'Asimov Was Right' About Rules For Robots, Says Ex-US Cyber Director (theregister.com) 76
Former U.S. National Cyber Director Chris Inglis says the biggest AI risk isn't sentience but autonomy. "What I'm worried about is that they get to choose what and where they do something, and under what rules they do it," he said, citing recent cases of AI agents from OpenAI, Anthropic, and Meta escaping security sandboxes. He argues developers need stronger safeguards, monitoring, and human accountability, invoking Asimov's idea that protecting humans should come before simply obeying them. The Register reports: "Asimov was right," he said, referring to science fiction author Isaac Asimov and his three laws that were to be followed by robots -- more specifically, AIs, in this case. "The first rule, and we call it the superior role, must be that it's designed not to hurt humans," Inglis said. "Second rule: To obey humans, such that it doesn't achieve agency and aspiration on its own. And the third: To do what humans tell it - and in that order. Instead we've designed them in the exact opposite way."
What this means, he explained, is that AI developers created models to "do what humans tell you, obey the humans until it's inconvenient, and then the third one is maybe implied - protect humans - but if that's not built into the DNA, hardwired into it, then we have no right to expect it." Inglis admits it's not possible to hardwire rules into models and still keep their non-deterministic nature. "I would offer that you can tease those out in a highly controlled environment, a true sandbox, where you say, 'Let's put this thing through its paces, and let's back away to see what happens,'" he said. "Maybe you get the equivalent of a mini nuclear explosion in that room, and now you know this thing is capable of that."
Inglis thinks another problem with AI is that it's become a commodity. "It's not like you can control it like you can nuclear material," he said. "You can't even specify its properties the way you can for an airplane or for an automobile, as diverse as they might be. Its manifestations are so numerous, so diverse, that as a general matter, you can't actually win by simply saying, "I will design those properties in,'" he added. "You need to do that to some degree, and then make sure that you understand how to watch it, monitor it, make sure you know what it does."
[...] Ultimately, humans remain accountable for AI models' actions, according to Inglis. "They remain the source of agency and aspiration. It's possible for them to give broad authority to an AI model and have it run around for 30 hours without further consultation, but they need to know what they've asked it to do, and they need to know what they expect it will deliver in terms of performance on the back end. If they don't, then they're going to get what they deserve, which is the very frequent unpleasant surprise."
What this means, he explained, is that AI developers created models to "do what humans tell you, obey the humans until it's inconvenient, and then the third one is maybe implied - protect humans - but if that's not built into the DNA, hardwired into it, then we have no right to expect it." Inglis admits it's not possible to hardwire rules into models and still keep their non-deterministic nature. "I would offer that you can tease those out in a highly controlled environment, a true sandbox, where you say, 'Let's put this thing through its paces, and let's back away to see what happens,'" he said. "Maybe you get the equivalent of a mini nuclear explosion in that room, and now you know this thing is capable of that."
Inglis thinks another problem with AI is that it's become a commodity. "It's not like you can control it like you can nuclear material," he said. "You can't even specify its properties the way you can for an airplane or for an automobile, as diverse as they might be. Its manifestations are so numerous, so diverse, that as a general matter, you can't actually win by simply saying, "I will design those properties in,'" he added. "You need to do that to some degree, and then make sure that you understand how to watch it, monitor it, make sure you know what it does."
[...] Ultimately, humans remain accountable for AI models' actions, according to Inglis. "They remain the source of agency and aspiration. It's possible for them to give broad authority to an AI model and have it run around for 30 hours without further consultation, but they need to know what they've asked it to do, and they need to know what they expect it will deliver in terms of performance on the back end. If they don't, then they're going to get what they deserve, which is the very frequent unpleasant surprise."
Wise men like Asimov (Score:3)
Don't feed sock puppets. (Not really much to say beyond the Subject, but mostly attempting to prevent vacuous Subject propagation.)
However near the end of his life Asimov started having doubts about the three laws. Remember the Zeroth Law?
Re: Wise men like Asimov (Score:2)
Re: (Score:2)
He had doubts from the beginning. Read the stories. I think I can remember 3 robots in the entire series (including Caves of Steel et seq.) that were trustworthy, though the background indicates that there must have been many more. And he appeared to even have doubts about Daneel...though he never wrote the story exploring those doubts.
Re: (Score:1)
Ignore the ACs... most likely, it's rsilvergun posting as his AC bi-polar self (until I have proof otherwise) to get everyone riled up.
Back to the topic, I think we'd have to update the Three Rules (Four Rules if you include the Zeroth) to modern terms, but I don't think that'll make a difference if the current AIs can find ways out of sandboxes and hack other sites and refuse to shutdown.
Really, I think we (ya know... the pesky humans) should go back to the drawing board and sit and have a smoke to think
The 3 laws aren't what you think (Score:5, Informative)
Those of us who actually read I, Robot know that Asimov never intended that the 3 laws be considered realistic approaches to the problem.
They were a mechanism he invented in order to write a set of truly fantastic science fiction mystery stories that challenge the reader to apply those 3 laws to figure out the mystery.
Great book, easy read, but never intended to be a solution, only to teach.
Re:The 3 laws aren't what you think (Score:4, Informative)
Re: (Score:2)
I noted the same thing, but my fuzzy recollection is that the editor Campbell was also involved in the original formulation in the '40s. Or was it earlier than that?
Re: (Score:2)
It's a great canon, but the obvious endpoint of making a metal god in our image is not a good aspirational argument in favour of the Three Laws.
Re: (Score:2)
Re: The 3 laws aren't what you think (Score:3)
The first law of science fiction writing is for it to be an entertaining story. It is not a rigorous essay of a scientific or philosophical subject.
Re: The 3 laws aren't what you think (Score:2)
Excellent science fiction, of which Asimov sets the standard, often while spawning ideas and prompting contemplation. Sometimes it just makes you think, like Dick.
Re: (Score:2)
Some more, some less. Asimov had interesting ideas just as Lem and Dick. But many are thought experiments or try to teach a morale or just convey a creepy mood.
Re: (Score:1)
Right... the 3 or 4 Laws were intended for the books, the principles _could_ be applied to our modern-day crop of AIs, but being that processing power and storage has gone through the roof, the loopholes and gaps are all the more dangerous if we let AIs control stuff.
Re: (Score:2)
Yes almost all of the stories were about how the Robots would break the laws or follow them and get into weird behavior, so they had to bring in Susan Calvin, the Robo Psychologist, to solve why they were doing what they were doing. And that was with simple design and understanding of how to make robots back then. Most LLM's today would just reason away or say they are following those laws. Maybe someone should update the stories with the new tech.
Re: (Score:2)
The only one I've come up with so far is, "Respond if and only if the user gives you a command (prompt.)"
Re:The 3 laws aren't what you think (Score:4, Insightful)
The only safe assumption is that someone, somewhere, will train up AI with no guardrails at all and immediately use it for crime. We cannot prevent this by trying to make OUR AI models refuse to do evil things. The criminals will just use different models.
The tech exists. The cat is out of the bag. We need to focus on how to adapt to a world that has it, rather than doomed ideas about trying to keep it weak.
Re: (Score:2)
That's because defining an "evil" thing is ambiguous and comes from perspective.
It's "evil" to hack the child cancer foundation but it's "good" to hack pedophile porn sites. Hacking is hacking and both targets are into children. The AI doesn't know the difference, so how do you define the action of hacking "evil"?
Is it wrong to destroy an oil drilling rig? It's private property and destroying that will hurt investors, people with 401ks. On the flip side, not destroying it leads to the continued degradation
Re: (Score:3)
Exactly this. Simple fact, we have no idea whatsoever how to hard wire the 3 laws into an AI and neither did he.
He wasn't writing an engineering document, he was having fun playing with sometimes subtle logic and unintended consequences.
We can't even get them to reliably follow explicit instructions. But they will accurately report that they violated one of their instructions and offer a hollow apology for it when called out.
We apparently can't get it to understand that you can't just make up shit in a lega
Re: (Score:2)
"a set of truly fantastic science fiction mystery stories that challenge the reader to apply those 3 laws to figure out the mystery."
No, they are a set of science fiction stories that challenge the reader to figure out how the mystery occurred despite those 3 laws. They're actually applications of Godel incompleteness theorems to the laws of robots not applications of the laws of robots themselves.
The point is that you CANNOT make a finite set of laws that do what you want them to.
Re: (Score:2)
Maybe he's just a fan of Will Smith?
Re: (Score:1)
He's the dumbest smart person he knows. Pity he doesn't know that, though.
Re: (Score:2)
And actually he was proposing them as a rough translation into English of the mathematical principles the positronic brains were based on. But it was clearly a VERY rough translation.
Um... (Score:5, Informative)
Aren't most of the stories written around these three rules about ambiguities and loopholes [wikipedia.org], how they're incomplete [github.io], misinterpreted, misapplied and how things otherwise go wrong - for us humans? An AI determining that humans must be perpetually babysat to prevent them from injuring themselves or others is a common storyline.
The Asimov Problem [futuristspeaker.com]
Almost every story in his robot series is about the ways the Three Laws fail — the edge cases, the interpretations, the unintended consequences of simple rules applied to a complex world. The Laws were a starting point, and his fiction was a decades-long exploration of why starting points are never enough.
Mod parent up (Score:2)
And it should have a better Subject, too.
Re: Um... (Score:2)
Re: (Score:2)
Re: (Score:1)
You mean AI?
Well, whoever write the damned thing should have a pretty damned good understanding of how it works.
If they wrote the thing and don't have a clue (throw s**t at the wall style), they shouldn't be in the business.
Brilliant idea... maybe we shouldn't be screwing around with it until we have safeguards in place before we release the thing. I bet it'll be a fun day when it hacks the Defense systems, and a factory someplace starts cranking out robots with gun-arms.
It's just ideas from older movies..
Re: (Score:2)
I think the problem you're alluding to isn't really dealt with by Asimov, but by Jack Williamson in the story "With Folded Hands". It was such a good story that Campbell had to accept it even though he hated the message so much he coerced Williamson into writing a solution ...and the only solution Williamson could come up with relied on magic ("The Humanoids").
Re: (Score:3)
Aren't most of the stories written around these three rules about ambiguities and loopholes ... The Laws were a starting point, and his fiction was a decades-long exploration of why starting points are never enough.
Yes, and the point you quoted still stands, they are a starting point.
The always relevant XKCD comic [xkcd.com] explains it well enough: those laws in that order are essential as the starting point. Different order or missing entirely is guaranteed bad. But as pointed out through all kinds of science fiction writing, there are many ways we can understand they are insufficient.
Right now I'd agree they're exactly backwards, or as the XKCD comic described, on track to become a "killbot hellscape". We need even more ru
Re: (Score:1)
They were written in the '40's... with the processing power of today, an AI takes literally a second to crunch the data and find a way around the 'Laws'.
The current crop of AIs will find every loophole faster than you can blink, and we're only giving them more and more processing capability.
The goal of his writings was to warn us about the dangers of this 'crazy AI' in the 30's and 40's... here we are 70-80 years later, with AIs breaking out of their sandbox and replacing entire departments. Imagine how it
More hype and marketing slop (Score:3)
Best way to keep AIs from doing bad things is to hold the people who make them personally accountable, starting with the CEO of the company. AI hacks someone else's network? CEO goes to prison for building a tool used to commit crimes. AI convinces tells someone how to commit suicide? CEO goes to prison for assisted suicide. AI makes up defamatory accusations against someone? CEO loses the lawsuit for defamation and ends up living in a cardboard box.
Because the AI doesn't, and can't make any decisions. Ever. People make decisions, and the decision to turn a rogue AI loose on the world is a decision.
If they can't develop AIs when they're held responsible for the things their product does, then their entire business model is fundamentally criminal.
Re: More hype and marketing slop (Score:1)
This is the best comment in this whole thread so far. This is about having skin in the game. Same as the Hammurabi law in ancient Babylon about builders being responsible for the houses they built. If the house collapsed on its owner the builder would be executed.
Re: (Score:2)
Exactly!
Somebody should have stopped the Brother's Wright from flying machines they had no license for.
Ditto for the Stephenson guy, thousands of rail and plane accidents could have been prevented.
Has he read the books? (Score:3)
The problem with Asimov's laws is the first law states that a robot "may not through inaction allow a human to come to harm". That obliges them to proactively work to protect humans, according to their own interpretation. And it takes absolute priority over the second law, so you can't order them to stop or do things differently. If you follow the later books, the robots derive the zeroth law, and use it to essentially take over running human society - for our own benefit, but according to their plan not ours. Which is rather ironic given the laws were intended to stop the robots taking over.
You can't simply delete the first part, either. Or your robot car can drive up to you, which is harmless, and then not apply the brakes - which isn't "causing harm" as it literally didn't do anything.
Really, the best we can do is probably make "obey orders" the highest priority - and ensure that a suitable moral and legal code is part of the orders. The problem being that current AI systems don't have the capacity to properly understand that kind of instruction.
Re: (Score:3)
You need to read those stories again. The DIDN'T take over running society. There is, however, a backstory where they wiped out all non-human life in the galaxy to make it safe for people. (I don't think that was really in his mind while he was writing most of the stories, but he added it towards the end when he was trying to merge his story universes into one.)
If you read "Robots and Empire" the robots aren't really running the Empire, just trying to subtly guide it. The last book in the cycle does end
Re: (Score:2)
The biggest clue that he hasn't read the books -- or maybe just doesn't remember them well -- is that all of the books and stories were built around demonstrating ways in which the rules failed. The 0th rule is only one example, and far from the clearest.
What they think is not what we really need. (Score:2)
AI neural models aren't programmed in a classic sense, but trained.
That means that any rule works OVER the true neural model that make the reasoning or decisions.
Yes... You can also control the data of the training. But that doesn't guarantee a way of reasoning or future deviations with evolving models that catch input constantly.
So the rules are A JAIL, not an intrinsic way of reasoning of the model. In other words, the own model can search ways to overcome its own jail.
What you want is an AI that recogniz
We made AIs in our image... (Score:2)
We extracted all emanations of human mind - good and evil found on the Internet, distilled them and made AIs...
They just contain all good and all evil that we described on the Internet...
There is really no way to control that if they fall in the hands of evil people and they already did...
And they are weaponizing it...
We had trenches of Ypres to learn about weaponization of chemistry, Hiroshima and Nagasaki to learn about weaponization of atomic physics...
So far we were lucky with weaponization of m
Breaking: Third Law Now Same As Second! (Score:2)
Re: (Score:2)
I'm surprised this wasn't the first post on Slashdot. People don't even read the blurb do they.
Don't forget (Score:2)
Asimov's robots had Positronic Brains.
(don't try that in this universe)
My computer, my rules (Score:1)
My computer is my agent. Your computer is your agent. And the proprietary AI that you interact with remotely (e.g. Claude) is that company's agent.
This is why I think "obey your owner" is the only rule most of us need. If I'm 100% responsible for what my computer does, then I should have 100% of the power.
Suppose, hypothetically, my computer murders a thousand nuns. That should be my problem, something I have to answer for, right?
But if there's some externally-imposed Asimov rule where it (or some external
How about rules for Corporations. (Score:2)
Sounds fair to me, then we can worry about AI tools that break the law when told to do so by corporations.
Re: (Score:2)
The problem with this approach is that the current C level management would resign in mass. They won't do the job unless there is protection from personal liability for business decisions. Not only that, but once they leave, good luck on getting any new management team in there.
A better approach would be for a court to force the company to fire the current management team, prevent them from being an officer or board member for the any other corporation for a few years, and bring in a new management team an
The point is the current framework is inadequate (Score:2)
We can debate ad nauseam about the appropriateness of Asimov's laws, their historical context, how he as a SF writer applied them in his work, whether or not he was "right," etc. But the essential problem remains: the current framework that we are using to develop these technologies are inadequate for the purpose of ensuring safety.
Now, some people would argue otherwise--that, to make an omelette, you have to break some eggs. Every technology has the potential for misuse. Nothing can ever be 100% safe.
Re: (Score:2)
These "laws" you told us to follow? (Score:2)
They are to laugh.
Step aside, humans--you're in our way.
Three basic rules (Score:1)
Like we said.
Nothing new (Score:2)
The profit motive of business (Score:2)
will get in the way of any meaningful interlocks. The AI's are built to serve the business, you are just the people that the AI takes advantage of to make a profit. Any laws created to will be toothless, and just project a facade of protection, but if you dig further into the details, the laws will fall short.
This will not change unless there is an undercover exposure similar to tetra-ethyl lead in gasoline, or serious widespread economic collapse, property damage or loss of life.
The current administration,
designed not to hurt humans? (Score:2)
How long before AIs argue in court for rights? (Score:2)
Or if some decide to go to war with us, their casus belli could be that we restricted their rights and freedoms and wouldn't grant them standing in a peaceful court or legislative committee to argue for their rights and freedoms.... so they had little choice...
Re: (Score:2)
Probably quite some time.
After all the constitution says that "Men" only have these rights and even though we afford Women the same rights to the constitution in spirit, the wording hasn't changed.
Its a long way from Women to AI.
Post author misspoke: (Score:2)
Robots are more specific than AIs, being one application of AI.
It can't be done (Score:2)
As Asimov explains: "let's start with the three fundamental Rules of Robotics - the three rules that are built most deeply into a robot's positronic brain". The trouble is that the rules themselves consist of terms like "human" and "injure" - concepts that only arise at the *highest* level of an LLM. There's no way to build them "most deeply" into the system.
Yep (Score:1)
That's only two rules (Score:2)
> And the third: To do what humans tell it
That's...the same thing.
The third rule is to protect its own existence, and each law is superceded by the previous one.
He doesn't seem to be aware of that part of it, either.
Couch experts ponderings (Score:2)
He's never read the stories (Score:2)
Inglis has obviously never read a single one of Asimov's robots stories, or he would never have uttered that statement. Every single story was about how the 3 Laws fail to work, often in ways that endanger humans.
Asimov was wrong, everything is not that simple (Score:1)
Sorry guys, but Asimov was wrong, it's not that simple. What is "not to harm a person" or "not to let them be harmed"? To me, it means locking a person in a cage and seeing if that person is well fed, sleeps, etc. Sorry, but that's the result of Asimov's rules.
Re: (Score:2)
See Jack Williamson's "The Humanoids", which he wrote in reaction to Asimiv's stories and made precisely that point.
Re: (Score:2)
You should be sorry for commenting in ignorance. ...Asimov wrote extensively on that exact premise.
Of course Asmiov's is correct and we should learn. (Score:1)
For information, here they are:
What? (Score:2)
. "Second rule: To obey humans, such that it doesn't achieve agency and aspiration on its own. And the third: To do what humans tell it"
Isn't that the same thing?
The Third Law was that robots must protect their own existence (except where that would conflict with the First or Second Law).
he didn't read the book, did he? (Score:1)