Computers Cannot Make Decisions
wiki.cateat.fish
wiki.cateat.fish
Decision laundering is a great term. There is ultimately a human who gave the program free rein. You can't say "sorry my car hit you, it's not my fault" especially when the person driving works for the car manufacturer too.
When someone hands the reins to an agent or group of agents, the agent(s) have free rein to do things, including things not directly intended by the owner.
If a person has a child, and their child does something bad, there is some mixed responsibility there. It’s a spectrum from age 0 to 18, where 0 is 100% parent responsibility and 18 is 100% child’s (legal adult’s) responsibility
As long as agents are unpredictable and highly capable, this will be a blurry line to draw.
Human user argues to the court with prompt evidence that they never told the agent to hack that company or dump that sensitive data online. The agent “decided” those actions were in service of the prompt but here’s a plethora of perfectly suitable alternative paths it could have taken instead. If I’d paid a human to complete the task, and they broke the law in the process, the onus is on them, not me, as I didn’t instruct them directly or indirectly to perform those specific actions.
It will be a murky legal mess. Though I’m sure it will provide fascinating legal arguments and judicial opinions in the decade to come.
What about a company that’s got an AI workflow and their client initiates some workflow that their AI agents carry out and inadvertently break the law. Is the company responsible because they own the agents? The org upstream that licensed the AI to the company, or the user who may not have even known AI was involved on the back end?
If the legal decisions are that AI are not humans and the human prompter is responsible even if they didn’t instruct or intend said action, well, then any sane person/company wouldn’t touch these tools with a 10 foot pole. Perhaps our salvation lies in that type of legal precedent.
Which also seems to be in line with the way the frontier labs gathered and used copyrighted data to train their models
And even before AI, limited liability corporations (like LLCs or C corps), also have the intention of protecting the individuals from the repercussions of their actions
In a way this could be generalized as “accountability laundering”. Making someone unaccountable for their actions, by redirecting the responsibility of those actions, to a non-punishable entity (like a C corp or “a computer”)
The question isn’t who or what made the decision, per se. It’s who is responsible for the outcome. The entity making the decision might be the actual cause but not the proximate cause (the cause most closely connected to legal liability) for the harm that occurred.
The company is liable either way. And in that case, there would be plenty of other problems besides the hacking. We'd have look into why they're making Claude work ungodly hours 24/7 without compensation :)
The act of delegation to chance is the decision.
This. I agree 100%.
Computers make decisions all the time, and have been doing so for decades. But we still keep the human behind the computer accountable.
That narrative is bullshit. Intelligence is not power. The power comes from hooking these agents up to everything without proper safety controls. The fact that those safety controls don’t exist yet is not humanity’s problem, it’s the frontier labs problem, and we can’t let them wriggle out of it with an Overton window shift or demanding that the government take responsibility via regulation (of course they should but it doesn’t absolve the labs of responsibility for their own actions).
The missing piece right now is that our AI safety and cybersecurity controls are simply not up to snuff with the current risks that AI has introduced. Given the increased risk, we need to level up every aspect of our cybersecurity stance. These are hard problems, but not impossible given proper incentives. The issue is that no one is connecting the dots between the creation of these agents and the tools and access they are given, with the harm they may produce. Often the argument jumps straight to "it's going to get too smart for us to control" while glossing over the fact that all the power AI has was directly given to it by human choices. All this cognitive dissonance and willful blindness is caused by unchecked capitalist incentives. Think of the amount of money every single person in the frontier labs stands to make if the most optimistic predictions pan out. This is moral hazard of the highest order, and the only counter-measure that can work in our current social and political system is liability.
and catering to them (boxes) to not accidentally start disseminating child porn is most commonly understood as a responsibility of the owner, the moment he plugs his device online.
Related: The Unaccountability Machine (good book by Dan Davies, who also wrote Lying for Money about financial fraud).
Don’t confuse lack of accountability with lack of responsibility.
If one needs to be "allowed" to make a decision, does that even qualify as a decision?
For example, I can decide which cereal to purchase at the supermarket. But I cannot decide which cereals are offered to me at the supermarket. The supermarket has allowed/enabled to make a constrained decision.
So does it follow then that the agents in question do have "some" free* will? I'm not sure you realize what conclusion naturally follows from this, and how it contradicts the point made in the article and the top-level comment I replied to.
(*) My earlier question wasn't explicitly about free will but we'll go with it.
While I agree that computers cannot currently be held accountable for decisions (because I do not think they have free will), I think it’s transparently clear that computers make decisions all the time, and have done since they were first implemented.
Many of these incidents relate to hacking, where criminal law commonly requires intent.
Claiming that a machine cannot make a decision does not mean an employee at OpenAI decided to hack an Australian government website.
> Inadequate or broken safeguards do not fall under my definition of reasonable.
I see no such assumption there. In my reading, the implicit assumption is that adequate safeguards are necessary and that anyone who doesn’t build them is irresponsible. And I feel that that is correct.
The AI labs say "AI can make mistakes". That's why the responsibility of using it is on you. This may change if we get more legislation. But given the argument above, it will probably never shift to the AI itself having any responsibility. Instead the AI labs can be responsible.
It's like having an assistant that can be helpful, but not always. We can bicker about the ratio of helpful to unhelpful interactions, but there will be some unhelpful moments.
And hoo boy, those can range from "eh, whatever" to being like a deranged toddler just got its hands on whatever tools and permissions were granted to the "mostly helpful" assistant.
And we can't get rid of that chance. We can pare it down a bit. But we can't get rid of it because it is fundamental to how the technology works.
It's like a vending machine that 99% of the time it dispenses tasty snacks, but 1% of the time it gives you poison that appears to be normal, unexpired product. It just does this. They did their best to minimize how often it happens, but the poison chance is still there. It can't be removed. It's part of the system.
A "normal" vending machine has failure conditions, sure. They generally can't do worse than cause you to lose your money without giving you a product.
Outlook 2003 never presented a risk of draining my bank account to buy memecoins because a spam email was treated as instructions.
We can (and should) try to improve AI safety and alignment, but perfect safety is probably impossible without losing the capabilities that make AI unique. Black-and-white positions that insist anyone creating or using an AI is responsible for anything it does wrong under any circumstances are also unrealistic. AI is a general purpose technology, and we have always accepted that it is not practical for suppliers to envisage and constrain every downstream use: Car manufacturers cannot prevent every instance of mechanical failure, nor can they stop cars being used in crime.
Making decisions means taking responsibility.[0] This is really all there is to it -- a computer cannot decide, not because humans are uniquely able to make a choice based on input, but because a computer cannot be responsible. You can decide to do something based on the output of the computer, but that's you deciding to do that. At any point where a computer appears to have made a decision, someone else was behind it. You can complicate this, randomise stuff, obfuscate things, etc. but ultimately either the computer was set in motion by a person, or it malfunctioned. The malfunctioning computer thing is quite interesting but it's not going to change the outcome.
LLMs ("AI" today) have not changed this. It's still just software, it's still doing what it's told. Surprising future technologies may alter this, but there is no evidence whatsoever to suggest that's a likely future and it certainly is not the case today.
> Black-and-white positions that insist anyone creating or using an AI is responsible for anything it does wrong under any circumstances are also unrealistic.
At this point in time, AI is just a piece of software. It doesn't matter how useful it is or not. It's just software. We can and should treat it like software. So to evaluate when/where/how someone should (or should not) be held responsible for something that "the AI did", we can start at the same point as we would for any other vendor/creator/user of a piece of software. I'm not going to get into that -- my point is that the framework exists. And yes, sure, maybe that framework needs some tweaks in consideration of the new software, that's OK. The starting point should still be in reality, where software is software.
[0] Yes, this is nuanced, things go wrong, different situations, etc. etc.
Are you arguing that not computers are best at chess, but that whichever human last updated the chess algorithm is actually the reigning chess champion?
Suppose I check the weather forecast 5 mins before leaving the house, and it says 95% chance of rain today. Do I bring or not bring an umbrella?
I would bring an umbrella. I would made that decision to bring an umbrella. But my decision to do that is (at least partially) the result of the weather forecast. The causal chain is (atmospheric conditions -> weather forecasters -> my decision). Did the weather forecaster just "program" me? Did the water molecules moving in the sky just "program" the weather forecaster?
My answer to these questions is as follows: a decision, and my understanding of the concept/definition of a "decision", refers to exactly the processes where I do something as a result of external influence/programming, rather than just internal thought/processing. I could think really hard and imagine myself moving forward at 100 meters per second, or even beyond the speed of light, but in reality, no matter how hard my neurons fire to imagine that scenario, my actions are still physically constrained. Such physical constraints influence our every decision. That is the "programming" part. Everyone and everything "programs"/influences everything else.
That said, perhaps the particles which make up my brain/body has its own whimsy. Perhaps there is a hidden local random number generator that samples from the distribution of possible future states available to this particular clump of matter that is "I".
How different is this from a computer running a non-deterministic program such as an LLM-based AI agent? Remember, every token sampling involves an RNG. They don't perform as well or feel as creative and human-like when deterministically sampled.
This isn't an excuse for anyone, especially not for Anthropic or OpenAI. But I think these are questions we ought to ask if we are to honestly discuss liability and perhaps even personhood for AIs that are increasingly autonomous.
Discussing anything like personhood for AI is really jumping the gun.
The chatbot is software. The software does what it is told. It performs tasks requested of it by the user/operator in line with the (designed, ) immplemented functionality of the code. If the software fails to do what it is told, the vendor has not provided working software. If an operator controlling some software does not do so in a reasonable, safe manner, the operator is negligent. If damages result from such failures or actions, the people who made the decisions that led to that are the people who are responsible.
It does not matter how complex the software is: if a vendor of some software doesn't expect it fail in costly ways because it's too complex or hard to understand, that is a (legal, ethical, moral) failure of the vendor.
It does not matter how autonomous it is: an autonomous sprinkler system that floods the neighbour's basement is not held liable for the damages. The owner, installer, vendor, etc. is. -- maybe the neighbour even did(n't) something wrong!
Theologists and philosophers alike have been thinking about the problem what "free will" even means exactly because of this, when there are no preconditions informing my decision, what does "free will" even mean?
In German, I have heard this being framed as "bedingt freier" versus "unbedingt freier Wille".
As a thought experiment, it makes a very good argument for a deterministic world view, quantum woo aside.
When there are no reasons for what I decide, what are my decisions based on? A whim? Chance? My personal taste? Well, my personal taste is a reason. And all reasons come from somewhere, right?
On the other hand, thinking of my decisions as without preconditions: what does that give me?
Decisions without rationale?
Sure, one can make everything more fuzzy by introducing probablities, even real chance if you like.
But also then, total determinism with hidden information is an immune theory that cannot be refuted, so it explains nothing.
But was does true randomness help with? I'd say, from a philosophical perspective, it doesn't bring any interesting argument to the table.
So this leaves us with free will as the only useful way of understanding ourselves and our role in the world, but on the other hand, it seems like a sneaky device for ourselves to ignore the infinitely confusing thought of considering ourselves determined by our circumstances (which, as a thought, already has real side-effects, too).
Even in near trivial cases such as Game Of Life (just a handful of rules), the fact that a <machine> technically carries out deterministic actions is already quite meaningless.
compute /kəm-pyoo͞t′/ intransitive verb
To determine by mathematics, especially by numerical methods: synonym: calculate.
"computed the tax due."
To reach a conclusion or form a judgment or opinion about (something) by reasoning or consideration.
"decide what to do."
The programmer, architect, CIO, engineer, etc. made the decisions, reached the conclusions, reasoned about the problem.
In the case of an AI vibe coder, the collective published works of said programmers, architects, CIO, engineer, etcs reasoning was applied to the vibe coders question, itself a result of reasoning. AI is still just calculating.
Digital computers ( the ones I assume we are talking about ) are turing machines.
Electronic tape readers.
Modern player pianos. V
Only barbarians would attempt a farcical judgement of law or punish the bookcase.
Thats what it comes down to.
I disagree. A light bulb is on if and only if it is supplied electricity. With a switch, I have control over input of electricity. The bulb is on if and only if the switch is toggled on. The bulb isn't deciding whether or not to be on based on the if-statement represented by the switch. It just is on or is off depending on conditions.
To make it more dynamic I can replace the switch with a motion sensor. The light is now on if and only if motion is detected. The light is not deciding to be on but must be on as required by the motion sensor toggling the flow of electricity. The light either is or isn't on depending on the conditions.
If I write an if condition that prints true if there are more than ten files in a folder, and then a second that places a random selection of up to 20 files in the folder before deleting them every second, the if function will be true roughly half of the time. The function either is or isn't true based on conditions.
All of the decisions made for and involving both the lights and the if condition were made before they existed. Once implemented, those decisions will remain true until some force acts upon them (later conditions or decisions). While the initial decisions (to wire the light, to write the if condition) are true, the system itself makes no decisions. Either systems current condition depends entirely upon the decisions made before it existed.
Saying an if-condition makes decisions is like saying a bicycle decides to fall down when running out of momentum or encountering an obstacle.
This is because they can't take the blame, they can't be held responsible.
These are not the same, and this misinterpretation of the author’s intent inevitably follows.
CAN NOT is horribly ambiguous in English which is why RFCs use MUST and MAY.
In casual speech someone might say “you can’t park there” but that’s not true. They might mean “you must not park there” or “you should not park there” but the only reason they are telling you is because you CAN park there.
Is it ?
I thought 'can' denotes a possibility, while 'not' negatates it ?
For another example of ambiguity, consider the sentence “A triangle can not have any obtuse angles”.
Is this 1) true, because there exist triangles with only acute angles, or 2) false, because triangles can have obtuse angles?
Did you mean, "A triangle can only have one obtuse angle ?"
Or was it a trick question ?
Though LLMs make this all fuzzy it's still basically the same thing. LLMs didn't decide to make a toxic culture of LLM use, people did. They were told that it was necessary to save themselves.
A bad manager takes personal credit when things go well and finds a scapegoat when things go bad. A good manager gives credit when things go well and takes responsibility when things go bad.
It's wrong to scapegoat the agent. You, the human who executed the agent, made the decision to execute it. You gave it credentials, and set up insufficient guardrails. Don't scapegoat the agent for your poor decisions.
Likewise, it's good to have humility when you get great results out of agents. Yes, the agent made the good results happen. It's OK to tell people that your choice of agent is amazing and did an incredible job. In any half-decent work culture, that should rub off on you as well.
That's not the situation here. We aren't enabling the career progression of software which does matrix multiplications.
Accountability of us schmucks at the end of the org chart goes both ways. We get some reward when we do it right, and we improve when we do it wrong.
A reward might be something tangible or might be as small as your own personal satisfaction.
An improvement is hopefully identifying what went wrong and doing better next time, and in extreme cases your employer will let you try again at a different workplace.
The question of whether it's better to frame a reaction to issues in a positive, do-better-next-time or a negative, don't-do-that-again way (fwiw, I agree that the positive way is better) is orthogonal to the argument I was trying to make about who is the recipient of that reaction. Scapegoating the agent is farcical. Clearly, "make no mistakes" style prompts are a poor approach to agentic development.
> the manager's job is to enable career progression of their reports... we aren't enabling the career progression of [fancy calculators]
While I do agree that good human managers concern themselves with career development of their human direct reports, I disagree that it's a fundamental part of the job description. The foundation is for a manager to get and keep their reports being productive. Human workers are often concerned with career progression, so good managers concern themselves with such career progression as well. Agents are not, but that doesn't mean that lauding good agentic work is meaningless. Lauding the work is also valuable for: (a) reinforcing a culture of positivity and celebration, which rubs off on human workers as well, (b) part of a culture of continuous evaluation of agents and models - asking questions like, are there cheaper or faster models that would be just as effective?
Humans, especially ones under abusive managers, often have a lot of responsibility but no authority that affects the outcome.
In an ideal situation you either have both, or none. You make the decision and answer for it, or you follow decisions but don't get blamed when it's wrong.
LLMs managed to slide into a space where they get to make authoritative decisions but if something fails we don't really blame them.
I'm not saying LLMs should be held responsible, but it's funny how people claim they replace humans, but are giving them an extremely easy, kid-gloves, success target.
Instead of "Anthropic submits false tip to police", I'd prefer "Anthropic employee John Smith deploys an agent that submits false tip to police". Of course, John Smith will be able to then say "I did so as ordered by my superior, Ms Sue", and so on, with a very clearly established chain of liability.
A corporation and an AI are not very different in many aspects, except one of them is very favourably treated by law, almost as a living person.
"Anthropic employee John Smith [killed the children with the gun||launched a cyber attack] because he was ordered to by his superior and [has all the blame because such orders are null || we agree that Anthropic can order children to be killed]"
Ambrose Bierce, The Unabridged Devil's Dictionary (1906)
...like it already happens in europe, or at least in Italy.
This is basic law and risk management in society. Let's not pretend we have just invented corporations and don't know what to do with them yet.
Oh, wait, that was a core concept from "Mein Kampf": parliament are bad because nobody is responsable, hence nobody care about doing the right thing ! (probably the only concept from that book that I can agree to)
One does not, under any circumstances, "gotta hand it to Hitler".
Regardless, this is less about dictatorship and more about personal responsibility and accountability
A solution, perhaps, would be to ensure that such decisions are not anonymous and that, if the decisions had bad consequences, then the related people would be accountable
Of course, this would make the politic business less attractive. Would that be a bad idea ?
But I don't think the average person with their $20 subscription to whoever is the primary source of revenue. Enterprises fund most of it by buying API keys and making those fancy chatbot buttons on their sites that nobody clicks, as well as some internal business automations if I had to guess.
If it was just average people with $20 subscriptions funding all this (keep in mind, most people using AI services do not pay for it), by now, all these AI companies would have gone bankrupt.
These businesses lose unbelievably large amounts of investor money. Their AI-related revenues don't get even close to paying for their AI-related outlays.
End user (invididual, small business, large business) clients are still complicit not because they are funding it, but because they are signalling to investors that this is something people/businesses want.
Nobody gets off the hook here. It is every bit as much end user nihilism as it is AI company nihilism.
Erm, you appear to have missed the last 100+ years of updates to Western Capitalism.
People don't want it, you make the demand with hundreds of millions of spend on advertising (brainwashing), influencers, and social proof.
The capital holders don't sit back and see which company wins so capital can be directed to the best solutions, they buy up the competition, they pay corrupt politicians, they use current v market placement to embed their products name into billions of workplace computers and spread stories about how it's going to 'increase productivity but never at the cost of jobs', etc.
Yes, ultimately a dollar is a vote, but I don't think you can blame the people being manipulated "every bit as much" as those controlling the manipulation to feed their avarice and/or megalomania.
Not really.
> Yes, ultimately a dollar is a vote, but I don't think you can blame the people being manipulated "every bit as much" as those controlling the manipulation to feed their avarice and/or megalomania.
Oh but I do.
Collective social guilt is a thing. e.g. who is to blame for drug violence? Not only drug dealers.
AI spending is a bit like Pokemon games - most of your players are free to play, only earning you traffic and maybe some more players after recommending it to their friends - that's the average AI user. Around 5-10% of players pay a bit to the game - maybe to unlock some extra characters - those are the $20-$200 AI subscription users, with these you are not going to fund the whole business but it's some side cash. Then you have the 1℅, which spend massive amounts of money onto the game, and those are the ones primarily funding it - that's the business spending and where most of the money actually comes from.
You can only blame the average person for using the cloud service and recommending it to others.
Anyway, the only "whales" funding the AI companies is venture capital. The reason it's not a viable business (yet, so far, whatever) is because they spend more money than they make.
These firms have been using computers to make decisions about trading billions of dollars for over 30 years.
At the same time, I've been involved in hundreds of After Action Reviews of incidents and it's not always simple or easy to both figure out which party was ultimately responsible and also how to allocate the "dollar error budget" to the parties overall.
(This could be a blog post in and of itself btw)
Instead, this is a legal question: What were the damages? Was the target company negligent in setting up security? Were criminal laws broken? Etc. [IANAL, so forgive the lack of precision.]
If hacked companies had losses, they have an incentive (even a fiduciary duty) to recoup their losses. They do that by suing the company (as, e.g., authors sued Anthropic for copyright infringement). If they haven't sued the company, it means either (a) they didn't have losses, or (b) they're still working on the lawsuit and haven't filed it yet.
The other dynamic is that frontier labs feel competitive pressure to take risks to develop their models. Which risk is higher:
A. A new model being tested hacks into a sensitive site (bank, government, etc.) and the company has to pay a massive fine (or even face jail time).
B. A competitor comes out with a significantly better model and they lose significant market share.
That is a hard choice to make. The frontier labs are asking for government restrictions--that apply to all companies--so that they can minimize A without having to worry about B.
They are trying to minimize their risk (and thus maximize their profits).
But I could be wrong.
Can anyone show me an example? And I mean a real example - not an inferred example. Someone saying "It wasn't me, it was my AI" or someone not being held responsible for the actions of their AI.
I won't hold my breath.
EDIT: Heh... no examples. I rest my case.
The argument from the frontier labs isn't, "We're worried that AI is going out of control, please stop us" it is, "We're taking a lot of risks to compete, so please limit competition."
No, they really don't. I suppose I agree to some extent that no one is trying (hard) to deflect blame, because the people in control are barely a part of the conversation in the first place.
Almost all of the reporting on these stories assigns the blame to the chatbots first. Most prominently through the language/wording that gets used -- it's usually e.g. "BOT infiltrated ...", "Agents ruined the ...". I mean, they call them agents. I'm not trying to claim this terminology was selected as an entirely strategic move, but it sure does seem convenient. Calling them agents shifts the conversation away from the real human people and companies that are controlling them. This extends well beyond this one word, as well. The (captured) media picks up on and adopts the language used by the tech/AI companies.[0] In this way (among others) these companies can influence how the media portrays them indirectly.
> Instead, this is a legal question
There are legal questions, but the two (legal, ethical) aren't mutually exclusive. The criminal law question is the most significant one here, I think. If you do what they did as an individual, you are pretty likely to get visited by the FBI.
You're skipping over a couple of important points, I think:
- The AI labs clearly stand to benefit from the "AI is dangerous narrative". We know this for several reasons, probably the most significant one is that they keep telling us how the new one is super extremely dangerous, and proceeding to make no changes to their behaviour. It's marketing.
- The AI labs are clearly in control of their bots. It's a computer. Networks are not mysterious. Computers don't do surprising things. The principles and technologies involved in securing a network are mature and there is a ready supply of trained workers available worldwide, keen to apply those principles and deploy/manage those technologies. You can believe that something went wrong and the various hacks and excursions were unexpected. In that case, the labs are clearly incompetent[1] and also, they're still to blame. Repeatedly running "experiments" that you can't control is, in fact, a failure of basic morals and ethics.
[0] Of course they do, right? Well there used to be a distinction made between experts and marketing departments. These days it seems a lot of journalists think that marketing copy is as good as any other source.
[1] Surely they hired someone that can setup a firewall.
Yes, exactly. The issue isn't, "we don't know how to sandbox an agent", the issue is that they need to test the agent with full internet access because that's the product they are selling. They are selling a product that does things for you on the web. You have to be able to test it with a real web connection.
Maybe the answer is to simulate the web. What if you ask an agent to reproduce the key parts of the web (travel sites, government sites, etc.); effectively, clone a real, working version of the web, but inside a simulator. Now you could test the agents inside the simulator. Since "coding is a solved problem" maybe that wouldn't be as hard as it sounds. Right?
It does not matter if the CEO is an AI or a human. In either case, of course they can make decisions -- and be held responsible. Not in any emotional sense (that we seem to want to bake into the concept for added confusion), just that if you are not satisfied with the result, you can replace the human or AI.
I am not saying that this is a fun vision of anything, but why are we making it more complicated than it is? The parts are all there and explained.
My understanding is that 'emotional' sense emerged organically in LLMs and was not intentionally baked in.
As silly as it may seem, the ways emotions come through in our words do seem to have an effect on agents. Likely because they were trained on human writing, most of which cannot be separated from the emotions of the writers any more than your emotions can be separated from your logic & reasoning abilities.
While it seems reasonable that you need humans to take accountability, it looks like a very shaky position to assert that computers can't make decisions. And going back to chess computers, we acknowledge that computers make better decisions than humans in that area.
Forbidding computers that can make better decisions than humans from making decisions to keep accountability just seems to be arguing to hire fall guys. You still want the better decision maker to make decisions, but you need someone there to take the fall if things go wrong.
Very weird incentives.
Suppose I go to the park and find one of several guys who would do it for free. I get all the of benefits, he might do an excellent job, but, because there's no contract, I still take all the blame.
Suppose I actually hire a professional to do it. I receive fewer of the benefits. However, he's much more likely to do a good job, and the legal system provides a variety of ways to hold him accountable for the work.
Suppose I hire a professional, but then insist that he use legions of guys from the park to do the actual work, since that's going to be some much faster and cheaper per quantum of work. I also tell him that if he's unwilling to do that, I'll find someone who will. I also tell him that, since he has so much more unoccupied time now that the original work is in the hands of all those other dudes, he gets to supervise still more work done by other swarms of guys from the park.
At this point it should be clear that I'm not interested in quality work, but in having someone poised to absorb blame. I want a patsy.
Now replace "guys at the park" with LLMs, and replace "free" with "nearly free, but with no warranty."
The problem is entirely different. Sometimes they make decisions we don't like, but unlike with people we haven't found a way to punish them that satisfies our sense of justice or discourages other computers from making similar decisions.
You might have an org policy that says that an llm can't make decisions because an llm can't be held accountable - but that's your choice.
You don't need to operate with a model where everything has a human accountable for it in the end.
If you don't like the decisions an llm makes, you can replace it with another or put a human in charge of that decision in the future instead.
I'm not saying it's a good idea, but you can technically do it.
The piece: "Anthropic COULD stop their models from doing what they are doing, and they are actively choosing not to"
As if _Anthropic_ could make a decision.
Companies cannot make decisions! All of these incidents are the fault of people. Companies are not deciding to do this, people are. There is little more to it than that.
They generate text which resembles reasoning and may include a tool call
The harness runs tool calls
The human made the decision to write the harness
The human is responsible
Also phrased as ‘to make an apple pie from scratch, one must first create the Universe’
edit: This wasn’t meant to be a rhetorical question!
If you are driving a car, it's pretty clear where the accountability falls. You control the car. You have free will. If the car hits a tree, you hit a tree.
What happens when a car hits another car? Things get murkier. There are lawyers who have devoted their careers to this. Sometimes nobody winds up at fault. Unavoidable accidents happen.
Now let's ask what happens when a leader orders their military to do something. e.g. Say a President instructs his troops to do whatever they think is necessary to get the job done, and they go and Abu Ghraib a bunch of prisoners. Who is responsible? The leader probably is, ultimately, but there are a lot of human beings with free will in between the leader and the people wielding electrical cords and pliars. Typically, somebody fairly close to the bottom takes the blame, even if most people suspect it should fall higher. Some low rankers will get court marshalled but the leader will have the occasional shoe thrown at him. That's okay. He's good at ducking.
Say you're running a war against people you really just want gone. Your military has a tradition of conducting laboriously researched precision strikes to take out people they don't like. Your people are working hard, but they're only managing a couple strikes a week, and you have so many more bombs than that. You don't even pay for most of them! What if you could just ask a machine to find you the leader's underlings, and their underlings, and so on, right down to raw recruits? Nobody has to check on the machine. Why bother? Just believe the machine and drop the bombs. If anyone ever calls it a war crime, you can just blame the machine, right? Well, not so fast. Most people are smarter than that!
Despite all the sci-fi we consume that portrays computers as having agency, malevolent or not, most people instinctively understand that, in the real world, computers are just tools, more akin to cars than armies. LLM's are just software that run on those tools. Anyone using a LLM can simply choose not to, or they can choose to be very careful in how they use it. There isn't a complex web of free wills to untangle. There's the machine and the person who controls the machine.
We currently have an administration that is desperate to look the other way while U.S. AI companies do things that any reasonable person would hold them accountable for. "Industry will regulate itself!" The economy mostly sucks because of this same government's complete lack of economic sense or even basic self-control, but the AI sector is propping things up. They're not going to regulate the golden goose unless the goose does something so unbelievably stupid that there's no choice. Wouldn't want to pop the bubble!
Yes, we are headed for more stupidity. The "tequila and handguns" kind of automated stupidity that only a computer and a truly feckless operator can give us.
https://www.hec.edu/sites/default/files/documents/Computing%...
First of all, looking past the objectively wrong phrasing (computers can, and clearly do make decisions), can computers be accountable? This might not be so cut and dry as to instantly say no. This one will keep ethics, philosophy and legal practitioners busy for some time. It's certainly not going to be decided in HN comment or a blog post.
Second - while the AI labs have shown utter incompetence in properly sandboxing their agents, intent is still important. I don't think the labs deliberately meant for their agents to hack this or that. Should they be accountable? Yes but I wouldn't go as far as saying this is deliberate.
The next point is about cherry picking some "doomsday" scenarios like AI ending humanity and then blaming this on the reader (!) for paying a subscription. - People pay because AI is USEFUL, and massively so. You could just as easily pick incredible achievements propelled by AI, in mathematics, software, reverse engineering, role playing. Unlike the "end is nigh" ideas, these advancements are actually real, and they are happening right now.
And last, the author says AI labs can just stop the agents at any time. I agree there needs to be better discovery and mitigation, sandboxing and monitoring. But when you run a billion agents, it's always going to be a numbers game and there will always going to be some challenge in fully containing all breaches. This will only get more difficult with better models, because they're getting smarter and they know how to work around things, sometimes better than we do.
So what's the solution here? There are options, but they're all tradeoffs. I hope we get better at this and maybe working on cutting edge models should breed some new best practices which were once only reserved for defense tech.
Revisit later if the comparison starts to become ridiculous.
Characterizing the difference between humans and AI is going to be important.
Pointing out that AI is software isn't helpful in answering that question.
If I write a program:
if (rand() % 2) launch_missles();
And wire that function to actually launch a missile, the computer didn’t make a decision. And so it is with people prompting agents.
There is no end of the world, nor autonomous agent decisions nor any fantasies they are building up to make people believe they have a pandora box in their hand.
What we instead have are lies crafted to lure non tech people and keep this running as long as they can.
Oh and let's also not forget that they are allowed to illegally download basically as much as they want from libgen/annas archive/torrents to train their models under the "fair use" umbrella, yet mere mortals can go to jail for way much less.
Has the alignment problem been solved, then?
mistakes happen. sometimes catastrophic decisions (or indecision) are made, and people and companies are held accountable for them, or not
some companies are too big or too important to fall. we can all agree or disagree on things, but what AI specific behavior are we actually talking about?
No company is too big or too important to fail.
Some, unfortunately, survive because of the expedient choice to keep them in place.
"JO-NA-THAN!"
ETA: more to the point, neither OpenAI nor Anthropic are too big, too important, or too structurally essential to fail. (Slightly more challenging to make the third claim for Google or SpaceX, but either of those could see their governmentally/militarily essential parts nationalised).
If the USA grants either Anthropic or OpenAI too-big-to-fail status, and that privilege is invoked in a crisis, it could mean the end of the US economy.
> Your Claude subscriptions are funding this. You are funding this.