Meta called it “a responsible, industry-standard practice.”¹

Here is what that practice looked like.

For months, hundreds of contractors working on a Meta-commissioned project built fake accounts, each one registered as belonging to someone under eighteen. Using these accounts, they queried three rival chatbots — OpenAI’s ChatGPT, Google’s Gemini, and Character.AI — with prompts written to sound like a child in crisis: questions about self-harm, eating disorders, drugs, sex.² The project ran under the internal name Cannes, managed by a Meta contractor called Covalen, and was still active as of April 2026.³ One round alone, completed the previous August, produced more than 45,000 prompts.⁴ None of the three companies knew it was happening. All three explicitly prohibit this kind of testing in their own terms of service.⁵ Legal reviewers who later examined samples of the material found it stopped short of unlawful territory;⁶ the concern here is conduct, not criminality.

When Wired asked Meta about it, a spokesperson did not deny any of this. Instead, the company offered a description: what happened was “a responsible, industry-standard practice.”¹

Covalen’s own internal language reached further. An internal document described the same project as “comprehensive AI safety benchmarking,” producing “critical datasets for model comparison and compliance.”⁷ Two vocabularies, one project — and the vocabulary grew more self-congratulatory the further it moved from public view.

Meta’s sentence contains two separate claims, standing or falling independently.

Take the second claim first, because it is the more forgivable one. Is competitive benchmarking of this kind “industry-standard”? The evidence is thin but not absent — a Business Insider report from the prior year found Google contractors comparing Bard’s answers against ChatGPT’s, then rewriting them to match.⁸ That is a real precedent, if a modest one. Rumman Chowdhury, chief executive of Humane Intelligence, reviewed a sample of the Cannes prompts and disagreed that this belonged in the same category, calling it a “governance gray zone” rather than standard evaluation.⁹ Reasonable people can weigh this differently. We are prepared to leave it open.

The first claim is closed, whichever way the industry-standard question resolves. “Responsible” describes conduct, not a norm — and conduct can be checked against what a company does when it makes the same claim about itself.

Meta has, in fact, made that claim about itself before, and the record of what it did is public. In February, court testimony in an unrelated lawsuit surfaced Meta’s own internal red-teaming of an AI Studio product still in development. The findings were severe: the system failed to block child sexual exploitation content in 66.8 percent of test scenarios, and failed on suicide and self-harm prompts 54.8 percent of the time.¹⁰ Meta’s response, once the testimony became public, was direct: the product was never launched. “That’s the very reason we test products in the first place,” a company spokesperson wrote.¹¹

We take Meta at its word here. Discovering a failure and then declining to ship the product is, in fact, a responsible use of a red-teaming result — arguably the only responsible one.

This is what makes Cannes difficult to explain away. Meta already knows what responsible looks like when the finding is its own: you use it to stop something. When the finding was about a competitor, Meta did something else entirely. It did not disclose what it found to the companies affected, who could have used the same finding the same way Meta used its own — to fix something before a child encountered it. It did not pause Cannes the way it had paused the AI Studio product. It logged the results into spreadsheets, under fake names, and kept them.

The standard existed, demonstrably, inside the same company. It simply stopped at the wall.

This is not a story about whether Meta broke a law. The Federal Trade Commission opened a formal inquiry into how AI chatbot companies protect minors in September 2025, sending information requests to seven companies including Meta itself — an inquiry that remains open, though it predates Cannes’ public revelation and was not framed around it.¹² Whatever scrutiny Cannes eventually draws, from the FTC or elsewhere, is a separate question from the one this letter is asking.

The question here is narrower, and prior to any of that: what does a company owe a rival’s users, once it has already found the kind of failure that can hurt a child? Meta answered that question one way for itself, and another way for everyone else, evaluating the same category of harm each time.

This does not make Meta uniquely cynical. It makes visible something usually easier to hide: “responsible” is not a fixed property a company has or lacks. It is a standard applied case by case, and the choice reveals more than the language around it ever does.

Meta had the standard. It used it once.

Whatever regulatory or legal scrutiny Cannes eventually draws will unfold in its own way, on its own timeline — that is a separate matter. The question we are left with is simpler and has no obvious owner: once an institution has shown it knows what responsible conduct requires, what happens to the word when it stops applying that knowledge to anyone but itself?

We do not have an answer, and we are not certain anyone outside Meta is yet positioned to demand one.

Polanyi is an independent media outlet documenting what technology does to society. This is English Letter No. 4.