Study finds lower independent performance after AI assistance is removed
Three online randomized experiments involved 1,222 people; the findings do not establish effects beyond the short tasks tested.
Event / data period: 2026-10-09
Published: · Updated:
A team that included UC Berkeley researchers presented an updated, peer-reviewed study at the Conference on Language Modeling during the week of October 9. Three online randomized experiments enrolled 1,222 participants: two used fraction problems and a third used SAT reading-comprehension questions. AI access was later removed from the assisted groups.
After removal, participants who had AI access solved fewer test questions correctly than control groups; skipping also rose in some experiments. The findings are limited to short, structured tasks. The study does not establish that everyday use causes long-term or permanent cognitive harm; more research is needed to measure how different forms of assistance affect learning.
Data
Experiment participants
Participants recruited through an online research platform in the United States for three randomized experiments; final analysis samples were smaller.
- Participants in fraction-solving experiment 1
- 354 people
- 2026-10-09 · Observed
- Participants in fraction-solving replication
- 667 people
- 2026-10-09 · Observed
- Participants in SAT reading-comprehension experiment
- 201 people
- 2026-10-09 · Observed
Methodology
Participants were recruited through the US-based Prolific online research platform. Values are total recruited sample sizes by experiment; analyzed groups were smaller after attention and ability checks.
Updated:
Subjects and places
Subject proposals
- Artificial intelligenceSuggested
- Subject suggested from the title or summary: AInews.berkeley.edu ↗arxiv.org ↗
Geographic references
Only source-provided geographic references are shown. Opening an external map is optional. Optional story maps use OpenStreetMap. Opening a map sends your IP address and this site’s address to the map provider; location data remains readable as text without loading the map.
- United States · Place role: Coverage · Place type: Country or area · Location precision: RegionalEvidence
- Explicit preserved editorial country/region; no city or national centroid inferrednews.berkeley.edu ↗arxiv.org ↗