Persistence and performance costs concentrate among participants who used the AI to obtain direct solutions, while hint-users resemble controls¶

Claim · All claims
Evidence · 2 studies (4 entries) · 1 causal r2, 1 theoretical r3 · q2 · i1 small
Currency · newest study 2026 · checked against its sources 2026-10-15 (automated, 4 of 4 entries) · not reviewed by a person

Subclaims¶

q2 i1 AI-usage groups did not differ at pretest but differed significantly at test in solve rate (η² = 0.039) and skip rate (η² = 0.014). → Grace Liu 2026 q2 i1 Direct-solution users showed lower test solve rates than controls, hint-users, and non-users. → Grace Liu 2026 (2) q2 i1 Direct-solution users also skipped more at test than controls and hint-users. → Grace Liu 2026 (3)

q2 i? Participants who used AI primarily to obtain direct solutions showed declining independent performance relative to their own pretest, whereas those who used it for hints and clarification showed improved performance. → Kuhn 2026

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

q2 · i1 · associational · r2

Cross-sectional analysis of the AI-assisted subset of Experiment 2, grouping participants by self-reported usage (direct answers 61%, hints 27%, no usage 12%). Pretest ANOVAs showed no group differences; test ANOVAs were significant for both measures.

"However, a one-way ANOVA test revealed a significant difference between the groups for both test solve rate (F(3, 581) = 7.89, P< 0.001, η2 = 0.039) and test skip rate (F(3, 581) =2.80,P=0.039,η 2 =0.014)."

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

q2 · i1 · causal · r2

Pairwise t-tests between usage groups in Experiment 2. Direct-solution users (mean 0.65) scored below controls (d = -0.36), below hint-users (d = -0.29), and below non-users (d = -0.66). The authors note these cross-sectional findings are not necessarily causal.

"Participants who used the AI to get answers directly demonstrated lower subsequent task performance (mean 0.65, s.d. 0.38) compared to participants in the control group (mean 0.77, s.d. 0.32; t(464) =− 3.77, P< 0.001; Cohen's d=− 0.36, 95% CI [−0.54,-−0.17])"

Grace Liu 2026 (3)¶

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

q2 · i1 · causal · r2

Same pairwise analysis: direct-solution users skipped more than controls (d = 0.20) and more than hint-users (d = -0.31 for hints vs direct). Self-reported usage was validated against chatlogs: 83.8% of self-reported non-users did not message the AI at all.

"Furthermore, participants who used the AI to get answers directly had a higher skip rate (mean 0.13, s.d. 0.30) than participants in the control group (mean 0.07, s.d. 0.23; t(464) = 2.14, P= 0.033; Cohen's d= 0.20, 95% CI [0.02, 0.39 ])"

Kuhn 2026¶

Kuhn, J., Gerjets, P., Trautwein, U., Greene, J. A., Malone, S., Vogt, P., & Fütterer, T. (2026). AI-Augmented Inquiry and Regulation in Hybrid Systems: A Control Allocation Architecture for Preserving Epistemic Agency in Hybrid Human-AI Cognition. https://osf.io/preprints

q2 · i? · theoretical · r3

The article reports, citing Liu et al. (2026), a randomized controlled trial in which usage patterns diverged: "participants with AI access who used it primarily to obtain direct solutions (61%) showed declining independent performance relative to their own pretest." The groups did not differ at pretest. No effect size is printed in this article.

"participants with AI access who used it primarily to obtain direct solutions (61%) showed declining independent performance relative to their own pretest, whereas those who used it for hints and clarification (27%) showed improved performance."

Studies bearing on this claim¶

When the support can do the task for the learner (AI assistants), does it cost independent performance? This table gathers the studies behind this claim and the 20 narrower claims that bear on it: 20 distinct studies (7 causal, 5 theoretical, 3 associational, 2 review, 2 design, 1 qualitative), q1–q4, 3 of 20 with an effect size coded. One row per study, codes as each entry records them; no effect is averaged and no verdict is drawn. Built by scripts/synthesize_claim.py from eval/synthesis/; a row that looks wrong is fixed on the claim it comes from.

Discussion¶

Related Claims¶

• 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 — related
• In a randomized experiment on fraction problems, participants who used an AI assistant solved fewer unassisted test problems and skipped more than control participants — related
• The majority of AI-assisted participants self-report using the AI primarily to get answers directly, a pattern validated against chatlogs — related
• AI assistance reduces persistence: persistence costs concentrate among learners who used AI for direct solutions, not hints — possibly the same claim (merge candidate)
• Unguarded answer-giving AI harmed unaided exam performance while a guarded version of the same model erased the harm (Bastani et al., 2025, as reported) — related
• Intensive AI assistance predicts strong task performance even with low engagement, but follow-up performance is substantially higher only when intensive AI use is combined with sustained task engagement — related
• Access to an LLM assistant during practice did not significantly lower subsequent unaided test performance — related
• Students use standard GPT-4 access as a crutch—copying answers rather than engaging with the material—which explains the harm to learning — related
• Nearly all students valued learning through effort while feeling drawn toward AI, and split into those who limited use and those whose use conflicted with stated values — related
• Adaptive spaced retrieval practice produced higher end-of-semester posttest performance than learner-directed AI study, but fixed spaced retrieval did not significantly outperform learner-directed study — related
• Dual-threshold instability: accelerated breakdown when internal generative engagement and genAI reliability both fall below task demands — related
• Students use the integrated AI assistant differently across writing stages: clarification early, verification late — related
• AI-assisted speed becomes a recalibrated performance baseline, making manual work feel slow and inadequate — related
• After workshops, students' active GenAI use concentrated in uncertain design phases and decreased or disappeared in more certain phases — related
• The equalizing effect of AI assistance operates through both content and writing subcomponents, with incomplete convergence in each — related
• Unscaffolded GenAI access did not significantly change pre-service teachers' self-regulated learning relative to traditional learning — related
• Developers who used AI assistance scored 17% lower on a subsequent comprehension assessment than developers who completed the same tasks without AI, per the cited controlled study — a narrower finding that bears on this claim
• AI-assisted performance during skill-building does not transfer to subsequent unassisted performance — a broader claim this one bears on
• AI-generated homework hints were positively associated with exam performance only when exam questions closely matched the homework content — related
• The Coach's hint budget was managed poorly: students split into early-spenders and savers, and within the Coach condition only those with a deliberate hint strategy scored higher — related
• Descriptive user-study patterns that the authors read as GPT output possibly scaffolding IVDA students and reducing CV students' performance, an interpretation by prior knowledge that the study did not measure — related
• Chat-log analysis indicates lower-education participants obtain substantial assistance from AI while higher-education participants use the tool somewhat more effectively across several margins, explaining why the gap narrows but does not disappear — related
• LearnMate2 improved learning outcomes and user experience compared to a baseline combining an online learning platform and an LLM in a user study (n=16) — related
• Written examination scores were low and confounded, providing a serious warning about students' unaided performance on foundational calculations — related
• Not requesting any hints in at least one problem is related to greater wheel-spinning, consistent with help avoidance — 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) — a broader claim this one bears on
• LLM-assisted self-study yields higher post-intervention exam performance than traditional resources among early-career knowledge workers — related