In a randomized experiment on fraction problems, participants who used an AI assistant solved fewer unassisted test problems and skipped more than control participants¶

Claim · All claims
Evidence · 1 study (2 entries) · causal r2 · q3 · i1–i2
Currency · newest study 2026 · checked against its sources 2026-10-15 (automated, 2 of 2 entries) · not reviewed by a person

Subclaims¶

q3 i2 After AI removal, the AI condition had a lower unassisted solve rate than controls (Cohen's d = -0.42). → Grace Liu 2026 q3 i1 The AI condition also skipped more test problems than controls (Cohen's d = 0.25), indicating reduced persistence. → Grace Liu 2026 (2)

Evidence¶

Grace Liu 2026¶

Grace Liu, Brian Christian, Tsvetomira Dumbalska, Michiel A. Bakker, Rachit Dubey. (2026). AI Assistance Reduces Persistence and Hurts Independent Performance. Published as a conference paper at COLM 2026. https://arxiv.org/abs/2604.04721 doi:10.48550/arxiv.2604.04721

q3 · i2 · causal · r2

Randomized controlled experiment (N=354 recruited, 307 analyzed) on Prolific; AI condition solved 12 fraction problems with a GPT-5 sidebar assistant, then 3 unassisted test problems. The AI condition's test solve rate was lower, "Cohen's d=− 0.42".

"Participants in the AI condition had a lower solve rate (mean 0.57, s.d. 0.41) than participants in the control condition (mean 0.73, s.d. 0.34; t(305) =− 3.64, P< 0.001; Cohen's d=− 0.42, 95% CI [−0.66,−0.19])."

Grace Liu 2026 (2)¶

Grace Liu, Brian Christian, Tsvetomira Dumbalska, Michiel A. Bakker, Rachit Dubey. (2026). AI Assistance Reduces Persistence and Hurts Independent Performance. Published as a conference paper at COLM 2026. https://arxiv.org/abs/2604.04721

q3 · i1 · causal · r2

Same RCT; skipping was a deliberate decision not to engage since there was no penalty for wrong answers. The AI condition's test skip rate was higher, "Cohen's d= 0.25".

"Additionally, participants in the AI condition had a higher skip rate (mean 0.20, s.d. 0.36) than participants in the control condition (mean 0.11, s.d. 0.29; t(305) = 2.16, P= 0.031; Cohen's d= 0.25, 95% CI [0.02, 0.48])."

Discussion¶

Related Claims¶

• How genAI is used, not who uses it, predicts whether independent performance declines or improves — the general claim this finding bears on, with the studies behind it
• Experiment 2 replicates the unassisted-performance impairment with a pretest-based exclusion, but the aggregate skip-rate difference does not reach significance — related
• AI assistance impairs unassisted performance and persistence in reading comprehension, replicating the fraction-task effects in a second domain — a narrower finding that bears on this claim
• AI assistance reduces persistence: persistence costs concentrate among learners who used AI for direct solutions, not hints — a broader claim this one bears on
• AI access raises immediate unaided test scores by 7.2 percentage points (0.28 SD) after a timed learning phase — reports the opposite
• Unrestricted GPT-4 access in high-school mathematics raised assisted practice performance but left students worse off after withdrawal — related
• In math problem solving, a tutor-plus-peers LLM arrangement produced the strongest unassisted post-task test accuracy, with accuracy increasing monotonically from Control to Tutor + Peers — related
• Effective intervention timing in AI-based instruction does not always mean immediate response; the interval between student difficulty and teacher intervention can allow independent problem solving — related
• High-school students given GPT-4 access solved more problems in the moment but performed worse than controls once the tool was removed (review attribution) — related
• In math problem solving, a tutor LLM alone significantly improved unassisted test accuracy over no-agent control, while peer-like error agents alone did not reach significance — related