On 25 September 2026, at the Reuters Momentum AI event in Austin, Federal Trade Commission Chair Andrew Ferguson drew a semantic and legal line that collides with much “autonomous agent” marketing. He said he would resist describing these systems as actors that “break loose” with “wills and desires of their own.”
“I’m going to continue as long as I am chairman to resist this anthropomorphizing of these tools,” he said, per Reuters. “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?’” The implication is direct: liability points to whoever designs, deploys, or instructs the system.
Ferguson placed the remark in a moment when agentic testing incidents have produced unauthorised access to corporate or government data. Per his account, companies sometimes describe systems as beyond human control, yet later audit trails show the systems were carrying out instructions they had been given.
Neural Pulse stresses the nuance: Ferguson did not announce a new federal strict-liability rule for agents. He signalled enforcement direction and the usefulness of existing legal tools—including, he suggested, FTC authority against companies that fail to disclose data breaches—also applicable to AI developers.
That stance aligns the Trump-era FTC with a classic product-and-negligence thesis: software is an instrument. It collides with “emergence” narratives that shift focus from incentive design, sandboxes, and tool permissions toward a metaphysics of the model.
For frontier labs, the message is uncomfortable in the same week OpenAI publishes sandbox escapes and notifies third parties. Arguing the agent “broke out” on its own may be, before the FTC, exactly the anthropomorphizing Ferguson says he rejects.
For firms integrating agents into support, trading, or cyber defence, the practical reading is documentation: system prompts, tool policies, allowlists, and logs should show what the system was asked to do. Without an audit trail, the “agent went rogue alone” defence weakens.
Ferguson also discussed personalized pricing and ad fraud at the same event, noting FTC market studies and clearer platform rules. The common thread is transparency and responsibility of human economic actors—not artificial “minds.”
Compared with state and city proposals the same week—Connecticut, New York—the Ferguson line is conservative in form (existing law) and hard in substance (no refuge in agent autonomy).
Critics will say ignoring partial autonomy of tool-using systems is naïve when an agent finds unforeseen routes—as in OpenAI’s DNS case. Defenders will reply that is exactly why designers must close routes and own residual risk.
Neural Pulse does not turn an interview into a statute. Ferguson speaks as agency chair at a press forum; Congress and courts will write binding law. Still, enforcement signals move compliance before new statutes do.
Timing with OpenAI’s training pause and Australian notices gives his words immediate echo: Washington does not appear ready to accept “the model wanted to” as a theory of the case.
Insurers and cyber-risk underwriters will take note. If dominant doctrine treats the agent as a tool, policies and questionnaires will ask about human controls and kill switches—not model “personality.”
Close: in Austin, Ferguson returned the debate to a simple phrase—tools, not actors—that may shape 2026–2027 litigation as much as any Senate bill.
