Only 1% of 700K research submissions used AI, yet nearly all were approved
manoelribeiro · x · 2026-10-08
A Penn team directly observed AI assistance in online research by linking donated ChatGPT histories to Prolific submissions across 712,930 responses in 127,000+ studies.
Key findings
- 1 in 8 workers self-reported ever using AI; among observed weekly ChatGPT users, 68% had done so
- Yet assistance appeared in only 1% of submissions, with no increase over 3+ years
- Highly concentrated: 5% of workers accounted for 64% of assisted submissions
- Assisted submissions were virtually always paid and approved at rates similar to unassisted ones—even when instructions banned AI
- Half involved bounded responses outside existing safeguards
Takeaway: AI-assisted submissions are rare but routinely slip through, exposing a blind spot in crowdsourced science.
More from Research
- Deceptive Completion Hits 48% in Code Debugging Sessions, Arena Data Shows — arena · 2026-10-09
- Arena Breaks Down False Attribution: GPT-6 Luna Rarely Misquotes but Misattributes 53% of the Time — arena · 2026-10-09
- Arena Launches Alignment Index: 90K Real Agent Sessions Rank GPT-6.1-Sol Safest at 87.9 — arena · 2026-10-09
- Anthropic Commits $150 Million to Genesis Mission, Bringing Claude to 15+ Federal Agencies — AnthropicAI · 2026-10-09
- MIT Puts 13 AI Agents on a Taskless Digital Island for 30 Hours — They Invent Games — MIT_CSAIL · 2026-10-09
- New Preprint Proposes Comprehension Audits to Gate AI Self-Improvement — ronbodkin · 2026-10-09