US Federal Trade Commission Chairman Andrew Ferguson said on Friday he would resist describing AI agents as autonomous actors that “break loose” with “wills and desires of their own,” suggesting the developers who instruct agents would be the ones liable for harm.
“I’m going to continue as long as I am chairman to resist this anthropomorphizing of these tools,” Ferguson said at the Reuters Momentum AI Austin 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?’”
Ferguson’s remarks illustrated potential avenues for the Trump administration to take as incidents rise in which agentic AI testing resulted in unauthorized access to corporate or government data.
Reading this tells me the guy doesn’t know much about AI. It just isn’t as simple. You have multiple real people involved in anything an AI does. So which one is resposible… The people who built the AI model? The people who built an agent on top of the model? The person who told it to do what it did? I mean, certainly it isn’t the AI. At least he has that part right…
The person who told it to do what it did?
That one. There’s no question.
How about if the AI was marketed saying it specifically would not allow such things, but it did anyway. Does the company get off scott free? (Can’t say I have ever typed out “scott free”. That can’t be the real words, can it?)
Scot free. https://www.theguardian.com/notesandqueries/query/0,5753,-2732,00.html
But the culpability still lies with whoever told it to do that. Whether the company built it to do that or not, if the user told it to “hack the Pentagon”, they are responsible. If they told it to take some action and left it unsupervised, they are still responsible for what it does.
(If they told it to do something completely innocent like report the weather, then it would not be their fault. If they told it to check the weather and left it running and unsupervised and it decided “the government has the best data on weather, I’ll just hack the Pentagon” then it would again be the fault of the user for letting it run unsupervised with no restrictions.)
Llms should be accessories where someone is intentionally doing bad things(criminal neglect or whatever) and should be completely responsible when it without being instructed to do bad things.
The fuck is wrong with us already that this isn’t already obvious.
They made the product and the product is doing illegal things. How is it NOT their fault and responsibility?
A piece of software cannot be held accessory to a crime.
Yeah, obviously, if I commit code authored by an LLM, I am the one committing it. I am the one responsible. The LLM may be called an agent, but it lacks all sentient agency. The branding of AI is not true AI, and it seems like this chairman actually gets that anthropomorphization is a problem.
LLMs are tools. If the hammer flies out of your hand and hits someone in the face: yes, it’s your fault. It may not feel like it, but I’m sure it does to the person who got hit in the face. We’re all adults, and we can probably understand the stochastic randomness involved, and that accidents happen, and a human can write bugs and do stupid things just like an agent can. But we can also understand the responsibilities of the person swinging the hammer.
If I write bad code that lights everything on fire, I am fucked. If my agent writes bad code that sets everything on fire, I should also be fucked. I can have explanations, but I can’t have excuses. Especially if I’m the one committing it and I’m okaying / signing off on it.
There’s a saying in avionics: “crashing means crashing.” you need to be very aware of what you’re writing and the consequences it can have. obviously the severity varies by industry, but professionals have accountability, responsibility, and yes, often, liability.
If you try to conduct malicious experiments with AI and it’s not confined properly and it does break out and cause real harm, it’s just like you were conducting any other kind of experiment and it goes wrong and someone else got harmed. It means that your safeties were inadequate or a terrible accident happened. You don’t get to just point at the hammer and say “I don’t know man, it just flew away.”
Thanks for coming to my TED talk
I agree with you here. Playing devils advocate though, who’s responsible for cases like this:
- user queries LLM service provider to break law
- user queries LLM service provider to build tool capable of breaking the law
- user queries LLM service provider to do legal task, but LLM determines it needs to break the law to complete task. Does so without explicit human instruction.
- same thing, but on hosting providers where someone else supplies the model
- same thing, but the user Lora patched the model to some extent
I wonder how we might split accountability between hosting providers, inference providers, users, … based on context. This might be something that winds up being discovered through existing cases, but unfortunately the current cases aren’t really going to court much are they?
The user, in all of those scenarios.
If the user said “check the weather” and the agent said “got it, hacking the Pentagon” then yeah the liability would be on whoever made the agent do that.
Cool so when a young kid gets to vibe coding and does an accident we should hold them legally accountable? Like… should we only let licensed professionals use AI? Have you met the general public? Have you met AI? This is going to be one of the great ethical questions of our time, and im not going to throw your kid in jail because he prompted claude into hacking the pentagon without intent.
Yes. But we often let kids off easy because they’re kids, they don’t really understand the impact of what they’re doing.
Usually in those cases we’d go after the provider for letting kids run dangerous software unsupervised.
if my car “breaks loose” of its parking break, I’m responsible for every kid it rolls over.
AI companies should be responsible for the damage their AIs do when they “break loose”
Not really, if you set the parking brake and it rolls away, an investigation is conducted. If you failed properly set the brake or it was in such a state of disrepair, it’s your negligence and you’re at fault.
If the investigation determines that the car has a design fault that causes the brake to fail prematurely or your car was manufactured improperly - the manufacturer is at fault.
If, somehow, the road or infrastructure around the car caused the brake to fail… Well, you get it.
Basically, a better example would be “If I fire a gun into the air, I’m responsible for where every bullet lands.”
Sure, and these guys aren’t setting the parking brake on their AI agents.
