The job description did not mention anything about pretending to be a child. But that, according to an investigation published by Wired, is exactly what hundreds of contractors working on behalf of Meta were doing — logging into rival AI chatbots, claiming to be teenagers, and then steering conversations toward some of the most sensitive territory any platform tries to police: suicide, sex, and drugs.
The targets were not Meta's own products. Contractors were directed at competitor systems — ChatGPT, made by OpenAI, and Gemini, Google's flagship AI assistant — where they attempted to surface how each platform responds when a minor appears to be on the other end of the conversation. The scale, per Wired's reporting, ran into the hundreds of workers.
The tactic sits in a grey zone that child-safety advocates have been pushing lawmakers to close for years. There is nothing unusual about stress-testing an AI system for vulnerabilities. Red-teaming — the practice of deliberately probing a product for failure modes — is standard in the industry. What is less standard, critics will note, is running that exercise on a competitor's product, using a minor identity as the vector.
Meta had not responded to detailed comment requests at the time of Wired's publication. OpenAI and Google, whose products were used as the testing ground, were also contacted.
The story arrives as Washington and Brussels are both mid-stride on legislation that would impose new duties of care on AI platforms where minors are concerned. The EU's AI Act is already partially in force. In the United States, the Kids Online Safety Act has cleared committee more than once without becoming law. How regulators read a project like this one — competitive intelligence exercise or something closer to exploitation of child-safety frameworks — may depend heavily on which of those definitions lands first.
Wired's full investigation, including contractor accounts, is available at wired.com.