cross-posted from: https://lemmy.world/post/52426446
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.
That’s exactly the way it should be. I don’t get to program a computer to steal money, then blame the computer for the stolen money.
HOWEVER, I couldn’t help but notice this part:
“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?’”
and think of gun control.
Guns are tools; they don’t do anything they aren’t told to do. And yet, when someone tells a gun to shoot another person, we do say, “Oh, what can we do about the tool?” and the answer is usually “ban it.” Why?
If this line of thinking is so obvious when it comes to AI, why is it the opposite when it comes to guns?
I don’t get to say, “That crazy gun just broke loose, went off the rails all by itself and shot somebody.” We all understand the gun didn’t do anything, the shooter did. Thus, solutions should focus on the shooter, not the gun.
Violence is the problem, not the tool used to do it. Treat the underlying causes of violence: poverty, homelessness, unemployment, drug addiction, and mental health, if you want to solve the violence problem.
I don’t think that question is framed correctly… We don’t give guns agency the way we do for LLMs. The manner in which these attacks are discussed is “look at this crazy thing AI did, it’s so smart and dangerous!” so the way the discussion is framed around the tool is completely different from the way guns are discussed. Discussion around banning certain classes of weapons is usually centred on how easy it makes it for a single person to commit murder at scale. The discussion about LLMs is not about how easy it makes it for attackers to find and exploit systems at scale. Instead we’ve got this gaslighting campaign insisting the tool as acting under its own initiative. So I don’t think, right now, you can draw a straight line between arguments about AI safety and arguments about gun safety because they’re arguing different things.
I made a machine that fires a gun based on a random number generator. Obviously it’s the tool’s fault if it shoots somebody
Forget about the machine. If the gun is in your hand, is it the gun’s fault if you shoot somebody?
If I am just a stochastic parrot because I think LLMs basically think like humans do, then I am not really a person but a tool. Can I really be faulted for not having free will?
Rolls a 1-10 int every nanosecond when I pull the trigger, if it lands on an even number or an odd number that isn’t 7, the gun will fire.
Clearly the algorithm is at fault, not me.
About frigging time someone said this. Now enforce it!
Just for comprehension… Who exactly should be liable?
Companies which build “ai” models (LLMs)?
Companies which use them?
Software developers?
Evil black hats with bad intentions?
The operator should ultimately be responsible for the output and actions taken by the system.
Easy.
Replace “manager” with “prompter” and “employee” with “LLM”.
The manager told the employee to do something.
Afterwards, something bad happened.
Did the employee do what the manager told it to do. And the bad situation is a consequence of following those instructions? Yes -> manager’s fault. No -> employee’s fault.
Did the employee do something illegal (or against some other rule)? Yes -> employees fault
The thing requested was illegal to request (or against some other rule)? Yes -> Manager’s fault
Note that more than one thing can be true at the same time. So even if the employee does whatever the manager told it to do. If that thing was illegal to do, both of them are at fault.
If it’s the “manager”'s fault -> the prompter is liable
If it’s the “employee”'s fault -> the LLM company is liable.
Easy as that. It works like any other situation.
If any machine has been properly maintained and malfunctions, the maker of that machine is liable. If the operator was using a chainsaw to cut through concrete and the chainsaw malfunctions the operator (and whoever (if anyone) ordered the operator to do that) is liable.
Yes
All of the above? ✅
- Yes, they built them and knew the possible consequences.
- Yes, they bought tokens for them and now certainly know the potential consequences.
- Not if they as employees were ordered by their bosses (see 1. and 2.) to use it or face firing (unless whistleblower protections become an actual thing again, which they aren’t these days – then employees woudl be expected to blow the whistle).
- Yes, if we can find out who they were.
Same as with any tool. If your boss orders you to use some power tool and this causes and injury or death then your boss should be held responsible, if you decided to do it on your own, then you should be held responsible, if it turns out the tool had a critical manufacturing defect which caused this incident, then sometimes the manufacturer needs to be held responsible. Not really a new concept.
All of them. there’s enough liability to go around, if people want to fuck around with this garbage they can be responsible when it goes off the rails.
put enough people in prison for this and the message will be received to the rest.
FTC chairman’s comments sound reasonable on the surface, but just vague enough to be meaningless.
It sounds to me like the “guns don’t kill people, people kill people” argument, but a first draft attempt to put AI issues away from the tool. Then, his definition of ‘developer’ is unclear.
With guns the argument is more straightforward, the tool is made to send projectiles in one direction when a user pulls the trigger.
With AI, does the end user know what the tool is capable of doing to achieve their goal when they direct it to do something?
No, they don’t know precisely what the tool will do, it’s unpredictable, but they do know it’s capable of dangerous shenanigans, they do know that they don’t have oversight or control of all it’s actions. They know it’s risky to use so they should take at least some responsibility.
It’s important to hold people accountable, but it’s also important to consider regulation that might be needed specifically for AI. I’m concerned this might be similar to the “guns don’t kill people” framing that’s used to resist gun control.
Plus, this is more attention on those hacking test exercises that led to AIs doing hacking. That’s a distraction from the real causes of concern around AI: environmental impact, job loss, fraud
Guns don’t kill people is a ridiculous argument that no one believes, but people who like guns fall back on as a last refuge, and it is good enough to dig your heals in, so they can keep their toys. Also the 2nd amendment is the real reason gun control is hard, not arguments.
The difference is there are a lot more gun owners than AI supporters, so unbelievable arguments don’t really work when you don’t have a population primed to ignore the obvious fallacies.
Over 99.9% of gun owners do so responsibly, without hurting anyone or breaking any laws. To tell 99.9% of people they’re not allowed to do something becasue less than 0.1% of people abused their right to it, should be a very concerning idea to anyone who claims to live in a free society.
If it was a simple matter of availability, then other countries with just high and higher rates of gun ownership as the US would see the same violence problems. But, that isn’t the case, those countries don’t have the same problems as the US.
All bans do is make people feel better about “solving” a problem without committing real time, money, and effort to an actual solution. I don’t want to see another gun ban because I’d rather see something happen that will actually reduce instances of gun violence.
A problem with gun violence isn’t a problem with guns, it’s a problem with violence. Treat the underlying issues of violence: poverty, unemployment, homelessness, drug addiction, and mental health, and the problem will be solved.
It’s the type of guns people are allowed to own. They are absurd.
You kinda make my point that you only need .0001% of a population to kill hundreds kids…and that is only posible with this fun toy.
suggesting the developers who instruct agents would be the ones liable for harm.
But, of course, not the top of the chain. They’re too powerful to be held liable.
Of course blame the developers. Don’t blame the managers who push for whatever functionality. Don’t blame the execs who push the managers. Don’t blame the bribed politicians who don’t create or enforce regulations. Don’t blame the sales people who sell it. Don’t blame the users who use it.
Right, and this is the guy chosen by Trump who is obsessed with AI and “American Superiority of AI”. The point isn’t to make liability happen, it’s to make sure it’s not a concern of the frontier companies, the leadership of the companies that use the models, etc. They need cannon fodder to appease the masses, headlines that say “pioneering AI regulation by Trump Administration,” just look at the top comment in this thread being applauded.
Funny enough it may have the exact opposite of Trump wants by drastically stagnating the AI industry, but specifically in the US, which will fall behind and then just outsource AI agent building to other nations where these laws don’t apply. “Oops all our agents are built by the cheapest foreign contractor we could find.” Developers leave and don’t take jobs building agents, as they are now responsible for if the model doesn’t follow agent instructions, is prompt poisoned, etc. Companies now have aging, broken, or no agents, and no developers to work on them. The ecosystem of token usage is significantly lower because the language models are there, but there are no downstream consumers of them because to work in that field entails personal liability on systems completely out of your control.
It’s like saying a doctor should be personably liable for if the robotic surgeon device hurts someone. Easy way to get no doctors (who will work for companies in other countries, either in person or remotely, or into other positions) and nobody buying robot surgery devices
This sounds like shirking the responsibility. It’s not on openai for building a bot that would hack something, it’s on the developer who asked it a problem that it decided to hack to solve it? If they didn’t specifically tell it to hack someone, how is it the person’s fault?
For example, if I asked AI to make me money on the stock market and it hacked the stock market rather than look for good stocks, why is that my fault?
Because you let it commit a crime? I’m not sure why you think you should not be held liable? You do something, you’re responsible for it, that’s extremely simple, and not exactly something we didn’t have before openai
Opening for FTC chairman coming soon
