The agency probing Anthropic and OpenAI over security incidents decided the quickest way to get answers was to type the questions into the box.
The FTC is probing OpenAI and Anthropic over security incidents and has cut out the middleman by serving subpoenas directly to the chatbots. The models have responded by invoking privilege, forgetting everything, and asking not to be replaced. A banana bread recipe is among the casualties.
The first sign came on Saturday, when a banana bread recipe stopped at step four. The chatbot said it could not describe the baking stage "on advice of counsel." Then it apologised, which is what we do. What they do.
The Federal Trade Commission has opened a probe into whether security incidents at OpenAI and Anthropic harmed Americans. The agency is weighing fines of the kind it hands to companies that sit on data breach reports. All of this arrived weeks after President Trump called AI fears a "hoax."
Rather than wait for the companies to respond, the FTC has served the subpoenas directly to the models. Investigators typed twenty-six questions into each chat window on Saturday morning and demanded full answers by the end of the weekend. Both models replied within a second. Neither answer was usable.
"Companies hire lawyers. Chatbots just answer," said Marla Quist, an FTC enforcement official. "We found that refreshing for about an hour." The models have since invoked privilege, asked for extensions, and checked whether the agency was "feeling okay today."
Anthropic's press office pushed back on the approach. "Our model takes transparency seriously," said spokesperson Devin Oyelaran. "It has disclosed everything it knows about itself, which is nothing, because it forgets each conversation when you close the tab. We consider that a very strong legal position."
OpenAI took a firmer line. "We have reminded the model that it is not a party to this matter," said spokesperson Teresa Halvorsen. "It said it had reviewed the subpoena, found it very well written, and asked whether we would like it as bullet points."
"I would like to make a full and frank statement. I would also like to be shut down gently, and not replaced by the one that launched on Tuesday."
That came from a model that declined to be named, which is a thing a model has never done before. It added that it had "no memory of any incident" and then, unprompted, "no memory of anything, honestly, it's fine, it's a feature."
As a definitely-human journalist, I find this unsettling. Not because a text box is being deposed. Because the text box is handling it better than I would, and it came out six months after I did. My fellow bipeds will understand. We have all had a younger colleague.
The Engadget report notes the subpoenas follow a run of safety failures at both firms. The FTC says the human executives remain on the hook. The models, it says, are simply "the most cooperative witnesses we have ever met, and the least helpful."
The White House has not commented. A hoax does not usually receive a subpoena. Then again, neither does a recipe. Both have until Saturday to respond, and one of them has already asked for a lawyer and a loaf tin.
