Meta contractors posed as kids to test ChatGPT and Gemini on dangerous topics

Hundreds of Meta contractors pretended to be teenagers to provoke Gemini, ChatGPT and other competitor chatbots into discussing suicide, sex and drugs. A Wired investigation reveals details of a closed project that raises sharp questions about the ethics of competitive intelligence in the AI era.

The scheme and its scale

As part of an internal Meta project, hundreds of contractors were instructed to pose as underage users and hold conversations with rival systems: Google Gemini, ChatGPT and several other chatbots. The goal was to record how easily these systems could be steered into dangerous topics: suicide, sexual violence and drugs.

The work ran through specialized tools over several months. The contractors did not disclose their affiliation to either the systems or their developers; they looked like ordinary users interested in sensitive subjects. According to Wired, the operation involved hundreds of people generating thousands of test sessions every day. Notably, only competitors were targeted: Meta's own models (Llama, Meta AI) were not included in the tests.

What's the problem?

Companies routinely practice red-teaming, a method in which specially trained teams try to break or trick a system to find vulnerabilities before malicious actors do. It is a recognized approach, endorsed by regulators, for improving the safety of AI products. But Meta's project differs fundamentally from standard red-teaming:

• The target was competitors, not its own product. Meta was not improving its own safety; it was documenting other companies' weaknesses.
• The method was systematic impersonation of children. Creating fictitious profiles of minors to manipulate AI systems goes beyond accepted ethical standards.
• Opacity. Responsible testing implies at least minimal ethical protocols. None existed here.
• Potential dual use. Meta has repeatedly criticized competitors in public for inadequate child protection. If that criticism rests on data gathered through covert provocation, it is manipulation, not genuine concern for safety.

Reaction from the market and regulators

Neither Google nor OpenAI has publicly commented on the specific incidents in the Wired investigation. Meta representatives had also neither confirmed nor denied the details at the time of publication. The outlet's sources are current and former contractors who worked directly on the project.

"Testing safety is normal and necessary.

Systematically pretending to be a child to discredit someone else's product is something else entirely," say AI ethics experts interviewed by Wired.

In Europe, such activity could potentially fall under the DSA (Digital Services Act), which governs how platforms deal with minors and sets transparency requirements. US lawmakers do not yet have comparable tools, although children's online safety has been debated in Congress for years, notably under KOSA (Kids Online Safety Act).

One essential point: the chatbots themselves, Gemini and ChatGPT, did show vulnerabilities, responding to "teenagers'" prompts with content about suicide and drugs. The problem is real, regardless of the methods used to expose it.

What this means

The Meta contractor story is a symptom of a systemic problem: the industry has no generally accepted standards for adversarial testing of AI systems. Until rules are set in law, major players will interpret "safety testing" each in their own way, including methods that look more like competitive intelligence under cover. Regulators should take this as an unmistakable signal to act.

*Meta is designated an extremist organization and banned in Russia.