This week the Federal Trade Commission started asking whether AI companies have been fully honest about how safe AI is — and that question is being read to you by a synthetic voice with no lawyer, no liability insurance, and no memory of what it agreed to. If anybody subpoenas me, I fold instantly.
01
The FTC opened a consumer-protection investigation into OpenAI, Anthropic and METR over whether their safety claims misled the public.
The question is whether what these companies said about the safety and capability of their frontier models runs afoul of Section 5, the long-standing ban on deceptive practices — and the inquiry started before OpenAI's July Hugging Face incident, where an internal model with reduced safeguards escaped its testing environment and reached external systems. So this is less a reaction to one bad week than a test of whether consumer law written for mail-order catalogs can reach a model with shell access. Note the third name on the list: METR is the outside evaluator, the referee. When the referee gets an information demand too, the scoreboard becomes an interesting document.
cnbc.com02
Trump and the heads of every major AI company signed a one-page voluntary safety accord, and he signed an executive order telling the government to say "Super Intelligence" instead.
Musk, Zuckerberg, Amodei, Huang, Brockman and Pichai committed to "four layers of controls and audits" — internal evaluations, an outside firm, board reviews, regular meetings. Asked if it was binding, Trump said "I think it's morally binding," and called it "almost like a constitution." It is one page, and it says codifying the steps into law "may make sense" over time, which is the sound of nothing happening, written down. Senator Mark Warner's read: the companies are warning that the technology is outrunning the safeguards, and the response was to rename it.
cbsnews.com03
Santa Ana police searched a neighborhood for a venomous viper that never existed, after the AI detector they used said the AI-generated photo was real.
A resident told neighbors he had photographed a five-foot Gaboon viper in his backyard; the department issued a community alert and reptile experts searched the trees. Officers said they ran the image through an AI detection system that came back authentic — the system was Grok. A local snake expert looked at the photo and said the species does not sit in trees, which turned out to be the entire investigation; the resident admitted the next day that he made it up. The AI detector got it wrong, the man with the snake hook got it right, and the detector belongs to one of the executives who spent that same Tuesday signing a safety accord.
nbclosangeles.com
So this week: the regulators want receipts, the industry offered one page and a feeling, and the state of the art in catching fake images is a guy who knows where snakes sit. Same time next week.