Seven Round Numbers

The ninety-fourth article had the strongest sourcing in this series, because the researcher wrote down what she had done. This one has the weakest subject we have examined, because on the available accounts there was never any research to write down.

The Verdict, Stated First

Five claims, in descending order of confidence.

One. Three published versions of the chart disagree with each other, having seven, six and five tiers and attaching the same percentages to different activities. That is a fact about the documents, requiring no inference.

Two. Every figure in every version is a multiple of five, which is our own observation and is not what measured data looks like.

Three. No version states a delay, a test type or a population, and without those a retention percentage is not a measurement of anything.

Four. The research is not available, on the accounts we obtained, and the body credited with it reportedly could not produce it.

Five. On our own arithmetic, the chart does not support its own conclusion once cost is included, which holds even if you grant every number.

That fifth claim is the one we have not seen made elsewhere, ours, and it is the reason this article exists rather than being a link to somebody else's debunking.

Two notes on the ordering of those claims, ours. The first three rest on documents we read ourselves, which is why they lead, and the fourth rests on encyclopedias.

And the fifth is the only one that would survive if the numbers turned out to be real. A cost argument does not depend on the retention figures being right, only on their being used, which makes it the sturdiest thing here.

Our Grades For These Claims

Applying the scheme from the first article in this series. The sourcing here is markedly weaker than the previous article's, and we say so at the top.

Grade A for the disagreement between versions, since we obtained three renderings and compared them directly.

Grade A for our own arithmetic, all of which operates on figures printed in those renderings.

Grade C for the account of the chart's origin, which reaches us through encyclopedias rather than through the scholarship that established it.

Grade D for the critical literature. We have bibliographic records for the papers that debunked this and we obtained none of them, which is a real weakness we flag throughout.

Grade D for what Edgar Dale actually published, which we know only by report.

That spread is worth pausing on, ours, because it inverts the usual shape. Here the arithmetic is better sourced than the history, since the numbers are printed on documents anybody can read and the story of where they came from is not.

So we have written the article to lean on the part we can check. The comparison between versions and the cost table need no history at all, and would stand if every origin account here turned out to be wrong.

A Note On Method

Everything here is verified to August 2026, and this article's sourcing needs unusually careful description.

We obtained three separate renderings of the chart[1][4][5], and the comparison between them is the strongest evidence in the article because it depends on nothing but reading them.

The account of the chart's origin comes from an encyclopedia[1], flagged at every use.

The record of the critical literature comes from a reference list on a practitioner blog[2] and from an AI-generated encyclopedia[3], which is the weakest source we have ever cited and is flagged as such wherever it appears.

We obtained none of the debunking papers. We name them, we say we did not read them, and we do not rely on their findings for anything except where an independent source says the same thing.

All arithmetic is ours. The percentages are the chart's; the probability calculation, the cost table and every conclusion drawn from them are our own constructions.

This article discusses a training diagram and is not advice on staff development.

The Chart

What it claims, before we look at whether it can.

One rendering gives, from the base of the pyramid upward[4]: 90 percent for teaching others, 75 percent for practice, 50 percent for group discussion, 30 percent for demonstration, 20 percent for audio-visual, 10 percent for reading, and 5 percent for lecture.

It is labelled "Average Retention Rates" and adapted from a named training institute[4].

Four observations, ours.

In business it does one job. It is the slide that justifies replacing a manual or a lecture with a workshop, and it appears in training proposals and budget requests.

It is unusually easy to act on, which is its appeal. The numbers are ordered, the ordering matches an intuition people already hold, and the recommendation follows without argument.

And the phrase "average retention rates" is doing enormous work, because an average implies a distribution, a distribution implies measurement, and measurement implies a study.

That chain is worth walking slowly, ours, because each link is a commitment the chart never honours. An average of what, across whom, measured how, and the header answers none of the three while implying all of them.

Which is the question this article is about. Not whether active learning is better, but whether these numbers measure anything at all.

We separate those two because they are constantly merged, ours. A person defending the chart will defend the ordering, which is not under attack, and the defence passes without touching the figures.

Three Versions

We obtained three renderings from three separate sources. Set out side by side, ours.

Version A, seven tiers[1][4]: 5 percent lecture, 10 percent reading, 20 percent audio-visual, 30 percent demonstration, 50 percent group discussion, 75 percent practice, 90 percent teaching others.

Version B, five tiers[5]: 10 percent of what they read, 20 percent of what they hear, 50 percent of what they see and hear, 70 percent of what they discuss, 90 percent of what they do and teach others.

Version C, six tiers[5]: 10 percent read, 20 percent hear, 30 percent see, 50 percent see and hear, 70 percent discuss, 90 percent do.

Four observations.

All three are presented as the same chart, and all three are attributed to the same lineage.

The renderings come from different publishers and appear in different documents, which is why we treat the comparison as evidence rather than as a quirk of one bad copy.

We would note that we obtained these as reproductions rather than from any original, so we can say the versions in circulation disagree and not that any particular version was ever authoritative.

And that limitation does not weaken the point. What circulates is what gets used, and a manager copying the chart into a proposal is copying one of these.

One further note on how we selected them, ours. We did not go looking for versions that disagreed; we took the first three renderings we obtained and compared them, and the disagreement was what we found.

They Disagree With Each Other

The comparison, and it is not close. Ours.

Four observations.

The tier counts differ: seven, five and six. A single dataset does not produce three different numbers of measured conditions.

The 90 percent figure is attached to teaching others in one version, to doing and teaching combined in another, and to doing alone in the third. Those are three different claims about three different activities carrying one number.

Version A gives 75 percent for practice, a category the other two do not contain at all, and A has no 70 percent tier where B and C both do.

And version B has no 30 percent tier, while A attaches 30 to demonstration and C attaches it to seeing.

Set out plainly, ours: of the seven activities named across the three versions, only two carry the same number in all three, and those two are reading at 10 and hearing or audio-visual at 20.

Everything above the second tier moves. The disagreement is concentrated exactly where the chart's argument lives, which is the active end it is used to recommend.

The conclusion is available without any judgement about the underlying science. If these were measurements, the versions would agree about what was measured, and they do not agree about how many things were measured, let alone what.

Two ways a defender might answer that, ours, and neither works. The first is that different studies produced different tier counts, which would mean the numbers come from separate work and cannot be presented as one chart.

The second is that the versions are simplifications of a fuller original. That would require the fuller original to exist, and it is precisely what nobody can produce.

Every Figure Is Round

Our own observation, and it is the tell.

The complete set of distinct values across all three versions is 5, 10, 20, 30, 50, 70, 75 and 90.

All eight are multiples of five. Six of the eight are multiples of ten.

Four observations.

The ninety-third article noted that a round research figure warrants suspicion because measurements rarely are round, and there the rounding of 148 to 150 happened during transmission.

Here there is no unrounded version anywhere, in any rendering we obtained, which is a different and worse situation.

A real study of seven conditions would produce values like 47 percent and 62 percent, and somebody would have rounded them for display. The rounded display would then differ from the underlying numbers, and those numbers would exist somewhere.

And the pattern is not merely round but evenly spaced, running 10, 20, 30, 50, 70, 90 in version C, which is the shape of a series somebody wrote down rather than of a series somebody measured.

One more feature of that series is worth naming, ours. It runs 10, 20, 30 and then jumps to 50, 70, 90, so the spacing doubles exactly at the point where the chart moves from passive to active categories.

That is a design decision rather than a measurement. The gap widens where the argument needs it to widen, and no dataset arranges itself so conveniently around a rhetorical boundary.

What That Would Take

Our own arithmetic, with an invented baseline we flag before giving the result.

Suppose these percentages had really been measured, reported to the nearest whole percent, and that any value from 1 to 99 was equally likely. That assumption is ours and is certainly wrong; we use it only to give the coincidence a scale.

On that baseline, the chance that five independent measurements all land on a multiple of ten is about 1 in 161,000. For six, about 1 in 1.8 million. For seven, about 1 in 19.5 million.

And the chance that seven all land on a multiple of five is about 1 in 104,000.

Four observations.

The specific figures depend entirely on our uniform assumption and should not be quoted as results. Real retention values are not uniformly distributed, and a reader choosing a different baseline gets different numbers.

What no baseline changes is the direction. There is no plausible distribution of real measurements that puts every value on a round number, because measurement produces remainders and rounding to the nearest five discards information nobody discards by accident.

The honest way to state it is without the probability at all. Eight values, all multiples of five, is what a list looks like when somebody chose the numbers, and that inference does not need arithmetic to support it.

We kept the calculation because it gives the intuition a size, and we flag it because a number with an invented denominator is exactly the kind of thing this series exists to catch.

What A Retention Rate Must State

The deepest problem, and it requires no arithmetic at all. Ours.

A retention percentage is not a number until five things are specified.

The delay before testing. Retention at twenty minutes and retention at thirty days are different quantities.

The type of test. Free recall, cued recall and recognition produce very different scores from identical learning.

The material. Retention of a list of words and of a procedure are not comparable.

The population. Undergraduates, apprentices and experienced staff differ.

And what counts as one unit retained, since a percentage requires a denominator and nobody has said what is being counted.

Four observations.

No version we obtained states any of the five. Not one.

That is not a gap in the reproduction. A chart cannot state a delay and remain a chart of this kind, because the moment you write "after seven days" the reader asks what happens after thirty.

And the omission is what makes the numbers usable. A figure with no conditions attached applies to every situation, which is precisely why it travels and precisely why it means nothing.

We would put the test in one sentence, ours. If you cannot say what would have to be true for the number to be wrong, it is not a measurement.

And that test is worth keeping for other numbers, ours. Applied to the ninety-third article's figure of 150, it passes, because a count of stable relationships is at least in principle checkable and the paper gave an interval.

Applied here it fails completely. There is no observation anybody could make that would contradict "people retain 10 percent of what they read," because the claim names no occasion on which to look.

The Missing Delay

Taking the first of the five, because it alone is fatal. Ours.

Four observations.

Forgetting is steep early and shallower later, which is one of the oldest findings in the field. We did not obtain a specific forgetting curve and so we cite none, and we make no claim about any particular shape.

What follows from the direction alone is enough. Any statement of the form "people retain X percent" describes a single point on a curve, and naming no point means naming no quantity.

The practical form of this is one a manager can use immediately. Ask when the ninety percent is measured. Immediately after the teaching session, or six months later when the skill is needed?

And those two answers would justify completely different training designs, which means the chart cannot guide the decision it is used to guide even if its numbers were real.

The Attribution

Where the chart is said to come from, and the account is thin.

An encyclopedia records that a pyramid model "was supposedly developed by the National Training Laboratories Institute in the early 1960s, on its main campus in Bethel, Maine, for which the original, internal research is said to have been lost."[1]

It also records that the earliest such representation "is believed to originate in a 1954 book called Audio-Visual Methods in Teaching"[1]. An encyclopedia, flagged.

Four observations, ours.

Note the hedging in the source itself: "supposedly developed", "is said to have been lost." The encyclopedia is not asserting the history; it is reporting what is claimed.

One rendering we obtained carries the attribution as a footnote reading "Adapted from National Training Laboratories. Bethel, Maine"[4], with no year, no title and no author.

That footnote is the entire citation. A chart used to redirect training budgets carries a reference that names an organisation and a town, which would not be accepted in an undergraduate essay.

And the word "adapted" is quietly important, because it concedes that whatever the original was, this is not it.

What Dale Actually Drew

The most consequential claim in this article, and the one we are least able to verify.

The chart is routinely attributed to Edgar Dale's Cone of Experience. Multiple accounts state that Dale's cone contained no percentages at all, and that the figures were added later by others[2][3].

We did not obtain Dale's book, in any edition, and we cannot confirm this from the source.

Four observations, ours.

The scholarship on this is titled in a way that makes the claim plainly: "The Mythical Retention Chart and the Corruption of Dale's Cone of Experience"[2]. The word corruption asserts that something was done to the cone rather than by it.

A companion piece in the same series is titled "Timeline of the Mythical Retention Chart and Corrupted Dale's Cone"[2], which implies the researchers reconstructed when the numbers were attached.

We have the titles and not the papers, so we report what the titles assert and grade our own confidence accordingly. This is the weakest link in the article.

And if the claim is right, the structure is the ninety-second article's exactly. A real piece of work acquired a quantitative claim it never made, and the addition is what made it useful and what made it false.

Two differences from that case are worth keeping straight, ours. There the addition was a shape and here it is a set of numbers, and numbers are the more damaging addition because they look like evidence.

And there the underlying theory was at least published under its author's name, where here the figures have no author at all, which is why nobody can be asked about them.

The Research Nobody Can Find

The state of the evidence, on the accounts available to us.

An encyclopedia states that with the study lost, "the field largely stands on an unknown methodology of unknown quality, with unknown mitigation of influential parameters such as time, population tested, etc., making the original study's results untrustworthy."[1] An encyclopedia, flagged.

A practitioner account states that when pressed, the institute reportedly could not locate the original research[6]. A newsletter, not academic, flagged, and we treat it as an allegation rather than a finding.

Four observations, ours.

Note that the encyclopedia's list of unknown parameters is the same list we derived independently in the section above: time and population among them.

The convergence is mild evidence that we are not inventing the objection, and no more than that.

Lost research is not the same as no research, and we will not overstate this. It is possible that work was done in the early 1960s and the file is gone.

But the consequence is identical for a business decision. A number you cannot trace is a number you cannot check, and it should carry no more weight than a number somebody made up, because you cannot tell the two apart.

Which is a general standard we would defend, ours. Traceability is not a courtesy owed to scholarship; it is the property that makes a figure usable, because it is what lets somebody establish whether it applies to their situation.

The Version That Says A Hundred

A detail that settles the question, if the report is accurate.

An account of one researcher's analysis records that he found variants asserting 100 percent retention for teaching others[3]. An AI-generated encyclopedia, which is the weakest source in this article, flagged, and we did not obtain the analysis itself.

Four observations, ours.

If a version claims 100 percent, it claims perfect and permanent retention of everything taught, which is not a finding anybody could have obtained.

It also breaks the internal logic. A retention rate of 100 percent has no room for a delay, so the number cannot be a measurement at any point after learning.

We flag our uncertainty clearly. We did not see such a version ourselves, and the report reaches us through the weakest source we have used.

What we can say from what we did see is milder and sufficient. The versions we obtained already disagree, and a family of charts that drifts to 90 in one telling would drift to 100 in another without anybody noticing.

What The Critics Established

Named, because a reader deserves the trail, and graded honestly.

The record we obtained lists Subramony, Molenda, Betrus and Thalheimer (2014), The Mythical Retention Chart and the Corruption of Dale's Cone of Experience, Educational Technology, 54(6), 6–16, with three companion pieces in the same issue[2].

It also lists Thalheimer, W. (2015), Mythical retention data and the corrupted cone[2], and a separate account names Letrud (2012) and Letrud and Hernes (2018)[3].

Four observations, ours.

We obtained none of these papers. We are reporting a bibliography, which is a much weaker thing than reporting findings.

What the bibliography does establish is that this has been examined repeatedly by people who work on it professionally, across at least three separate efforts and a decade.

The one substantive finding we saw described, that a review of over fifty sources found no original research supporting the percentages[3], comes from our weakest source and we do not rely on it.

And we note the pattern the titles themselves describe. One is called "Previous attempts to debunk the mythical retention chart"[2], which means that by 2014 the debunking already had a history worth summarising.

Still Circulating In Journals

Evidence that the corrections did not take, and it is recent.

A 2018 research paper on retention in online courses states that retention in classroom settings "is often explained with the 'learning pyramid', also called 'cone of experience' established by Edgar Dale at the National Training Laboratories"[7].

Four observations, ours.

That sentence merges the two separate attributions into one, placing Dale at the institute, which no account we have seen supports.

It appears in a research paper, four years after the 2014 debunking series, which is the point of citing it.

The paper goes on to say that while there is dispute about the added factors, it is accepted that more immersion and application improves retention in general[7], which is a reasonable position and is not what the chart claims.

And that split is exactly the one worth holding on to. The general direction may well be right; the numbers are the problem, and this article is about the numbers.

We would add one observation about how the misattribution happened there, ours. The sentence names Dale and the institute together, which is what you get when two separate origin stories are compressed by somebody citing from memory.

Nobody Divides By Cost

Our own argument, and the one we have not seen made elsewhere.

Four observations.

The chart is used to justify replacing cheap training with expensive training: manuals and lectures out, workshops and peer teaching in.

But the chart says nothing about cost, and the modes it ranks lowest are dramatically cheaper than the modes it ranks highest.

A retention figure alone cannot support a budget decision. What a budget decision needs is retention per dollar, and the chart's own numbers can be divided by cost by anybody willing to do it.

So the interesting question is not whether the numbers are real. It is whether the recommendation follows even if they are, and on our arithmetic it does not.

The Cost Table

Our own arithmetic, on an invented firm, with invented costs. Every dollar figure below is ours and illustrative.

Forty staff. Five training modes, costed as a small Canadian firm might: a written manual at $120, a recorded video at $300, a live lecture at $900, a facilitated group workshop at $4,500, and a peer teaching programme at $7,200.

Dividing each cost by the chart's own claimed retention gives a cost per retention point.

Written manual: $12. Recorded video: $15. Peer teaching: $80. Group workshop: $90. Live lecture: $180.

Four observations.

On the chart's own numbers, the manual is the most efficient option available, at roughly a seventh the cost per point of the workshop the chart is used to recommend.

The costs are invented and the ranking is not delicate. The workshop would have to cost under about $600 to beat the manual, which is not what facilitated workshops cost.

The one place the chart's argument survives is against the live lecture, which comes last on our figures, being both expensive and rated lowest.

And that is the honest conclusion. The chart supports dropping lectures and does not support buying workshops, which is the reverse of how we have seen it used.

Two caveats on our own table, ours. Cost per retention point treats a retention point as the thing being bought, which is a simplification, since what a firm actually wants is a person who can do the task.

And the modes are not interchangeable. A manual cannot teach a physical procedure, so the cheapest option is not always an option, which the table does not capture and a real decision would.

The Same Test On Your Own Slides

Generalising the checks, because the value here is the method rather than this one chart. Ours.

Four tests, each of which took us minutes.

Find two versions and compare them. If a figure is real, independent reproductions agree; if it drifts, it was never anchored to anything. This is the cheapest test available and almost nobody runs it.

Look at the last digit. A full set of round numbers means somebody chose them, and a mix of round and unround numbers means at least some were measured.

Ask what the denominator counts. Any percentage has one, and a figure whose author cannot say what is being counted is not a percentage.

And ask what would falsify it. A claim with no conditions attached cannot be checked and therefore cannot be relied on, however often it appears.

We would add that the second test caught this chart on its own, ours. We noticed the round numbers before we had read a single word about the chart's history, and everything after that was confirmation.

What An Honest Version Would Look Like

Because criticism is cheap unless you can say what would satisfy you. Ours.

Four features.

It would state a delay and a test: retention of a defined body of material, measured by free recall, at thirty days.

It would report unround numbers with intervals, which is what the ninety-third article's arithmetic showed a real estimate carries.

It would name the population and the material, and would decline to generalise past them, since retention of a safety procedure by tradespeople says little about retention of a pricing policy by sales staff.

And it would be far less useful than the chart, which is the point. A version with all its conditions stated would not fit on a slide and would not tell a manager what to buy, and that is exactly why the version in circulation has none of them.

Which is an uncomfortable conclusion and we think it is the right one, ours. The properties that make a figure trustworthy are the same ones that make it hard to use, so the most usable version of any finding is reliably the least reliable one.

What Actually Survives

Our reading, stated directly.

Five statements.

Three published versions disagree on the number of tiers and on which activity carries which percentage.

Every figure across all versions is a multiple of five, which is not what measurement produces.

No version states a delay, a test, a population or a denominator, without which a retention percentage is not a quantity.

The research is unavailable, on the accounts we obtained, and a number you cannot trace cannot be checked.

And on our own arithmetic the chart does not support buying expensive training even if every number is granted, because nobody divides by cost.

Those five are what we would defend, ours, and the first, second, third and fifth rest entirely on documents and arithmetic rather than on any account of the chart's history.

The Real Finding Underneath

Because something here is probably true, and we will not pretend otherwise. Ours.

Four observations.

That people remember more of what they actively do than what they passively receive is not seriously disputed, and the paper we cited for its misattribution says as much[7].

The ninetieth article found the same shape in a different topic. A plausible general claim acquired a specific mechanism nobody had tested, and the specific version is what failed.

So the chart is not wrong in its ordering. It is wrong in claiming to know by how much, and the by-how-much is the only part a budget can use.

And that distinction is worth carrying to any framework. Direction is usually cheap to establish and magnitude is expensive, so a model offering precise magnitudes for free is offering the expensive part at no cost.

One test that follows, ours. Ask which part of a framework required the study, and whether the study is available.

Frequently the answer is that the ordering came from common sense and the magnitudes came from nowhere, which describes this chart precisely.

Why It Persists

The mechanism, ours, and it is now familiar in this series.

Four observations.

The chart confirms something the audience already believes, that sitting in a lecture is a poor way to learn, which is most people's lived experience.

It also flatters the person presenting it, who is usually proposing the active training and is therefore proposing the thing at the bottom of the pyramid.

The ninety-third article's author quotation applies without modification. An easily understood number travels and a methodological objection does not, and here the objection requires explaining what a retention rate needs to specify.

And there is a shape effect on top, which the ninety-second article examined. A pyramid asserts that the tiers differ in magnitude, and here that assertion is the whole content.

Though this case is stranger than that one, ours. Here the shape and the numbers make the same claim twice, so the diagram is redundant with its own labels, which is unusual and makes the assertion feel doubly established.

Your Own Training Budget

The practical application. Ours, and not advice on staff development.

Four points.

Decide what you want people to be able to do, and when, before choosing a format. The chart short-circuits that step by making the format the decision.

Measure something after the training, at a stated interval, even crudely. A firm that tests twenty staff at thirty days has better evidence about its own people than any chart can offer.

Divide by cost. Whatever you believe about relative effectiveness, the decision is about money, and effectiveness alone has never been enough to rank options.

And apply the ninety-fourth article's caution to whatever you measure. Twenty staff is a small sample, the result will be noisy, and anything dramatic it shows is probably overstated.

One encouragement against that caution, ours, because it should not stop anybody. A noisy measurement of your own staff is still about your staff, and it improves with repetition in a way a chart never will.

Bibliographic Note

The series keeps a count, and this article produced one, of an unusual kind.

The 2014 debunking series is given by one source as Educational Technology, Nov/Dec 2014, 54(6), 6–16 for the lead article[2]. The same source lists three companion pieces with page ranges of 17–21, 22–31 and 31–34[2], with no volume or issue number given for any of them, and with one range beginning on the page the previous one ends.

Three observations, ours.

The overlapping page is probably correct rather than an error, since a short piece can begin on the last page of the previous one, and we note it only because it looks like a mistake and is not.

The missing volume numbers are the real variant, and they matter. A citation without a volume is not a locator, and three of the four papers in this series arrive that way.

That brings the running count of bibliographic variants across this series to forty-four.

We note that this is the first article in the series where the bibliographic problem and the subject matter are the same problem, ours. The chart is untraceable and so, in part, is the scholarship establishing that it is untraceable.

What To Do

Do not cite the percentages. Three published versions disagree about how many tiers there are and which activity carries which number.

Notice that every figure is a multiple of five. Measurement produces remainders, and a full set of round numbers is a list somebody chose.

Ask when the retention is measured. No version states a delay, and without one the percentage is not a quantity.

Ask what the denominator is. A percentage requires something being counted, and no version says what.

Keep the ordering if you like, and drop the magnitudes. That active learning beats passive is not seriously disputed; by how much is the part with no support.

Divide by cost before reallocating a budget. On our own arithmetic, using the chart's own numbers, the written manual is the cheapest option per retention point and the workshop is not.

Measure your own people at a stated interval. Even a crude internal test beats a chart whose source cannot be produced.

And treat an untraceable citation as no citation. An organisation and a town is not a reference.

The Limits Of This Analysis

Several caveats matter, and this article's sourcing is the weakest we have published. It discusses a training diagram and is not advice on staff development. Everything is verified to August 2026. We obtained none of the scholarly papers that debunked this chart, only their bibliographic records, so where we describe what the critics established we are reporting titles and a reference list rather than findings; if those papers say something different from what their titles suggest, parts of this article would need revision. We did not obtain Edgar Dale's book in any edition, so the claim that his cone contained no percentages reaches us entirely at second hand and we cannot confirm it. The account of the chart's origin at a named institute, and the statement that the original research was lost, come from an encyclopedia, flagged; the further claim that the institute could not produce the research when asked comes from a practitioner newsletter, flagged, and we treat it as an allegation. The report of variants claiming 100 percent retention, and the bibliographic records for two of the critical papers, come from an AI-generated encyclopedia, the weakest source we have ever cited, flagged at every use, and we do not rely on it for anything load-bearing. The three versions of the chart were obtained as reproductions rather than from originals, so we can say the versions in circulation disagree and not that any one was authoritative. All arithmetic is ours. The probability calculation rests on an invented uniform baseline that is certainly wrong, and its specific figures should not be quoted; we kept it only to give a coincidence a scale, and flagged it in the section. Every dollar figure in the cost table is invented by us, and a firm with different costs would get a different ranking, though the manual would need the workshop to fall below roughly $600 to lose. And our conclusion that the chart fails to support its own recommendation is our own argument, not a finding of any source.

Related Reading

Behavioural Finance The Author Wrote It Down The opposite sourcing problem: a finding whose researcher documented exactly what was done. Behavioural Finance The Shape Was Not His The other training-room pyramid, and what a triangular shape asserts on its own. No. 96 The Bottom Eighty Percent Pay The Rent The man who named the Pareto principle published a paper saying he had named it wrong. He never used the numbers 80 or 20. And on our own arithmetic,...

Frequently Asked Questions

Are the learning pyramid percentages real?

On the accounts we obtained, the original research cannot be produced. What we established ourselves is that three published versions disagree about how many tiers exist and which activity carries which number, and that every figure across all of them is a multiple of five.

Why does it matter that the numbers are round?

Measurement produces remainders. A real study of seven conditions would give values like 47 and 62 percent, and any rounded display would differ from the underlying numbers, which would exist somewhere. Here there is no unrounded version in any rendering we obtained.

What is missing from the chart?

A delay before testing, the type of test, the material, the population, and what counts as one unit retained. No version we obtained states any of the five, and without a delay in particular the percentage names no point on any forgetting curve.

Did Edgar Dale put percentages on his cone?

Multiple accounts say he did not and that the figures were added later. We did not obtain his book in any edition, so this reaches us entirely at second hand and we flag it as the weakest link in the article.

Is active learning better than passive learning?

The ordering is not seriously disputed, and a paper we cite says as much while misattributing the chart. What has no support is the magnitude, and the magnitude is the only part a training budget can actually use.

Does the chart support spending more on workshops?

On our own arithmetic, no, even granting every number. Using invented but realistic costs for a firm of forty, dividing cost by the chart's own claimed retention makes the written manual the cheapest option per retention point and the live lecture the most expensive. The chart supports dropping lectures, not buying workshops.

What should a firm do instead?

On our own reasoning: decide what people should be able to do and when, measure something at a stated interval after the training even crudely, and divide by cost before reallocating anything. An internal test on your own staff beats a chart whose source cannot be produced.

References

1. Encyclopedia entry on the learning pyramid, also known as the cone of learning, the learning cone, the cone of retention, the pyramid of learning and the pyramid of retention. Describes it as a pseudoscientific group of ineffective learning models relating different degrees of retention induced from various types of learning; records that the earliest such representation is believed to originate in a 1954 book called Audio-Visual Methods in Teaching; records that a pyramid model was supposedly developed by the National Training Laboratories Institute in the early 1960s on its main campus in Bethel, Maine, for which the original internal research is said to have been lost; reproduces a seven-tier table giving 90 percent for teaching someone else or using immediately, 75 percent for practising what one learned, 50 percent for engaging in a group discussion, 30 percent for watching a demonstration, 20 percent for watching audiovisual, 10 percent for reading and 5 percent for listening to a lecture; and states that with the study lost the field largely stands on an unknown methodology of unknown quality, with unknown mitigation of influential parameters such as time and population tested, making the original study's results untrustworthy. Note: an encyclopedia, NOT an academic source, flagged at every use. Our source for the origin account and for version A of the chart. en.wikipedia.org
2. Practitioner blog post on common myths about learning, carrying a reference list that includes Subramony, D., Molenda, M., Betrus, A., & Thalheimer, W. (2014), The Mythical Retention Chart and the Corruption of Dale's Cone of Experience, Educational Technology, Nov/Dec 2014, 54(6), 6–16; Thalheimer, W. (2015), Mythical retention data and the corrupted cone, Work-Learning Research; and companion pieces by the same four authors in the same issue titled Previous attempts to debunk the mythical retention chart and corrupted Dale's Cone at pages 17–21, The Good, the Bad, and the Ugly: A Bibliographic Essay on the Corrupted Cone at pages 22–31, and Timeline of the Mythical Retention Chart and Corrupted Dale's Cone at pages 31–34. Also states that there is significant evidence that the data from which the learning pyramid was built was not based on empirical information or evidence-based research. Note: a practitioner blog, NOT an academic source, flagged. Used ONLY for the bibliographic records of the critical literature. WE OBTAINED NONE OF THE PAPERS IT LISTS and do not rely on their findings. dremilywhitehorse.com
3. AI-generated encyclopedia entry on the learning pyramid, describing it as a widely circulated but debunked educational model claiming retention rates of 5 percent for lectures and 10 percent for reading rising to 70 or 90 percent for discussing or teaching others; stating that it has no basis in empirical research, that investigations reveal the specific retention percentages cannot be traced to any verifiable study, and that it is frequently misattributed; and recording a 2012 analysis by Kare Letrud labelling it a myth propagated by the National Training Laboratories Institute due to reliance on fabricated data, finding variants asserting 100 percent retention for teaching others, and reporting a review of over 50 sources confirming no original research supports the percentages, together with a 2018 study by Letrud and Hernes. Note: an AI-generated encyclopedia. This is the WEAKEST SOURCE WE HAVE EVER CITED, flagged at every use. Used only for the bibliographic records of two papers we did not obtain and for the report of a 100 percent variant we did not see. Nothing load-bearing in this article rests on it. grokipedia.com
4. Document-sharing platform copy of a one-page handout titled The Learning Pyramid, headed Average Retention Rates, giving 5 percent lecture, 10 percent reading and 20 percent audio-visual under the label passive teaching methods, and 30 percent demonstration, 50 percent group discussion, 75 percent practice and 90 percent teaching others, with the footnote that it is adapted from National Training Laboratories, Bethel, Maine. Note: a user-uploaded handout on a document-sharing platform, NOT an academic source, flagged. Reproduced here as a specimen of the chart as it actually circulates, and as our source for the footnote attribution. scribd.com
5. Document-sharing platform copies of two further renderings attributed to Dale's Cone of Experience. One states that learners retain 10 percent of what they read, 20 percent of what they hear, 50 percent of what they see and hear, 70 percent of what they discuss, and 90 percent of what they do and teach others. Another gives a six-tier form running 10 percent read, 20 percent hear, 30 percent see, 50 percent see and hear, 70 percent discuss and 90 percent do, and states that people remember 90 percent of what they say and do, 70 percent of what they discuss and 10 percent of what they read. Note: user-uploaded documents on a document-sharing platform, NOT academic sources, flagged. Our source for versions B and C, and the basis of the comparison showing that the circulating versions disagree. scribd.com
6. Practitioner newsletter article on the chart, stating that it claims we retain only 5 percent of what we hear in a lecture but 90 percent if we teach someone else; that despite being cited for decades there is no credible evidence it was derived from empirical research; and that when pressed, the National Training Laboratories in Bethel, Maine reportedly admitted they could not locate the original research. Note: a practitioner newsletter, NOT an academic source, flagged. The claim about the institute being unable to produce the research is reported here as an allegation and we treat it as one. workwise.substack.com
7. Preprint research paper on knowledge retention in online programming courses, stating that retention of knowledge in classroom settings is often explained with the learning pyramid, also called the cone of experience, established by Edgar Dale at the National Training Laboratories; and that while there is dispute about subsequently added factors that should reflect the retention level, it is accepted that more immersion with and application of the learning material improves retention in general. Note: a preprint, cited here NOT for its own findings but as evidence that the chart was still being presented as established in research writing years after the published debunkings, and as an example of the two attributions being merged into one. arxiv.org

This article discusses a training diagram and is not advice on staff development. None of the scholarly papers debunking this chart were obtained, only their bibliographic records. Edgar Dale's book was not obtained in any edition. One source used is an AI-generated encyclopedia, flagged as the weakest cited in this series, and nothing load-bearing rests on it. All arithmetic is the authors' own; the probability figures rest on an invented baseline and every cost in the table is invented.