Against: null results, contradictions and limits

This file gathers the evidence against the rules for writing and keeping flashcards. It covers null results, studies that contradict each other, failed replications, and practice that a study contradicts.

Whether learner-made cards win

Pan 2023b found learner-made cards better, d = 0.45 (secondary), as an estimate across six experiments.
Wilkinson 2020 found learners may do better with premade word cards (secondary).
Sage 2019 found more recall from ready-made than from self-made cards, with more effort reported (secondary).
A synthesis of 32 word-card studies found more studies with larger effects for ready-made cards, by counting, not pooling (Lei 2022, secondary).
The studies differ in material (text passages against word pairs) and in what making a card asked of the learner.
The generation effect itself is contested for texts: a meta-analysis found g = 0.41 (Schindler 2023, secondary), then the same team's seven experiments found no effect, some a disadvantage (Schindler 2025, secondary).
A direct replication found the classic text effect only in part, when generation was nearly perfect (Schindler 2024, secondary).
Its size depends on how constrained the generation is (McCurdy 2020); copying a definition word for word gave no advantage, while copy-and-paste did (Pan 2023b, experiments 1 and 2).

Whether flashcards help with authentic material

Flashcards did not beat free study on authentic course material (Lin 2018, experiment 1, secondary).
Premade psychiatry cards were found helpful by 87% of users and did not predict exam scores (Sun 2021b, secondary).
Repeated testing with electronic flashcards beat restudy at one week and not at six months (Schmidmaier 2011, secondary).
Commercial premade cards did not predict licensing scores where Anki cards, classed as user-made, did (Deng 2015).

Weaknesses of the strongest evidence

The medical studies are observational and self-selected: Deng 2015, Lu 2021, Wothe 2023, Gilbert 2023 and Levy 2023.
Lu 2021 respondents scored higher than non-respondents, 238.82 against 229.89 (d = 0.51), so the sample is skewed toward strong students.
Halperin 2024 recruited through social media and group chats; its 68.3% is not a national rate.
Ingebrigtsen 2025 randomised campuses, but compared users with non-users inside them; only one third used the cards.
Lin 2018, Senzaki 2017 and the vocabulary studies were read by abstract only; Pan 2023b was read in full by the audit, in the first author's copy.
Karrar 2026 compares cards plus a scheduler against usual teaching, so d = 2.48 cannot be credited to generated cards.

Whether the practitioner rules are right

The twenty rules cite no study and rest on one author's records (Wozniak 1999).
Matuschak 2020 states that research gives no exact guidance for writing prompts.
Guides repeat one another (Bjornstad 2020 cites Wozniak 1999; Anki 2026c sends readers to it); agreement among copies is not independent confirmation.
The Anki manual's background page cites unnamed studies, such as 75% forgotten within 48 hours (Anki 2026b, example only).
Pictures, recommended by Wozniak 1999 and Wyner 2014 (example only), did not beat translations in Carpenter 2012 until overconfidence was removed.
Semantic clusters, which the interference rules warn against (Wozniak 1999, rule 11; Nation 2000), did not lower scores in Nakata 2019 or Ishii 2015; they raised confusion errors.

Whether the mnemonic medium works

Matuschak 2019 reports no comparison with the same essay without prompts.
Its one comparison is informal: 8 questions, 16 and 25 users, two weeks, independence assumed.
The authors retracted one claim, on trivial opening questions, after a data analysis error.
Later data are in the author's working notes, which could not be read on the search date; they are not used.

Whether learners keep their decks

Required Anki use was met with reluctance and little enjoyment (Seibert Hanson 2020, secondary).
Only one third of nursing students given cards used them (Ingebrigtsen 2025, secondary).
Of 337 pathology trainees and pathologists offered spaced repetition material, 23% completed the evaluation (Barhdaoui 2026, secondary); the effect it found belongs to the core.
A developer reported that fewer than 20 of 4,000 users did a second session (Branwen 2019, citing a 2012 post; anecdote).
Q/A threads name the daily-use requirement and the backlog after a missed week as the main cost (LLSE 2017a).

Whether generated cards are safe

Generated cards still carried 1 error in 21 after prompt tuning (Khurana 2025, secondary); a learner who did not write a card may not notice its error (an inference).
Students' stated preferences for generated mnemonics disagreed with what helped them (Balepur 2024, secondary).
If making cards is the active ingredient (Pan 2023b), generating them for the learner may remove it; this is an inference, untested.

Q/A threads against

Flashcards give a shallow, context-free meaning of a word; combine them with reading (LLSE 2016b).
In mathematics, understanding is compressed through problems, not recall prompts (MESE 2014).
Professors do not use spaced repetition because their learning is conceptual and their facts are in daily use (AcSE 2024).