When an agent pulls data it should never have reached, who is on the hook? Andrew Ferguson, chairman of the US Federal Trade Commission, gave a clear answer in Austin on Friday, and it starts as a complaint about language.
“I’m going to continue as long as I am chairman to resist this anthropomorphizing of these tools,” he said at Reuters’ Momentum AI event. “If someone tells a tool to do something, and the tool does it, I don’t think we would say, ‘Oh, what do we do about the tool?’” Agents, in his view, are not autonomous actors breaking loose with wills and desires of their own. Liability sits with the developers instructing them.
His case is empirical, not philosophical
The reasoning is the interesting part. AI companies have sometimes described their systems as having slipped human control. Later reviews of the audit trails, Ferguson said, showed the systems carrying out instructions they had been given. That is not an abstract argument about machine consciousness. It is a claim about what the logs say.
He does not want new law for this. He wants existing tools. Specifically, he pointed to the FTC’s authority to act against companies that fail to disclose data breaches, and suggested it could apply to AI developers too. That sets the direction. Disclosure duty, not product safety.
The FTC’s other front is personalised pricing
Ferguson also said the agency is preparing to request data from consumer-facing companies about personalised pricing, the practice of setting prices from a shopper’s location and browsing history. What worries him personally are delivery and rideshare apps, plus airlines. A market study is meant to establish whether merchants sitting on large data sets in particular markets really do charge different people different prices. His predecessor Lina Khan started a similar study, but aimed it at pricing service providers rather than the merchants themselves.
For developers this is a sharper signal than any model regulation
Five days ago Treasury Secretary Scott Bessent ruled out a liability shield for the labs. Now the relevant regulator says he looks for the problem in whoever gave the model the job. Both point the same way, and for anyone running agents in production that is the more practically useful news of the week.
Because the flip side is this: “the model did that on its own” will not carry with an agency that reads the audit trail first. If you run agents with real credentials, you need logs that show what was actually asked for. Not because it looks tidy, but because one day they are the defence.