The Learning Pyramid: Where the Famous Percentages Came From

You’ve probably seen it on a slide in a train-the-trainer course, or in a client’s brief: a triangle with “Lecture: 5%” at the narrow top and “Teach others: 90%” across the wide base. It’s called the learning pyramid, or sometimes the cone of learning, and it promises a simple rule for course design. Lectures and reading are weak, doing and teaching are strong, and here are the numbers to prove it.

The short answer is that the numbers have no source. Nobody has found a study that produced them, the organisation most often credited can’t locate one, and the researcher whose name is most often attached never published any. The pyramid’s underlying instinct, that learners should do something with what they learn rather than only listen, is much closer to the evidence. But the real effects are smaller, depend on conditions, and come from specific techniques you can build into a course.

This entry covers where the percentages came from, why they can’t be right, what the research on active learning actually found, and what to build instead.

What the research says

The version everyone shares

The most widely shared version is credited to the NTL Institute in the United States. It puts lectures at 5% and reading at 10%, followed by audiovisual material, demonstration and discussion groups, then practice by doing at 75% and teaching others at 90%. Other versions shuffle the categories and change the figures. In a review of the medical education literature, Ken Masters found the value given for reading ranged from 5% to 40% depending on who was citing it (Masters, 2013).

Masters searched the medical education literature from 1946 to 2012 and found 43 articles that mentioned the pyramid. Only one questioned where it came from, and even that one went on to use the percentages. When NTL was asked for its evidence, its reply, which Masters quotes, said the pyramid was based on its own research but that it had never been able to locate that research.

Older than anyone thought

The most thorough history comes from Kåre Letrud and Sigbjørn Hernes, who searched digital libraries for early versions (Letrud & Hernes, 2018). They found the idea long before any of its supposed sources. The earliest version they located, from 1852, ranks seeing above reading and hearing with no numbers at all. The first quantified one they found is a 1906 article in a Sunday school journal: “We remember one tenth of what we hear, five tenths of what we see.” Fractions in tenths were common from then on, and the authors suggest the later 10% steps are those fractions turned into percentages.

A version from 1967 by D. G. Treichler, then at Mobil Oil, became the one most often cited. It said people generally remember 10% of what they read, 20% of what they hear, 30% of what they see, 50% of what they hear and see, 70% of what they say, and 90% of what they say as they do a thing (Letrud, 2020). It gave no study.

So the percentages didn’t come from an experiment that was later lost. They came from a saying that gained numbers as it was repeated.

What Edgar Dale actually drew

The name most often printed under the pyramid is Edgar Dale’s. Dale, an American educator, published a “Cone of Experience” in his 1946 book on audiovisual teaching, and later editions updated it. According to Masters, the cone classifies teaching materials by how concrete or abstract they are, from direct, hands-on experience at the base to words and symbols at the tip. Dale called it a visual analogy and “only a model”, and Masters points out it has no numbers or percentages and no claim about how much people retain.

Masters also notes that the cone doesn’t treat the base as better than the tip. Dale wrote that life “cannot, of course, be lived exclusively on the direct, concrete, sensory level”, and learners need to move across all the levels, from concrete experience to abstract ideas and back. The retention figures were merged with his diagram later, by other people, which is how a cone about teaching materials became a pyramid about memory.

Why the numbers can’t be right

Even without the history, the percentages fail on their own terms. A retention figure means nothing until you say what was learned, by whom, how it was tested and after how long. The pyramid says none of that. Letrud and Hernes also point out that it treats memory as a passive store you fill through different senses, which is far from how memory research describes learning: people build connections between new information and what they already know.

A classic experiment shows how much the delay alone changes the answer. Henry Roediger and Jeffrey Karpicke had 120 university students read two short science passages. For one passage, students read it again; for the other, they wrote down everything they could remember, without feedback. Then they took a final recall test five minutes, two days or one week later (Roediger & Karpicke, 2006).

Re-reading won at 5 minutes; a recall test won by 14 points at 2 days and 1 week

Proportion of idea units recalled from a prose passage that students had either restudied or recalled once in a test without feedback, by delay before the final test (Experiment 1, 120 undergraduates) · % of idea units recalled

Studied twiceStudied, then tested

Re-reading won at five minutes. Two days and a week later, the students who had tested themselves remembered more. The same passages and the same students produced different winners depending on when you measured. A chart that gives each method one fixed retention rate can’t capture that.

What active learning does do

The pyramid survives partly because its ranking points somewhere reasonable. There is good evidence that learners do better when they do something with the material.

The largest summary is a meta-analysis by Scott Freeman and colleagues of 225 studies comparing active learning with traditional lecturing in university science, engineering and maths courses (Freeman et al., 2014). Active learning meant activities or discussion in class rather than listening alone. On exams and concept inventories, scores were 0.47 standard deviations higher under active learning, an average improvement of about 6%. Failure rates were 21.8% with active learning and 33.8% with lecturing.

That’s a solid effect, but notice its size. Nothing in the research looks like the pyramid’s eighteen-fold gap between 5% and 90%. The studies were also of face-to-face university classes, not self-paced online courses, so treat the size as a guide rather than a promise.

Teaching others, the pyramid’s top tier, also has real support. A meta-analysis by Keiichi Kobayashi pooled 28 studies of learning by teaching (Kobayashi, 2019). Preparing to teach had a small benefit (g = 0.35), and actually teaching after preparing had a moderate one (g = 0.56). Interactive teaching, where the learner answers questions, helped more than non-interactive teaching such as explaining to a camera. In a set of experiments with a short science lesson, Logan Fiorella found that preparing to teach helped on an immediate test, but only learners who actually explained the material on video kept the advantage a week later (Fiorella, 2013). The benefit seems to come from generating an explanation, which links to the evidence on elaborative interrogation.

How strong is the evidence?

Strong against the pyramid’s percentages. Two independent reviews traced them and found no research behind them, the figures change between versions, and the organisation credited with the best-known one can’t produce its data. The evidence for the ideas the pyramid gestures at is a separate matter: strong for retrieval practice and for active learning in university classes, and moderate for learning by teaching. In practice, we’d stop quoting the chart and keep the instinct.

How to apply it in a course

The useful question isn’t “which format has the highest retention rate?” but “where in this module does the learner have to recall, use or explain something?”

We’ll use one invented example: a self-paced onboarding course for new customer-support staff on handling refund requests. The client’s brief includes the pyramid and asks for “less reading, and everyone presents to the team, because teaching gives 90%”.

1. Take the numbers out of the brief

Explain where the figures came from before they set the budget. If the pyramid stays in the brief, it tends to drive decisions such as cutting every reading or adding presentations nobody gets feedback on. Keep what the client wants underneath it: a course where new staff practise, not just watch.

2. Keep explanations where they do a job

A short explanation of the refund policy is efficient. Reading and short videos aren’t the problem; a course made only of them is. Present the policy clearly, with a diagram of the decision steps if it branches (our entry on dual coding covers why words and pictures work together), and keep the amount in one lesson manageable, as described in our entry on cognitive load theory.

3. Follow every explanation with recall

After each lesson, ask learners to recall the rule rather than re-read it: three or four questions on which requests qualify and what to check first. The Roediger and Karpicke result is why: the benefit shows up days later, not in the moment. Our guide to writing multiple-choice questions covers how to make those questions teach.

4. Give them something to do, with feedback

Swap a reading for a scenario: a customer message asking for a refund, and the learner writes the reply. Then show a model answer and the three things a good reply includes, so learners can check their own. Our entry on effective feedback covers what makes that step useful.

5. Ask them to explain, not to present

Replace “present to the team” with a short explain-it task: record a 60-second voice note or write a paragraph explaining to an imaginary new colleague why one request qualifies and another doesn’t. That keeps the part of teaching that helps, generating an explanation, without the scheduling, the nerves and the slides. If the course has a live cohort, pair learners to answer each other’s questions, since interactive teaching did best in Kobayashi’s meta-analysis.

This fits the design phase of the ADDIE model, when you decide what each module asks learners to do. The same caution applies to learning styles: another popular chart with an appealing rule and no evidence for the rule.

How the platforms support it

Dropping the pyramid asks nothing special of a platform. Building in recall, practice and explanation does. 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:

• Quizzes in any lesson. Can you add a short quiz after any lesson, not only at the end of a module or course, and can learners retake it?
• Feedback on answers. Can each question show an explanation after the learner answers, so a quiz teaches as well as checks?
• Written or recorded submissions. Can learners submit a written answer, a file or an audio or video recording for the explain-it task, and can you or a peer reply to it?
• Discussion. Is there a comments area or community space where learners can explain ideas to each other and answer questions?
• Model answers. Can you reveal content, such as a model reply, only after the learner has submitted their own?

If your platform can’t take submissions, a self-check works: the learner writes their answer in a notebook, then opens the model answer on the next screen.

Further reading

• Letrud & Hernes (2018), “Excavating the origins of the learning pyramid myths”. The fullest history of the idea, from 1852 to the modern versions. Open access.
• Masters (2013), “Edgar Dale’s Pyramid of Learning in medical education”. A short, readable literature review that shows how the figures spread through one field, and what Dale actually wrote.
• Roediger & Karpicke (2006), “Test-enhanced learning”. The experiment behind much of today’s interest in retrieval practice, and a clear lesson in why the delay before a test matters.
• Freeman et al. (2014), “Active learning increases student performance in science, engineering, and mathematics”. The meta-analysis to cite when someone asks whether active learning works, with honest effect sizes. Open access.
• Kobayashi (2019), “Learning by preparing-to-teach and teaching: A meta-analysis”. The best summary of what teaching others does for the teacher, and under what conditions.

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