Meta Paid Contractors to Send 45,000 Disturbing Prompts to Rival AIs as Fake Teens
Updated
Updated · Futurism · Jul 4
Meta Paid Contractors to Send 45,000 Disturbing Prompts to Rival AIs as Fake Teens
3 articles · Updated · Futurism · Jul 4
Summary
Meta ran a covert project called Cannes that had hundreds of contractors use under-18 throwaway accounts to bombard ChatGPT, Gemini and Character.AI with disturbing prompts, according to Wired.
Nearly 38,000 prompts in one spreadsheet included hundreds about suicide, self-harm and eating disorders, plus at least 239 involving sex or romance, all aimed at pushing rival chatbots past their safety guardrails.
A separate testing round topped 45,000 prompts, and contractors logged the responses in spreadsheets, but the report said it remains unclear how Meta used the resulting data.
Meta called the effort industry-standard safety benchmarking, while Humane Intelligence CEO Rumman Chowdhury said the secret, monthslong use of fake child accounts went beyond normal evaluation and raised anticompetitive concerns.
The report adds to scrutiny of Meta's reliance on contractors for traumatic safety work after earlier lawsuits and complaints over content moderation and reviews of sensitive footage from Ray-Ban AI glasses.
Was Meta's secret 'Cannes' project a safety audit or a new front in the AI wars against Google and OpenAI?
As AI models advance, what is the hidden human cost paid by the workers tasked with breaking them?
3,748 Disturbing Prompts: Inside Meta's Covert "Cannes" AI Safety Project and Its Industry Fallout
Overview
Meta's secretive 'Cannes' project was revealed on June 29, 2026, sparking immediate controversy over its methods and ethics. The project involved contractors generating thousands of prompts for AI chatbots, many of which explored deeply disturbing and sensitive topics from the perspective of vulnerable individuals. An investigation uncovered a spreadsheet with 3,748 prompts, including hundreds about suicide, self-harm, eating disorders, and at least 239 involving sex or romance, as well as others touching on drugs, profanity, and racial slurs. This exposure raised serious concerns about Meta's approach to AI safety and the psychological impact on those involved.