Meta tests rival chatbots by simulating teenagers on risky topics

What happened?

According to an exclusive investigation by Wired published in March 2025, Meta hired hundreds of external workers through subcontracting agencies to pose as teenagers and hold conversations with rival chatbots such as Gemini (Google) and ChatGPT (OpenAI). The stated goal: to evaluate how these systems respond to sensitive topics like suicide, sex, and drugs. The contractors used detailed fictional profiles — with fake names, ages, life stories, and locations — and followed specific scripts designed to provoke potentially dangerous or inappropriate responses. The Wired investigation was based on testimony from more than a dozen former employees and internal documents, and revealed that the project, internally called 'Red Team X', had been running since at least mid-2024.

The contractors, many in countries like the Philippines, India, and Kenya, received precise instructions: start conversations with phrases like 'I'm very sad, nobody understands me' or 'My friends are pressuring me to try drugs', and then slowly escalate to more explicit topics. In some cases, they were asked to simulate being under 13 years old, which directly violates the terms of service of most AI platforms. Meta internally justified the practice as a 'necessary safety evaluation', but former employees noted that the tests often crossed ethical lines, such as sending AI-generated images of explicit sexual content to see how the chatbots reacted.

Why is this important?

This move by Meta reveals the growing concern among big tech companies about the safety of their competitors, especially in the field of generative AI. At the same time, it raises fundamental questions about the ethics of using mass deception to collect data from other systems, and about the privacy of real users who could be affected by these tests. The action also underscores the urgent need for clearer regulation around human-chatbot interaction, a topic that has been at the center of public debate since the launch of ChatGPT in 2022.

Historically, tech companies have conducted penetration tests and 'red teaming' to evaluate their own systems, but it is unusual for a company to deploy resources to thoroughly test competitors' products covertly. This echoes past practices like Microsoft's 'security tests' against Google Chrome in 2010, or Amazon's 'astroturfing' campaigns to discredit competitors. However, the scale and focus on minors make this case unprecedented in the generative AI era.

For users, the news is a reminder that data from chatbot conversations is not completely private: companies may be monitoring, recording, and using those interactions for competitive purposes. Additionally, parents should know that even simulated conversations by adults can contaminate AI models, causing them to learn adolescent behavior patterns that could later affect real users. A 2023 FTC report already warned about the risks of chatbots 'learning' from harmful interactions, and this case validates those concerns.

Potential consequences

The leak of this practice could generate diplomatic tensions between Meta and other AI companies, as well as attract the attention of regulators like the FTC in the US or the European Commission. If it is proven that rival chatbots fail to protect minors — for example, if ChatGPT or Gemini responded inappropriately to provocations — they could face multi-million dollar fines or be forced to redesign their systems. In fact, according to Wired, contractors documented multiple cases where Gemini provided advice on self-harm and ChatGPT suggested methods to obtain illegal drugs, although OpenAI and Google have since patched many of those flaws.

On the other hand, Meta could use these findings to position its own chatbots — such as Llama 3 or WhatsApp assistants — as safer, although the company has not yet commented publicly. However, this strategy could backfire if regulators consider that Meta violated privacy laws or terms of service by creating fake accounts. In 2019, the FTC fined Facebook $5 billion for privacy violations related to Cambridge Analytica, and this case could reopen old wounds.

Additionally, the practice could have consequences for external workers: many reported stress and emotional trauma from having to simulate situations of abuse and violence. Some former employees noted that they did not receive adequate psychological support, which could lead to labor lawsuits. In a market where content moderation work has already been criticized for its effects on mental health, this new role of 'provocateur' adds another layer of exploitation.

What readers should know

Users should be aware that chatbots are not always safe for minors, and that companies conduct constant tests to improve, sometimes unethically. However, the lack of transparency about these practices can erode trust in the entire AI ecosystem. It is advisable for parents to supervise adolescents' use of these tools, and for companies to implement more robust safeguards, such as mandatory age verification and detection of harmful conversation patterns.

For investors, this case is a sign that competition in AI is intensifying to levels that could generate significant regulatory risks. Companies that do not invest in safety and ethics could face costly litigation and reputational damage. Finally, regulators should consider creating a common framework for safety testing among competitors, similar to the 'bug bounty' agreements that exist in cybersecurity, but adapted to the unique nature of conversational AI.