Death by PowerPoint: The Two Most-Quoted Facts Are Both Invented

There is a genre of talk and blog post about slides that opens with a retention statistic and closes with the Space Shuttle. Both of its load-bearing facts are wrong, and the real evidence underneath is more interesting and much more conditional than either.

Some slide-design effects are real and measurable. The popular version of the science is considerably stronger than the science.

The retention percentages are from a 1967 trade magazine

You have seen them: people remember 10% of what they read, 20% of what they hear, 30% of what they see, and so on up a tidy ladder.

They trace to D. G. Treichler (1967), writing in Film and Audio-Visual Communications. Treichler was an employee of Mobil Oil. The article cited no research at all.

The numbers are frequently attached to Edgar Dale’s Cone of Experience (1946). Dale’s cone contained no numbers, was not derived from research, and Dale explicitly warned readers not to take it literally. The percentages were welded onto it by other people later.

This is the single most repeated fake statistic in the entire genre. If a presentation about presentations opens with it, that is a useful signal about the rest of the presentation.

A second one to know: “presentations with visuals are 43% more persuasive,” usually attributed to a Wharton or 3M study. We could find no primary document. Every trail ends in a marketing blog citing another marketing blog.

The Columbia claim, precisely

The Columbia Accident Investigation Board report (2003), Volume 1, Chapter 7, page 191, says this, verbatim:

“At many points during its investigation, the Board was surprised to receive similar presentation slides from NASA officials in place of technical reports. The Board views the endemic use of PowerPoint briefing slides instead of technical papers as an illustration of the problematic methods of technical communication at NASA.”

The Board called it “an illustration of” a communication problem. It did not say PowerPoint caused the loss of Columbia.

There is a further wrinkle worth knowing. The report credits Edward Tufte on the same page as the analyst of the Boeing slide it discusses. So the CAIB is not independent corroboration of Tufte’s argument - it is a document that drew on it.

And Tufte’s own essay, The Cognitive Style of PowerPoint, is analytical and rhetorical rather than experimental. It is a close reading of artefacts plus design theory, self-published through his own imprint. A published review notes that “he provides no data on the relative effectiveness of this tactic in teaching.” It is valuable as criticism and useless as an effect size.

What the research does support

The real evidence comes from multimedia learning research, and it is genuine.

Mayer and Fiorella (2014), in the Cambridge Handbook of Multimedia Learning, summarise the principles with box scores:

Principle	Tests supporting	Median effect size
Coherence	23 of 23	d = 0.86
Redundancy	16 of 16	d = 0.86
Signaling	24 of 28	d = 0.41
Spatial contiguity	22 of 22	d = 1.10
Temporal contiguity	9 of 9	d = 1.22

The redundancy principle is the one that matters most for slides, and it is narrower than the popular reading: it concerns identical on-screen text added to a narrated animation. Putting your script on the slide and reading it out is worse than either alone.

Two cautions belong with that table immediately.

A 100% hit rate across 16 tests is a publication-bias signal, not a strength claim. These are box scores of tests the authors selected, not a systematic review with a search protocol and a funnel plot.

It is a handbook chapter by Mayer, edited by Mayer, summarising Mayer’s own research programme. That is not misconduct, but it is not an independent synthesis.

Independent meta-analyses exist and are more sober.

Ginns (2005) meta-analysed the modality effect across 43 independent effects and N = 1,927: between-subjects d = 0.72, 95% CI 0.52-0.92. Ginns (2006) meta-analysed spatial and temporal contiguity across 50 effects and N = 2,375: d = 0.72 [0.61, 0.82] and d = 0.78 [0.63, 0.92].

Noetel and colleagues (2022), in Review of Educational Research, pooled 29 systematic reviews, 1,189 studies and 78,177 participants, finding 11 design principles with significant positive effects on learning.

And the corrective

Cromley and Chen (2025), in Educational Research Review, meta-analysed 181 studies and 591 effects from 92 of Mayer’s own articles, 1990-2022.

Overall g = 0.37 - less than half the median effect sizes the handbook chapter advertises - significantly moderated by all nine moderators tested, including a small decline in effect size per year.

There is a sharper corrective still, hiding inside Ginns’s own meta-analysis. The modality effect was strongly moderated by pacing: system-paced d = 0.93 versus self-paced d = -0.14, z = 7.83, p < .001.

Self-paced is where slide audiences actually live. When people can read at their own speed, the flagship effect disappears and reverses.

And the modality principle has a published double replication failure. Tabbers and van der Spoel (2011) ran two experiments (N = 60 and N = 70). Their conclusion, verbatim: “The double failure to replicate the modality effect and the lack of evidence for the different channels explanation raises doubt on its theoretical rationale and limits its practical use as a design principle.”

Nobody has tested bullet points

This is the finding that ought to end most of the genre.

We could locate no controlled experiment that isolates bullet points as the manipulated variable.

The nearest thing is Garner and colleagues (2011), an ASEE conference paper testing Michael Alley’s “assertion-evidence” slide format against conventional topic-subtitle slides. N = 111, randomised from five course sections.

The results are a trade-off, not a win:

• Immediate essay on complex concepts: 8.01 (SD 4.37) against 9.24 (SD 4.24), t(109) = -1.78, p = .078 - not significant
• Delayed multiple choice on complex concepts: 1.84 (SD 0.84) against 2.17 (SD 0.81), t(104) = -2.11, p = .038
• Recall of secondary statistics: 69% for bullet slides against 39% for assertion-evidence, F(1,105) = 28.29, p < .000

The bullet slides won decisively on recall of the details they contained and the assertion-evidence slides omitted.

And the two conditions differed in words per slide (30.4 against 19.3), graphics coverage (54.5% against 100%), slide count (11 against 10) and animation (0 against 9). With four things changed at once, nothing can be attributed to bullets as such.

“Bullet points harm learning” is an extrapolation, not a finding.

The theory has a falsifiability problem

Cognitive load theory underpins most of this advice, and its most serious published criticism is worth quoting.

de Jong (2010), in Instructional Science: “the fact that cognitive load is composed of three different elements that are ‘good’ (germane), ‘bad’ (extraneous), or just there (intrinsic) means that every outcome fits within the theory post-hoc.” And: “Overall, results from these studies give inconsistent results, raising doubts whether students can themselves distinguish between different types of load.”

A framework that accommodates every result is not making predictions.

What this leaves you with

Do not read your slides aloud while the audience reads the same words. This is the redundancy principle, it is the best-supported thing in the topic, and it is the one piece of advice that survives every deflation above.

Put the words near the picture they describe, not in a caption block elsewhere. Spatial contiguity pools around d = 0.72 across 50 effects.

Expect a modest effect, not a transformation. The best estimate across Mayer’s own corpus is g = 0.37, and it has been declining over time.

If your audience controls the pace, most of the narration advice stops applying. That is what d = -0.14 means.

Stop quoting the retention ladder and the shuttle. One is a 1967 trade article citing nothing; the other says “an illustration of,” not “caused.”

For what happens when the same material is read on a screen instead of paper, see screen reading vs paper. For what playback speed does to comprehension, see watching lectures at 2x speed. For a typography intervention that did not survive testing, see Sans Forgetica.

Limitations of this article

We did not obtain the redundancy-specific effect size from the 2025 meta-analysis. Its full text is embargoed, and redundancy appears in neither the “largest” nor “medium” effect lists in the abstract. So the sharpest possible comparison - Mayer’s d = 0.86 against an independent pooled estimate for the same principle - is not available to us.

We did not obtain per-principle effect sizes from Noetel and colleagues (2022), only the review-level counts.

We could not read three of the four papers in the 2011 modality-effect debate issue, and report only the one we could.

We could not verify the sample size of an Alley study from 2006 or a follow-up from 2013, and have therefore cited neither.

A critique of the assertion-evidence work exists as a self-archived university working paper, not a peer-reviewed comment. Its argument - that Alley compares his design against deliberately poor control slides - is worth knowing and is not peer-reviewed. We note it here rather than citing it as evidence.

Michael Alley originates the assertion-evidence method, is an author on essentially all supporting studies, and maintains a site promoting it. Originator-evaluates-own-method.

Tufte self-publishes the essay discussed here.

One process note. An automated read of de Jong (2010) initially returned four fluent quotations with page numbers impossible for an article printed on pages 105-134. Re-reading the PDF produced different, correct quotations at pages 119 and 125. Those are the ones printed above. We record this because it is the failure mode this programme exists to catch.