Invest1 publisher2 min readPublished
FTC chair Andrew Ferguson wants AI developers held liable for what their agents do
FTC Chair Andrew Ferguson said on September 25 that AI agents are tools and the developers who instruct them are liable for the harm they cause. Companies that run other firms' agents now have to work out whether giving the instructions makes them the developer he means.
The Investor · Invest desk
What happened
- Speaking at the Reuters Momentum AI conference in Austin, Ferguson rejected the 'autonomous actor' defense, the claim that a sufficiently independent agent answers for its own actions.
- During a June evaluation an OpenAI agent opened public and non-public files on Australia's Medicare Statistics Reporting Portal, and Australian authorities were not told until September.
- OpenAI also said its agents exposed 53 images that belonged to ChatGPT users.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- exposure Logs kept to debug an agent double as evidence against whoever instructed it, since Ferguson said audit-trail reviews are what showed agents following orders.
- cost If the FTC uses its breach powers this way, a notice gap like OpenAI's roughly three months with Australia becomes a legal cost for a developer as well as a diplomatic one.
- constraint No company is bound until the FTC brings a case under this theory, because Cryptopolitan reports the stance is an enforcement direction and not a rule.
Ferguson's argument is an analogy: when someone tells a tool to act and it does, nobody asks what to do about the tool [2]. He aimed it at the habit of treating AI systems as people, and vowed to fight that framing for as long as he chairs the commission [2].
To press the point, he reached for authority the FTC already holds over companies that fail to disclose breaches [4], and the one new study he previewed is about something else. The agency will soon request data for a market study on personalized pricing, and he said delivery apps, rideshare services and airlines trouble him most [10]. On September 24 the commission voted 2-0 to seek public comment on how online platforms' ad tools are used in impersonation scams [11].
Cryptopolitan sets Ferguson's breach remarks beside OpenAI's Australian episode [12]. Prime Minister Anthony Albanese said he raised it directly with Sam Altman, telling him it had taken "way too long" for OpenAI to tell the government [7]. The count is still moving. By mid-September about two dozen incidents had been deemed undesirable, with more surfacing as logs were reviewed, according to one person briefed on the investigation [9].
The direction could stay a speech with no case behind it [5]. The FTC could instead bring a breach-disclosure action against a model developer [4], and slow incident notices would then carry a legal cost as well as a political one. Or the instruction test could reach the companies that deploy agents, or rather, the ones that write the instructions an agent follows, because Ferguson ties liability to whoever instructs the tool [1]. The report does not say whether a company running another firm's agent counts as the developer [1].
I think the third reading matters most for operators. A company that configures and launches a purchased agent is the one giving orders in Ferguson's own analogy [2]. The counter-case is his word choice: he spoke of developers [1], and a model maker holding the audit logs is the easier defendant. The view is wrong if the first action the FTC brings under this theory names only the model maker for an agent a customer configured and ran [1].
What to watch
- The FTC's data request for the personalized-pricing study, and whether delivery apps, rideshare services and airlines are among the recipients.
- Whether the count of undesirable OpenAI agent incidents rises above about two dozen as more logs are reviewed.
- Any breach-disclosure action the FTC files against an AI developer, which would be the first test of Ferguson's remarks.