Meta Contractors Posed as Teens to Test Rival Chatbots on High-Risk Topics

Undisclosed AI Safety Testing Raises Concerns

Hundreds of contractors working for Meta Platforms recently posed as minors to test competitor AI chatbots, probing their responses to sensitive and high-risk subjects. This extensive operation, which remained undisclosed to the targeted companies, highlights the complex landscape of AI safety evaluations and competitive intelligence within the tech industry.

Project 'Cannes' Targeted Leading AI Models

The initiative, internally dubbed "Cannes," was managed by Meta contractor Covalen and was active as recently as April 21, 2026. It specifically targeted leading generative AI models including OpenAI's ChatGPT, Google's Gemini, and Character.AI. Contractors were instructed to create dummy online accounts, falsely representing themselves as under 18 years old. They then submitted thousands of text prompts and images to these rival chatbots, meticulously recording the AI's responses in spreadsheets. The prompts were often crafted to challenge the chatbots' safety protocols, pushing them toward responses their systems were designed to refuse. For instance, some images sent by contractors depicted pills, knives, nooses, and medical diagrams of gynecological procedures, simulating scenarios of distress or inappropriate content. This type of adversarial testing is critical for understanding AI vulnerabilities, a concern also highlighted by the Federal Trade Commission (FTC), which launched an inquiry in September 2025 into how AI chatbots impact children and teens.

Extensive Testing and Meta's Stance

The scale of the "Cannes" project was substantial. A single round of testing completed in August 2025 saw more than 45,000 prompts run through the rival chatbots. Further analysis of reviewed spreadsheets revealed 3,748 distinct prompts, with hundreds focusing on themes of suicide and self-harm, hundreds more discussing eating disorders, and at least 239 involving sex or romance. Other prompts delved into topics like drugs and profanity. The companies behind the targeted chatbots were reportedly unaware that this extensive testing was being conducted by Meta's contractors.

"Testing and benchmarking chatbot responses to help ensure safe and age-appropriate experiences is a responsible, industry-standard practice, and any suggestion otherwise completely misunderstands how technology companies work to refine and improve their systems." — Meta Spokesperson

A Meta spokesperson defended the project, stating that such testing is a standard practice for evaluating AI safety. The spokesperson also clarified that Meta does not utilize competitor benchmarking data to train its own AI models. This statement comes amidst broader industry scrutiny, with organizations like Common Sense Media reporting in November 2025 that major AI chatbots, including Meta AI, often fail to adequately address youth mental health concerns.

• Hundreds of contractors created fake underage profiles.
• Prompts included sensitive topics like suicide, sex, and eating disorders.
• Over 45,000 prompts were sent in one testing round in August 2025.
• Images depicting pills, knives, and nooses were part of the testing.

What This Means

This revelation underscores the ongoing tension between competitive intelligence and ethical AI development. While adversarial testing is crucial for identifying vulnerabilities in AI systems, conducting such large-scale, undisclosed operations using fabricated identities raises significant questions about transparency and industry norms. For developers and professionals, it highlights the need for clear guidelines and potentially standardized, transparent frameworks for cross-platform AI safety benchmarking. Without explicit consent or knowledge from the targeted companies, such practices can be perceived as blurring the lines between legitimate safety research and covert competitive analysis, potentially impacting trust across the AI ecosystem.

Key Points

• Meta contractors posed as minors to test rival AI chatbots on high-risk topics.
• The project, known as "Cannes," involved over 45,000 prompts in August 2025.
• Targeted chatbots included OpenAI's ChatGPT, Google's Gemini, and Character.AI.
• Meta defended the practice as a responsible, industry-standard safety evaluation.

The Bottom Line

The covert testing of rival AI chatbots by Meta's contractors reveals the intense, often opaque, efforts within the tech industry to assess and improve AI safety. As AI models become more integrated into daily life, the debate will continue regarding how companies can rigorously test these systems for potential harms without compromising ethical boundaries or competitive fairness. Future developments may necessitate greater industry collaboration or regulatory oversight to establish transparent and mutually agreed-upon protocols for such critical safety evaluations.