C2 weakened
When learners outsource practice work to an AI tool that supplies finished answers or solutions (homework-outsourcing / unprotected answer-giving use), practice scores go up but learning measured later without the tool is lower. Mere access to an answer-capable chatbot does not by itself establish that loss: in a higher-education RCT, students given off-the-shelf GenAI access still scored somewhat higher on an unaided knowledge test one week later, with no sign of a loss (about 0.21 SD in the current version of the paper, not statistically significant at the usual 5 % level, p = .068; a working paper), and the gain held best among students who used AI mainly to explain concepts and disappeared among those who had it draft finished text (Contractor & Reyes 2026). The loss is also not shown, and was reversed, when the AI's finished output is itself a visible worked example of the skill being practised: adults who practised rewriting cover letters with an AI tool that instantly handed back a revised letter later wrote better without AI than those who practised alone, immediately and one day later (Lira et al. 2025, preprint).
Bounds: Evidence in the bank: Bastani et al. 2024/2025 (upper-secondary maths RCT, Turkey: answer-giving GPT Base +48 % correct answers during practice, −17 % on the closed-book exam compared with pupils who never had GenAI; kill-condition test survived 2026-09-27; correction 2026-10-26: the closed-book exam was the third part of the same 90-minute session, right after practice and a brief teacher review of the answers, so it measured learning minutes later, not days later; the authors state that their analysis focuses on short-term rather than long-term learning and defer long-term learning to future work; the control arm had no laptops; checked in the authors' version stored in the project, the PNAS 2025 text was not compared). The claim's "measured later" rests on studies with longer delays, such as Strömberg, Lei & Wu 2026 and Bergh et al. 2026; Strömberg, Lei & Wu 2026 (secondary-school DiD in one Chinese county, conference paper, results may be revised: homework scores +18 %, monthly closed-book tests −20 % within six months, entrance exams −18 to −24 % with the full penalty visible after about two years); Fan et al. 2025 (university lab, second-language essay: ChatGPT raised essay scores but not knowledge gain or transfer; process maps showed fewer orientation/evaluation links than a human-expert condition — the authors call this metacognitive laziness); Barcaui 2025 (HE RCT, Brazil: unrestricted ChatGPT as study aid → lower surprise retention at 45 days vs traditional study, d = 0.68; effect held after controlling for study time; kill-condition test survived 2026-09-27). Kreijkes et al. 2026 (England Year 10 RCT: LLM-only study of informational text → lower 3-day literal retention d=0.44, comprehension d=0.38, free recall d=0.21 vs note-taking; kill-condition test survived 2026-09-29). Bergh et al. 2026 (HE CS experiment, arXiv: ChatGPT during coding → higher task scores but lower immediate and 48 h recall without the tool, g≈0.71–0.77; kill-condition test survived 2026-09-29). The loss is concentrated in "homework outsourcing" use (about 80 % of AI users in Strömberg et al.); users who keep study time similar to non-users show small losses. The claim is about learning measured without the tool; performance with the tool goes up. Positive pooled effects in ChatGPT/chatbot meta-analyses (Deng et al. 2025, Doo & Park 2026, Wu & Yu 2024, Fan et al. 2026) are recorded as challenges, because the bank notes that most of their outcomes were not measured without the tool. Contractor & Reyes 2026 (arXiv:2607.08849 / IZA DP 18792; full text verified 2026-10-03): AI-allowed undergraduates scored +0.27 SD on immediate unaided knowledge tests and +0.27 SD (~5.1 pp) one week later without AI (n=211/204); corrected 2026-10-16 to arXiv v2 (4 Oct 2026): immediate +0.28 SD (7.2 pp, p = .021, n = 210), one week later +0.21 SD (3.9 pp, p = .068, n = 205), augmentation users one week later +0.24 SD (p = .095), mixed users +0.28 SD (p = .220), automation users' advantage gone; augmentation (explain) users retained gains while automation (draft) Session One quality gains faded; scopes the living claim to outsourcing / unprotected answer-giving rather than mere off-the-shelf access. Lira, Rogers, Goldstein, Ungar & Duckworth 2025 (arXiv:2502.02880 v4, preregistered, Prolific adults; full text verified 2026-10-05; preprint): an AI tool that instantly rewrote the learner's cover letter according to taught writing principles raised practice quality (d = 1.01) and also unaided writing on a new letter without AI (d = .38 vs practice without AI; one day later d = .41; N = 2,238); practising without AI was no better than no practice (d = .05); a replication found d = .34 (24 h d = .29) and that merely seeing the AI-revised letter helped as much as practising with the tool; AI practice also beat personalised editor feedback (d = .20). Authors note that in maths or programming a final answer does not reveal the procedure; scopes C2 away from practice in which the AI's finished output is itself a worked example of the target skill. Kazemitabaar et al. 2023 (CHI, peer reviewed; 69 novices aged 10–17, Python with or without OpenAI Codex, each AI-assisted authoring task followed by a manual modification task): no loss on unaided tests — immediate authoring 61.3 % vs 62.9 % (p = .838), one week later 59.1 % vs 49.8 % (p = .262) and modifying 47.5 % vs 34.8 % (p = .092); single small study, generated code is itself a worked example and practice included a forced manual step; kill-condition test survived 2026-10-16 (not replicated, not the homework-outsourcing pattern). Felt-learning caveat (added 2026-10-27 from candidate K52; not a claim): learners' own sense of how much they learned can fail to track the unaided result — in Bastani et al., pupils with the answer-giving AI did not rate their learning or exam performance lower even though the closed-book exam fell; Fan et al. 2025 likewise found no intrinsic-motivation difference despite large essay gains and flat knowledge/transfer.
Attack record: 7 attacks: 5 survived, 2 narrowed · Evidence links: 28 linked sources (14 supports, 12 limits, 5 challenges)