AI-Assisted Submissions in Online Research Are Rare and Highly Concentrated but Routinely Approved
A · Strong evidence University of Pennsylvania and HPI Potsdam study that directly observes AI use by linking donated ChatGPT histories to 712,930 Prolific submissions, a stronger method than the self-reports and detectors used in earlier estimates.
Summary based on the full text·Original: about 20 minutes to read
In short
An arXiv preprint (7 October 2026) by Neil K. R. Sehgal, Manuel Tonneau, Dunigan Folk, Lyle Ungar and Sharath Chandra Guntuku (University of Pennsylvania, Hasso Plattner Institute). It measures how often participants on Prolific, a major online research platform, use ChatGPT to answer studies. Many workers have done it at least once, but AI assistance appears in only 1% of submissions and is concentrated in a few people. Researchers almost never catch it: assisted work was paid as often as unassisted work, even when AI use was banned.
Key numbers
- 1.00%share of 712,930 Prolific submissions with observed ChatGPT assistance
- 64.1%share of assisted submissions coming from the top 5% of workers (21 of 408)
- 99.6%approval rate for AI-assisted submissions (99.70% for unassisted)
- 60 of 60submissions paid despite instructions explicitly prohibiting AI use
As reported by the source.
English·Web page·opens arxiv.org
What it reports
- Survey of 2,500 workers: 11.31% said they had ever used AI on a Prolific submission when asked directly; an anonymous list experiment estimated 14.24%. About 63% used ChatGPT at least weekly.
- Among 408 weekly ChatGPT users whose histories were linked, 68.4% had used ChatGPT on a study at least once, but only 13.5% did so on 2% or more of their submissions.
- Assistance appeared in 7,100 of 712,930 submissions (1.00%, 95% CI 0.72%-1.32%), with no upward trend from December 2022 to February 2026.
- The top 5% of workers produced 64.1% of assisted submissions, the top 10% 78.2%; excluding the top 5% cuts the rate to 0.40%.
- Use was localised: 40% of assisted submissions had a single ChatGPT exchange.
- 52.6% of assisted submissions involved bounded answers (choices, ratings), outside the scope of the platform's text-based AI detector; 55.1% involved subjective or personal responses.
- Approval rate: 99.60% for assisted submissions, 99.70% for unassisted; all 60 submissions with explicit AI bans were paid. A one-standard-deviation rise in daily ChatGPT use doubled the odds of assistance (OR 2.09).
How they got it
Two studies of US, UK and Canadian Prolific workers. A survey with a randomised direct-question versus list-experiment design, then a linkage of previously donated ChatGPT histories to platform submission records for 408 users, covering more than 127,000 studies. A validated classifier flagged ChatGPT exchanges that overlapped in time with a submission and materially contributed to the response.
Limits
- Only ChatGPT histories were observed; use of other assistants or agents is not captured, so 1% is a lower bound for these workers.
- The linked sample is weekly ChatGPT users who agreed to donate histories, which may differ from the wider pool.
- One platform; results may not hold for less curated panels.
Why it matters for companies
Companies increasingly rely on online panels for market research, employee-style surveys and human feedback for AI evaluation. This study suggests contamination is low on average but concentrated in a few heavy users, so screening for those individuals removes most of it. It also shows that bans in instructions have no enforcement effect when payment is not tied to compliance. Any team buying human-generated data should ask suppliers how they detect assistance on bounded answers, not only open text.
Summary based on the full text. Read the original for the full argument and data.Read the original on arxiv.org