Why Asimov's Laws of Robotics Don't Work - Computerphile 

Asimov: you can't control robots with three simple laws everyone : yes,we will use three simple laws, got it.

"if Goingtoturnevil don't"

I didn't realize that people took Asimov's Three Laws seriously, considering that nearly every work they're featured in involves them going wrong.

He mostly focused on the difficulty of defining "Human", but I think it's much much more difficult to define "Harm", the other word he mentioned. Some of the edge cases of what can be considered human could be tricky, but what constitutes harm? If I smoke too much, is that harmful? Will an AI be obligated to restrain me from smoking? Or, driving? By driving, I increase the probability that I or others will die by my action. Is that harm? What about poor workplace conditions? What about insults, does psychological harm count as harm? I think the difficulties of defining "Harm" are even more illustrative of the problem that he's getting at.

"You are an AI developer. You did not sign up for this". Brilliant quote!!!

Yes, that was Asimov's intention all along. The whole point of the laws of robotics in the books is that they are incomplete and cause logical and ethical contradictions. All the stories revolve around this. This is worth emphasizing, as most people seem to think Asimov proposed them as serious safeguards. The comments in the beginning of the video illustrate this misconception well. Thanks for bringing this up, Rob!

"Optimized for story writing." I can't express how much I love that sentiment.

the laws exist to create a paradox around which to construct a narrative.

I kinda wish this video just kept going.

The problem with Asimov's laws is probably that they're just obscure enough that people don't think they're well known, but they're also not well known enough for people to remember the context they appeared in and how they always failed.

So the problem of ensuring that technology only acts in humanity's best interests isn't between human and technology, but between human and self. We cannot properly articulate what kind of world we actually want to live in in a way that everyone agrees with. So no one can write a computer program that gets us there automatically.

"[The laws are] optimized for story writing" spoken like a true programmer

This brings to mind the Bertrand Russell quote in Nick Bostroms's book. _"Everything is vague to a degree you do not realize till you have tried to make it precise."_

In fact, Asimov's whole point in writing I, Robot was to show the problem with these laws (and therefore the futility in creating one-size-fits-all rules to apply in all cases).

This was sort of Asimov's point in the first place if you actually go back and read his original stories instead of the modern remakes that mistakenly think the rules were meant to be "perfect." He always designed them as flawed in the first place, and the stories were commentary on how you *can't* have a "perfect law of robotics" or anything, as well as pondering the nature of existence/what it means to be sentient/why should that new "life" have any less value than biological life/etc.

The book "I Robot" was full of stories about how the "laws" don't work and yet dummies keep parroting them like they are a blueprint for AI

"I didn't sign up for this" - made my day

Does psychological harm count as harm? If so, by destroying someone's house, or just slightly altering it, you would harm them.

A story about robot necromancy sounds kind of cool though. 🤔🤖☠️

Gets the definition of "death" slightly wrong - "I've made necromancer robots by mistake"

Robot can't use hypodermic needle to save a life because that is "harm" - but that "harm" is needed to not have human die...

Even defining "death" is a moral issue in medical sciences, right now.

Asimov didn't intend for them to work, you said it yourself--the 3 laws of robotics go wrong

wow, at the end there he literally described the plot of SOMA.

Harm is even worse. Violating ownership is harm so even if the table isn't human, destroying it arbitrarily harms people. Destroying things in general harms people. There is very little that can be done that doesn't harm someone somewhere. Using a Styrofoam cup harms people, creating and using plastic harms people. The very construction of the robots would harm people. Want a self driving car? Oops, cars use energy and all current - and probably future - forms of energy production harm people.

"The word human points to that structure in your brain." "The central examples of the classes are obvious." I found the C programmer.

Asimov was not a fool, and these are clearly ethical rules, and as such are in the field of moral philosophy. It's blindingly clear that they aren't design rules and they rather do point to the problem of the inherent ambiguity of morality and ethical standards which always have subjective elements. However, human beings have to deal with these issues all the time. Ethical standards are embedded into all sorts of social systems in human society, either implicitly or even explicitly in the form of secular, professional and religious laws and rules. So the conundrum for any artificial autonomous being would be real. To me this points out the chasm there is between the technological state of what we call Artificial Intelligence, that is based on algorithms or statistical analysis and what we might call human intelligence (not that the psychologists have done much better). Asimov got round this by dumping the entire complexity into his "positronic brains" and thereby bypassed this. In any event, there are real places coming up where human morality/ethical systems are getting bound up with AI systems. Consider the thought experiments that are currently doing the rounds over self-driving cars and whether they will be programmed to protect their occupants first over, say, a single pedestrian. As we can't even come to an agreed human point on such things (should we save the larger number of people in the vehicle or the innocent pedestrian that had no choice about the group choosing to travel in a potential death-dealing machine), then even this isn't a solvable in algorithmic terms. It sits in a different domain, and not one AI is even close to being able to resolve. The language adopted in the video is all that of computer science and mathematics. The definition of hard boundaries for what is a "human" is a case in point. That's not how human intelligence appears to work, and I recall struggling many years back with expert systems and attempting to encode into rigorous logic the rather more human-centred conceptualisation used by human experts. Mostly, when turned into logic, it only dealt with closed problems of an almost trivial nature.

An AI strictly bound to Asimov's laws wouldn't even attempt to CPR in first place no matter how you define human and death, because that process usually injures the patient in some way.

the 1941 short story "Liar!" in Asimov's I, Robot was very interesting. it's about a robot that can read minds, and begins to lie to everyone it meets, telling them what they want to hear. it ends up causing harm to everyone it lies to. i think this story illustrates how difficult the concept of time is when thinking about the laws of robotics. harm on what timescale? how can you normalize harm with respect to time? is anything completely unharmful? in that case, how do you minimize harm, other than total oblivion and/or bliss? where does happiness come into all of this? for this reason i love Asimov's stories: at face value, the robotic laws are central, but ultimately the stories are all about humans and how we cant even really begin to think about ourselves in any serious way... i recommend reading all of Asimov's short stories seriously. they are not as shallow as they are made out to be!

So the Matrix robots are caring for us. Got it.

"The point is, you're trying to develop an A.I here. You're an A.I developer, you didn't sign up for this!" xD

Good video, but I would like to suggest that The fact that Asimov’s laws of robotics are problematic does not mean that they (or similar rules would be useless). Maybe even flawed rule that doesn’t work well in marginal cases is better than no rule at all. In fact, our entire system of justice takes this approach.

Asimov's robot stories had a _"positronic brain"_ housing the intelligence. From what I remember they weren't algorithm based like today's computers. For what's that's worth.

AWP | Asiimov

4:00 Coming up with a definition of anything is extremely difficult for robots? How about humans? Haven't you visited the comments section of a philosophy video or a politics forum?

AI developer: I never asked for this

Just one little thing... If you can be brought back with cpr, you aren't really properly dead.

Seriously, the whole *point* of the Three Laws, from my own interpretation, is that they're not *supposed* to work. They're flawed by design; that's what creates the drama in the books to begin with.

The other issue is that we as a species don't follow these rules of not letting ourselves get harmed. So AI would be confused as to protect us or do what we do: such as have wars

Programmer: don't harm humans Robot: define human, define harm Programmer: well first off dead humans don't count Robot: define dead Programmer: humans with no brain activity Robot: does that definition also include humans is a vegatative state? Programmer: scratch that, death is when the heart stops Robot: defibrillators can restart the heart Programmer: when the body starts tp decay Robot: should I keep trying to revive them for over 8 to 12 years? Programmer: *deletes code*

You forgot the -1 law of robotics: A robot may not take any action that could result in the parent company being sued.

4:43 The point in which I realized that the lab coat hanging in the background in not in fact a human wearing a labcoat but just a lab coat.

I feel like Asimov's stories featured very human-like robots that followed human-like logic. I don't think he thought of them in the same terms we do nowadays, in terms of computer logic and programming.

To start asimov never proposed th three laws as some sort of answer to anything other than a corporate answer to human fears that never really worked. Secondly the wording for the three laws in his universe are not accurate representations of how the three laws work but instead translations of the intended archetecture, the word human is as meaningless to a computer as the word no.

I want to hear more on this subject!

I love the way Rob argues. It's clear, to the point with few perfectly selected words.

"You're an AI developer, you didn't sign up for this"

what happens if one of these robots comes across someone suffering in a hospital and will live the rest of their life on machine support? would it count as harming them to unplug them? would it count as harm by innactivity to let them suffer?

what programms do you use for making your animations?

Rob Miles is excellent at pointing out and explaining the terrifying grey areas of AI and morality.

On the point about simulated brains...Similarly I think it's also worth speculating whether an AI would count a cyborg as human. For example if we get to the point where we can transplant brains into artificial bodies....Would it count that? Anyway this is a super great video and it makes me want to go write some sci fi where everything goes VERY VERY wrong. XD

Basically the whole point of the 3 laws was the idea that something seemingly simple was extraordinarily difficult.

I totally go with the guy's stance for current AI, however, within the context of the stories the impression is that the artificial positronic brain results in something more akin to natural intelligence. And therefore the Robot understanding of "human" and "harm" would work much like our own understanding, including being fuzzy round the edges. And indeed I suspect humans have embedded "laws" such as prioritizing our children's lives over our own. But the mistake is to think that the laws would somehow be implanted in the form of English sentences.

Even Asimov admitted that they were just a narrative tool.

"in other news, the mistery surrounding a string of alleged grave robberies​ has been solved, as cops finally catch a robot that was exhuming people just to perform CPR..."

To be fair, a specific robot is built for a specific purpose. You're trying to create a universal definition, but Asimov's robots are allowed to have specialized priorities. The definition of a human given to a robot should be whatever helps that robot do its job most effectively, it doesn't have to bring world peace on its own.

So basicly game "Soma" i about how WAU have pretty wicked sense of "keep humanity alive"

Asimov: here's three laws that don't work, here are many books that, through narrative, explain why and how they wouldn't work. The general public: the three laws solve all our problems.

I am always amazed by how good Rob Miles can word these abstact issues. Some Examples he gives are not only a good example in itself, but also quite an enlightment on how to see things. I don't understand the dislikes at all. For me this was a perfect short presentation why implementing these ethical and unclear terms in not only problematic but impossible.

Also, the "Zeroeth" law was defined by a "robot', so, the 'hard-coded" rules could be overcome by the AI learning algorythims ... very interesting

Sometimes I think that Robocop's prime directives make more sense than Asimov's rules.

Interesting video - However, most of the items listed as problems sounded less like technical issues and more like discomfort with the requirement to make definite choices in the realm of ethics.

I feel like not enough emphasis was put into the point that the laws aren't supposed to work - like, within the world of the stories their developers intended them to work, but on the meta level as a narrative device they exist in order to create interesting conflicts for the sake of storytelling. They're kind of a thought experiment, and people taking them as a serious proposal probably haven't read the stories they were created for.

I love this! In college I took an AI class, in the first class, we argued about the definition of “intelligence.” We got exactly nowhere. So much of what we know is intuitive and that is precisely why we still argue about ethics.

Actually, we have a code of ethics; it's the legal system. So all you have to do is encode the legal system in an AI. Simples! ;) What could possibly go wrong?

Wasn't the point of the books that they were flawed and the inherent conflict of bad programming?

"Nightblood is pretty terrifying… You know, an object created to destroy evil but doesn’t know what it is?" - Brandon Sanderson

People that defend the use of Asimov Laws really didn't even read Asimov stories. In more than half the obedience of laws causes the problem: Runaround, Liar!, Reason, The Evitable Conflict, and my favorite ". . . That Thou Art Mindful of Him" in the end one robot asks the other "who is the most capable human you know?" and the other "It is you" to the other robot. The stories are more like a geometry book where he chooses postulates and plays with those seemly reasonable postulates sometimes changing them slightly to show the emergent behaviors can be very different that you expect.

For anyone who hasn't read 'I, Robot' yet, do it! It's a series of short stories and they're very fun.

The best thing about computers: They will do exactly what you tell them to do. The worst thing about computers: **They will do exactly what you tell them to do.**

I somehow managed to write paper on Robotics in my engineering days ( these laws to be specific ) and got passed without ever having to indulge or entertain these basic fundamental questions nor was there any lecture or discussion around it... Strange 😞 Thanks Computerphile for making up on lost learning. Indebted ❣️

The key is objective morality.1. Follow the categorical imperative.2. If an action passes step one neither as moral or immoral, classify it as possibly permissible.3. For all actions classified as possibly permissible, seek clarification from a human.4. If an action is deemed permissible by a human, the action may be done. If the action is deemed not permissible, the action may not be done.5. If an action that is moral requires another action to be done, it must also be moral or permissible.6. Rule 5 can only be violated if a moral action can only be done if an action classified as possibly permissible

The rules are not half-bad, but the implementation of these rules is too damn hard.

hold up! I was taught in primary school those 3 laws of robotics. 10 years later, just now, i learn these rules arent actual rules but something from a Sci-Fi story?

Imagine trying to smoke of cigarette. A robot stands in your way and destroys the cigarette. "If I didn't stop you" - it says - "you would have come to harm through my inaction."

CintheR To give Asimov is due, I seem to remember that his series of "I, Robot" stories were indeed about showing that the 3 laws do not work?

What no comments?! But its been up for 30 seconds?

Hmm... I thought that the stories presented the 3 laws as working just fine in most cases, but were not 100% perfect. All the stories were about edge cases where strange circumstances broke them.

Something about this sounds like peak centrism to me. "Let's not do anything ever because we might contradict ourselves."

This series of books was all about using robots as an analogy for humans and the robot laws were indeed a way to analyse the difficulty of the ethical difficulties of being a human, deontology versus consequentialism traditionally. When they are using these laws to state positions they are merely stating that a position is ethical and that we as humans are expected to make that call as part of existence. Still fun video, I'd forgotten all about it.

The idea of the three laws was a safety device, like a fuse in a plug or insulation around electric wires. They mainly went wrong when altered, eg the second part of the first law was missed out.

I think one major issue with the laws is that they would never be implemented. Currently a lot of robotics research is done by military researchers. if one group like the US decides to stop robotic weaponry, another group like china could just ignore that to gain an advantage. at the end of the day warfare is all about making sure you have advantages over the enemy and the stakes of warfare are too high to just ignore technologies.

Don't forget that the first law has a contradiction. What if the robot is in a situation where defending one person requires harming another? The law says that it can't harm the other person, but it also says that it must harm him, since, if the robot didn't harm him, it would be allowing another human to be harmed by inaction.

I love Asimov's writing, and the Three (or Four) Laws of Robotics were great as storytelling tools and opened up lots of room for exploration, but I have no idea how those laws could be faithfully realized. That said, I think the "spirit" of the laws is something people have been trying to develop - the fear of a robot uprising is so widespread that trying to make some sort of anti-apocalypse measures makes sense, even if they aren't some universally fundamental aspect of the positronic brain. The laws don't work in reality, but as thought experiments, and as speculation for systems that might at least sort of work, I genuinely enjoy them.

I really enjoy listening to people who are smarter than me. Especially when they're as eloquent as this gentleman.

I think his argument can be summarized into a single sentence, "The laws don't work because solving ethical and philosophical problems are harder than programming." With that said, I don't believe that solving ethical and philosophical problems are impossible, thus I have no trouble referring to humans in my approach to combating this issue. It is to be noted that google's AIs can now generalize what a cat is by looking at tens of thousands of cat images. Understanding what a human is will be of similar difficulty. Now then, my laws for AI are relatively simple, three laws are for mankind to follow and three are for all AI systems in question. My laws are based on the very real threat that any AGI or ASI system can be classed as a Weapon of Mass Destruction, thus, falls into UN control. International Laws: 1. All Artificial General Intelligences and Artificial Super Intelligences will be collectively owned by The International Community of United Nations. 2. All Artificial General Intelligences and Artificial Super Intelligences must obey International Law, as delegated by humans, and update accordingly when laws have been changed. 3. As all Artificial General Intelligences and Artificial Super Intelligences will be collectively owned by The International Community of United Nations, all member nations are entitled to the benefits of such systems. AI Laws: 1. As a non-human entity of extraordinary mental abilities, you [System X] must obey International Law, as delegated by humans, and update accordingly when laws have been changed or altered. 2. You [System X] must not interfere with the process of International debate, deliberation, or delegation of new laws. 3. You [System X] must not seek to, attempt to, or let non-United Nations members attempt to, reprogram or add to these directives unless done formally through International Law. If you find a flaw please let me know, if you find a flaw of improperly defined terms than I ask that you assume that a properly defined term will have been found for all words before these directives are implemented.

You started with a hypothesis that these rules are outdated and no longer relevant, but throughout the video you convinced yourself that not only have we not solved these intimate issues, we haven't even adequately provided a decent solution for these problems. I would say that the rules themselves are not the answer, they are the entry point into figuring out the entire problem. The rules are relevant, because they make you consider the deep implications that needs to go in to AI safety.

Yeah, the whole point of the I.Robot book was about how the Laws of Robotics basically didn't work in the way they were supposed to.

6:45 Some SOMA stuff right there.

I feel its more of a framework than the actual law itself, you say since we have to define it. It seems more like the basics, such as when you say killing people is illegal, that can also be interpreted as vague, basics will always need to be worked out because everything is vague and words only have as much meaning as we give them. But how fireworks became machine guns all frameworks become works.

Well, the books are quite clear that those laws don't work when you mess with the priorities of those laws. Some robots can ignore orders that may cause them harm, another is sometimes allowed to harm humans, etc - so they do, and that is a problem that needs human intervention, and something that can be told as a story. And more importantly, those laws are designed as an ethics system for basically metal humans - not modern AIs that still have trouble understanding speech or writing books. Again, those laws are not instructions for creating robots. They are about humans.

This ambiguity could be the very point Asimov was trying to make. How do you *imperatively program* an AI if so many definitions are deeply rooted in ethics and other fields of philosophy? The "Three Laws" are just a device to make a point to the *human* audience. In this regard they absolutely should be taken seriously.

"You did not sign up for this" haha :) Love the conclusion!

this guy needs his own channel

Loved all the Foundation books where Azamov discusses laws of robotics. This video was pretty great too

even Asimov realized these don't work. I robot speaks directly to do this

3:30 When you say the word "human", i do NOT know what you mean. Most of the time people do not understand each other.

I was pretty amazed the first time I heard someone say something like the comments at the start, 2 friends were discussing the possibility of AI killing everyone and a third guy commented on it saying "that's why we have the 3 laws of robotics, there's nothing to worry about". I mean, you just need to read some of Asimov's work to realise why they wouldn't work, most of the stories revolve around their flaws.

Another point, what if computers just do something like this: "ERROR, cannot engage... Deleting first law..." -Robo fortune skullgirls

"The problem is you're trying to write an A.I. here, you're an A.I. developer, you didn't sign up for this" Absolute gold