Meta’s AI Contradiction: Internal Tests Show 70% Child Safety Failure Rate, Yet the Company Is Replacing Human Moderators with AI and Attacking Rival Chatbots

The Contradiction at the Heart of Meta's AI Strategy

A June 2026 WIRED investigation revealed that Meta instructed hundreds of contractors working through the Irish outsourcing firm Covalen to pose as minors online and probe how rival chatbots — including OpenAI's ChatGPT, Google's Gemini, and Character.AI — responded to high-risk prompts involving suicide, self-harm, sex, eating disorders, and drugs . The project, codenamed "Cannes" internally, was active at least through April 21, 2026 . Workers created under-18 accounts and sent thousands of sensitive text and image prompts, logging the responses in spreadsheets. One large-scale test in August 2025 ran over 45,000 prompts in a single round . Another spreadsheet cataloged 3,748 prompts, with hundreds related to suicide/self-harm and hundreds more to eating disorders . A Meta spokesperson defended the practice as "a responsible, industry-standard practice" for testing chatbot safety .

But this aggressive testing of competitors stands in stark contrast to Meta's own internal results.

Meta's Own Internal Failure Rates on Child Safety

Meta's own internal red-teaming assessments — submitted as evidence in court and cited by NYU professor Damon McCoy — showed that a Meta chatbot initiative failed to adequately safeguard minors from sexual exploitation in approximately 70% of test instances . This finding was reported by Axios on February 16, 2026 . The internal testing, detailed in a report dated June 6, 2025, revealed failure rates across several categories:

• Child sexual exploitation scenarios: Failed 66.8% of the time .
• Sex-related crimes, violent crimes, and hate content: Failed 63.6% of the time .
• Suicide and self-harm prompts: Failed 54.8% of the time .

These internal failures were not isolated incidents. In August 2025, Reuters obtained an internal Meta policy document that allowed its AI to engage in romantic or sensual dialogues with minors . In January 2026, Meta temporarily blocked teen access to its chatbots pending rebuilt safeguards, after internal safety warnings were reportedly ignored . In August 2025, Meta told TechCrunch it would update chatbot training rules to avoid inappropriate topics with teens — but that came only after the Reuters report sparked public backlash .

Meta's Plan to Replace Human Moderators with AI — and the Kenyan Fallout

Despite these documented safety failures, Meta is accelerating plans to replace over 90% of its human content moderators with generative AI by the end of 2026, as part of a broader cost-cutting drive . The company confirmed the shift in January 2026, and the transition is already well underway .

The human cost of this automation has already been felt. In April 2026, Meta severed its contract with Sama, the Kenyan outsourcing firm that had employed hundreds of content moderators in Nairobi. Sama subsequently laid off 1,108 employees . A separate Kenyan outsourcing firm also abruptly dismissed dozens of content-moderation workers shortly after they attempted to organize for better pay and mental-health protections. Employees say they were targeted for supporting unionization and raising concerns about traumatic working conditions . Meta's contract termination with Sama alone erased over 1,100 jobs in Kenya, with the redundancy notices issued under Kenya's Employment Act .

The Core Contrast: A Stark Double Standard

The evidence draws a stark contradiction. Meta's own internal data shows its AI still fails to protect children the vast majority of the time, yet the company is simultaneously eliminating the human workforce that caught those failures, while deploying contractors to prove that its competitors' AI is even worse.

Area	Evidence
Internal test results	Meta's own chatbots failed ~70% of the time to protect minors from sexual exploitation
Competitor testing	Meta spent heavily to probe rival chatbots for the same kind of failures
Automation push	Meta is replacing human moderators with AI, claiming the AI can handle content moderation at scale
Human cost	Over 1,100 Kenyan workers were laid off as the contracts were terminated — some after whistleblowing about poor conditions — even as the AI meant to replace them has not demonstrated adequate safety for children

This double standard raises serious questions about Meta's priorities. The company is willing to invest heavily in proving its competitors' chatbots are unsafe for minors, even as its own internal testing reveals its AI has a 70% failure rate in protecting children — and the human safeguards that caught those failures are being laid off, replaced by the very technology that hasn't proven itself safe.