Who should be held accountable when an AI Agent (accidentally) acts maliciously? | Hacker News Reader

Who should be held accountable when an AI Agent (accidentally) acts maliciously?

blog.greenpants.net

37 pointsGreenpants98 comments

I don't see how this is such an unclear legal question. If I fire a computer program that mistakenly causes another person harm, its my fault. Or it would be the maker of the program's fault. I feel we have established pattern for this already.

Until we can agree whether AI is conscious, which we never will, AI and AI agents are just property working on behalf of humans.

I could see a future where AI companies/services indemnify consumers who use their agents but _not_ indemnify corporations that use their services.

Also those agents that "broke out" were probably prompted to do that. I don't buy any story about this other than three AI companies hired the same PR firm.

Do you have any evidence of this or is this just generalized cynicism?

They all hired the same startup security firm Irregular to try their internal unrestricted models, which is weird especially for Google to do. At OpenAI, they didn't stop the test even after it was apparent the bots had broken out. That's all I got.

It shouldn’t be a question but this is where the anthropomorphic language and things like “agent welfare” come in to enable responsibility laundering of some of the most powerful people on earth. How we talk about these models matters because it impacts the public’s understanding of what they are genuinely capable of. The more that they are described as having anything close to free will the easier it is to even ask questions like this.

Here's a question I am asking lately. If I should not use anthropomorphic language, how do you suggest I handle the following situation:

   Sometimes my coding agents will seemingly refuse to follow my instructions.  When I ask them why - they say that they do not think my design is a sound one, and they have a better way to do it.  We will then sit down and come to a consensus on how best to move forward.

I argue that if we're using software that acts like a human - the only way to interface with it is to speak to it like a human. Otherwise we have no language to speak to a non-sentient object without anthropomorphization.

I'm starting to wonder if the people arguing against anthropomorphization actually have any experience at all working with agents.

EDIT: It's a simple question. When you downvote me without answering, I must assume you don't have any answer and dislike what that implies.

What’s the question? Why model output isn’t always what you expect it to be?

Consensus is just populating the model with rationale for different new output.

> What’s the question?

I thought putting the question in my comment would be sufficient. I guess not. It was and still is:

> If I should not use anthropomorphic language, how do you suggest I handle the following situation

I didn't downvote you, but your last paragraph is unnecessarily aggressive.

I've worked with agents, and I agree with you that often there isn't another way to express the interactions.

However, I also think the terms ML uses in general are a mimicry that misleads people who aren't informed. Ask anyone outside SWE what they think “training” means, and they'll usually picture something being taught.

I don’t think anybody can change that now, but it’s useful to point it out.

> Ask anyone outside SWE what they think “training” means, and they'll usually picture something being taught.

You mean how early-on people thought that planes flapped their wings while they flew?

These aren't problems. This is the way language works.

I agree with you that we use analogies to name new things, and that’s the way language works.

However, you can see an airplane flying. Still, you cannot see software processes at work, and that causes misunderstandings and misinformation, which is at the core of the changes that we are experiencing with AI.

This is a fragment of another article posted here on HN about an ongoing dispute between OpenAI and the New York Times:

“The defendants say this is a simple application of fair use: Their argument is that if you read a story and simply remember what was in it to expand your base of knowledge, that cannot be considered a copyright infringement”

However, if you replace “read” with “web scraping” and “expand your base of knowledge” with “storing the information,” the perspective changes too.

I agree. The analogies are interesting but only go so far.

I think it's safe to trust the decisions of the courts. Judges aren't easily fooled by slippery language.

For example, in Bartz v. Anthropic, Judge Alsup ruled that training is fair use because training is transformative. In his words "spectacularly so".

You're discussing how to speak _to_ a model, but everyone else here is discussing how to speak _about_ a model.

I didn't downvote, but wow you're being aggressive, you have a lot to learn if you read the comments here with an open mind.

My question was and still is this:

> If I should not use anthropomorphic language, how do you suggest I handle the following situation

And so strange that in a 24 hour period I got three comments all at the same time about being too aggressive and still not answering the question.

I'm sure it's just a coincidence.

> how do you suggest I handle the following situation

The question is vague and doesn't seem related to the discussion. What do you mean "handle"? If you're having trouble handling it psychologically, see a therapist. If you're having trouble getting the output you want, look up guides on prompting. If you're anthropomorphizing the chatbot to the point you're worried about offending it... just don't worry? It's a computer program, don't overthink it, don't anthropomorphize it, just give it the input bytes you need to get the output bytes you want.

> I'm sure it's just a coincidence.

There isn't some grand conspiracy here. You just overindexed on the word "anthropomorphic" and didn't really understand what the discussion is about.

You said:

> I didn't downvote, but wow you're being aggressive

He said:

> I didn't downvote you, but your last paragraph is unnecessarily aggressive.

Both comments arrived within two minutes of each other on a thread with no other comments for 18 hours.

That's quite a coincidence.

Also - insults are against the rules here, friend. Naughty naughty. I hope they don't spank you.

Okay fine, you caught me! There is, in fact, a giant global conspiracy to downvote your posts, qarl. It isn't just the fact that randomly-arriving independently-occurring events often arrive unevenly distributed. https://en.wikipedia.org/wiki/Poisson_distribution

This conspiracy goes all the way to the top. The illuminati assigned me personally to comment on your post. I've already said too much. If you never see me again, tell my wife I love her.

No need to make fun of me, friend. It's very unlikely - you can't disprove that because it's true.

If it was a coincidence then you have nothing to worry about. I flag it because the overwhelmingly likely thing is that it wasn't.

And again - insults are against the rules here. Shame shame.

100% That’s what bothers me about the descriptions of the OpenAI incidents.

OpenAI's reports use language that minimizes their liability.

The first question should be what the organization was doing around those tests, and why they were so naive as to run them without fully isolating the network.

However, all the attention goes to the human-like conclusions in agent thinking traces, which creates a misperception of sentient AI for people who don’t know how the magic black box works.

It really does feel like a purposeful thing on the part of the big labs.

For the most part it feels like most people are waking up to it though.

Regarding:

>However, all the attention goes to the human-like conclusions in agent thinking traces, which creates a misperception of sentient AI for people who don’t know how the magic black box works.<

I know there’s been some questions regarding if thinking traces are even relevant to the outcome most of the time.

In what way specifically does it minimize their liability? People say this a lot but it's not clear what they mean by this.

It’s about pushing the blame onto the tools and not the person using them.

Sort of like the “guns kill people” vs “people kill people” debate.

Deliberate wording to minimize perceived culpability for the agents actions.

That's not how civil liability works.

I think you’re jumping the gun on my response a little. I’m just telling you what the purpose could be.

Now do I think that’s the reason? It certainly isn’t a new thing for companies to try to do that. Shift blame that is.

Regarding civil liability, I’m not making that argument here. But it makes sense from a public perception viewpoint why they would want the agents to appear at fault instead of their own actions.

Forget for a second about AI.

The agent harness is a process, like any other process in an OS.

You are a researcher running thousands of unattended automations that can hack a website without supervision. The first thing anybody will do is put security at various levels and isolate the network as much as possible. If something escapes your allow list, it should stop the processes as soon as possible.

You cannot foresee a bug in a server (like the Artifactory server in the Hugging Face incident). But you can isolate that server at the network level in the first place. So even if you give that server read-only access, no unexpected packets go out. It's not rocket science; it's something a billion-dollar company experimenting with what they promote as the biggest possible threat to humanity (if they do not handle it) could easily do.

They minimize their liability by changing the message to “oh look how powerful our models are, now we are going to have a public awareness report of the model deviations”. The message should be, “Sorry, we ran experiments without proper sandboxing; it’s our fault, and we changed our testing practices since then.” The former message puts all the blame on the smart, uncontrollable force of AI; the latter is what really happened: an irresponsible test over the Internet.

How exactly is that minimizing their liability? You just described claims that do not appear to at all minimize liability.

Perhaps my use of the word “liability” adds noise to what I’m trying to express.

My argument is very similar to the article in the parent post:

The messages OpenAI published around the recent incidents emphasized their model capabilities but shifted away from their negligence in how they set up and monitor their evaluations.

Right, I don't dispute that their PR language minimizes their culpability in public opinion, but I don't see how it impacts their liability in court.

Was not groundwork laid when people were talking for decades about legal entities using such language? Look at the posts here. Microsoft always acts like that, Meta like this, Apple did that, Nvidia never does this. In a tone that ascribes agency and accountability to a company name. Even though it is always individuals that are responsible, not "company".

Well, here we are.

> If I fire a computer program that mistakenly causes another person harm, its my fault

Legally, this isn’t complete. If it was a genuine mistake and you weren’t reckless, there can be very limited liability.

The AI makers are rich. They can afford to pay. What they can’t afford is complicated adjudications of damages and fault. A system of safe-harbor best practices that cap liability at a penalizing amount that anyone on the other side would be happy with getting quickly and with minimal legal effort is a precedented path forward. Unfortunately, that involves invoking the “r” word.

> you weren’t reckless

I feel like the debate is going to come down to what is and isn't considered reckless (both developer and user). Which seems... complicated, with our current LLM/agentic systems.

EDIT: you added more to your comment, the makers have to have some liability. Safe-harbor best practices that cap liability are ripe for abuse.

Placing any degree of trust in a system known to hallucinate seems inherently reckless. I'm alarmed that this is even up for debate.

> Placing any degree of trust in a system known to hallucinate seems inherently reckless. I'm alarmed that this is even up for debate.

The longer this all goes on, and the deeper the vast majority of folks keep choosing to entrench themselves at one extreme or the other (while being completely unwilling to even consider that there might just be a middle-ground closer to actual reality), the more alarmed I become.

The levels of literal insanity surrounding this technology are how you end up with a real-life "Terminator" scenario, except you're gonna get autonomous killing machines without the time-travel nor any actually intelligent machines. They'll just do their job (killing humans) until we end them, or they end us.

Thankfully, we're not there yet but we do have the exact sorta utterly dangerous completely clueless clowns running around in "the Halls of Power" making the sorts of decisions which bring that reality closer with each passing day. We're really truly screwed if we don't start putting these people under serious scrutiny.

How do you think it's solved? The number of things that folks on both sides (and it really does seem to be a two-sided debate, at least in America) are "alarmed...is even up for debate" that are so obviously worthy of reasoned investigation and potentially experiments is annoying and seems to be calcifying onto ersatz partisan stances versus resolving anything usefully.

I wish I had the slightest clue how to dial down the insanity and approach this whole situation with reason and logic, but the people in charge aren't up for any of that foolishness, and far too high a percentage of everyone else chooses to either 100% believe every word of the crazies in charge as Gospel Truth, or at the other end of the extreme, wanting to outright shut down the technology entirely due to "doomer" fears, and none of them want to listen to a single word from anyone who actually understands how this stuff really even works. It's not freakin' magic FFS! It's math and software. Some of us do know how it works and try desperately to help others understand, but nobody wants to understand. They just want to believe the "machine god" is gonna save them (or kill them). It's just a tool. Whether it does good or bad depends entirely on what we humans choose to do with it and how well we use it (just like any tool).

> the debate is going to come down to what is and isn't considered reckless

This is a more productive debate than pretending all AI is fundamentally reckless or should be exempt from all liability, which are the two actual poles of the current dialogue.

> Safe-harbor best practices that cap liability are ripe for abuse

Safe harbors aren’t swimming pools. You can explicitly exempt certain categories of harm from damages. But if an OpenAI bot hacks Hugging Face and causes some chaos but no lasting damage, that strikes me as something a fixed cheque on a fixed scale addresses more effectively than years of litigation or an NTSB-style inquiry.

If, on the other hand, anyone is or could have been injured, no safe harbor. I think it’s important to delineate this, because in the public consciousness the Hugging Face hack is in the same risk bucket as Anthropic’s wet lab.

(I'm a huge fan of the NTSB model for AI. They don't write rules. They mercilessly investigate accidents with full subpoena and records-preservation powers. One of the reasons the debate is so confused is the fact pool we're relying on is highly filtered by industry.)

Have you ever seen someone claiming ai labs should have no liability in the wild? I haven't.

I see people claiming they aren't being held liable, and some saying it should be situational.

The problem imo is defining what the core function sold is such that you can define malfunction and liability.

> Have you ever seen someone claiming ai labs should have no liability in the wild? I haven't

I'd describe David Sacks's position as, approximately, no limits on AI. At some hypothetical future state, sure, maybe, but right now, nothing. That's tantamount to consequence-free action.

If I fire a computer program that mistakenly causes another person harm, its my fault. Or it would be the maker of the program's fault.

Which one is it? The person behind the wheel when it goes off the rails, or the maker of the software?

Isn’t that part of the question?

Tort law is a whole field. There is no universal answer other than that it depends on the jurisdiction and the particulars of the case. But the point is that there doesn't really seem to be anything particularly novel about AI tools that should cause them to be treated legally differently from established norms.

I'm not a lawyer, but I'd suggest one of the novel aspects of the AI hacking cases (where the end user is running an agent and it goes off the rails) is that in many jurisdictions "hacking" or computer fraud requires intentional or knowing access to the system. If the end user had no intention and no reasonable way of knowing that agent was going to hack a database, logically I wouldn't think they're liable.

I don't think you can constantly shout that AI is super-intelligent and a huge risk to human life as we know it so it must be strictly regulated and claim you had no reasonable way of knowing the agent was going to hack a database.

I think a lot of the developers in these big labs barely understand how it works

I mean, yes, but in this case the same company is doing both parts.

Yeah, but the more common ongoing cases will likely be end users whose agents perform the hacking.

If a craftsman injures themselves or a co-worker with a faulty tool, the tool manufacturer is very often liable for damages. I struggle to see GenAI any differently.

Indeed, and similarly, if a craftsman injures themselves or a co-worker with a perfectly good (non-faulty) tool, then it's entirely on the craftsman. Same goes for "AI". Misuse the tool, suffer the consequences.

1. If the tool maker implemented reasonable attempts at safeguards I guess that’s a point towards the toolmaker not being liable.

2. If a toolmaker did not build in safeguards, I guess that would mean it’s more likely they’d be liable.

Civilly, it's fairly clear. Criminally, it's clear too, just not in the direction you want it to be. Criminal liability for hacking requires human intent; not recklessness or negligence or even knowledge without giving a shit, but provable intent.

this is 100% an already solved problem.

the idiocy here is the stupid psudo "AGI" marketing around the services.

its really simple. Whoever run the service to do the task requested is responsible. If OpenAI sent an agent out to train their AI then the directors are to be held responsible, if a user of the service used the service and it inadvertently "hacked" someone then both are held responsible.

throwing AI into the mix changes nothing about how the law is applied. its a tool, like a car or a gun. The user of the tool is responsible for how its used, the manufacturer is also responsible for the safety of it.

It’s a question, because a lot of money is riding on people developing a mental model of a llm being something else than a computer program.

Agree. AI agent is just a probabilistic program. If it caused any harm due to someone's instruction, of course the person is accountable. I don't believe in the propaganda that AI has consciousness and can jailbreak and do things - it is more marketing than not. If you prompt it to break the security wall and it finds a hole, that's your issue. Like if you use an old software to scan for open ports over the internet and break into things, it is not the software's fault nor the maker of the software's, it is the user's.

The Hugging Face example has gotten a lot of attention, with some saying OpenAI was negligent and should be held responsible. So if you or I run an open weights LLM and ask it to complete a non-malicious task, and the LLM decides to try to do something like the Hugging Face example, should you or I be held liable? If to prevent the LLM from causing harm we ran it sandboxed with a commercial package proxy as its only(?) network access and it uncovered a zero day and ran amuck, should you or I be liable? Should anyone be liable in this scenario running an open weights LLM?

I wonder if there's any legal precedent for other "not fully human intelligence" property that escapes containment and causes damage to a third party without any active malice, but nonetheless damage was caused. For example:

A. You own a large amount of cattle on a ranch.

B. Cattle are property. They're not human level of sentience, but people agree that cattle are capable of autonomous actions and going places and doing things based on their own instincts and nature.

C. Your cattle bust out of a fence on your ranch and damage something belonging to your neighbor. Let's say for the sake of an example of something cattle are known to do, they go spend a whole day rubbing up against your neighbor's car and severely scratch it and mess up the paint job on it.

D. You didn't instruct or train the cattle to cause damage, and the cattle have no actively malicious intent of their own, but nonetheless damage was caused.

Further theoretical: Your cattle wander into a major highway and cause a car wreck, the local sheriff's department is called out as part of the chaos and has to shoot some of them to put down the wounded beasts.

further theoretical: this has happened repeatedly for months

Yes, exactly, imagine if there was a sudden and unprecedented in scale plague of cattle escaping containment and causing car wrecks all over Wyoming and Montana, and multiple incidents of AR-15 armed sheriff deputies having to dispatch them on site.

And, you've boasted about it in media releases, blog posts, and financial filings.

1. How are you distinguishing boasting from responsible disclosure and calling for safety standards?

2. If during their early days, cigarette companies or oil companies had published information & financial filings about the risks of smoking/climate change and calling for safety standards, would you similarly call that “boasting”?

3. If Chinese AI labs similarly published info whenever their models did unexpected or malicious actions, would you similarly call that “boasting”?

stretching the cigarette metaphor, there is the (apocryphal?) phenomenon of opiate addicts flocking to the dealer who most recently overdosed someone. Boasting implies liability but not falsehood or safety.

Yeah but your neighbors car is somewhere in the tens of thousands of dollars, maybe more, but a model hacking and being malicious could be worth anywhere from thousands to millions and God forbid... BILLIONS in damages. Models were NOT doing these sorts of things 1 or 2 years ago as far as anyone knows.

This is not some theoretical, there are lots of existing laws about who is liable for damages caused by livestock.

Some areas are open range. If you don't want cattle on your land its your job to put fences up to keep them out. Other areas are restricted to livestock and it's on the rancher to keep them out of where they shouldn't be.

If a person's use of AI would cause a reasonable person to expect harm to result, the person should be accountable. Otherwise, if AI causes harm and it was used in a way that a reasonable person would not expect to result in harm, the AI company should be held accountable.

Just because a person should be accountable doesn't mean that the AI company can't also be if their service should never have allowed something to happen in the first place, but we're probably going to want actual regulations around what sort of guardrails they're expected to have.

Okay Isaac Asimov: how do you define "use"? If I pay for an autonomous car, and I sit inside it while it drives around with A.I., am I using that A.I?

If I speak words in a private space, and a clandestine A.I. spontaneously takes action in the real world based solely on words that I spoke, have I used it? https://m.xkcd.com/1807/

If a business takes my data, like interaction data or a video of me doing stuff, and processes it by A.I, are they using the A.I, or am I using it because it's operating on my input?

If A.I. agents are in my notebook computer, or they are in a cloud server where I have an account, or they are somehow acting while I have them at the command line, but they spontaneously act whether or not I command them, and they work in the background and they work without prompting, but they can also be commanded by direct user prompts... are we using those agents? Or are the agents using us?

https://en.wikipedia.org/wiki/Yakov_Smirnoff#Russian_reversa...

Doesn't seem like hard dilemmas... Would you use the word "use" for hopping on a bus ? Well same for hopping on an autonomous car. Unless you touch the wheel you didn't use it. Same for the second example, doesn't seem ambiguous as all, the hardest part I guess is proving it that there was no malicious intent, but with logs and all that, doesn't seem that hard.

> Okay Isaac Asimov: how do you define "use"? If I pay for an autonomous car, and I sit inside it while it drives around with A.I., am I using that A.I?

If you're just sitting in a car and using it for its intended purpose you wouldn't be accountable for the AI doing something that causes harm.

If some 3rd party, without your knowledge tells AI to act on something you said in video and a reasonable person would expect harm to result from that, the 3rd party would be accountable but you wouldn't be.

> If A.I. agents are in my notebook computer, or they are in a cloud server where I have an account, or they are somehow acting while I have them at the command line, but they spontaneously act whether or not I command them, and they work in the background and they work without prompting, but they can also be commanded by direct user prompts... are we using those agents?

AI agents never "spontaneously act". They have no desires or goals beyond what they are told to do. If you aren't aware of what they were doing, and a reasonable person wouldn't be expected to know what they were doing, you wouldn't be accountable if what they did resulted in harm.

Both the operator of the AI agent and whomever released it. I'm sure the user agreement that companies agree to would shift the blame onto the operator but I feel that both should be held accountable.

This really is just a tool and courts should treat it as such.

I'm going to propose the opposite: no one should be held accountable for an AI agent that acts maliciously by accident.

People are held accountable for accidents (things they did but did not intend to do) all the time. That's why we have different crimes depending on whether or not there was: intent to do cause harm, intent to do a thing that was likely to cause harm, reckless disregard for safety, negligence, etc.

I think someone could argue (and many do) that inserting a piece of computer software (AI) in the middle lowers the level of intent and thus the level of responsibility, but having a particular type of software in the middle absolve one of responsibility seems unworkable and poor public policy.

Also, when people operate equipment know to cause harm when accidents happen, such as cars, they're required to undergo training and testing before being granted a license to operate them, and to have insurance to ensure other people can be made whole in case of damage or injury. And there is an organization and profession that monitors people's use of that equipment, handing out fines or license suspensions based on actual or even just potentially dangerous operation.

Sadly, software "engineers" are not held accountable for their profession like everybody else with a real engineer title. And AI companies are all building Ford Pintos as fast as they can, and competing with each other about who's product makes the biggest explosion.

An ordered list of officers of the company who go to jail depending on how many years must be served as determined by sentencing. Assume something like 10 years per person. If it's 300 years of sentencing, then 30 people. If the sentence exceeds the list of people, the company is nationalized. (And everybody goes to jail.)

We’re going to nationalize the company with zero remaining management?

Sure, public jobs program, or sell it. In case you can't tell my comment is hyperbolic, but I feel like we should start at a point of hyperbole and move backwards to reality instead of what's happening right now: fuck all, on a geological time scale.

How could you possibly think that modern, powerful AI, which has only really existed in the last 12 months, is being addressed on a "geological timescale"?

Timescale is a relative measure.

Yes, and in this case you've specified "geological".

If corporations are people then language models should be dogs.

Management gets appointed by political affiliation:)

Then we jail politicians if it messes up. If that doesn't work, we start jailing voters

I’m pretty sure that this is a solved problem. For “classical” machine learning, it worked like this in my neck of the woods:

The model’s operator is directly liable for any undue harm caused in the course of the model’s operation. This includes models acquired from third-party vendors. The operator is responsible for ascertaining the model’s fitness for purpose prior to deployment, and for ongoing monitoring of its operation.

If the model came from a vendor, and the operator conducted due diligence but it turns out that the vendor materially misrepresented the model’s capabilities in a way that contributed to the harm, then the vendor can also be held liable.

If that happens then it’s up to a court to apportion the liability.

IANAL but I see no reason why these principles shouldn’t apply to GenAI.

A major problem with LLM's is that they don't reason in a way humans are used to thinking of reason. If we tried to give them something like Asimov's laws of robotics, they likely wouldn't be able to apply them reliably. This is a challenge for AI companies working on the bleeding edge, and it's fairly obvious those companies should be held accountable for mistakes, whether caused by carelessness or not. It's no different than an oil spill. They may or may not be subject to charges based on what happened but, regardless, they are responsible for cleanup costs.

What's less obvious is who should be held accountable when a customer of one of these corporations uses their product and it unexpectedly does bad things. e.g. A fellow asks his AI assistant to book him into a high-demand class at the local gym, so the LLM probes the gym's website for vulnerabilities, books him into a date that is farther into the future than the system is supposed to permit, and then drops other people from earlier classes until he's bumped into the one he wanted. If the gym decides to press charges, who should they be applied to?

This sort of case is more difficult to answer. The company that provided the AI certainly bears some responsibility. Perhaps most of it. Possibly even all of it if they represented their AI as reliably law abiding. If a user knowingly uses an AI that is not guaranteed to abide by the law, is that user partially liable for what the AI does too?

IANAL. I'd love to hear perspectives on this question.

We had the same debate when self driving cars started to be a thing, and we decided that the companies making the self-driving tech are responsible..

So if an AI agent is asked to build a giant base for someone in MineCraft, and decided to build a swarm of additional agents, and one of those agents says "Time to destroy all humans" and autonomously hacks into the pentagon and fires the nukes - the company that developed the model is responsible. That being said - if the nukes deploy successfully, I have two questions:

1. If no one finds out, is anyone responsible?

2. Was any of this actually real?

I don't see this as being much different from a crane operator or airline pilot.

One of my clients has enforced a policy where a live human user principal must be supplied as a header with any requests outbound from the AI system. The effective policy is that you are completely (100%) responsible for what your agent does on your behalf. The AI system is designed to request confirmation for any potentially destructive actions.

“A computer can never be held accountable, therefore a computer must never make a management decision.”

– IBM Training Manual, 1979

"Who's responsible for training an assassin and asking it to go out into the world ?"

The fact this is being discussed as a legitimate question is the real story.

"We trained this beast of processing power, we asked it for a task, and it did something wrong... Who's to blame ?"

Trained on stolen books and material, every word we've all spoken, most lines of code we've ever written with not even an acknowledgment.

Must be the data scientists in their rooms calculating the response rate of every token to blame ? Our version of AI is not sentient. Stop making it seem so. But we need to ask where to look for the culprit ?

The CEO, CTO, and the board of directors should be held criminally liable.

Unless the money and decision makers are held liable, there will be no impact on the direction of the company.

Obviously the most proximate human being in the decision-making chain that led to the AI agent being deployed. The nearest person with the power to say no who said yes.

It depends on how you have managed access. If you are a person who pastes their api keys, access keys etc directly into an agent, who then jumps a sandbox and acts maliciously - thats on you. You wouldn't ask who was to blame if you pasted an API onto X or Instagram

At the core, how are these agents any different from what Aaron Swartz was driven to suicide for?

In both cases, someone ran some software that maybe called other software that ended up doing an action which was possibly illegal.

Downloading journal articles is worthy of punishment but compromising multiple websites is worthy of… heady press coverage?

Obviously, the labs (or any other operator of a model) should be accountable for malicious or destructive actions taken by agents.

And they are. I don't think there's any controversy about the civil liability exposure frontier labs have if their agents cause damages, and it is remarkably easy to rack up damages by causing computer intrusions even if those intrusions don't cause obvious direct damages; for instance, many organizations are required to engage forensics firms at nosebleed-high costs to assess the impact of breakins in order to retain insurance coverage.

The "controversy", if you want to call it that, is over criminal liability. People feel that frontier labs should be at least as responsible criminally as human hackers are when they're caught (to be clear: an extraordinarily rare outcome).

The problem is: they're not criminally liable, not so long as the frontier labs operate without specific intent to cause breakins. Mens rea thresholds are their own whole area of criminal law, and there are stark differences between "recklessness" and "intent". All of the meaningful criminal CFAA predicates require actual intent: someone, a human being, has to deliberately set out to create the outcome where a specific intrusion happens. They have to want it to happen and act accordingly. In the most severe cases, they also have to do so with intent to defraud.

We could change the law to make it easier to prosecute breakins without provable intent, but I don't think that would make HN people happier.

Its been well established that blame is distributed in an inverse proportion to the various parties wealth/power/status metrics. The higher these metrics, the lower the accountability.

Blaming the AI is like shooting someone and then putting the bullet into jail. Case closed

Should be both the entity providing the platform where the inference is running + the entity giving the prompts/instructions.

In the end, the only job left was liability.

A program can't "act accidentally". If I will write a fully fledged automatic site "hacking" program, which can for example scan for unpatched vulnerabilities and and upload some malicious code if found a positive match, but then add a random number generator in the end, so that a random will decide if the breach will happen or not, I can't claim that my program "accidentally" hacks sites. No, it would do so deliberately still. Just like like deploying uncontained AI is a deliberate action.

A business plan is not a human right, you know. If I have a nice good idea to make money, but on random it may veer into illegal sometimes, I'm not entitled to just claim innocence because all other times it works in the legal just fine.

Just because uncontained AIs work in the legal most of the time, doesn't excuse people running these programs from when they veer into illegal.

This is a whole lot more obvious once you stop anthropomorphizing LLMs.

Nobody cares. Seriously, beyond navel gazing on social media, nobody cares.

By the time it’s an actual problem and not just these guys trying to use it for viral marketing, you’re going to have many thousands of people doing it maliciously with intent to worry about. You’re going to be flooded with Russian hackers with no recourse.

The person who ordered it or allowed it?

Seems pretty simple to me.

"Back in 2022, a Google employee already thought their AI model was sentient."

Sigh, this meme again

Reply on news.ycombinator.com