The Expertise Reversal Effect: When Help Starts to Hurt

Every course designer has met the learner who already knows half the material. They sit through a worked example of something they’ve done a hundred times, click through the narrated walkthrough, and either skim or leave. It’s tempting to treat that as a motivation problem. The research says it’s a design problem with a name: the expertise reversal effect.

The short version is that the support which makes a lesson work for a beginner can make it work worse for someone who already knows the basics. This entry explains what the research found, including a large 2025 meta-analysis that most explanations online haven’t caught up with, then shows how to build a course that serves both kinds of learner, and what to check in your course platform.

What the research says

The effect grew out of cognitive load theory, the account of how a small, forgetful working memory limits what learners can take in at once. Slava Kalyuga, John Sweller and colleagues named it and set out the early evidence in a 2003 overview (Kalyuga, Ayres, Chandler & Sweller, 2003).

The definition

Kalyuga’s fuller 2007 review gives the plainest definition: the effect is “the reversal in the relative effectiveness of instructional methods as levels of learner knowledge in a domain change” (Kalyuga, 2007). A method that beats another for novices can lose that advantage, or fall behind, for learners who know more.

Why help can hurt

The explanation is about redundancy. A beginner has no stored knowledge to guide them through a new task, so instruction has to supply that guidance from outside: every step written out, every part of the diagram explained. An experienced learner already carries that guidance in long-term memory, in what psychologists call schemas (connected bundles of knowledge that can be handled as one unit). Give them the external version as well and they now have two sources saying the same thing.

They can’t simply ignore the redundant one. Kalyuga’s review describes the extra work as cross-referencing: learners spend working memory checking the external guidance against what they already know, and that effort doesn’t add to learning (Kalyuga, 2007). In cognitive load terms, the same explanation that reduces load for a novice adds extraneous load for an expert.

Expertise reversal effect examples from the research

Several of the early studies used technical trainees, the kind of learners many workplace and vocational courses serve. Kalyuga’s 2007 review summarises them (Kalyuga, 2007):

• Diagrams with text. Trainees learning from electrical wiring diagrams did best when the explanations were embedded in the diagram. As they gained knowledge, the text that had been essential became redundant.
• Narration over diagrams. With detailed narrated explanations of a type of mechanical engineering diagram, the narration helped at first. As learners became more knowledgeable, a diagram-only group outperformed the group that heard the narration.
• Worked examples versus problem solving. Worked examples (problems shown with every step solved) beat problem-solving practice for trainees learning to program logic controllers, but the advantage disappeared as trainees gained experience.

A later study took the question into a computer-based module, closer to an online course. Reisslein and colleagues taught 185 introductory engineering students to analyse electrical circuits using three example-based sequences. On near-transfer questions, students with low prior knowledge did best when each worked example came before a practice problem, while students with high prior knowledge did best when they tried the problem first and saw the example afterwards. The pattern didn’t show up on far-transfer questions, which is a useful reminder that individual studies are noisy (Reisslein, Atkinson, Seeling & Reisslein, 2006).

What the 2025 meta-analysis found

Individual experiments only take you so far. In 2025, Tetzlaff, Simonsmeier, Peters and Brod pooled 60 experimental studies, published between 1992 and 2024, with 176 effect sizes and 5,924 participants (Tetzlaff et al., 2025). Each study had to randomly assign learners to instruction with more or less assistance, and to measure their prior knowledge. The authors defined assistance broadly, as any change to the instruction or task intended to make processing easier.

The headline results:

• Novices learned more with high assistance. For learners with low prior knowledge, high assistance beat low assistance by d = 0.505 (95% confidence interval 0.260 to 0.750).
• Experienced learners learned more with low assistance. For learners with high prior knowledge, the direction reversed: d = −0.428 (95% confidence interval −0.647 to −0.209).
• The effect isn’t symmetrical. Giving novices support mattered more than taking it away from experts.

The authors conclude that the effect is robust across a wide range of contexts. They also note where it’s less certain: there were very few studies with primary-school children, and the evidence was less clear for humanities and language learning than for other fields. For novices, the effect was larger in higher education and secondary school than in the small primary-school sample (Tetzlaff et al., 2025).

How strong is the evidence?

Strong, for the overall pattern: a meta-analysis of randomised experiments, built on three decades of studies, finds it on both sides. Three limits matter for course design.

First, the results varied a great deal from study to study (heterogeneity was around 90% on both sides), and the authors say the moderators they tested don’t explain all of it. Second, “novice” and “expert” are relative: most studies split learners into higher and lower prior knowledge on one topic, not beginners and professionals. Third, the asymmetry means the cost of over-supporting an expert is real but smaller than the cost of under-supporting a novice. Popular summaries that say worked examples “harm experts” overstate it. The fairer reading is that support should be generous at the start and should not be compulsory forever.

How to apply it in a course

The design principle follows directly from the meta-analysis: give novices the support they need, then give everyone else a way past it. Here’s how we’d build that into a self-paced course, using one invented example: a four-module course teaching office staff to build and check budget spreadsheets. Some learners have never written a formula; others use spreadsheets every day but want to tidy up their habits.

1. Build the beginner’s version first

Because helping novices matters more than freeing experts, start with the fully supported path. Module two, on lookup formulas, opens with two fully worked examples: a real budget sheet, each step of the formula written out and narrated, with the explanation placed next to the cell it explains. Don’t strip this back to suit the experienced learners. You’ll give them their own route in step 3.

2. Check what learners already know, quickly

You can’t adapt to prior knowledge you haven’t measured, and a long pre-test at the start of every module would drive people away. Kalyuga and Sweller’s first-step method is a faster option: learners see a task for a short time and write down the first step they’d take towards solving it. A beginner writes the first move; someone more expert may skip straight to a later step or the answer. Validation studies found scores correlated with traditional tests at .72 to .92, while cutting testing time by factors of 2.8 to 4.9 (Kalyuga, 2005). Those were small, early studies, so treat the method as a well-founded idea, not a finished instrument.

In a course platform, a short multiple-choice version works: show three partly completed formulas and ask “What would you do next?”, with options that range from the first step to the finished result. Our guide to writing multiple-choice questions covers how to write distractors that reveal what learners actually know.

3. Give experienced learners a faster route

Use the check to open a different path, not to lock people out. Learners who get all three questions right in module two skip the worked examples and go straight to a practice set of three lookup problems. Anyone who struggles there can step back into the worked examples. Everyone ends at the same module assessment, so the course still confirms the skill.

4. Make explanations available, not compulsory

Much of the redundant support in online courses is narration and on-screen explanation that every learner has to sit through. Where you can, move detail into optional layers: a “Show the steps” link under a solved example, a hint button on a practice question, a transcript instead of narration that repeats the slide. Novices open them; experienced learners don’t have to.

5. Fade the support as the course goes on

Within a module, learners become more expert as they go, so the support should change too. Renkl and Atkinson’s fading approach starts with fully worked examples, then removes worked steps one at a time until learners are solving whole problems on their own (Renkl & Atkinson, 2003). In the spreadsheet course, that means two fully worked lookups, then one with the last step left for the learner, then one with the last two steps left, then a problem from scratch. Kalyuga’s review recommends exactly this direction: detailed guidance early, removed as learners progress (Kalyuga, 2007).

6. Watch for the signs

Your analytics will tell you where the balance is off. Experienced learners skipping or fast-forwarding the same worked example, or high scores on a module quiz alongside complaints that the module is slow, suggest support that has become redundant. High drop-off at the first unsupported problem suggests you faded too fast. Formative assessment throughout the course gives you the data to judge both.

An example of how we’ve designed for a mixed cohort at Avondale University, is creating case studies that apply to different levels of understanding or contexts. In a postgraduate Governance course, for instance, we knew that a significant number of our students were leaders across the South Pacific, which means working not only in developed and mature business environments like Sydney, but working in Fiji, The Solomon Islands and Vanuatu, where there are a number of cultural as well as practical differences to how business is conducted. So we developed a ‘choose your case study’ segment across the whole business course, where we created case studies from a Pacific Islands lens, and another case study from a traditional western business lens, allowing learners to choose which context applies to them.

This sits naturally in the design phase of the ADDIE model, when you analyse who your learners are and decide how lessons will differ for them.

How the platforms support it

Adapting to prior knowledge depends on a few platform features: a quick check that can change what a learner sees next, the ability to make content optional, and practice questions with feedback. We haven’t yet published our reviews of Thinkific, LearnWorlds, Teachable and Kajabi, and features change often, so check these in your platform’s current documentation, on the plan you’re paying for:

• Branching on a quiz result. Can a short quiz score send learners to different lessons, or unlock a skip route?
• Optional lessons. Can you mark a lesson as optional, so experienced learners can skip it without breaking their progress or their certificate?
• Free navigation. Can learners move ahead within a module, or is everything strictly sequential? Strict sequence suits novices and frustrates experts.
• Hints and explanations on questions. Can practice questions show an explanation after an answer, or a hint before it?
• Lesson-level analytics. Can you see where learners skip, rewind or drop off, so you can spot redundant or missing support?

If your platform can’t branch, a simpler version still works: put the check at the top of the module, and tell learners in plain words which lesson to jump to if they got every question right.

Further reading

• Kalyuga, Ayres, Chandler & Sweller (2003), “The expertise reversal effect”. The paper that named the effect and gathered the early evidence. Nine pages.
• Kalyuga (2007), “Expertise reversal effect and its implications for learner-tailored instruction”. The fullest review, including the first-step method and early adaptive-learning experiments.
• Tetzlaff, Simonsmeier, Peters & Brod (2025), “A cornerstone of adaptivity”. The meta-analysis. Open access, and the best single source on how big the effect is and where it holds.
• Renkl & Atkinson (2003), “Structuring the transition from example study to problem solving”. The case for fading worked examples, step by step.
• Kalyuga (2018), “Expertise reversal effect and its instructional implications”. A short version for teachers, from the researcher himself, in the Chartered College of Teaching’s journal.

For the rest of the principles behind good course design, browse our research library.