Syntes: Liu (2026): about ten minutes of AI help raised assisted performance but lowered unaided performance and persistence (preprint)

Källa: Liu, G.; Christian, B.; Dumbalska, T.; Bakker, M. A.; Dubey, R. (2026 (arXiv preprint, v5)). AI Assistance Reduces Persistence and Hurts Independent Performance. arXiv:2604.04721 (v5). PREPRINT, not peer reviewed (the preprint rule). https://arxiv.org/abs/2604.04721

Verified content (abstract (arXiv v5 page); the 2026-10-06 nightly run also assessed it, checked 2026-10-13)

• Preprint flag (the preprint rule): arXiv preprint, not peer reviewed, five versions so far (v1–v5). If any nightly verdict ever relies on it, the ledger entry must be flagged (preprint) and added to "Preprints to recheck".
• Design (abstract): a series of randomised controlled trials, N = 1,222, across tasks including mathematical reasoning and reading comprehension.
• Findings (abstract): AI assistance improved short-term performance, but afterwards people performed significantly worse without AI and were more likely to give up. The effects appeared after brief interactions (about 10 minutes).
• Authors' mechanism (stated as a hypothesis): AI conditions people to expect immediate answers, so they do not work through challenges themselves.
• Limits: no delayed test of learning is reported in the abstract; participant population not described in the abstract (age tag is an assumption: adult online participants).

Why it is in the library

Causal evidence for the mechanism behind C2 (answer-giving AI raises assisted scores but lowers unaided work). The 2026-10-06 nightly run declined it as an attack because there is no delayed test, so it supports C2's mechanism without testing its later-learning clause.