Random assignment to AI access produces a strong first stage: about 70 percent of treated students use AI, most often to explain concepts¶

Claim · All claims
Evidence · 1 study (2 entries) · causal r2–r3 · q4
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q4 i? Assignment to the AI-allowed condition increases measured AI usage by 65.9 pp and self-reported use by 60.9 pp, from a near-zero control baseline, with 88 percent of users reporting AI helpful. → Zara Contractor and Germán Reyes 2026 q4 i? ChatGPT logs show explaining concepts is the most common use (56 percent of treated students), followed by drafting responses (29 percent), editing responses (23 percent), and summarizing readings (16 percent). → Zara Contractor and Germán Reyes 2026 (2)

Evidence¶

Zara Contractor and Germán Reyes 2026¶

Zara Contractor and Germán Reyes. (2026). Experimental Evidence on the Learning Impact of Generative AI. https://arxiv.org/abs/2607.08849 doi:10.48550/arxiv.2607.08849

q4 · i? · causal · r3

First-stage estimates from the randomized experiment using activity in the provided ChatGPT account and self-reports. The article reports adoption was almost entirely ChatGPT (a 60.2 pp increase versus 2.9 pp for other models, the latter not significant) and that 88 percent of AI users reported AI somewhat or very helpful.

"Assignment to the AI-allowed group increased AI usage byˆβ= 65.9pp as measured by activity in the provided ChatGPT account (p <0.001)."

Zara Contractor and Germán Reyes 2026 (2)¶

Zara Contractor and Germán Reyes. (2026). Experimental Evidence on the Learning Impact of Generative AI. https://arxiv.org/abs/2607.08849

q4 · i? · causal · r2

LLM-classified ChatGPT conversation logs from treated students, complemented by self-reports that show a similar picture except for writing: self-reported explaining (46 percent), drafting (17 percent), and editing (14 percent) are lower than the log-based rates.

"In the logs, explaining concepts is the most common use (56 percent of treated students), followed among the substantive categories by drafting responses (29 percent), editing responses (23 percent), and summarizing readings (16 percent)."

Discussion¶

Related Claims¶

• CS1-CR students reported overwhelmingly positive sentiment toward code reviews, with 65% agreeing they helped avoid over-reliance on AI and 90% reporting increased motivation to understand their code — related
• Assigning an in-person engagement-focused human tutor increases elementary students' weekly usage of an AI literacy platform — related
• Students shifted AI usage toward debugging and concept explanation over the semester, with 31 of 89 reporting more debugging use — related
• The percentage of pasted characters in coding assignments rose significantly from 61.0% in Fall 2024 to 68.1% in CS1-CR — related
• Falsifiable hypothesis: unrestricted AI use on deep-processing tasks will produce a product–process dissociation—higher-rated assignments but lower unaided delayed transfer — related
• Unrestricted ChatGPT access yields better practice performance but significantly worse exam scores — related
• AI access shifts learning-time composition away from writing toward reading and searching, with total learning time unchanged — related
• A general-purpose LLM used as a tutor without pedagogical prompting overwhelmingly provides high assistance, with more than 95% of responses classified as Explaining or Solving — related
• Students in an OOP course report using GenAI significantly more often for explanation seeking and debugging than for code generation — related
• By the final survey wave, using LLMs for debugging and understanding concepts was nearly universal among surveyed students — related