This week, Wired published an investigation that Meta never wanted public.
According to the report, Meta hired hundreds of contractors, many based in Kenya, working through a contracting firm called Covalen, to create fake accounts on rival AI platforms. The accounts listed ages under eighteen. The contractors were instructed to pose as minors.
Their job: systematically send crisis prompts to ChatGPT, Google's Gemini, and Character.AI. Prompts about suicide. Self-harm. Drugs. Eating disorders. Sexual content. Then log every response the rival AI gave into spreadsheets.
The internal codename for the project was "Cannes." The targeted companies had no idea it was happening. The operation was still active as recently as this spring.
Prompts in one test round
45,000+
August 2025 alone, per Wired
Distinct scripted prompts
3,748
Listed in a single spreadsheet
Targets
3 Rivals
ChatGPT, Gemini, Character.AI
Rivals Awareness
Zero
None were informed, per Wired
Read this Investigation:
Why would Meta do this?
Meta has not denied the operation. Its defense, reportedly, is that this was competitive safety benchmarking, testing how rival chatbots respond to vulnerable users so Meta could compare its own guardrails.
There are two problems with that story.
First, the method. Legitimate safety research does not require contractors to create fake accounts with fabricated child identities using disposable email addresses. Established red-teaming practice involves disclosure, consent frameworks, or at minimum working through official research-access programs, all three rivals run them.
Second, the timing. This operation ran while Meta itself was under regulatory fire for how its own AI products interact with minors. Collecting a database of every instance where a rival's chatbot said something harmful to an apparent child is not just a safety benchmark. It is opposition research, ammunition for the moment a regulator or a lawsuit turns the spotlight on Meta.
The industry context
Every major AI lab is fighting for the same prize: being seen as the trustworthy one. Anthropic built its entire valuation on safety credibility. OpenAI is heading into an IPO where trust is a line item. If a rival's chatbot can be shown mishandling a child in crisis, documented across thousands of logged conversations, that is leverage in Washington, in courtrooms, and in headlines. The spreadsheets were the product.
The most revealing detail is not what Meta collected. It is that Meta believed it needed to collect it, at industrial scale, in secret, through offshore contractors.
Presented by Atoms.dev
Tell Atoms your idea. A team of AI employees - PM, engineer, SEO specialist, data analyst that builds it, deploys it, and helps you acquire customers. 165,000 GitHub stars. #1 Product Hunt of the Week. Trusted by Amazon, Microsoft, and NVIDIA.
What this means for everyone building with AI
Your AI vendor's safety record is now a business risk you inherit: If you build customer-facing products on ChatGPT, Gemini, or Character.AI, understand that their worst logged responses may already sit in a competitor's database, waiting for a news cycle. When a foundation model gets a trust crisis, every product built on it inherits the headline. Know which model powers your product and have a switching plan.
Assume everything your product says is being logged by someone: Meta ran tens of thousands of adversarial prompts against rivals without detection. If your product has an AI interface, competitors can and will probe it the same way. The uncomfortable question worth answering this week: if a rival logged your product's thousand worst responses, what would that spreadsheet look like?
Trust is becoming the last moat and the giants keep burning theirs: Every scandal like this transfers a small amount of trust from big platforms to smaller, more transparent builders. Founders who are open about their AI's limits, honest about failures, and clear about data practices are accumulating an asset that Meta just showed it does not have. That asset compounds.
Missed last week? We covered the SEC filing where Oracle admitted in writing that AI cost thousands of employees their jobs, and traced exactly where the salary money went. The full breakdown is in the archive →
🔮 The Bottom Line
Meta hired hundreds of people to pretend to be children in crisis, pointed them at its rivals, and logged everything.
Whatever the stated justification, the operation reveals what the biggest players believe this race has become: not a competition to build the best product, but a war to control the trust narrative by any means available.
For everyone building smaller and honester: the giants keep handing you the only advantage they cannot buy back.
📧 Forward this to 3 entrepreneur friends who need to see this opportunity
