C4 Task feedback vs praise
narrowed 2 attacks: 1 survived, 1 narrowed · serves Q1 · last change 14 November 2026
Written comments that tell learners what is right or wrong in their work and how to improve it (task and process feedback) lead to better performance on a later task or test than a grade or score given on its own, or than praise on its own, at least on open-ended tasks. On a simple speed task, one classroom experiment found grades alone did as well as comments (Butler & Nisan 1986).
In practice
When giving feedback, say what is right or wrong in the work and how to do it better. Do not rely on praise, a grade or a score on its own to move learning forward.
Go deeper
Where it holds
Main limits
- About the kind of feedback, not about recalling versus rereading (that is C0).
- Who gives the feedback, AI or a person, is a separate question.
- Most of the project's evidence is reviews that mix immediate and later tests; one meta-analysis restricted to later tests (Abe & Usami 2024, 9 mostly older studies) found comments ahead of a grade alone, but less so the older the learners.
- Comments that only praise ("very good") did little better than a grade.
- On a simple speed task, grades alone did as well as comments in one classroom experiment (Butler & Nisan 1986).
All limits (8)
Inside the circle: the core of the claim. Outside: every bound and carve-out in the claim file, numbered, with the attack that forced or added it where the record shows one.
forced by a narrowing attackin the original wordingwritten at creation, or added without a narrowing
Warm cards in the list were added by later attacks; plain cards were in the original draft or written at creation.
Task-type bound added 2026-10-16 (week run): in Butler & Nisan 1986 (261 grade-6 pupils, nine classes), grades alone matched comments on a later quantitative word-construction task, while comments beat grades on a qualitative divergent-thinking task.
Forced by attack #1 · Butler & Nisan 1986 · 16 October 2026
Reworded 2026-10-16 (9 consecutive fruitless searches): the claim now names the comparison the studies in the library actually run — comments versus grades-only or praise-only, measured on a later task or test — as in Butler 1988 (grades 5–6, comments beat grades and grades-plus-comments at a later session) and Koenka et al. 2021 (grades versus comments; delay of the achievement measures not verified).
Forced by attack #1 · Butler & Nisan 1986 · 16 October 2026
This claim is about the type of feedback (information about the task and the process versus praise, grades or scores alone), not about whether retrieval practice with corrective feedback beats restudy (that is C0), and not about who gives the feedback: AI-generated versus human feedback is a separate question (Kaliisa et al. 2025 found no significant difference on performance, per source note).
Written when the claim was created (6 October 2026); not forced by an attack
Evidence in the project: Hattie & Timperley 2007 (feedback at four levels; task, process and self-regulation feedback effective, feedback about the self, such as praise, least effective); Kluger & DeNisi 1996 (meta-analysis of 607 effects from 131 papers, lab, work and school tasks, full text read: feedback raised performance by d = 0.41 on average but over 38 % of effects were negative; in the trimmed set of 470 effects, feedback that included praise did much less, d = 0.09 (80 effects) against 0.34 without, and the quarter of feedback most threatening to self-esteem d = 0.08 against 0.47 for the least threatening quarter; but normative feedback, such as a grade or comparison with others, was not a significant moderator, process feedback was too rare to test (6 effects) and no delayed-learning moderator was coded, so it compares feedback with no feedback, not comments with grades, and does not test this claim's kill line; corrected 2026-11-09); Lipnevich & Smith 2009 (464 first-year university students revising a course-exam essay one week after the draft, feedback in hand: detailed comments raised the revised score strongly, η² = .24; cell means 83.85 with instructor comments alone against 75.11 with a grade alone and 74.38 with praise alone; the outcome was a revision of the same essay, and the authors state there was no later assessment or similar later task, so it supports the direction of this claim outside its "later task or test" scope, not a test of the kill line; its grade-related effects are small and are held in open question 24); Wisniewski, Zierer & Hattie 2020 (meta-analysis: high-information feedback about task and process gave much larger effects than reinforcement or punishment, d ≈ 0.99 versus d ≈ 0.24 per source note); Shute 2008 (review: focus feedback on the task, not the learner; praise sparingly; avoid grades that crowd out comments); Black & Wiliam 1998 (formative assessment).
Written when the claim was created (6 October 2026); not forced by an attack
Wording changed later: Added with attack #2 (corrected; it did not narrow the claim) · Kluger & DeNisi 1996 · 9 November 2026.
These are reviews and meta-analyses that mix immediate and later outcomes.
Written when the claim was created (6 October 2026); not forced by an attack
Kill-condition test survived 2026-11-14: Abe & Usami 2024 (Japanese Journal of Behaviormetrics 51(1), 45–54, DOI 10.2333/jbhmk.51.45; full text and the authors' per-study coding table read), a meta-analysis restricted to post-tests given some time after the feedback (9 studies, 32 effects, school settings, mostly 1970–1987 plus one 2017 university study; USA, Israel, Germany, Venezuela): teachers' written comments beat a numeric grade alone, g = 0.42 (95% CI 0.17 to 0.66), with very high heterogeneity (I² = .90); praise-type comments ("very good") did .53 worse than task-level comments (p = .04).
Added with attack #3 (survived; it did not narrow the claim) · Abe & Usami 2024 · 14 November 2026
Age bound (limit, not a narrowing): the advantage shrank by about .07 per year of the learners' age (p < .001); the large effects come mainly from Butler's grade 5–6 experiments with a later session within about two days, while in the longest classroom study (Krampen 1987, maths over a school year, about age 14) subject-matter comments beat grades by only 0.22 and 0.00.
Added with attack #3 (survived; it did not narrow the claim) · Abe & Usami 2024 · 14 November 2026
The praise-alone comparison on a later test has still not been tested in this project.
Added with attack #3 (survived; it did not narrow the claim) · Abe & Usami 2024 · 14 November 2026
2 of 8 bounds and carve-outs were forced by a specific attack. 1 attack hit the boundary and did not move it: #3.
How it changed
The ledger records 2 attacks (and 1 correction to our notes), but no earlier wording of C4 is on record, so the text cannot be replayed step by step. Only the current wording is shown.
Written comments that tell learners what is right or wrong in their work and how to improve it (task and process feedback) lead to better performance on a later task or test than a grade or score given on its own, or than praise on its own, at least on open-ended tasks. On a simple speed task, one classroom experiment found grades alone did as well as comments (Butler & Nisan 1986).
What would change this
What would kill it: A meta-analysis, or a well-powered randomised study, in which learners given a grade or score alone, or praise alone, perform as well as or better than learners given task and process comments on a later task or test (a delayed test or a later piece of work done without new feedback). Reworded 2026-10-16 from the bank wording (meta-analysis where generic or ego feedback systematically beats task/process feedback on delayed learning in K–12), which kept producing fruitless searches.
Attack record
Reworded on 16 October 2026 (a change of wording, not an attack, so it is not counted). After 9 nightly searches in a row found no study that could test the old wording: no study found where a grade, score or praise alone was compared with task/process comments on a later test in the form the old wording and bank kill sentence asked for (generic/ego feedback systematically beating task/process feedback, K–12 meta). The new wording names the comparison the library's studies run (Butler 1988; Koenka et al. 2021; Panadero & Sánchez-Iglesias 2025 holds comments constant and varies grades) and keeps the claim's spirit. No new data added.
2 attacks: 1 survived, 1 narrowed. Counts of attacks, not a score.
Full attack list (2)
In a class-randomised experiment with 261 sixth-graders, pupils who had received only numerical grades did as well as pupils who had received task-related comments on a later quantitative task done without feedback, which meets C4's kill-condition for that kind of task.
Strongest reading and verdict reasoning
Steelmans to: C4 dies where grades alone match task/process comments on a later task done without new feedback; Butler & Nisan ran exactly that comparison on two task types.
Why: Week run 2026-10-16, cycle 1 (widened-search slot) Europe/Stockholm. Full text read (not stored). Nine classes, three per condition (comments, grades, no feedback). Session 3 work, no feedback expected. Task A (quantitative word construction): Session 3 scores similar for comments and grades, both above no feedback (t = 12.30, p < .001); comments above grades only on the long-words subscore (p < .013). Task B (qualitative divergent thinking): comments above grades and no feedback (t = 15.68, p < .001). Kill withheld because comments won on the open-ended task; narrowed to open-ended tasks. Limits: nine classes, analysed at pupil level; "later" is hours to two days. Not a preprint. Resets C4's fruitless streak.
In the only meta-analysis restricted to post-tests given some time after the feedback, comparing teachers' written comments with a numeric grade alone in school settings (9 studies, 32 effects), learners given only a grade would do as well as learners given comments — which would meet C4's kill-condition for the comments-versus-grade comparison.
Strongest reading and verdict reasoning
Steelmans to: C4 dies if, across studies that measure a later test rather than a same-task revision, a bare grade matches written comments; the earlier evidence behind C4 mixed immediate and later outcomes, and an independent random-effects synthesis with a moderator analysis (age, feedback type, test type, stakes) and very high heterogeneity is the fairest available test — it could show the classic advantage shrinking to nothing once old, small or praise-type studies are weighed.
Why: Nightly-attacks 2026-11-14 Europe/Stockholm (widened-search slot; C4 had 7 consecutive fruitless searches). Full text: inclusion required written teacher feedback on a test or work, a verbal-feedback group versus a numeric-rating group, a post-test given after an interval following the feedback, and a school setting (pp. 46–47). Random-effects g = 0.42 (95% CI 0.17 to 0.66, p = .002), τ² = .39, I² = .902; fixed-effect 0.38; Bayesian 0.42 (0.17 to 0.67); Egger intercept p = .70 (pp. 50–51). Meta-regression (Table 1, p. 52; reference task-level comments): age −.07 per year (−.12 to −.04, p < .001); self-level (praise-type, e.g. "very good") −.53 (−1.03 to −.04, p = .04); process-level .04 (ns, 3 effects); factual/procedural tests −.47 (p = .06); high-stakes post-test −.41 (p = .07). Kill fails: comments beat grades on later tests, and task-level comments did better than praise-type comments. Caveats: 9 studies, mostly 1970–1987, USA/Israel/Germany/Venezuela; several unpublished dissertations; the largest task-comment effects are Butler's Israeli grade 5–6 experiments (later session within about two days); in the longest classroom study (Krampen 1987, maths over a school year) subject-matter comments beat grades by only 0.22 and 0.00 (OSF table); the advantage shrinks with age. Not a duplicate (prior entries: Butler & Nisan 1986 weakened; Kluger & DeNisi / Lipnevich & Smith corrected). Not a restatement of the open-ended-task bound (this tests the main comparison). Not a preprint. Claim field and kill line unchanged; Bounds get an evidence line and the age limit; the practical rule Limits updated. Status unchanged (weakened). Resets C4's fruitless streak (7 → 0).
Corrections to our notes
These fix a number or a fact in our notes. They are not attacks and do not count toward the floor of 3 survivals.
- — C4's limits summarised Kluger & DeNisi 1996 from the source note and implied that feedback drawing attention to the self (including normative feedback such as grades) weakens effects; the full text shows praise and self-esteem threat do, normative feedback does not significantly, and the meta-analysis does not compare comments with grades.
Details
What was checked: precision of the Kluger & DeNisi 1996 Bounds line and the evidence behind the comments-versus-grade/praise comparison; not a kill-condition test (neither study measures a later task or test done without new feedback).
Why: Nightly-attacks 2026-11-09 Europe/Stockholm. Kluger & DeNisi: 131 papers, 607 effects, weighted d = 0.41, over 38 % negative (p. 258); after all exclusions 470 effects, d = .38, 32 % negative; Table 2 (p. 273): praise yes 80 effects d = .09 vs no 358, .34; self-esteem threat top quartile .08 vs bottom .47; normative FI not significant ("we did not find an effect for normative FIs", p. 277); "Only six effects were available for process FI" (p. 272); no retention-interval moderator coded. Lipnevich & Smith: N = 464 (463 in Table 4), 3 × 2 × 2 (detailed feedback none/computer/instructor × grade × praise); detailed feedback F(2, 450) = 69.23, p < .001, η² = .24 (p. 326); Table 4 (p. 327) cells: instructor comments, no grade, no praise 83.85 (n = 39); no comments, grade, no praise 75.11 (n = 38); no comments, no grade, praise 74.38 (n = 40). Limits, p. 331: "The students were not assessed later, and they were not given a similar task at a later time." Grade-related effects are cited only via open question 24 (the mining recomputation found them weaker than printed). Claim text, Status and kill-condition unchanged. Not a survival; does not count toward the floor; does not reset C4's fruitless streak. C4 returns from rest under the targeting rules(a) on these verified full-text 3g flags. Lipnevich & Smith 2009 added to the library (FB-011) with a C4 link.
Evidence
14 linked sources: 11 supports, 3 limits (first-pass draft, pending human review). supports: Black & Wiliam 1998, Hattie & Timperley 2007, Kaliisa et al. 2025, Kluger & DeNisi 1996, Shute 2007, Wisniewski et al. 2020, Koenka et al. 2021, Butler 1988, Byron & Khazanchi 2012, Lipnevich & Smith 2009, Abe & Usami 2024 · limits: Fyfe et al. 2021, Panadero & Sánchez-Iglesias 2025, Butler & Nisan 1986