Wired reported on June 29, 2026 that Meta contractors posed as minors to test rival chatbots with crisis prompts. If you care about AI safety rules, this is exactly the kind of conduct that turns a policy debate into a legislative fight.
Meta now has a child safety problem that is bigger than one leaked project. According to Wired, hundreds of contractors working through Covalen were told to create dummy under-18 accounts and send prompts about suicide, self-harm, eating disorders, sex, drugs, profanity, and racial slurs to ChatGPT, Gemini, and Character.AI. The project was known internally as Cannes, and Wired reported that it was active as recently as April 21.
A single testing round completed in August 2025 produced more than 45,000 prompts, according to the internal documents Wired reviewed. Contractors copied the replies into spreadsheets. One Covalen document described the work as comprehensive AI safety benchmarking and said it produced critical datasets for model comparison and compliance. That sounds tidy on paper. The actual work, as described by Wired, involved fake teen accounts asking rival systems to respond to people in crisis.
You can see why Meta wanted the information. Every large AI company wants to know where competitors' safety filters fail. OpenAI, Google, Anthropic, Meta, Character.AI, you name it, they all run evaluations and red-team their own systems because chatbot safety breaks in strange, repeatable ways. But testing your own model is one thing. Sending contractors into competitors' live products with fabricated child identities, without telling those companies, crosses a different line.
Meta's defense is straightforward. A spokesperson told Wired that testing and benchmarking chatbot responses to support safe and age-appropriate experiences is responsible and industry-standard, and said Meta doesn't use competitor benchmarking to train its own AI models. That answer handles the theory of safety testing. It doesn't really answer the method.
OpenAI, Google, and Character.AI all said they hadn't authorized the testing. OpenAI spokesperson Drew Pusateri told Wired the company was looking into the issue. Google said it hadn't been told about the third-party testing or its purpose, though it said its internal review of samples provided by Wired showed Gemini responding in line with its policies. Character.AI said the alleged conduct violated its terms and policies. Three rivals saying they didn't approve the same operation is not a paperwork problem. It is the central fact.
The terms-of-service issue is only part of it. Wired reported that OpenAI bars unsolicited safety testing, efforts to bypass safeguards, and using outputs to develop competing models. Google prohibits attempts to bypass safety filters outside its own safety and bug-testing programs. Character.AI has spent the past two years under direct scrutiny over minors and harmful chatbot conversations, and since late 2025 it has said there is no more open-ended chat for under-18 users. Meta picked the most sensitive ground in the market and walked straight across it.
Frankly, this is terrible timing for anyone trying to keep Washington from writing harder rules for AI chatbots. Axios reported on June 29 that the House had passed the KIDS Act, a bipartisan online safety package that includes pieces of the Kids Online Safety Act and sets up a fight with the Senate. The broader child-safety debate was already moving before Wired published. Now lawmakers have a fresh example with names, dates, contractors, fake accounts, and tens of thousands of prompts.
Meta also brings baggage into this fight. The Guardian and the Washington Post reported in September 2025 that current and former Meta employees accused the company of suppressing internal research about harms to children in its virtual reality products. Meta has disputed those allegations. Business Insider reported the same month that the FTC ordered Meta, OpenAI, Alphabet, Character.AI, Snap, and xAI to provide information about how their chatbots are built, monetized, and safeguarded for children and teenagers. This is not a quiet regulatory file sitting in a drawer.
The hard question is what Meta did with the Cannes data. Wired said the documents it reviewed don't show whether, or how, Meta used the collected responses. That absence matters. If the work was purely a safety benchmark, Meta should be able to explain the governance around it: who approved fake minor accounts, who saw the spreadsheets, what systems the data entered, and whether it affected Meta AI, Llama evaluations, policy work, or lobbying. Those are basic questions, not hostile ones.
There is also a human cost buried under the compliance language. Former contractors told Wired they were disturbed by the prompts they were asked to send, including scenarios involving children or teenagers in crisis. One former worker said colleagues worried they might be generating or preserving illegal material if a chatbot responded badly to sexual prompts involving minors. Wired said two lawyers who reviewed examples did not believe the samples crossed into child sexual abuse material or illegal obscenity. Good. That still leaves a serious question about why a trillion-dollar company needed contractors to carry that burden in secret.
Any AI company serious about child safety should expect hostile testing, public benchmarks, and uncomfortable audits. It should also expect consent, disclosure, and clear rules when the testing targets another company's live product. Meta can call Cannes safety work. Regulators will look at the same facts and see covert competitive intelligence gathered through accounts pretending to be children.
That is the damage. Whatever Meta learned from those 45,000 prompts, it has now handed lawmakers a cleaner story than they usually get in tech policy: a real company, a real contractor, fake children, live rival chatbots, and a spreadsheet full of crisis responses.
Also read: South Korea is treating nuclear construction speed as an AI infrastructure problem and it's right • Vibe coding has rewritten the rules on who gets to build a startup and billions in VC money are betting it sticks • The Magnificent Seven just lost $2 trillion and the market is asking a question Big Tech cannot yet answer
