The claims

All claims

• C0 Retrieval beats restudy narrowed

  Challenged 12 times, still holds in a narrower form ·

• C1 What counts as learned narrowed

  Challenged 7 times, still holds in a narrower form ·

• C2 Answer-giving AI in practice alive

  Challenged 4 times, still holds as written ·

• C3 Paper vs screen, deep reading narrowed

  Challenged 4 times, still holds in a narrower form ·

Rules for learners and teachers

Each rule comes from one claim and is no stronger than its current wording and bounds.

1. Once something has been learned, spend review time recalling it from memory instead of rereading it. Check every answer against the correct one, and keep going until each item has come back correctly more than once.

   Only within these limits (7)

   • The payoff is on a later test, not one a few minutes away.
   • It does not rescue material that was never learned in the first place.
   • The practice questions have to match what must be remembered later.
   • Multiple-choice practice without feedback is excluded.
   • Low-stakes practice, not high-stakes testing.
   • The claim does not say the advantage keeps growing the longer you wait.
   • Does not transfer to novel meaningless abstract visuals without preexisting meaning to hang on.

   From C0: Retrieval beats restudy narrowed 12 attacks: 6 survived, 6 narrowed

2. Don't judge recall practice by a quiz a few minutes later. On an immediate test, rereading for the same time can come out ahead; the advantage of recalling is on a delayed test.

   Only within these limits (2)

   • The claim does not say how long "delayed" has to be (open question 1).
   • Same conditions as R1: learned material, matching practice, repeated correct recall with feedback.

   From C0: Retrieval beats restudy narrowed 12 attacks: 6 survived, 6 narrowed

3. Count something as learned only when it can be brought back later, without looking at the material again, from the same kind of prompt it will be needed for. Having read it, seen it, or followed it while it was in front of you does not count.

   Only within these limits (4)

   • "Learned" is relative to the prompt: success with one kind of question does not prove it for a different kind.
   • Many classic retrieval-practice studies did not check this item by item; this is a stricter standard than they verified.
   • With corrective feedback, retrieval can still beat restudy on a later test even when first-pass recall during practice is experimentally low (~22%); the "learned" bar is stricter than what that contrast needs (Gupta, Pan & Rickard 2022).
   • Without practice feedback, the retrieval-over-restudy advantage shrinks when practice success is experimentally low (~25–29%, d=0.43) versus high (~76–81%, d=1.71); it does not appear equally (Racsmány, Szőllősi & Marián 2020; kill-test survived).

   From C1: What counts as learned narrowed 7 attacks: 3 survived, 4 narrowed

4. While practising, don't use an AI tool that hands over finished answers or solutions. Practice scores go up, but less is learned when it is later measured without the tool, so check learning without the tool.

   Only within these limits (3)

   • Tutor modes that hold back direct answers are outside this claim.
   • Evidence so far covers secondary school to university (including England Year 10 LLM-vs-notes and HE programming with ChatGPT); not yet shown for younger pupils or in Swedish classrooms.
   • Performance with the tool goes up; the claim is about learning measured without it.

   From C2: Answer-giving AI in practice alive 4 attacks: 4 survived

5. When a continuous informational text must be understood in depth by readers beyond early primary school, paper gives somewhat better understanding than a scrolling screen on average. If it is read on screen, don't trust the feeling of having understood it; check. Do not treat that average as a rule for first-grade beginner readers of short texts.

   Only within these limits (11)

   • The average gap is small.
   • The gap is clearest under time pressure; with self-paced reading two meta-analyses found no significant difference (open question).
   • For narrative (story) text the difference is about zero or not significant.
   • Much of the gap is tied to scrolling; with paginated screens the differences are small or uncertain.
   • On tablets and e-readers the disadvantage is about half as large, and for schoolchildren it was not significant.
   • First-grade beginners: no screen-inferiority main effect in Florit et al. 2025; tablet can match or beat paper on some superficial measures.
   • Does not cover navigating and evaluating sources online.
   • Digital texts with added features (interactive e-books, AR, reading supports) are a different comparison: for children aged 1–8 story-congruent enhancements can beat paper, and interactive digital texts beat paper in pooled studies up to middle school.
   • Much of the evidence for the average is from studies up to 2017 and from university students; the newest meta-analysis (studies to 2022) finds a clear paper advantage only for university students.
   • Norwegian VET (Foss et al. 2026): whole-sample screen inferiority on informational laptop reading; clearer among lower performers.
   • Italian grade 7 lengthy informational texts (Ronconi et al. 2025): no medium main effect on literal/inferential comprehension with scrolling desktop; calibration not tested; kill-test survived.

   From C3: Paper vs screen, deep reading narrowed 4 attacks: 3 survived, 1 narrowed

Claim map

Questions at the top, claims below. Arrows show that one claim depends on or relates to another. Each claim shows its attacks as dots, oldest first.

survived narrowedone dot per attack, oldest first

Where it's weakest

1. C2 Answer-giving AI in practice: 4 open challenges, 4 attacks
2. C0 Retrieval beats restudy: 2 open challenges, 12 attacks
3. C3 Paper vs screen, deep reading: 0 open challenges, 4 attacks
4. C1 What counts as learned: 0 open challenges, 7 attacks

Where the next attack should aim →