Satisfaction Ratings Predict Learning at .02. Four Myths About the Post-Training Feedback Form.

A post-training feedback form measures trainee reactions: how satisfied learners were, and how useful they expect the content to be. It is not a learning measure. Meta-analysis puts the correlation between satisfaction-type reactions and immediate learning at .02, and between reactions of all kinds and learning at roughly .08. The form measures experience, not capability.

That gap matters because the feedback form is the most widely collected piece of evidence in corporate learning, and in most organisations it is the only one. A programme that averages 4.6 out of 5 gets renewed. One that averages 3.8 gets redesigned. Neither number carries information about whether anyone can now do anything they could not do before. Here are four beliefs that keep the form in a job it was never able to do.

Myth 1: A session rated 4.6 out of 5 taught people something

Alliger, Tannenbaum, Bennett, Traver and Shotland (1997) meta-analysed the relations among training criteria in Personnel Psychology (full text). Affective reactions — enjoyment, satisfaction with the trainer, "I would recommend this course" — correlated .02 with immediate learning across 11 correlations. Pooling every type of reaction across 23 correlations lifts it to about .08. Against on-the-job transfer, reactions overall sat at .13, and affective reactions alone at .07.

A weak correlation is the charitable reading. The relationship can also run backwards. In Deslauriers, McCarty, Miller, Callaghan and Kestin (2019), 149 introductory physics students at Harvard experienced both active instruction and conventional lecture in a crossover design, with content and instructors held constant. Active instruction raised scores on the test of learning by 0.46 standard deviations — and lowered students' feeling of learning by 0.56 standard deviations. Learners who rated the instructor as more effective at teaching reported a feeling of learning 0.51 SD higher, which is to say the rating tracked the fluency of the delivery rather than what was retained.

Higher education has been round this loop already. Uttl, White and Wong Gonzalez (2017), in Studies in Educational Evaluation, re-analysed the multisection studies behind the long-standing claim that students learn more from highly rated instructors. The earlier correlations, they found, were driven by small samples and publication bias; the prior meta-analyses accounted for at most 1% of the variability in learning. Their own meta-analysis of all available multisection studies found ratings and learning unrelated.

Myth 2: "I can now apply this at work" is a learning measure

Adding a self-rating of capability to the form feels like an upgrade from a satisfaction item. It is not one. Sitzmann, Ely, Brown and Bauer (2010), in Academy of Management Learning & Education, meta-analysed what self-assessed knowledge actually correlates with:

• Objective cognitive learning — corrected ρ = .34, from 137 correlations and 16,951 learners
• Reactions and satisfaction — ρ = .51, from 105 correlations and 28,612 learners
• Motivation — ρ = .59, from 47 correlations and 9,534 learners
• Self-efficacy — ρ = .43, from 32 correlations and 3,720 learners

Self-assessment sits closer to how learners felt than to what they knew. The authors' conclusion is blunt: self-assessments are better understood as affective measures, and a knowledge test is what answers a knowledge question. A five-point "I can now apply this" item is a satisfaction question wearing a capability label — and it carries all the incentive problems of a rating collected in the room, from the trainer, with the learner's manager cc'd on the report.

Myth 3: The four levels are a chain, so reactions are an early warning

The widely used four-level model is a taxonomy of things you can measure: reaction, learning, behaviour, results. The causal reading — that reaction feeds learning, which feeds behaviour, which feeds results — is an assumption practitioners added, and Alliger and Janak (1989) tested it thirty years after the model was published. They examined the three assumptions implicit in how it was being used — that the levels ascend in value, that each causes the next, and that they are positively intercorrelated — and found the supporting evidence thin.

The practical consequence is specific. If you are monitoring reaction scores as a leading indicator of behaviour change, the 1997 figure for that link is .07. You will not see a transfer problem coming in a reaction score, and a high one will not warn you that nothing is being applied.

Myth 4: If reactions don't predict learning, scrap the form

Not every item on the form behaves the same way. Alliger and colleagues separated affective reactions from utility reactions — judgements of relevance and expected use, as in "this applies to my job" or "I expect to use this within a month." Utility reactions correlated .26 with immediate learning and .18 with transfer. The Blume, Ford, Baldwin and Huang (2010) transfer meta-analysis points the same direction: utility reactions carry a modest relationship to transfer, affective reactions none.

Honesty about the size of that evidence matters in an article about measurement. The utility figures rest on six and three correlations respectively — enough to tell you which items to keep, not enough to treat a utility score as proof of anything. Use them as triage.

Practically, that means rewriting the form rather than deleting it:

1. Collapse enjoyment to a single summary item and stop reporting it as a programme score.
2. Ask which specific task the learner intends to apply the content to, and when.
3. Ask what will stop them — workload, tools, manager, approval, missing data.
4. Ask whether the content matched the work as it is actually done, not whether it was interesting.
5. Keep the logistics items. Room, platform, timing and pre-work are the questions the form answers well, and they are fixable.
6. Label the output "experience data" in every report, so nobody downstream reads it as evidence of capability.

The three instruments, and what each is allowed to claim

Reaction form — measures experience

• Captures satisfaction, facilitator quality, logistics, perceived relevance
• Legitimate uses: operational fixes, trainer coaching, spotting disengagement early
• Not admissible for: learning claims, programme renewal decisions, anything described as ROI

Knowledge or skill test — measures learning

• Same blueprint administered before and after, with items written from the job's task list rather than the slide deck
• Legitimate use: whether measured capability moved, and for whom
• A short pre/post pair is usually the cheapest instrument in the programme — on AssessAll it runs on pay-as-you-go credits at ₹30 or US$0.50 per attempt, with no seat licence to justify first

Behaviour evidence — measures transfer

• Manager observation against anchored scales, work samples, or scored scenarios administered 60 to 90 days after the programme, when the novelty has worn off
• Legitimate use: whether the job changed
• AI-graded scenarios make this affordable at a scale where observation alone is not, provided the scoring rubric comes from the same task analysis as the test

None of this requires abandoning the feedback form. It requires demoting it. A reaction score tells you whether a programme was tolerable to sit through, which is worth knowing and cheap to collect. It does not tell you whether anyone learned anything, and the evidence that it might — assembled over three decades — tops out around .08.

The takeaway: if the only number your learning function reports is a satisfaction average, you are measuring the room, not the capability. Add one short pre/post test built from the job's tasks, and you will know more after a single programme than the feedback archive has told you so far.