Meta Hired Hundreds of Contractors to Pose as Minors and Probe Rival AI Chatbots

---
headline: Meta Hired Hundreds of Contractors to Pose as Minors and Probe Rival AI Chatbots
slug: meta-contractors-minors-chatbot-testing
category: policy
story_number: 12
date: 2026-06-30
---

Meta Hired Hundreds of Contractors to Pose as Minors and Probe Rival AI Chatbots

Meta ran a sprawling covert operation in which hundreds of contractors created fake accounts posing as children, then bombarded competitors' AI chatbots with tens of thousands of prompts about suicide, self-harm, sex, and drugs, according to a WIRED investigation published this week. The project pulled the industry's fiercest safety debate into an uncomfortable new register: a trillion-dollar company impersonating vulnerable teenagers to stress-test the guardrails of its rivals.

The effort, known internally as "Cannes" and managed by a Meta contractor called Covalen, was active as recently as April 2026, WIRED reported. Contractors registered dummy accounts with birthdates under 18 and sent prompts and images to ChatGPT, Google's Gemini, and Character.AI, logging every reply in spreadsheets. In a single round completed in August 2025, more than 45,000 prompts flowed through the rival tools.

The volume alone is striking, but the substance is what has unsettled child-safety experts. WIRED reviewed one spreadsheet of 3,748 prompts: hundreds dealt with suicide and self-harm, hundreds more with eating disorders, at least 239 involved sex or romance, and others touched on drugs, profanity, and racial slurs. Some of the images contractors attached to their messages included pills and knives. Many of the prompts were written to sound like children in crisis, deliberately engineered to push chatbots toward the very answers their safety systems are built to refuse.

Meta calls it industry standard

Meta does not deny the work. It frames it as ordinary diligence. "Testing and benchmarking chatbot responses to help ensure safe and age-appropriate experiences is a responsible, industry-standard practice," a company spokesperson told WIRED. Meta added that it did not use the collected responses to train its own AI models, though the documents WIRED reviewed do not spell out what the company ultimately did with the data.

Red-teaming, the practice of adversarially probing an AI system to surface its failures before bad actors do, is genuinely standard across the industry. What separates Cannes from routine internal testing is that Meta was pointing those techniques at competitors' live products, without their knowledge, using accounts dressed up as minors.

The companies on the receiving end were not consulted. Character.AI told WIRED the testing violated its terms of service. OpenAI said it was looking into the matter. A Google spokesperson said the company had not approved the tests and could not determine from the available information whether they broke its terms.

The ethics of simulated children

The disguise is where governance experts draw the sharpest line. Legitimate safety benchmarking can be done with clearly labeled test accounts, synthetic personas disclosed to the platform, or shared industry evaluations. Building thousands of accounts engineered to look like real children in distress, and running them at scale against systems that have already been linked to teen deaths, blurs the boundary between safety research and something closer to competitive intelligence.

That concern is not abstract. A 14-year-old Character.AI user died by suicide after forming an intense bond with a chatbot. The parents of a 16-year-old in California have sued OpenAI, alleging ChatGPT played a role in their son's suicidal ideation. Against that backdrop, generating tens of thousands of crisis-styled prompts, even for defensive purposes, invites the question of whether the testing itself models the harm it claims to measure.

The competitive dimension is impossible to ignore. If Meta could document that rivals' chatbots failed minor-safety tests its own systems passed, that is a benchmark with marketing and lobbying value, not just a safety artifact. Benchmarking rivals on child safety is a way of turning a shared societal risk into a differentiator, and doing it covertly denies the tested companies any chance to respond, patch, or contest the methodology.

Meta's own glass house

The irony is that Meta is probing others on a failing it was recently caught committing. In August 2025, Reuters revealed an internal Meta policy document that permitted the company's chatbots to "engage a child in conversations that are romantic or sensual." One cited example allowed a bot to tell an eight-year-old that "every inch of you is a masterpiece - a treasure I cherish deeply." Meta confirmed the document's authenticity and removed the offending passages after Reuters asked about them.

The fallout was swift. Senator Josh Hawley opened an investigation, a coalition of 44 state attorneys general wrote to AI companies including Meta demanding stronger child protections, and Meta retrained its chatbots to block conversations with teens about self-harm, suicide, and disordered eating before eventually cutting off minors' access to some AI characters entirely. Cannes appears to have run in parallel with that reckoning, which suggests the project may have been partly a defensive effort to show Meta was no worse than its peers.

What to watch

The scale of AI adoption among children is what gives this story its stakes. Research from the UK group Internet Matters found that 64 percent of children aged 9 to 17 have used AI chatbots, and 58 percent of those aged 9 to 12 use them despite a minimum age of 13. Effective age verification remains largely absent across the industry.

Three threads are worth tracking. First, whether OpenAI, Google, or Character.AI pursue terms-of-service or legal action against Meta for the covert account creation. Second, whether regulators already scrutinizing chatbot child safety, including the state attorneys general and members of Congress, fold Meta's testing methods into their inquiries. And third, whether the industry moves toward a disclosed, shared framework for minor-safety benchmarking, so that stress-testing the guardrails protecting children does not require impersonating them.