s editorial standards and corrections policy are published in full. More about Clay.

Teacher Emergency Toolkit — practical resources, real classroom support. Shop on TPT.Teacher Emergency Toolkit — practical resources, real classroom support. Shop on TPT.