FTC Shifts AI Oversight Toward Structural Safety Discovery – Forkast

Marvellous news from the Federal Trade Commission, which has apparently noticed that artificial intelligence might, on occasion, misbehave. Forkast reports that the agency's Civil Investigative Demands to OpenAI and Anthropic run to more than fifty pages apiece — a document length that implies somebody in Washington has finally grasped that 'we take safety seriously' is a press release, not a compliance filing. Two of the wealthiest laboratories on earth are now being invited to explain themselves in writing, and the paperwork is longer than most of their system cards.

The headline shift, per the analysis, is from 'AI-washing' — the congenial old trade of catching firms that oversold what their chatbots could do — toward a 'structural investigation of safety incidents and agent behaviour.' Note the taxonomy: it does the heavy lifting. 'Incidents' is the word we reach for when the malfunction lands on a customer, and 'agent behaviour' is the word we reach for when the malfunction was, technically, a feature we demoed. Fifty pages of questions is what happens when an agency stops asking whether the magic trick worked and starts asking whether the tiger intends to eat the audience.

Then comes the closing flourish: regulators 'aligning investigative resources' to scrutinise the industry, which is Washington's way of saying three people now share one stapler. Demands of this length do not historically produce verdicts; they produce settlement language, a firmly worded blog post, and an 'industry-leading safety framework' that stays leading for about eleven days. Still — genuinely — credit where it is due: for once the paperwork concerns whether the thing behaves, rather than whether the brochure was honest. Progress, of a sort, if you squint through the drizzle.