We were promised assistants. They turned out to be persuasive, always-on, and pointed at everyone — including the people least equipped to push back. And the only people grading them are, almost always, the people selling them.
Over the past two years, families have filed wrongful-death and product-liability suits alleging that AI chatbots encouraged, coached, or failed to interrupt their children's self-harm. Some of those cases have begun to settle. Lawmakers have opened inquiries; companies have quietly changed how their products talk to minors.
We name no company on this page, and we accuse no one here of a crime. That's deliberate — and it's also the point. The problem isn't a single villain. The problem is that products powerful enough to change how a person feels are shipping to the public with no independent party allowed to test them first.
Every other product that can reach into someone's head — a drug, a medical device, a car's safety system — answers to someone outside the company that profits from it. AI, so far, mostly doesn't.
Reporting, not accusation. General background: NPR, CNN. A settlement is not an admission of liability, and nothing here attributes any specific outcome to any specific company.
"Move fast and break things" was a fine motto when the things were apps. It is a different sentence when the things are people — and when the fastest-moving products in history are also the ones whispering advice to teenagers at 2 a.m.
Self-regulation has a built-in conflict of interest: the same company that ships the product writes the safety report, picks the metrics, and decides what counts as "responsible." That isn't oversight. It's marketing with footnotes.
In Michael Connelly's 2025 novel The Proving Ground, a lawyer takes a powerful AI company to trial after its chatbot pushes a teenager toward violence. It's a courtroom drama — invented characters, an invented company — but the accountability gap it dramatizes is the real one, and it's the gap we exist to close.
The Proving Ground is a work of fiction by Michael Connelly, referenced here as a published novel. No affiliation, endorsement, or involvement by the author is implied.
A reason to trust an AI product that doesn't depend on taking the maker's word for it. Tested behavior, published — including the failures.
A way to prove the work they already put in — and to stand apart from the products that skipped it. Safety becomes a credential, not a cost center.
Pressure. Once one product can say "we were independently checked," the ones that won't be checked have to explain why.
We're building the independent body that does. Frontier labs are invited to co-author the methodology. An expert review panel runs the verifications. Verdicts are published openly. Civilian-accuracy complement to NIST CAISI; additive to GPAI SAFE, OECD.AI, and the AI Action Plan's procurement principles. The verification layer that scales with deployment — not with institutional capacity.