The "no AI" clause in your research instructions does nothing. I used to assume it was weakly enforced. I now think it is pure decoration.

Last week a paper gave me the number. 99.6% of AI-assisted submissions to Prolific get paid. In the subset of studies that explicitly prohibited AI, 100% got paid. 60 of 60.

The head of insights who told her panel no AI

A workshop in London last Tuesday. The head of insights at a mid-market consumer brand walked the room through a 2,000-respondent segmentation deck. Clean quadrants. Confident clustering. She named the panel source: Prolific.

Then the line that stopped me. "We instructed respondents not to use AI in the open-text questions." She said it as if it mattered.

I asked how the vendor enforces that. She paused. The platform's terms repeat it. The study instructions repeat it. Nothing else.

The deck was going to the board on Thursday.

What the research says

Neil Sehgal and colleagues at Penn and the Hasso Plattner Institute ran the first audit I have seen that links behaviour to submission. Their cohort: 408 Prolific workers who donated three years of ChatGPT exchanges and three years of submission records. They counted 712,930 submissions across more than 127,000 studies, December 2022 to February 2026.

One in eight surveyed workers reported ever using AI on a study and 68% of observed workers had done so.

The gap is roughly five-fold. The survey about AI use is itself contaminated by social-desirability bias. If you survey your own panel, you under-measure by about the same order.

The second finding worth sitting with is the concentration. 5% of workers produced 64% of the AI-assisted submissions. Gini 0.856 against a baseline of 0.575. Twenty-one people doing it on a lot of studies. Spread unevenly. The kind of pattern a random QA sweep misses.

The last finding is the one for the exec. 99.6% of assisted submissions were approved. All 60 submissions in studies that explicitly prohibited AI were approved. Each one-SD increase in a worker's daily ChatGPT use raised their odds of assistance by OR 2.09.

The respondents most practiced at using AI are the ones you most want to catch, and the platform cannot see them. The approval queue is not detecting the thing you most need it to detect. Writing the prohibition into your brief is a signal to yourself, not to the respondent.

Your research function cannot name the 5% contaminating your panel. The "no AI" line in your instructions buys you one thing: the ability to tell the board you tried. The board gets the same number either way.