The Hallucination Defense
niyikiza.com
niyikiza.com
> That’s the defense. And here’s the problem: it’s often hard to refute with confidence.
Why is it necessary to refute it at all? It shouldn't matter, because whoever is producing the work product is responsible for it, no matter whether genAI was involved or not.
AI is just better because no one can actually explain why the thing does what it does. Perfect management scapegoat without strict liability being made explicit in law.
The software world has been very allergic to getting anywhere near the vicinity of a system like that.
Licensure alone doesn't solve all these ills. And for that matter, once regulatory capture happens, it has a tendency to make things worse due to consolidation pressure.
AI is worse in that regard, because, although you can't explain why it does so, you can point a finger at it, say "we told you so" and provide the receipts of repeated warnings that the thing has a tendency of doing the things.
And when sub-agents or third-party tools are involved, liability gets even murkier. Who's accountable when the action executed three hops away from the human? The article argues for receipts that make "I didn't authorize that" a verifiable claim
If you don't want to be responsible for what a tool that might do anything at all might do, don't use the tool.
The other option is admitting that you don't accept responsibility, not looking for a way to be "responsible" but not accountable.
Had it worked then we would have seen many more CEOs in prison.
Such so magic forcefield exists for you, though.
In practice, almost everyone is held potentially or actually accountable for things they never had a choice in. Some are never held accountable for things they freely choose, because they have some way to dodge accountability.
The CEOs who don't accept accountability were lying when they said they were responsible.
[ $[ $RANDOM % 6] = 0 ] && rm -rf / || echo "Click"
# terrible code, never use ty
def cleanup(dir):
system("rm -rf {dir}")
def main():
work_dir = os.env["WORK_DIR"]
cleanup(work_dir)
At what point is it negligent?
Though rm /$TARGET where $TARGET is blank is a common enough footgun that --preserve-root exists and is default.
Without quotes it becomes: `rm -Rf /home/user/documents and settings /bin`
0: https://github.com/MrMEEE/bumblebee-Old-and-abbandoned/commi...
If one decided to paint a school's interior with toxic paint, it's not "the paint poisoned them on its own", it's "someone chose to use a paint that can poison people".
Somebody was responsible for choosing to use a tool that has this class of risks and explicitly did not follow known and established protocol for securing against such risk. Consequences are that person's to bear - otherwise the concept of responsibility loses all value.
Frankly, I think that might be exactly where we end up going. Finding a responsible person to punish is just a tool we use to achieve good outcomes, and if scare tactics is no longer applicable to the way we work, it might be time to discard it.
It's scary that a nuclear exit starts looking like an enticing option when confronted with that.
I'm not claiming to have all the answers about how to achieve that, but I am fairly certain punishment is not a necessary part of it.
Twelve–year–olds aren't capable of dealing with responsibility or consequence.
That value proposition depends entirely on whether there is also an upside to all of that. Do you actually need truth, meaning, responsibility and consequences while you are tripping on acid? Do you even need to be alive and have a physical organic body for that? What if Ikari Gendo was actually right and everyone else are assholes who don't let him be with his wife.
What if I hire you (instead of LLM) to summarize the reports and you decide to email the competitors? What if we work in the industry where you have to be sworn in with an oath to protect secrecy? What if I did (or didn't) check with the police about your previous deeds, but it's first time you emailed competitors? What if you are a schizo that heard God's voice that told you to do so and it's the first episode you ever had?
"Three months later [...] But the prompt history? Deleted. The original instruction? The analyst’s word against the logs."
One, the analysts word does not override the logs, that's the point of logs. Two, it's fairly clear the author of the fine article has never worked close to finance. A three month retention period for AI queries by an analyst is not an option.
SEC Rule 17a-4 & FINRA Rule 4511 have entered the chat.
No, it's trivial: "So you admit you uploaded confidential information to the unpredictable tool with wide capabilities?"
> Who's accountable when the action executed three hops away from the human?
The human is accountable.
----
A computer can never be held accountable
Therefore a computer must never make a management decision
Now companies are all about doing bad all the time, they know they're doing it, and need to avoid any individual being accountable for it. Computers are the perfect tool to make decisions without obvious accountability.
That's an orthodoxy. It holds for now (in theory and most of the time), but it's just an opinion, like a lot of other things.
Who is accountable when we have a recession or when people can't afford whatever we strongly believe should be affordable? The system, the government, the market, late stage capitalism or whatever. Not a person that actually goes to jail.
If the value proposition becomes attractive, we can choose to believe that the human is not in fact accountable here, but the electric shaitan is. We just didn't pray good enough, but did our best really. What else can we expect?
It really doesn't. That falls straight on Governance, Risk, and Compliance. Ultimately, CISO, CFO, CEO are in the line of fire.
The article's argument happens in a vacuum of facts. The fact that a security engineer doesn't know that is depressing, but not surprising.
That's a very subtle guinea pig joke right there.
You'd lose that lawsuit in a heartbeat.
I don't know what kind of drivers education you get where you live but where I live and have lived one of the basic bits is that you know how to park and lock your vehicle safely and that includes removing the ignition key (assuming your car has one) and setting the parking brake. You aim the wheels at the kerb (if there is one) when you're on an incline. And if you're in a stick shift you set the gear to neutral (in some countries they will teach you to set the gear to 1st or reverse, for various reasons).
We also have road worthiness assessments that ensure that all these systems work as advertised. You could let a pack of dogs loose in my car in any external circumstance and they would not be able to move it, though I'd hate to clean up the interior afterwards.
Who's liable?
I think this would be a freak accident. Nobody would be liable.
> Who's liable?
You are. It's still your dog. If you would replace dog with child the case would be identical (but more plausible). This is really not as interesting as you think it is. The fact that you have a sentient dog is going to be laughed out of court and your neighbor will be in the docket together with you for attempting to mislead the court with your AI generated footage. See, two can play at that.
When you make such ridiculously contrived examples turnaround is fair play.
But you would be liable for civil damages, because that does not. There are multiple theories for which to establish liability, but most likely this would be treated as negligence.
Everything an AI does is downstream of deliberate, albeit imperfect, training.
You know this, you rig it all up and you let things happen.
They correlate, but we must be careful not to mistake one for the other. The latter is a lower bar.
There was a notorious dark web bot case where someone created a bot that autonomously went onto the dark web and purchased numerous illicit items.
https://wwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwww.bitnik.or...
They bought some ecstasy, a hungarian passport, and random other items from Agora.
>The day after they took down the exhibition showcasing the items their bot had bought, the Swiss police “arrested” the robot, seized the computer, and confiscated the items it had purchased. “It seems, the purpose of the confiscation is to impede an endangerment of third parties through the drugs exhibited, by destroying them,” someone from !Mediengruppe Bitnik wrote on their blog.
In April, however, the bot was released along with everything it had purchased, except the ecstasy, and the artists were cleared of any wrongdoing. But the arrest had many wondering just where the line gets drawn between human and computer culpability.
If I build an autonomous robot that swings a hunk of steel on the end of a chain and then program it to travel to where people are likely to congregate and someone gets hit in the face, I would rightfully be held liable for that.
That idea is really weird. Culpa (and dolus) in occidental law is a thing of the mind, what you understood or should have understood.
A database does not have a mind, and it is not a person. If it could have culpa, then you'd be liable for assault, perhaps murder, if you took it apart.
At worst, you would confer liability to the platform, in the case of some sort of blatant misrepresentation of capabilities or features, but absolutely none of the products or models currently available withstand any rational scrutiny into whether they are conscious or not. They at most can undergo a "flash" of subjective experience, decoupled from any coherent sequence or persistent phenomenon.
We need research and legitimate, scientific, rational definitions for agency and consciousness and subjective experience, because there will come a point where such software becomes available, and it not only presents novel legal questions, but incredible moral and ethical questions as well. Accidentally oopsing a torment nexus into existence with residents possessed of superhuman capabilities sounds like a great way to spark off the first global interspecies war. Well, at least since the Great Emu War. If we lost to the emus, we'll have no chance against our digital offspring.
A good lawyer will probably get away with "the AI did it, it wasn't me!" before we get good AI law, though. It's too new and mysterious and opaque to normal people.
We as a society, for our own convenience can choose to believe that LLM does have a mind and can understand results of it's actions. The second part doesn't really follow. Can you even hurt LLM in a way that is equivalent to murdering a person? Evicting it off my computer isn't necessarily a crime.
It would be good news if the answer was yes, because then we just need to find a convertor of camel amounts to dollar amounts and we are all good.
Can LLM perceive time in a way that allows imposing an equivalent of jail time? Is the LLM I'm running on my computer the same personality as the one running on yours and should I also shut down mine when yours acted up? Do we even need the punishment aspect of it and not just rehabilitation, repentance and retraining?
For most crimes, this is circular, because whether a crime occurred depends on whether a person did the requisite act of the crime with the requisite mental state. A crime is not an objective thing independent of an actor that you can determine happened as a result of a tool and then conclude guilt for based on tool use.
And for many crimes, recklessness or negligence as mental states are not sufficient for the crime to have occurred.
A good example might be this. In one case, a driver's brakes fail and he hits and kills a pedestrian crossing the street. It is found that he had not done proper maintenance on his brakes, and the failure was preventable. He's found liable in a civil case, because his negligence led to someone's death, but he's not found guilty of a crime, so he won't go to prison. A different driver was speeding, driving at highway speeds through a residential neighborhood. He turns a corner and can't stop in time to avoid hitting a pedestrian. He is found criminally negligent and goes to prison, because his actions were reckless and beyond what any reasonable person would do.
The first case was ordinary negligence: still bad because it killed someone, but not so obviously stupid that the person should be in prison for it. The second case is criminal negligence, or in some legal systems it might be called "reckless disregard for human life". He didn't intend to kill anyone, but his actions were so blatantly stupid that he should go to prison for causing the pedestrian's death.
I'm not sure it will actually work that way in real courts. At least not consistently.
Yeah that's exactly the I think we should adopt for AI agent tool calls as well: cryptographically signed, task scoped "warrants" that can be traceable even in cases of multi-agent delegation chains
> Agent Trace is an open specification for tracking AI-generated code. It provides a vendor-neutral format for recording AI contributions alongside human authorship in version-controlled codebases.
Now if you did all of that, but made a bug in implementing the filter, then you at least tried and wasn't negligible, but it's on you.
Ah, I get it. So the token can be downscoped to be passed, like the pledge thing, so sub agent doesn't exceed the scope of it's parent. I have a feeling, that it's like cryptography in general -- you get one problem and reduce it to key management problem.
In a more practical sense, if the non-deterministic layer decides what the reduced scope should be, all delegations can become "Allow: *" in the most pathological case, right? Or like play store, where a shady calculator app can have a permission to read your messages. Somebody has to review those and flag excessive grants.
The LLM can request a narrower scope, but attenuation is monotonic and enforced cryptographically. You can't sign a delegation that exceeds what you were granted. TTL too: the warrant can't outlive its parent.
So yes, key management. But the pathological "Allow: *" has to originate from a human who signed it. That's the receipt you're left holding.
You're poking at the right edges though. UX for scope definition and revocation propagation are what we're working through now. We're building this at tenuo.dev if you want to dig in the spec or poke holes.
Sure, But I generally speaking want my agent to send out emails, so I explicitly grant email reading and email writing. I also want to it to pay for invoices but with some semantic condition.
Then I give it the instruction to do something that implicitly requires only email reading. At which point is the scope narrowed to align my explicit permissions granted before with implicit one for this operation? It's not really a problem cryptography is helpful for to solve.
Should it be the other way around maybe -- only read permission is granted first and then it has to request additional permissions for send?
You keep a coarse cap (e.g. email read/write, invoice pay) but each task runs under a narrower, time-boxed warrant derived from that cap. Narrowing happens at the policy/UX layer (human or deterministic rules), not by the LLM. The LLM can request escalation ("need send"), but it only gets it via an explicit approval / rule.
Crypto isn't deciding scope. It's enforcing monotonic attenuation, binding the grant to an agent key, and producing a receipt that the scope was explicitly approved.
For a single-process agent this might be overkill. It matters more when warrants cross trust boundaries: third-party tools, sub-agents in different runtimes, external services. Offline verification means each hop can validate without calling home
If you gave a loaded gun to a five year old, would "five-year-old did it" be a valid excuse?
Purely organically of course.
In weird way, that's actually true. It's a highly- (soon to be fully-) autonomous giga-swarm of the most complicated nanobots in existence, the result of investments over hundreds of thousands of years.
That said, we don't really get to choose which ones we own, although we do have input on their maintenance. :p
It's negligence, pure and simple. The only reason we're having this discussion is that a trillion dollars was spent writing said scripts.
Courts of law have already found that AI interactions with customers are binding, even if said interactions are considered "bloopers" by the vendor[1]
[1] https://www.forbes.com/sites/marisagarcia/2024/02/19/what-ai...
Engineers are accountable for the actions they authorize. Simple as that. The agent can do nothing unless the engineer says it can. If the engineer doesn't feel they have control over what the agent can or cannot do, under no circumstances should it be authorized. To do so would be alarmingly negligent.
This extends to products. If I buy a product from a vendor and that product behaves in an unexpected and harmful manner, I expect that vendor to own it. I don't expect error-free work, yet nevertheless "our AI behaved unexpectedly" is not a deflection, nor is it satisfactory when presented as a root cause.
I think there is no distinction between the two for liability purposes.
The new and interesting part is that while we have incentives and deterrents to keep our human agents doing the right thing, there isn't really an analog to check the non-human agent. We don't have robot prison yet.
The workflow of starting dozens or hundreds of "agents" that work autonomously is starting to gain traction. The goal of people who work like this is to completely automate software development. At some point they want to be able to give the tool an arbitrary task, presumably one that benefits them in some way, and have it build, deploy, and use software to complete it. When millions of people are doing this, and the layers of indirection grow in complexity, how do you trace the result back to a human? Can we say that a human was really responsible for it?
Maybe this seems simple today, but the challenges this technology forces on society are numerous, and we're far from ready for it.
When orchestrators spawn sub-agents spawn tools, there's no artifact showing how authority flowed through the chain.
Warrants are a primitive for this: signed authorization that attenuates at each hop. Each delegation is signed, scope can only narrow, and the full chain is verifiable at the end. Doesn't matter how many layers deep.
People want to use a tool and not be liable for the result.
People not wanting to be liable for their actions is not new. AI hasn't changed anything here, it's just a new lame excuse.
> It shouldn't matter, because whoever is producing the work product is responsible for it, no matter whether genAI was involved or not.
| *Who authorized this class of action, for which agent identity, under what constraints, for how long; and how did that authority flow?*
| A common failure mode in agent incidents is not “we don’t know what happened,” but:
| > We can’t produce a crisp artifact showing that a specific human explicitly authorized the scope that made this action possible.
The point is "explicitly authorized", as the article emphasizes. It's easy to find who ran the agent(article assumes they have OAuth log). This article is about 'Everyone knows who did it, but did they do it on purpose? Our system can figure it out'
If the user claims that they only authorized the bot to review files, but they've warranted the bot to both scan every file and also send emails to outside sources, the competitors in this case, then you now have proof that the user was planning on committing corporate espionage.
To use a more sane version of an example below, if your dog runs outside the house and mauls a child, you are obviously guilty of negligence, but if there's proof of you unleashing the dog and ordering the attack, you're guilty of murder.
With no amount of detailed logging makes "the AI did it" a valid excuse.
It's just a tool.
It's like blaming a loose bolt in a Boeing 737 on "screwdriver did it".
On the contrary, if they just owned up to it, chances are I wouldn't even write them up once.
The "Hallucination Defense" isn't a defense because end of the day, if you ran it, you're responsible, IMO.
And if a professional wants to delegate their job to non-deterministic software they're not professionals at all. This overreliance on LLMs is going to have long-term consequences for society.
1. I vaguely recall that in the early days of Windows, the TOS explicitly told users they were not to use it for certain use-cases and the wording was something like "we don't warranty windows to be good for anything, but we EXPLICITLY do not want you to use Windows to do nuclear foo".
I expect that if the big LLM vendors aren't already doing this, they soon will. Kind of like how Fox News claims that some of heads on the screen are not journalists at all but merely entertainers (and you wouldn't take anything they say seriously ergo we're not responsible for stuff that happens as a result of people listening to our heads on the screen).
2. IANAL but I believe that in most legal systems, responsibility requires agency. Leaving aside that we call these things "agents" (which is another thing I suspect will change in the marketing), they do not have agency. As a result they must be considered tools. Tools can be used according to the instructions/TOS, or not. If not - whatever it does is down to you. If used within guidelines - it's the manufacturer.
So my conclusion is that the vendors - who have made EPIC bets on getting this tech into the hands of as many folks as possible and making it useful for pretty much anything you can think of - will be faced with a dilemma. At the moment, it seems like they believe that (just like the infamous Ford bean counters) the benefits of not restricting the TOS will far outweigh any consequences from bad things happening. Remains to be seen.
Read the terms and conditions of your model provider. The document you signed, regardless if you read or considered it, explicitly removes any negative consequences being passed to the AI provider.
Unless you have something equally as explicit, e.g. "we do not guarantee any particular outcome from the use of our service" (probably needs to be significantly more explicitly than that, IANAL) all responsibility ends up with the entity who itself, or it's agents, foists unreliable AI decisions on downstream users.
Remember, you SIGNED THE AGGREMENT with the AI company the explicitly says it's outputs are unreliable!!
And if you DO have some watertight T&C that absolves you of any responsibility of your AI-backed-service, then I hope either a) your users explicitly realize what they are signing up for, or b) once a user is significantly burned by your service, and you try to hide behind this excuse, you lose all your business
One in which you sell yourself into slavery, for example, would be illegal in the US.
All those "we take no responsibility for the [valet parking|rocks falling off our truck|exploding bottles]" disclaimers are largely attempts to dissuade people from trying.
As an example, NY bans liability waivers at paid pools, gyms, etc. The gym will still have you sign one! But they have no enforcement teeth beyond people assuming they're valid. https://codes.findlaw.com/ny/general-obligations-law/gob-sec...
"But the AI wrote the bug."
Who cares? It could be you, your relative, your boss, your underling, your counterpart in India, ... Your company provided some reasonable guarantee of service (whether explitly enumerated in a contact or not) and you cannot just blindly pass the buck.
Sure, after you've settled your claim with the user, maybe TRY to go after the upstream provider, but good luck.
(Extreme example) -- If your company produces a pacemaker dependent on AWS/GCP/... and everyone dies as soon as cloudflare has a routing outage that cascades to your provider, oh boy YOU are fucked, not cloudflare or your hosting provider.
Sure, if someone from GCP shows up at your business and breaks your leg or burns down your building, you can go after them, as it's outside the reasonable expectation of the business agreement you signed.
But you better believe they will never be legally responsible for damages caused by outages of their service beyond what is reasonable, and you better believe "reasonable outage" in this case is explicitly enumerated in the contact you or your company explicitly agreed to.
Sure they might give you free credits for the outage, but that's just to stop you from switching to a competitor, not any explicit acknowledgement they are on the hook for your lost business opportunity.
Sure, but not all liability can be reassigned; I linked a concrete example of this.
> But you better believe they will never be legally responsible for damages caused by outages of their service beyond what is reasonable, and you better believe "reasonable outage" in this case is explicitly enumerated in the contact you or your company explicitly agreed to.
Yes, on this we agree. It'd have to be something egregious enough to amount to intentional negligence.
Or perhaps scapegoating at scale.
https://news.ycombinator.com/item?id=43877301 - 398 comments
https://news.ycombinator.com/item?id=41891694 - 308 comments
It is the burden of a defendant to establish their defense. A defendant can't just say "I didn't do it". They need to show they did not do it. In this (stupid) hypothetical, the defendant would need to show the AI acted on its own, without prompting from anyone, in particular, themselves.
Licensed professionals are required to review their work product. It doesn't matter if the tools they use mess up--the human is required to fix any mistakes made by their tools. In the example given by the blog, the financial analyst is either required to professional review their work product or is low enough that someone else is required to review their work product. If they don't, they can be held strictly liable for any financial losses.
However, this blog post isn't about AI Hallucinations. It's about the AI doing something else separate from the output.
And that's not a defense either. The law already assigns liability in situations like this: the user will be held liable (or more correctly: their employer, for whom the user is acting as an agent). If they want to go after the AI tooling (i.e., an indemnification action) vendor the courts will happily let them do so after any plaintiffs are made whole (or as part of an impleader action).
Use AI if being 80% right quickly is fine. Otherwise if you have to do the analysis anyway because accuracy is critical, there's little point to the AI - its too unreliable.
1. A person orders an AI agent to do A.
2. The agent issues a tenuo warrant for doing A.
3. The agent can now only use the tool to perform A.
The article is about that 'warrant' can now be used in case of an incident because it contains information such as 'who ordered the task' and 'what authority was given'.
I get the idea. This isn't about whether a person is responsible or not(because of course they are). It's more about whether it was intentional.
However... wouldn't it be much easier to just save the prompt log? This article is based entirely on "But the prompt history? Deleted."(from the article) situation.
But warrants aren't just "more audit data." They're an authorization primitive enforced in the critical path: scope and constraints are checked mechanically before the action executes. The receipt is a byproduct.
Prompt logs tell you what the model claimed it was doing. A warrant is what the human actually authorized, bound to an agent key, verifiable without trusting the agent runtime.
This matters more in multi-agent systems. When Agent A delegates to Agent B, which calls a tool, you want to be able to link that action back to the human who started it. Warrants chain cryptographically. Each hop signs and attenuates. The authorization provenance is in the artifact itself.
At execution time, the "verifier" checks the warrant: valid signatures, attenuation (scope only narrows through delegation), TTL (authority is task-scoped), and that the action fits the constraints. Only then does the call proceed.
This is sometimes called the P/Q model: the non-deterministic layer proposes, the deterministic layer decides. The agent can ask for anything. It only gets what's explicitly granted.
If the agent asks for the wrong thing, it fails closed. If an overly broad scope is approved, the receipt makes that approval explicit and reviewable.
Eg a list of transactions that isn't AI generated where the only actions that actually move money must operate on the data displayed in the human designed page.
A human looks at this and says yes that is acceptable and becomes reasonable for that action.
So while that should happen, it won't. They'll just add an extra layer of AI to do the verification.
This way, they can avoid being legally blamed for stuff-ups and instead scapegoat some hapless employee :-) using cryptographic evidence the employee "authorized" whatever action was taken
> A computer must never make a management decision, because a computer cannot be held accountable.
AI is just branding. At the end of the day it's still just people using computer software to do stuff.
There is going to be a person who did a thing at the end of the day -- either whoever wrote the software or whoever used the tool.
The fact that software inexplicably got unreliable when we started stamping "AI" on the box shouldn't really change anything.
but the person who turned it on can
simple as