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- The capacity figure teachers are usually given is out of date.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.

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.

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.

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.
- 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.
- 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.
- 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.
- 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.
- 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.



















