Practice vs. Lecture: What the Research Really Says About Learning People Skills

"Lectures deliver 5-10% retention. Simulation-based practice delivers 75-90%."

That sentence sits on the Conflict Campaign homepage. I signed off on it, and for longer than I'd like to admit, I believed it the way you believe anything that appears in enough slide decks. Then I went looking for the study behind it. There isn't one.

So here's the honest version. The famous retention percentages – 5% for lecture, 75% for practicing by doing, 90% for teaching others – come from the "learning pyramid," a chart with no traceable research behind it. The numbers are folklore. The underlying claim survives anyway: rigorous meta-analyses consistently find that practice-based learning with feedback and structured reflection produces large, measurable gains in interpersonal skills – Hedges' g = .72 for negotiation training, d = 0.79 for after-action reviews – while lecture-heavy training runs headlong into the transfer problem that sinks most workplace training. The direction is right. The precision is fake.

I found this out the slow way. Before Adam White and I designed Operation Aetherfall, our first scenario kit, we synthesized more than 200 studies on instructional design – transfer, skill decay, simulation fidelity, debriefing. The pyramid's percentages never appeared in any of them. A statistic that famous, missing from the actual literature, bothered me enough to trace it back to its source.

Where does the 75-90% retention figure come from?

The percentages come from the learning pyramid, a chart usually attributed to the National Training Laboratories (NTL) Institute in Bethel, Maine, dating to the early 1960s. Lecture retains 5%, reading 10%, audiovisual 20%, demonstration 30%, discussion 50%, practice by doing 75%, teaching others 90%. You've seen it. It hangs in training rooms, opens conference talks, and anchors vendor pitches – including, until this audit, ours.

Trace it backward and the trail dissolves. The pyramid's shape borrows from Edgar Dale's Cone of Experience, published in his 1946 book Audio-Visual Methods in Teaching. Dale's cone classified learning experiences from concrete to abstract – and carried no percentages, made no retention claims, and came with Dale's explicit warning not to read it as a rank order. By the 1969 edition he was frustrated enough to spend six pages addressing the misconceptions. The earliest documented appearance of the percentages themselves is a 1967 magazine article by D. G. Treichler, an employee of Mobil Oil, who offered no citations and no methodology. Michael Molenda, the Indiana University professor who spent years hunting the origin, traced a possible earlier version to WWII-era visual-aids training at the Aberdeen Proving Ground – and found that no archive, military or academic, contains the supporting research.

And NTL? When researchers wrote to ask for the original study, the institute's reply – reproduced in Lalley and Miller's 2007 review – has become a small classic of the genre: "Yes, we believe it to be accurate – but no, we no longer have – nor can we find – the original research that supports the numbers."

Twenty-plus years of product work has put me in a dozen roadmap reviews where somebody defended a forecast with "we believe it's accurate, we just can't find the analysis." It didn't go well for them either.

Is any part of the learning pyramid true?

The pyramid fails in two ways: the percentages aren't real, and the fixed hierarchy isn't either. When Lalley and Miller reviewed the research on each method in 2007, they found every one of them – lecture included – produced retention, with none consistently superior across contexts. Daniel Willingham's critique is the simplest: retention depends on the material, the learner, the delay before the test, and a dozen other variables, so assigning a fixed percentage to a teaching method is a category error. Subramony, Molenda, Betrus, and Thalheimer, who published the fullest debunking in 2014, traced the numbers to "folkloric maxims" circulating since the early 1900s and noted that the percentages appear in dozens of contradictory permutations – a tell that no single dataset ever produced them.

The roundness should have been the giveaway. Real studies produce values like g = .72 and d = 0.79. Folklore produces 5, 10, 20, 30, 50, 75, 90.

What survives the audit is a weaker, better claim: for skills – things you do rather than things you know – practice with feedback has strong evidence behind it that lecture alone doesn't. That claim doesn't need a pyramid. It has meta-analyses.

What does rigorous research actually support about practice-based learning?

Practice-based learning has large, replicated effects on interpersonal skills – when the practice comes with feedback and structured reflection attached. The honest comparison of experiential learning vs. lecture retention isn't a pair of tidy percentages. It's a set of effect sizes from studies someone can check.

What the evidence supports	Source	Result as the research states it
Negotiation skills respond to training	Meta-analysis of 17 studies on interest-based negotiation training	Large effect (Hedges' g = .72); over 80% of trained professionals report long-term workplace use
The debrief drives the learning	Keiser & Arthur's 2020 meta-analysis of 61 studies, 3,000+ participants	After-action reviews improve training outcomes at d = 0.79
Debriefs lift performance broadly	Tannenbaum & Cerasoli's meta-analysis	20-25% performance improvement (d = .67)
If-then plans carry skills back to work	Sheeran, Listrom, & Gollwitzer's 2025 meta-analysis of 642 tests	d = .66 for behavioral outcomes
Transfer is the real bottleneck	Training transfer literature	Only 10-15% of trained skills transfer to the job by default
Game mechanics alone don't teach	JMIR meta-analysis of gamified tasks	More motivating, but no better on measured performance

Two details in that research change how you should read the table. First, the practice has to involve genuine stakes. Harvard's Program on Negotiation documented that simulations with genuinely competing interests produce more transferable skills than scripted role-plays, because participants must develop real strategies rather than perform predetermined ones. Second, what drives transfer is psychological fidelity – whether the exercise evokes the same cognitive and emotional processes as the real situation – not whether it looks like your office. Reviews in medical education found cognitive skill transfer tracks the simulation's cognitive fidelity and is independent of its physical realism, and Davies and Krame's 2024 study found novices actually learn better in lower-fidelity environments that isolate the key demands. That finding is the reason a fantasy tabletop scenario can teach real conflict skills instead of being a gimmick.

The audit cuts both ways, and the last table row deserves a beat. A JMIR meta-analysis found gamified tasks were more motivating but no better on measured performance – engagement isn't learning. In the CPP Global study, only 27% of trainees reported feeling more comfortable with conflict after training. And one study of 234 participants found conflict training reduced willingness to confront among high-avoidance individuals – the people who most needed it. Practice-based formats fail too, mostly when the practice floats free of feedback, reflection, and follow-up.

Why does practice beat lecture for people skills – when it does?

Interpersonal skills are performances, not facts. A lecture can hand you the concept of opening a hard conversation gently; it can't give you the experience of your pulse climbing while you try it on a colleague who's pushing back. Skill lives in the doing, and lecture formats contain no doing.

But the practice itself isn't where most of the learning happens, and this is the part the pyramid's fans and critics both tend to miss. The simulation literature is blunt: the debriefing phase accounts for more learning than the simulation itself. Practice generates the raw material – decisions made under pressure, the moment a negotiation turned – and a structured debrief converts that material into something a participant can name and reuse. The fair comparison was never lecture vs. practice. It's lecture vs. a practice system: authentic dilemmas, feedback, guided reflection, and a concrete plan for the workplace.

We saw this in both pilot runs of Operation Aetherfall. The game produced the stories, but the debrief is where players connected what their characters did under pressure to what they do in standups and code reviews – and it ran long both times because nobody wanted to stop. That conversation, not the dice, is the mechanism.

Should Conflict Campaign keep quoting the 75-90% figure?

Not without this article attached. The homepage line gets the direction of the evidence right and dresses it in false precision, and now that I've traced the number, "everyone quotes it" isn't a defense I'm willing to make. Ours needs a rewrite, and it'll get one.

Here's the version I'd defend in a room of learning scientists: practice with feedback and a structured debrief produces large, replicated gains in interpersonal skill – effects in the d = .66 to .79 range across meta-analyses covering hundreds of tests – and lecture alone reliably fails to change behavior. It's less quotable than 75-90% vs. 5-10%. It's also checkable, and that difference is the whole point. Anyone can verify d = 0.79 across 61 studies. Nobody could ever verify the pyramid.

What does this mean for choosing a training format?

Ignore retention percentages entirely – no fixed percentages exist for any format, whoever's quoting them. Evaluate a workshop, course, or scenario kit (ours included) by whether it contains the components with measured effects: What do participants actually do during the session, and are the stakes real or scripted? Who gives feedback, and when? How is the debrief structured, and how much time is protected for it? What happens in the 90 days after – because interpersonal skills decay substantially within 30 to 90 days without reinforcement, and if-then plans made at session close are the best-evidenced bridge across that gap.

Lecture still has a job inside that system: concepts, shared vocabulary, the theory that makes practice legible. Where the telling belongs in the sequence is its own design question. What lecture can't do is carry the session alone, and no pyramid percentage was ever going to settle how much of it you need.

We put a folklore number on our homepage because it was tidy and everyone else was already quoting it. The real research turned out to be better than the folklore – less symmetrical, more conditional, and verifiable. A training claim is a promise to somebody's team, and promises should be made with numbers someone can check.

The pyramid was never load-bearing. The practice is.