The post-ChatGPT response-time divergence on AI-susceptible items vanishes entirely under proctored assessment¶

Claim · All claims
Evidence · 1 study (2 entries) · causal r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q2 i? In non-proctored ALEKS PPL assessments, response times on AI-susceptible topics declined 1.11% per quarter relative to graph-based topics (11.6% cumulative). → Rismanchian 2026 q2 i? In proctored assessments, the post-ChatGPT slope and cumulative effect are effectively zero and statistically indistinguishable from zero. → Rismanchian 2026 (2)

Evidence¶

Rismanchian 2026¶

Rismanchian, S., Uzun, H., Matayoshi, J., Cosyn, E., & Kurd-Misto, E. (2026). Faster Completion, Less Learning: Generative AI Reduced Study Time on Math Problems and the Knowledge They Build. arXiv preprint. https://arxiv.org/abs/2605.21629 doi:10.48550/arxiv.2605.21629

q2 · i? · causal · r2

Ramp models on the ALEKS PPL placement-assessment panel of 12,190,662 response-time observations across 70 items, split by supervision condition. Non-proctored estimate "β=− 0.0112, p< 0.001", an 11.6% cumulative reduction.

"In non-proctored assessments, response times for AI-susceptible topics decline by 1.11% per quarter relative to graph-based topics after ChatGPT's release (β=− 0.0112, p< 0.001)"

Rismanchian 2026 (2)¶

Rismanchian, S., Uzun, H., Matayoshi, J., Cosyn, E., & Kurd-Misto, E. (2026). Faster Completion, Less Learning: Generative AI Reduced Study Time on Math Problems and the Knowledge They Build. arXiv preprint. https://arxiv.org/abs/2605.21629. doi:10.48550/arxiv.2605.21629

q2 · i? · causal · r2

Identical fixed-effects ramp models estimated on the proctored subset of the ALEKS PPL response-time panel (about 4.5 million proctored responses) show no post-ChatGPT divergence. The authors treat this as a falsification test ruling out general efficiency gains.

"In proctored assessments, the estimated post-ChatGPT slope and cumulative effect are effectively zero and statistically indistinguishable from it"

Discussion¶

Related Claims¶

• The post-ChatGPT learning-time decline follows an age gradient: largest in high school (31.3% cumulative), small in middle school (9.0%), and undetectable in Grade 5 — related
• College students' learning time on AI-susceptible word problems declines 2.80% per quarter relative to graph-based problems after ChatGPT's release, cumulating to 26.9% over eleven quarters — related
• The same retention estimator applied to non-proctored assessment yields an opposite-signed 85% increase in odds of correct response on AI-susceptible items — related
• On randomly assigned proctored retention items, odds of correct response on AI-susceptible items decline 25.0% cumulatively after ChatGPT's release — related