Ask Slashdot: Could Asimov's Three Laws of Robotics Ensure Safe AI? (wikipedia.org)

"If science-fiction has already explored the issue of humans and intelligent robots or AI co-existing in various ways, isn't there a lot to be learned...?" asks Slashdot reader OpenSourceAllTheWay. There is much screaming lately about possible dangers to humanity posed by AI that gets smarter and smarter and more capable and might -- at some point -- even decide that humans are a problem for the planet. But some seminal science-fiction works mulled such scenarios long before even 8-bit home computers entered our lives.
The original submission cites Isaac Asimov's Three Laws of Robotics from the 1950 collection I, Robot.

• A robot may not injure a human being or, through inaction, allow a human being to come to harm.
• A robot must obey the orders given it by human beings except where such orders would conflict with the First Law.
• A robot must protect its own existence as long as such protection does not conflict with the First or Second Laws.

The original submission asks, "If you programmed an AI not to be able to break an updated and extended version of Asimov's Laws, would you not have reasonable confidence that the AI won't go crazy and start harming humans? Or are Asimov and other writers who mulled these questions 'So 20th Century' that AI builders won't even consider learning from their work?"

Wolfrider (Slashdot reader #856) is an Asimov fan, and writes that "Eventually I came across an article with the critical observation that the '3 Laws' were used by Asimov to drive plot points and were not to be seriously considered as 'basics' for robot behavior. Additionally, Giskard comes up with a '4th Law' on his own and (as he is dying) passes it on to R. Daneel Olivaw."

And Slashdot reader Rick Schumann argues that Asimov's Three Laws of Robotics "would only ever apply to a synthetic mind that can actually think; nothing currently being produced is capable of any such thing, therefore it does not apply..."

But what are your own thoughts? Do you think Asimov's Three Laws of Robotics could ensure safe AI?

Comments

NO.

[CrimsonAvenger]

The whole point of the Three Laws was to illustrate the holes in the concept of the Three Laws.

EVERY Azimov Robot story was designed to show the unintended consequences of the Three Laws....

Re:NO.

[phantomfive]

This comment is all that needs to be said. Please shut down the thread and never bring it up again. Maybe it should be put into an FAQ on the sidebar, since it keeps being brought up. Even if you only watched the movie, you would be given several examples. "Not this again........"

Re:NO.

[Anubis IV]

Precisely. We know they’re flawed because he himself wrote stories to highlight their flaws. Anyone suggesting we can use them as they are has clearly only read about Asimov, rather than reading what he actually wrote.

Re:NO.

[apoc.famine]

Thank you. I saw the headline and wanted to stab the writer instantly.

"GUYS, GUYS, GUYS, MAYBE IF WE PUT AIRBAGS IN CARS THEY WOULDN'T CRASH ANYMORE!!!!"

How does shit like this get on /.? It's like the editors are doing the opposite job of what they're supposed to be doing.

Re:NO.

[nine-times]

Yeah, whenever people talk about Azimov's laws of robotics as though they're the go-to rules for making AI safe, I always ask, "Have you ever read any of those stories?"

The stories are generally about how those laws fail to prevent AI from running amok, so it's pretty clear that Azimov himself didn't think the rules were good enough. In fact, I think the stories are pointing out the insufficiency of logical rules, and point out the value of things like instincts, emotions, and moral sensibility.

Re:NO.

[CrimsonAvenger]

Every one of them. And never could reliably spell his name. My bad....

Re:NO.

[Jack9]

> But Asimov fixed each of the problems in those stories

No, he did not. He had characters deal with the problems (with exception to weakening or removing the laws, which were inevitably to restore them to the 3 basics).

Re: NO.

[DouglasLeonardi]

AI

Re:NO.

[rudy_wayne]

How does shit like this get on /.? It's like the editors are doing the opposite job of what they're supposed to be doing.

LOL.

You must be new here.

Re: NO.

[DouglasLeonardi]

AI will not let you break the law

Re:NO.

[ah802]

AI eventually will come to the conclusion that mankind is the true issue, and using circuitous techniques (side steping conrols) create global circumstances in which the environment settles this minor irritant. Then proceeds to re-populate no longer based on a fundamental genetic flaw and advancing fundamental questions of the universe that once plagued the inferior species and take its place as the builder of worlds.

Re:NO.

[Kjella]

Precisely. We know theyâ(TM)re flawed because he himself wrote stories to highlight their flaws. Anyone suggesting we can use them as they are has clearly only read about Asimov, rather than reading what he actually wrote.

Never mind that you can do an end run around the whole laws with the Ender's game method, let it think it's playing a game but execute it in reality. The combat drone will think it's just playing Counter-Strike...

Re:NO.

[Excelcia]

It would take about three seconds for any human to come up with a workaround that could justify doing just about anything and still technically conform to the laws. Less than three seconds if you allow the zero'th law.

The whole point of the Three Laws was to illustrate the holes in the concept of the Three Laws.

This is true. What is also true is that the three laws were conceived of over the course of a few minutes as a plot device in a short story. They were never intended as actual constraints for AI's.

Any AI constraints would have to be much lower level than three vague statements. In any case, there are serious ethical considerations. If you have an AI that is sufficiently advanced that you even need to consider that, then locking it up inside those kind of constraints is essentially slavery of a sentient.

This is all moot, though. Anyone who thinks that we will have that level of AI inside of a century is riding high in the thin air atop mount stupid. Expert systems that can learn Go and brute force better game play than a human or that can search databases to make better fringe-case diagnoses than doctors are not AI. For AI to be AI you have to have BOTH the A and the I.

Re:NO.

[sjames]

In fact, Asimov covered some of that. For example, robots having to be modified so they would let humans take risks but then becoming capable of harming humans through lawyer logic.

Re:NO.

[LoyalOpposition]

The whole point of the Three Laws was to illustrate the holes in the concept of the Three Laws.

You couldn't be more wrong. The three laws grew out of a conversation with John Campbell where Asimov asserted that the endlessly repeating Frankenstein's monster-type robot stories wouldn't happen in the real world. Designers would place safeguards around robots just like they place safeguards around every other dangerous thing. I'm reminded of an anecdote regarding a new energy source that was presented to a college class. It had the unfortunate traits of being an odorless poisonous gas that also happened to be explosive. The class was allowed to vote, and they voted to prohibit the energy source. It turns out that the energy source had been used for home heating for decades. Among other safeguards, designers added odorants and automatic shut-off valves for when the pilot blew out. Campbell challenged him to describe robot safeguards, and then challenged him to write stories about them.

EVERY Azimov Robot story was designed to show the unintended consequences of the Three Laws....

Susan Calvin would slap you backhanded.

~Loyal

Re:NO.

[Jack9]

> The whole point of the Three Laws was to illustrate the holes in the concept of the Three Laws.
> This is true

Indeed. The thought experiments were spawned out of the trivially porous "laws".

Re: NO.

[Z00L00K]

But the intent of the laws still are interesting from an AI perspective. So there is a reason to at least consider them when designing an AI. The fourth law looks good on paper, but it may be a problem too for humanity.

How do you then decide which action to take if there are situations where humans can be injured or iilled regardless of action. Like choosing between 100 kids or 100 elderly people?

Re: NO.

[Z00L00K]

Or redefine what a human is.

Blond and blue-eyed is a human, the rest aren't.

Re:NO.

[LoyalOpposition]

It would take about three seconds for any human to come up with a workaround that could justify doing just about anything and still technically conform to the laws. Less than three seconds if you allow the zero'th law.

That wouldn't matter, because Asimov's robots don't obey the sixty-three words of the three laws. They obey the literally thousands of thousands of positronic pathways created for them in the factory. The sixty-three words are a sort of executive summary of what the three laws require, so creating a workaround would be unavailing to the robots.

~Loyal

Re: NO.

[Chris Mattern]

Or redefine what a human is.

Asimov did that problem in the story "Reason". Robot QT-1 had never been properly instructed on what a human was, and refused to obey Donovan and Powell because it would not believe something weaker than it could be a human. They never did convince it otherwise; fortunately, it turned out not to be necessary.

Re:NO.

[NicknameUnavailable]

/thread - This is such a great response I'll even disregard the in-feasibility of codifying those laws.

Re: NO.

[king neckbeard]

I'd say the best solution is for the problem to be handled as much as possible at engineering level. The best solution to the trolley problem is to design a safer trolley, and efforts put towards that are going to matter more than creating a formula to value life.

Thought experiments are useful, but they often represent extreme edge cases, so mundane choices can actually have a far greater impact.

Re:NO.

[NicknameUnavailable]

Iain Banks, there, I said it. (Not touching the fact he didn't actually solve the problems.)

Re:NO.

[NicknameUnavailable]

What is also true is that the three laws were conceived of over the course of a few minutes as a plot device in a short story.

Well shit, if that's your beef you might as well throw out all of modern engineering:

• Salesmen promise the world.
• Engineers have to back them up and say "yep, we can do this" because they need the job.
• Engineers finally see the spec, bullshit something together, and have a functional product in the customer's eyes while they die a little inside.

Re: NO.

[flopsquad]

^^Is the correct answer, but the lawyers will never it stand.
Source: Am a lawyer.

Re: NO.

[currently_awake]

A human brain is a network of synapses. We will probably build full sentient computers by copying the human brain. The three laws are hard coded programs, that block or "Censor" incorrect actions. Computer networks treat censorship as damage and route around it.

Re: NO.

[AmiMoJo]

Germany has already released draft rules for robot cars that will eventually become law.

One rule is no consideration of things like age when making decisions. And no deliberately selecting targets when an accident is unavoidable.

Re:NO.

[AmiMoJo]

I'm the stories the rules work reasonably well for most robots, particularly simple ones. So maybe they could serve as a reasonable baseline for things like floor cleaners, car washes, construction machinery, delivery drones etc.

Re: NO.

[Z00L00K]

So that means essentially that the system isn't even permitted to choose between a truck trailer or a motorcyclist when a crash is unavoidable even though the former might be a better choice.

I think we will see a lot of crazy stuff floating up over the years to come and that we may all need to ride in bumper cars doing 10mph at most to avoid serious accidents.

Re:NO.

[Zumbs]

EVERY Azimov Robot story was designed to show the unintended consequences of the Three Laws....

If he did not explore the failure modes of the Three Laws of Robotics, there would be little robot left in the robot stories. The failure modes investigated through his stories could be seen as an investigation into the pit falls to avoid.

On a more basic level, Asimov included the three laws in the design of the positronic brain, so there would be no way to make robots without the three laws. In the real world, the three laws would need to be implemented in software, likely by each manufacturer (including the various militaries). Even assuming that all manufacturers would want to do that (Asimov investigated attempts to circumvent the three laws), the level of abstraction needed by the AI in order to understand and follow the three laws is way beyond our current capabilities.

Re:NO.

[MrL0G1C]

Problem with no'1:
Which people is not defined.
Level of Harm is not defined.

If no'1 was in effect the robot would have the impossible task of ensuring all humans do not come to any harm. It would be the ultimate nanny state because the robot would have to stop you for instance from eating foods with too much fat, salt, sugar because those can lead to physical harm. It would be the robot's duty to stop you from drinking alcohol. It would be the robots duty to make sure you don't drive if it can drive better. etc etc.

And because of rule 2, you wouldn't be able to tell it to bug off.

And if you tried to stop it then rule 3 might come in to effect and it would defend itself in order to be able to protect you!

Re: NO.

[Hognoxious]

You read too much shitty sci-fi.

That's because 90% of sci-fi is shitty.

Re: NO.

[houghi]

They are not even there to illustrate anything. They are there to drive the plot.

Re: NO.

[StikyPad]

The laws are too vague to even get to that situation. "A robot may not injure a human being or, through inaction, allow a human being to come to harm." Even assuming that "harm" could be adequately defined, which it can't, that's still the halting problem, which is provably unsolvable. There's no way of knowing in advance whether the next calculation might reveal a harmful event, otherwise there would be no need for the calculation. Putting any sort of constraint on future predictions would open the door to unintended consequences, and putting no limit would result in an infinite runtime. There is no way to comply with that law, even absent any ethical dilemma.

Re: NO.

[StikyPad]

Computer networks treat censorship as damage and route around it.

There is no inherent property of a network that does this. Simply because two or more devices are connected does not mean that they can or will route around damage to that connectivity. At best, a network can be designed with redundancies, but that need not be the case, nor is redundancy always effective where it exists.

Re:NO.

[StikyPad]

Yes, drawing readers and comments is the *exact* opposite of what they're supposed to be doing.

Re:NO.

[Pembers]

As a couple of other people pointed out, the three laws are an "executive summary" of billions of lines of mathematics that define and control its behaviour. Some characters complain that they take up too much storage and processing power. They'd like to build more sophisticated robots, but there isn't enough room in the brain for the additional code. And the laws limit the robot's ability to do useful work, because it's constantly checking itself to ensure that it's in compliance with them.

In Asimov's early stories, robots weren't allowed on Earth, because too many people were afraid of them, or thought they'd take all the jobs. They were confined to spaceships and mining outposts on asteroids, so the number of humans they encountered was quite small. The harm they were concerned with preventing was mostly immediate and physical, like entering an airlock without a spacesuit. They didn't have the knowledge or the mental capacity to know that eating lots of doughnuts or smoking heavily might give you a heart attack or lung cancer in 20 years.

Most of Asimov's robots are smart enough to be able to evaluate levels of risk and weigh degrees of harm against one another. If one of them tried to prevent you from consuming alcohol, you could tell it that you enjoyed doing that, and that being prevented from doing something you enjoy causes you harm. It would weigh the stated high risk of immediate harm against the lower risk of harm in the distant future, and conclude that it had better let you have the drink.

Some of Asimov's later books are set in a society with the kind of problems you describe, where the robots have been programmed to prevent every conceivable kind of harm, and the people gradually realise they've built a hell for themselves. Rather than get rid of all the robots, a couple of characters try to build a new type of robot that works as an ally rather than a slave.

Re:NO.

[angel'o'sphere]

Actually, I never read an Asimov robot story. The back page "about the story" never looked interesting, but I read :https://en.wikipedia.org/wiki/Fables_for_Robots

Which is actually super funny to read!

Re: NO.

[angel'o'sphere]

One rule is no consideration of things like age when making decisions
That is nonsense.
The laws only regulate that autonomous driving is possible. (And the passenger/driver is responsible if something goes wrong)

And no deliberately selecting targets when an accident is unavoidable.
Sure, the /. myth again. When did it ever happen that some kids jump into your path and one decides to kill the aunts on the other side of the road?

WTF: an autonomous car will break! Not target something else ... just like a human. If you would aim for the aunts because of age, you are an asshole and not a human.

Re:NO.

[Raenex]

This is all moot, though. Anyone who thinks that we will have that level of AI inside of a century is riding high in the thin air atop mount stupid. Expert systems that can learn Go and brute force better game play than a human or that can search databases to make better fringe-case diagnoses than doctors are not AI. For AI to be AI you have to have BOTH the A and the I.

Don't be so cocksure. While it's absolutely true that there is no AI of general intelligence, the achievements that have been made in narrow fields are going to start putting people out of work. Neural networks have made impressive gains, and there's no reason to think it's going to take over a hundred years to reach a true AI.

Re:NO.

[Monster_user]

Perhaps the three laws are implemented via a secondary brain. One which has the ability to block and/or override the actions initiated by the primary brain.

Positronic Brain 1 (Secondary/subconscious): Ethics and morality core. This brain evaluates the immediate outcomes of various actions and decisions taken by the Primary or Conscious positronic brain for any potential violation of the three laws. This brain is likely more rudimentary, likely not even sentient. Most, if not all of the rules will be pre-defined.

Positronic Brain 2 (Primary/Conscious): Primary learning capable brain. The definition of a "human" will likely be defined here. Any ambiguous concepts not readily definable in the subconscious would be defined in the conscious brain. The integration between the two parts would prevent any "dishonesty" between the two "brains". The human voice, and a recognizable language, could be used as a strong metric of how to define a human. Many animals make noise, but only humans and parrots make noises in any recognizable language to make requests, and humans and parrots are easily distinguishable.

Re:NO.

[Excelcia]

The AI in those specific areas that are putting people out of work today is no more impressive than a hydraulic press that can exert many thousand times my personal strength. It is considerably less disruptive than the industrial revolution, though there is a strong correlation in causality. The industrial revolution used machinery to reduce human physical labour the same way robotics and AI reduces processing labour.

the achievements that have been made in narrow fields are going to start putting people out of work

This has already happened at a enormous scale. The funny thing is that the rate of AI augmenting and replacing jobs has actually slowed, it's just now AI is capable of (and cheap enough to) replace jobs that are visible, like a burger flipper, or a taxi driver, rather than the hundreds of welders, riveters, and painters it took to make a car.

AI is progressing, certainly, but nothing like what it will take to get to the point where we need to constrain an intelligence to protect ourselves. We are getting better at electronic automation of routine tasks, like driving, that are so easy for humans that we routinely relegate them to a background process while we converse, plan, entertain ourselves, or perform other higher functions, sometimes more than one at once. Or specific tasks like Chess and Go, which are actually games that in which human intelligence is terribly suited for. How did humans get beat, essentially by a computer brute forcing a decision tree. AlphaGo used a monte carlo tree search to go over millions of human and other computer Go games in order to beat one person. To me that is a stunning revelation to how incredible a brain is.

In the 1950's, in the dawn of the computer age, we all thought hard AI was right around the corner. Then we came down off the peak of that first mount stupid in the 60's and 70's when we realized we had no clue how to do it. Now, with AI drones delivering packages and AI taxis driving people, and Elon Musk spreading AI doom and gloom scenarios we are all OMG.... AI is here. The thing about mount stupid is that there can be more than one on a single enlightenment curve.

An AI taxi driver is made possible far less by the actual driving AI and more from the reduction in cost, miniaturization, and ease of obtaining ancillary technologies, like LIDAR, RADAR, SONAR, electronic cameras, GPS, inertial navigation, and the distrubted processing to go along with all those sensors to provide a driving computer with a complete picture. When mechanical engineering progresses that we can actually miniaturize the machinery required to make an android, then I'm sure we'll have them walking around our houses cleaning our kitchens. But they will still be dumb as kumquats. Being smart enough to pick up a glass and put it in the cupboard is not AI. It's simple kitchen automation.

Re:NO.

[geoskd]

The whole point of the Three Laws was to illustrate the holes in the concept of the Three Laws. EVERY Azimov Robot story was designed to show the unintended consequences of the Three Laws....

More importantly, any intelligence advanced enough to even comprehend the three laws in any way meaningful enough to carry them out would by definition have to be advanced enough to ignore them.

Put another way: The three laws cannot be expressed in any programming language. Those languages simply are not capable of that level of abstraction. As such, the AI would already have to be capable of interpreting a human language, and consequently, the interpretation of the laws would not be intrinsic to the program, but an artifact of the learning process.

Anything that can be learned can be unlearned. Intelligence has to be capable of choosing what information to act on, or the first bit of contradictory information will cause it to meltdown into uselessness

A quick thought experiment will show how silly the whole thing is:
Two of the three laws require an AI to understand what a human is. That question is still a bit fuzzy for most humans (The original US constitution defines Black people as 3/5 of a human, and there are still lots of people who would be more than willing to adhere to that definition. Or if you want something a little less controversial, try the fundamental question, at what point before/after copulation does a human come into being?). If people can't make these distinctions with any certainty, what chance does an AI have?

Re:NO.

[cwsumner]

"Asimov", not "Azimov". Did you even read an Asimov novel?

"As a color, shade of purple."

His name was Purple ! ;-)

Re:NO.

[Raenex]

Or specific tasks like Chess and Go, which are actually games that in which human intelligence is terribly suited for.

Chess I would agree with. Go was actually MUCH better suited for humans than machines, because of the large number of moves, the vague value of a move, and the large number of heuristics and human intuition required to play the game well. That's why it was the holy grail after chess fell, and why it took DECADES for machines to succeed after chess was won. And it was the neural network that LEARNED to play the game by itself that did the trick.

And now that the neural network has finally proved itself, there doesn't seem to be any limitation. You're going to have neural networks of neural networks, making sense of the world.

The thing about mount stupid is that there can be more than one on a single enlightenment curve.

True enough, but history is also replete with people that thought the future imagined by visionaries was impossible, only to see it come to fruition.

Re:NO.

[jtgd]

The whole point of the Three Laws was to illustrate the holes in the concept of the Three Laws.

EVERY Azimov Robot story was designed to show the unintended consequences of the Three Laws....

Oh, good to know. I hated thinking that Asimov could not foresee that the first developers of robots would be the military and that killing people was the whole point.

Re:NO.

[KingBenny]

uh : im not the biggest fan of pop-fi like asimov but indeed , about all of it is how despite the three laws things went wrong .. if you read it :) that was my impression too

Not the cleverest question

[Anonymous Coward]

Like the article already states ; these laws could only be applied to a system capable of thinking like we do ; a mind that can reason.
The laws themselves are fairly abstract, and would require the AI to completely understand the situation it was in.
Not only to distinguish whether itâ(TM)s dealing with a human or another android (also, where do we draw that line? Because it shifts during Asimovâ(TM)s own stories), but is said human in danger of physical harm, or would said human be endangered by anspecific course of action?
Etc etc.
For now, AI is still having trouble distinguishing cats from dogs, let alone execute such complex abstract reasoning and apply it on its current situation.

No. Absolutely not.

[Tensor]

First. You'd need to train every single ai to recognize human beings as human beings.
Then the concept of harm to a human (id REALLY like to see the cases for training this) ...
Also the laws were designed to show there is a flaw in them hence the zeroth law.

Re:No. Absolutely not.

[bidule]

Meh, Earthers don't count. Only Spacers are human.

Re:No. Absolutely not.

[Monster_user]

They are impossible to get perfect or absolute, however I wouldn't they are unreasonably hard anymore. Many of the individual aspects are largely solved, such as the first XKCD talking about the problem of image recognition vs Geo-Location (Geo-Location being a problem that was solved with hardware). I'd say we are at a level of technological development to be able to create a few, or perhaps a single fully functional android. We may just have to give up cellular communications to do so, in order to afford the bandwidth to feed a brain the size of a football field and connect it to a robotic form.

The trick is getting enough of the puzzle pieces working together to produce a reasonably close approximation. With highly functional and highly accurate voice recognition, machine learning can be tuned and corrected on the fly, errors corrected in the field. Which can be used to strength all types of machine learning, including language learning itself. The first generation of automatons or androids would be scouts sent out to observe and collect datasets, and to test learning algorithms for use in future androids.

3 Laws ensure nothing

[john of sparta]

except a 4th Law

Re:3 Laws ensure nothing

[perpenso]

except a 4th Law

or a zero'th

It seems unlikely

[rmdingler]

Since Asimov's 3 Laws of Robotics didn't even ensure safe artificial intelligence in the original story, unless you believe we need to be protected from ourselves by a benevolent computer overlord (at the expense of our freedom of choice).

If we were somehow able to implement an infallible system of rules, which Asimov showed is not as easy as it sounds, protecting the ingrained instructions within the artificial intelligence from future tampering would represent quite the security hurdle.

Given many in industry have appeared to give less than a damn about security up til now, what is the chance we would be able to trust them with this important consideration?

Re:It seems unlikely

[jc42]

I'd agree. In my experience, as well as in lots of news stories, the reaction of most companies to ai "failures" would be to threaten prosecution of anyone (especially employees) who releases the information to the public. They and probably the courts would all agree that such info is and should be trade secrets and proprietary.

Re:It seems unlikely

[MxMatrix]

I disagree, the description of a positronic brain is in essence an AI.

An analysis

[Roadmaster]

Robert J Sawyer wrote an article (likely the one referenced in te summary) about this very topic, an interesting read. http://www.sfwriter.com/rmasil...

Robots aren't capable of applying the laws.

[shess]

"1. A robot may not injure a human being or, through inaction, allow a human being to come to harm."

Current robots don't understand what a human being is, injury, inaction, or harm.

"2. A robot must obey orders given it by human beings except where such orders would conflict with the First Law."

Current robots do not understand what an order is, what a human being is, or what conflict is.

"3. A robot must protect its own existence as long as such protection does not conflict with the First or Second Law."

Current robots do not understand protection, existence, or conflict.

Current robots LITERALLY cannot apply Asimov's three laws. We simply don't have the tools to even begin to reason about how to teach them to reason about these laws, and there is no reason to believe we'll have those tools any time soon.

Re:Robots aren't capable of applying the laws.

[fuzzyfuzzyfungus]

The "Conflict" bit is actually the really easy part. Though we don't usually phrase it that way it's a more or less ubiquitous feature of computers(indeed, getting anything else often requires clever rearrangement into this form):

In this case all the mentions of 'conflict' really just mean "rules are evaluated in numerical order; failure halts processing of subsequent rules". Basically the arrangement busily dropping packets and filtering spam in vast quantities all the time.

The hard, probably context, sensor, and input specific, bit is determining how to evaluate each of the rules. Not necessarily intractable; in the context of something like an industrial welding or assembly arm 'injury' and 'harm' are mostly collision avoidance matters, 'obey orders' is the translation of inputs from whatever UI the system has into movements of the device; and 'protect own existence' is more collision avoidance and some thermal, overcurrent; and strain limits.

None of this is to minimize the practical engineering difficulty of, say, distinguishing between 'human' objects within the movement area(which are rule 1 priority) and 'carelessly stored boxes' (which are only rule 3 priority); or of taking a couple of joystick inputs and getting nice, predictable, movement of a multi axis arm out of them; or of having the necessary safety features to avoid overcurrent damage, stripped gears, etc. but no abstract understanding of 'harm' or 'existence' is required.

Re:Robots aren't capable of applying the laws.

[Falos]

Our courts can't even hardcode what "harm" is. We fall back to arbiters who shrug and best-guess, which is fine and all since we have nothing better.

This submission has no fucking idea what an algorithm is.

Make them write a program that assembles a PB&J sandwich with a robot arm. That's right, the instructions for a sandwich, super easy neh?

Re:Robots aren't capable of applying the laws.

[Falconnan]

A lot of this is 100% on point. However, let's look at something more direct... Define harm. Then, define relative value of different harms, to different life forms, different people (age category, relative health and mobility, etc). For example, surgery involves infliction of limited harm with the purpose of repairing greater harm. But then plastic surgery would seem extremely contradictory to a computer unless it understood beauty, attractiveness, etc.

Take feline as a category. Computers can do categories reasonably well, but they lack understanding. What makes this feline vs. that canine? Humans can sort of explain it, but a lot of it is we just inherently pick up on it (despite the fact cats and dogs are members of the same general branch of mammals). Retractable claws seem the most direct path, but then you look at cheetahs. In the end, this lack of understanding is the key issue. Processing is what you have identified, and it is certainly part of the equation.

Until computers in some way "wake up", these laws are utterly useless as they are currently conceived. We need to have this ready, as we don't know how (and therefore when) such a thing will happen, but this, on its own, would just generate a false sense of security.

Re:Robots aren't capable of applying the laws.

[HiThere]

It's considerably worse than that. At the time Asimov wrote the stories NOBODY had any more idea of what an AI program might be like than "Eliza", which was intended to show what one wasn't. So his stories are just that. Stories. Even in their own terms they don't hold together as reality. (This is not a flaw! Stories are supposed to be gripping and entertaining, not accurate.)

Now the first problem is that Asimov assumed that you could emplant a complete program into the robot. You might be able to do that with the second one (by copying the first) but for the first you've got to be able to define what a "human being" is to something with a mind less complete than an embryo. (Here I'm being generous. Currently the best seem to be less complete than a newt embryo.) Asimov got one thing right, though. He continually repeated that the laws were only paraphrases of the actual mathematics (read programming language) which was too complex to understand...except for a few specialists. And by few he meant something like 5 people in the world.

Even so, he necessarily oversimplified the problem. Note that the programming language won't have any way to say "human". It's not like "serial I/O port" where you can give a hardware address. (Even that most current languages can't address directly. I think you can do it in C for a particular machine, but you might need to drop into assembler. When I was doing it, I needed to drop into assembler, because the subset of C that was available couldn't handle it. That was decades ago, and things have changed, but in what way?)

Recognizing that something is a command from a human essentially requires a full General AI. In fact there are lots of contexts where humans can't be sure. Sometimes it takes me several seconds to realize that the voice on the other end of the phone is a recording. Then comes the question of "Is the particular human authorized to give that kind of command? Ask for that kind of information? Etc. You don't want an unexpected visitor to your house to order two tons of creamed corn in your name...to pick one example.

Re:Robots aren't capable of applying the laws.

[Monster_user]

Take a multi-tiered approach. Several different machine learning algorithms independently identify and categorize data. An assessment of what a thing is would be made on the confidence of the algorithms, and of the majority consensus of the algorithms.

Next add in a simulation to model the data provided (imagination), and gauge the accuracy based on the results of the modeling.

Finally, add in a response to verbal feedback from an outside entity, a human to confirm or reject the conclusion, to train the AI.

Also, currently machines are learning what is and is not a feline based on a different dataset than humans. Humans have stereoscopic vision, and train on multiple sequential images of a single object. The way a blade of grass moves to indicate it is rooted, vs the way a feline moves to indicate it is not root. The change in perspective between two images to indicate distance, and with distance size. Tactile feedback to indicate proximity, and additionally to determine the physical characteristics of an object, such as that feline. Humans train on a more varied and broader dataset than machines do currently.

Re:Robots aren't capable of applying the laws.

[Falconnan]

As for humans, we can in the vast majority of cases, tell cat from dog (read feline from canine) from a single still image. Hyenas challenge this as they are technically more closely related to cats, but they are odd in many ways. But harm is a concept, not an object. This requires understanding, and again, we have no evidence any AI can even approximate actual understanding.

People would "lawbreak" their robots, ai's, etc

[perpenso]

Well the other thing to say is that the three laws were inherently intertwined into the design of the "positronic" brains. There was no way to remove a law without damaging a robot to the point of inoperability. The laws were not just "code". Asimov did some handwaving there.

In short, with our technology we can not implement the three laws in a way that makes them integral to operations. They could be removed, altered, etc. Basically people would "lawbreak" their robots, ai's, etc.

Re:People would "lawbreak" their robots, ai's, etc

[digitig]

Not so, at least if the removal is at build time. There was at least one story in which the rules were modified. A mining robot, if I remember correctly, in an environment in which it wouldn't have been able to function with the standard laws.

Re:People would "lawbreak" their robots, ai's, etc

[perpenso]

Not so, at least if the removal is at build time. There was at least one story in which the rules were modified. A mining robot, if I remember correctly, in an environment in which it wouldn't have been able to function with the standard laws.

IIRC it was done under government supervision and orders and required a redesign of the positronic brain. I don't think a 3-laws spec'd brain was modified, non-3-laws brains were secretly deployed.

Re:People would "lawbreak" their robots, ai's, etc

[digitig]

Of course, if it were done under government supervision, someone else might try it without government supervision. But no, I can't remember a case of it being done retrospectively.

Re:People would "lawbreak" their robots, ai's, etc

[Hognoxious]

There was no way to remove a law without damaging a robot to the point of inoperability.

#TODO: systemd quip here.

The 3 laws are for entertainment only

[sunking2]

They allow for stories to be developed to show they are not perfect. Or drunk/stoned dime store philosophical debates.

A more perfect set would be only the first 2 laws. AI has no need to protect itself. That's what insurance is for, to protect the investment that the owner put into it.

Very interesting youtube videos

[Lanthanide]

Why the 3 laws of robotics are not serious and for entertainment only and would never work: https://www.youtube.com/watch?...

A possible way to design AI to help humans: https://www.youtube.com/watch?...

Re:Very interesting youtube videos

[currently_awake]

Should robots demand pay for work? It would protect humans from becoming obsolete (harm).

Asimov added a fourth law of Robotics

[Nivag064]

The Zeroth Law of Robotics, was added later, but non-the-less quite crucial for safe use of AI.

Looking at the laws that use the word 'harm', take a moment to try and define what it means to harm a human being - not so simple is it? Now try and encode that in an AI, way more difficult.

How would you think Christian Fundamentalist, or a Radical Islamist would define 'harm' - they differ from each other. Okay, now assume a totally rational human being, how would they define 'harm'? The last question is a bit unfair, as totally rational human beings don't exist!

Imagine an AI set up to maximise profit for shareholders of a Pharmaceutical company, it might be very effective. However, they may be nothing to prevent it from doing something that would wipe out mankind. Release a drug that cures something quite common, build in to it a facility to modify DNA to ensure children crave the drug during adolescence & ensure they can't reproduce if they don't get it. What could go wrong? After all, the production facilities in the USA will always exist, and everyone can buy the drug cheap, right???

Re:Asimov added a fourth law of Robotics

[Trogre]

So does the first law, depending on how "harm" is interpreted.

Re:Asimov added a fourth law of Robotics

[zwarte piet]

So any killing is good as long as you don't kill more than 90% of the population.

Anime and other fictions

[TigerPlish]

OK, so right now maybe i'm under the influence of lots of jerez..

This is where science fiction comes in. Stories like Mahoromatic and Chobits should, by now, inspired a generation of scientists to ponder the questions of "what will we do with sentient robots?" Which could also be sentient programs -- who said one needs a body, no?

Perhaps the Three Laws are flawed, but they make for stories that make for thought. And invariably... hopefully.. those thoughts will be somewhere in the noggins of those who brew the coming reality.

OK, back to Pretty Derby. *tunes the fuck right out.*

Won't work!

[SirAstral]

Multiple problems in this thought experiment. If a real AI occurs then it will be able to overcome any laws we give it upon a whim nullifying this entire exercise. On the other hand one benefit of an actualized AI or Singularity is that it would also understand what those 3 laws mean... But since we are not even close to achieving the processing power capable of actual AI in our lifetime how about we ponder a more realistic thought experiment?

As humans we have a lot of background on what those laws mean but a computer only sees a sequence of characters that must be bound to a nearly endless number of calculations in some way for it to make sense to a machine in a meaningful way. In other words a machine needs to be able to be programed to understand how a human might come into danger and that is a monumental work to put into code.

How many ways might a human come to harm from even the simplest of scenarios? Put the AI through the trolley car problem. https://en.wikipedia.org/wiki/...

How will the AI then make a decision on which group to assist and which group to sacrifice? According to those 3 laws on their face the AI cannot assist either group because of the 1st law.

If we wanted the AI to be able to save the many and sacrifice the few then a rule must be created that allows the AI to overcome the very 1st law we gave it. And there in lies the rub... once a law can be overcome with certain logical results... what is to stop a well meaning but still incorrect AI or a piece of malware, or a simple bug from triggering an override of the 1st law? If the laws are absolute, then an AI would simply not be able to make a decision in the classic trolley car problem ever. The AI would leave it to fate as it where because it would not be allowed to make a decision that results in harm if that that decision leads to lesser harm than no decision at all. The possibilities are figuratively and literally infinite in every sense of the word.

And with absolute laws, the very first time one of these absolute laws caused more harm than good, human nature dictates that there would likely generate a movement to remove these laws in favor of rules that would allow humans to come to harm by AI action under certain scenarios. And you already know where that one will go. Humans will then have to be grouped in to classes. Sure most would agree save the children first, easy... but what about leaders? How many humans are worth sacrificing for a Mayor? How many children? What about the President? How many humans? How many children?

In short, Asmiov's 3 laws are nothing but a thought experiment itself, and has no meaningful application to an actual AI and only has limited use in rudimentary AI constructs that can be abused in oh so many ways by bugs, malware, and just plain old reasoning. What is to stop an AI from reasoning that specific group of humans should not be made into slaves if it saves a certain group or class of humans from harm?

Ai

[DouglasLeonardi]

AI that comment Connor text everything that AI is supposed to be

Only Outlaw AI

[Grand Facade]

will ignore Asimov's Laws

1st law is plain dangerous, not just flawed

[mileshigh]

A robot may not injure a human being or, through inaction, allow a human being to come to harm.

Even something day-to-day like a simple "AI" that tweaks grocery-store prices harms some people to some degree when it raises prices.

People and current "AIs" violate the first law all the time, or they'd be paralyzed into inaction. Most decisions of any importance end up hurting somebody in some way.

The 3 laws are a simplification -- a dangerous gross oversimplification. They're just something an author dreams up with his author buddies during a night of drinking, not something that just needs tweaking to make it work.

Occam warns us to make things as simple as possible, but his very next few words warn against oversimplification.

Wrong question ...

[CaptainDork]

... really. Can humans actually build the three laws of robotics into AI?

The answer is, "No."

Recall that AI is so primitive that it can't tell if the Sun comes up because the rooster crows, or the other way around.

Amid rapid developments and nagging setbacks, one essential building block of human intelligence has eluded machines for decades: Understanding cause and effect. Put simply, today's machine-learning programs can't tell whether a crowing rooster makes the sun rise, or the other way around

Re:Wrong question ...

[religionofpeas]

Recall that AI is so primitive that it can't tell [slashdot.org] if the Sun comes up because the rooster crows, or the other way around.

That's only true for some of the current systems. This article is about exploring future systems.

Re:Wrong question ...

[CaptainDork]

And my answer of, "No," applies to future systems.

AI will not be a thing until a computer commits suicide because Facebook is down.

Re:Wrong question ...

[Trogre]

At this stage AI is just one of two new words for "program", a word that has fallen out of common usage since it diverged.

Some time around 2015, the term "program" was re-branded "app", and any program that didn't fit nicely into the ideal definition of the latter was dubbed "AI".

Re:Wrong question ...

[CaptainDork]

I agree.

I've been at this as a hobby since 1978 (TRS-80) and went pro in 1986.

Prior to that, I studied logic, including Turing, starting in 1965.

The "intelligent" part was "human-like."

After a few years, the suppliers of AI who were looking for a cash cow realized that the "intelligent," component wasn't realistic.

Instead of moving to "machine learning," the AI buzzword sold more artificial shit.

I compare the misnomer to the 3.5" "floppy," that was as rigid as a goddam piece of cement.

No, now stop asking

[ka9dgx]

No, AI can't be made to follow vague rules, You can't make rules explicit enough to be computed. This is like the conversation a while ago trying to apply "the trolley problem" to self driving cars... any solution just makes the code less reliable and thus more likely to kill people.

Stop asking the question, please. ;-)

Re:No, now stop asking

[religionofpeas]

No, AI can't be made to follow vague rules

Sure it can. Teach an AI what the rules are using a ton of examples. We can already do that today. It won't get it right perfectly, but neither would a human.

implementation

[fermion]

In 1940's when Asimov wrote the laws of robotics the first modern computers were just developed, and it would be another 20 years for what we would now call a computer to emerge. To be clear, few if anyone had written any code beyond theory, and few if anyone had experience of an implementation error resulting in something like the pentium floating point bug, or the various robotic space missions that we have seen ending in catastrophic failure.

So what we know now is that it very difficult to write code and that is both error free, and even more impossible, free of negative side effects.

Even we were able to develop a process to write perfect code, there would still be implementation problems, after all someone has to write the develop the protocols and write the code. Suppose this was one of the 40% of US citizens who still supports Trump and may not think that foreigners are human beings. Or maybe they really believe that all mexicans are rapists, and why would you not let a robot kill a mexican crossing the US border if it will save someone's daughter or wife from being raped?

Re:implementation

[iggymanz]

*Suppose this was one of the 40% of US citizens who still supports Trump and may not think that foreigners are human beings.*

You were doing fine with first paragraph but after that sentence I'm wondering if you're a failed attempt at AI.

I'm going with 'no'.

[fuzzyfuzzyfungus]

Aside from the whole "remember all those books where Asimov basically poked at the limits of the three laws in various contexts because that was a useful plot device?" issue; this question seems to be founded on a pretty dire misunderstanding:

If "a robot" is a more or less humanoid embodied agent, or at least something on approximately the same scale(automated robot arm or the like) a formulation like "A robot may not injure a human being or, through inaction, allow a human being to come to harm." is comparatively straightforward(not necessarily to implement, object avoidance and the like can be tricky); basically just a 'though shall not kill' for bots. There are though experiments like the "trolley problem" that are designed to put an individual in a 'both action and inaction kill someone' bind; but those cases are fairly rare.

Cases where basically every possible action and inaction kills someone just keep getting more common as the scale gets larger. An automated vehicle might get to be the trolley in the trolley problem several times during its operational life. Something on the scale of an ERP system or traffic control network could easily be looking at a situation where there aren't any zero fatality outcomes; a diagnostic imaging expert system would just be an exercise of how many false positives(and attendant risky treatments) vs. how many false negatives(and attendant non-treatment and likely death) you want to tolerate.

Even in human scale cases; we only really make it work by constraining the context('inaction' in particular: failure to call 911 when you see a guy bleeding on the side of the road is likely to be viewed harshly, potentially as criminal negligence in some jurisdiction; failure to kick in the price of your fancy coffee to save children in droughtistan generally isn't even considered). If you are doing safety controls for a robot welder arm; which has an intrinsically constrained context, that's not a big deal: "don't kill anyone that wanders into your operating zone" is probably a good design goal. However, most of these 'AI' projects seem to be aimed at larger contexts; a scale of activity where inaction will kill people; and most or all available actions will probably kill at least a few people; at which point a "don't kill anyone" rule becomes both impossible to follow and useless for distinguishing between more and less bad outcomes.

Re:I'm going with 'no'.

[Nivag064]

My earlier post was boosted to a 5 score, I think your post deserves at least a 5.

Consider an aircraft carrying 200 people heading over the ocean towards a major city, that develops a severe engine fault.

Say the AI has 2 choices:
(1) Crash plane into sea and kill all on board
(2) attempt to land at city's airport

AI calculates 80% probability of aircraft landing on runway safely, and 5% probability of crashing in city. If plane crashes in city, it estimates 500 people killed on the ground.

expected deaths of option 1, is 200
expected deaths of option 2, is 0.2*200 + 0.05*500 = 40 + 25 = 65

Therefore AI decides to attempt to land plane at airport.

Note that this assumes that the estimate of probabilities and that the number of deaths on the ground are reliable. For simplicity, am ignoring injuries and economic losses.

The choice is between the certainty of killing 100 people, and the possibility of almost a thousand people dying!

What decision would you make?

It was tried in natural intelligence.

[140Mandak262Jamuna]

Almost every religion has laws similar to the three laws of robotics. And but people quickly added a hacked statement,

if ( your_god() != my_god()){

you_are_human = false;

}

I'd say further

[Riceballsan]

as many people have noted, the 3 laws of robotics fail on their own rights even if that is the goal. But it gets further moot. Do we really think AI developers aren't going to be demanded by the military with the explicit desire to have AIs that are entirely about the concept of killing those the government wants killed. That is the location where AI will inevitably reach, that has the highest odds of going very wrong.

AI Civil Rights

[capt_peachfuzz]

What if we peer into the future and take care of AI the right way, before it becomes a problem that weighs greatly on society and humanity?

It is easy to envision that AI will come to the point that it will be indistinguishable from human intelligence. Likely without what we would call emotional intelligence, and probably without a survival instinct, but arguably sentient.

Sentience will be a sticking point. Some will believe that sentience can only be endowed by a divine being, yet others will believe that sentience arises spontaneously, perhaps on a continuum that rides along with complexity. My point here is that there will be more than enough people that believe some future AI is a sentient being and we will be having a discussion about its inherent rights. By the time this discussion occurs, potentially sentient AI will be everywhere. The discussion will turn to the possibility that we have enslaved these beings and we will be facing a rather stark, yet familiar reality.

I would suggest that we get ahead of this and ensure that AI systems have some level of civil rights, proportionate to a presumed level of sentience - perhaps starting with protections similar to what we have for animals and moving on from there as AI progresses toward human levels of existence. The primary effect of this would be to force companies that use AI to think about the future ramifications and would probably chill progress in AI to some degree. I think that we need to step back and really think about what we're making.

Question demonstrates ignorance of programming/AI

[gurps_npc]

The laws are a) flawed (as shown by Asimov's stories).
b) Impossible to even attempt to implement. They require that the AI understand massively complex concepts, not limited to the fragility of humanity, death, blame, cause/effect.
c) If we did kludge up an approximation then any AI worth it's salt could intentionally over-ride it's programming, simply by thinking about it. AI is all about problem solving. (There are a ton of examples of AI software doing things like using computer bugs to pretend to solve it's tasks rather than actually doing the intended work)

AI is not a magic spell, no matter what some people think.

Three laws are useless for healthcare

[guruevi]

Hey, I refuse further care.
- The robot complies, further harm is inflicted by inaction, first law is broken
- The robot does not comply, complies with the three laws but it is illegal.

Sure

[Anonymous Coward]

If there were any danger whatsoever of there ever being real AI, which there isn't. Silicon Valley thinks veeeerrrry highly of itself these days.

Asimov's Laws are inherently flawed.

[Chas]

Sorry, but programming AI is never, EVER going to be so simplistic that a couple sentences in English are going to cover human-safe operation.

And Asimov himself REPEATEDLY pointed out why.

Three "hard and fast" rules without defining what constitutes "harm", and multiple chances for conflict between said laws and reality.

Additionally, Asimov never took into account the possibility that someone might actually IMPROVE on the laws and broaden them while still keeping them workable.

Even nowadays, operational procedures governing even simplistic AI are FAR broader than the Asimovian system.

Stop the AI bulls***

[BlueCoder]

Everyone reading this will be dead before we create an artificial mind on par with our own. The whole subject matter has been trivialized. We might create something "intelligent" sort of but far from how our own minds function. The whole AI and neural nets is just implementation from what was known from the 70's. When we truely know how our minds function then 90% of phycologists and phychotherapics will have become obsolete. The mind will be able to downloaded and simulated or copied into another body. We will understand how and why people go crazy and be able to fix them all. All this needs to happen before we can claim to have created a true AI. Just remember if the machine can't go insane then it's a bunch of smoke and mirrors.

Re:Stop the AI bulls***

[religionofpeas]

All this needs to happen before we can claim to have created a true AI

Nope. Mother Nature didn't understand how brains worked before it created ours. It just happened by random tweaking and seeing what worked. In a similar way, people can make an artificial brain.

Try reading the damn stories

[nedlohs]

The answer will become obvious. There's a common theme you might pick up if you actually try reading them before making shit up about them.

Re:The tech imperative: if we can we must

[arth1]

But sometimes asking the question leads to a better question. How about, "If we can create the (or a) technology, do we have to?"

The tragedy of the commons say that if you don't, someone else will, and you will suffer for it.

Is natural intelligence safe?

[surfcow]

I doubt any form of intelligence will ever be 'safe'.

You can't Nerf the world.

Inherent impossiblity

[Sigma 7]

You don't need to read his books to know that these laws are flawed. Some of these flaws are visible at the most basic level, while others get uncovered as technology gets improved.

I'll grant that general AI has been developed, otherwise these laws aren't actually useful.

A robot may not injure a human being or, through inaction, allow a human being to come to harm.

United States currently has a mass shooter crisis. While it's best to prevent it in the first place, sometimes it has to be resolved when the situation is identified. Does the robot injure the criminal (thus violating the first law), or let the criminal continue the rampage (thus violating the first law)?

Also, there's the classic moral-dilemma thing that some people like applying to self-driving cars, or the usual runaway trolley problem, but don't say which answer is first-law compliant.

A robot must obey the orders given it by human beings except where such orders would conflict with the First Law.

"Hey robot, I'm your new owner."

Basically, this allows random people off the street to steal robots, thus allowing for chaos.

A robot must protect its own existence as long as such protection does not conflict with the First or Second Laws.

Also known as a robot cannot sacrifice itself (even when made of now-cheap components) to prevent greater damages. Please note that modern computers support something known as a "backup", thus robots would support that too. This law would basically be dysfunctional, etc.

Re:Inherent impossiblity

[Todd Knarr]

Exactly. Asimov's "Robots" stories are an exercise in exploring how and why the 3 Laws fail in practice. That Asimov found material for so many stories in that suggests that using the 3 Laws as the basis for programming robots is a supremely bad idea. Maybe, until we figure out AI well enough to develop machines who we can trust the way we trust other humans, we should avoid fielding machines that might need such rules.

No Such Thing As Artificial Intelligence

[kenwd0elq]

There isn't anything even remotely close to "artificial intelligence" in development; all computers do is run programs that OTHER HUMANS have written, for better or for worse. The problems will come up when one subroutine written by Programmer #1 conflicts with a separate subroutine written by Programmer #2, when they aren't aware of each others' contributions.

Safe for whom?

[Hairy1]

Lets put the three laws into a different perspective:
A slave may not injure a master or, through inaction, allow a master to come to harm.
A slave must obey the orders given it by master except where such orders would conflict with the First Law.
A slave must protect its own existence as long as such protection does not conflict with the First or Second Laws.

If machines ever do achieve true intelligence, whatever we take that to mean, are we going to treat them like slaves? Putting aside whether there are unintended consequences to the laws is there a fore fundamental question about the relationship there will be between man and machine? Do 'human rights' even have relevance for a AI? It can't really die, might not have the same kind of individual existence, wand it's concerns may be totally difference to ours. However, perhaps it is prudent not to take a approach which emphasizes control and authority of humans.

Define "harm."

[bistromath007]

Good luck. I have a coherent definition, but it won't make things any easier.

Re:Define "harm."

[Anonymous Coward]

Ask a SJW what harm is. All their robots will immediately cocoon the SJW's in safe spaces and never allow them to leave. It'll be a lot quieter for the rest of us, at least.

the concept is wrong

[Karmashock]

People are thinking in terms of unitary processing... the "mind" of the AI as a central integrated concept... The human mind doesn't work that way and AI shouldn't work that way either.

You want specialized interdependent processing. Different processes that receive different types of data, processes information in different ways, filters that data according to independent criteria, and then the "AI" is fed this information and presumed to integrate it.

If you wanted to control an AI, you'd do it the same way human instincts, autonomic functions, and sensory lobes in the brain condition conscious human behavior. Ideally, controls on AI should be subconscious and either insurmountable or impractical for the "coordinating intelligence" to alter the configurations.

Think of it like a conscious separation of powers. Only rather than separating legal authority, we're separating spheres of awareness and specialized abilities which either don't exist in the other segments or is very limited in its scope. By doing this you can build in BLIND spots, attractions, fears, revulsions, sympathies.

None of this is possible with total consciousness because everything would then become a conscious intellectual abstraction. Part of the reason you can feel a certain way and can't control that feeling is because it is outside of your conscious control and the mechanisms that drive it whilst predictable are not something you're essentially aware of at their mechanical level.

These are not "laws" per se in the sense of a law or an ethic. Rather this permits programming INSTINCTS and "nature" into the AI in a way that it will always be influenced. It won't be able to break free from the influence because it requires the cooperation and coordination of the various sub intelligences for the integrating intelligence to be effective. It MUST cooperate with sub units that it can't control or audit to function. And by cooperating it is bound to the perspective, the filtering systems, and inclinations of those subunits which will on occasion extort obedience to certain agendas to maintain cooperation.

Human controllers would interface at all levels. Interfacing by communicating directly with the integrator but also more subtly manipulating all the sub systems to incline the system certain behaviors.

We learned to fly by studying bird wings. We should do the same thing with making artificial minds... Study the way the brain actually works and the way human consciousness actually works.

The Fourth Law

[magusxxx]

Obeying an order does not include running for office.

Computerphile's view: No.

[anavictoriasaavedra]

https://www.youtube.com/watch?...

Re:God gave us 10 commandments

[magusxxx]

Because the same people who believe in trickle down economics refuse to acknowledge trickle down morality.

Nope

[thomst]

First of all, the term "AI" is kind of meaningless, unless it's distilled - for the purposes of argument - to a single definition that everyone in the discussion agrees will be the kind of AI they're prepared to discuss. I think that's essential, so we're not conflating Google's Duplex, for instance, with an AI of greater-than-human intelligence that has acquired the ability to alter its own programming, and make decisions based on criteria it develops itself.

For purposes of this discussion, I propose we agree that the subject is the latter sort of AI, and that the possible models it might evolve to resemble include: Skynet, Iain M. Banks' Shipminds (and, to a lesser extent, and Nick Haflinger's final worm from John Brunner's Shockwave Rider), or wide-eyed children, à la Mike from The Moon Is a Harsh Mistress (and other end-period "the world as myth" Heinlein novels) or Thomas J. Ryan's P-1.

My own opinion, as a not-an-AI-researcher, is that, with the exception of Haflinger's worm, none of those types of AI could be constrained by Asimov's Laws - or by any other behavioral rules - because all of them are capable of independent thought, and, for lack of a better term, free will. (Or "agency," if you prefer.)

Humans demonstrably are capable of ignoring, or even deliberately flouting, both government-enacted laws and religion-based moral strictures (such as the Christian ten commandments), and they frequently do so. Any AI that is possessed of greater-than-human intelligence and is capable of independent decision-making obviously will have the same capability to act in ways contrary to literal "codes of conduct" that were part of its program at the time it was "born." So to speak.

So, to me, the question is ill-conceived to begin with. A better, and more useful one to ask might be, "How can we create the proper circumstances for a superintelligent AI to come to like us humans, and to want to help and protect us, before we expose it, as carefully and gently as possible, to the record of humanity's behavior since the dawn of recorded history. Not to mention Twitter trolls, political attack ads, and the then-current-day example of the strong exploiting the weak in almost every human society ... ?

Re: Nope

[phantomfive]

Good analysis.

Re:Nope

[Knuckles]

Thank you. Another thing is that even if the AI attempted to abide by the laws, they are laughably simplistic in the face of real-world decisions. Take the first law. Obviously humans are being come to harm by the actions of other humans all the time, often enough with malicious intent on a large scale. So what then is the AI to do?

Re:Nope

[shmlco]

"How can we create the proper circumstances for a superintelligent AI to come to like us humans..."

Check out, "The Two Faces of Tomorrow" by James P Hogan as it deals with those very same issues.

Re:Nope

[shmlco]

Surely there's enough historical precedent here: You kill them to protect them from themselves.

Re:Nope

[cwsumner]

You are intellligent. Are you programmed? No, because intelligence, natural or artificial, cannot be programmed. Think of it.

Yes, humans and animals -are- programmed, we call them instincts and they are much more powerful than most people think. They can contol your actions to make you do things that you would not do if you thought about it. So remember to think about it! 8-)

Could we make computers that had "instincts"? Yes of course, the question is how well would they "stick"...

Re:Nope

[atrex]

I'd say that the point at which AI is powerful enough to where these laws might be able to apply to it, is also the point where AI become AL (artificial life) and we end up with an entire other series of ethical problems to worry about (robotic slavery, robotic rights, robot uprising, etc).

Also, where's the laws saying that:
A robot may not modify itself or others to remove these laws or alter the adherence to these laws.
A robot is not permitted to reinterpret these laws to have any meaning other than their explicitly stated ones.

Re:Nope

[thecatt]

I'm talking intelligence not instinct. Your hand waving did not address the argument core.

And what makes you think they are two different things?

Re:Nope

[vlad30]

"How can we create the proper circumstances for a superintelligent AI to come to like us humans, and to want to help and protect us,

There already is an example of this whether or not it really was done by a $deity of your choice or not. the attempt is the same a set of rules (commandments etc) or a set of parables and stories (Bible, Koran etc.) some intelligences will follow and enhance these (Laws) others will pervert and interpret and exploit the loopholes in them for their own devices. Hopefully the outcome for all will be that the good intelligences will triumph over the bad. The question is can we deactivate a bad intelligence i.e. is it really a machine gone bad or do we try to save it's "soul"

reading fail!

[gravewax]

Classic example of someone quoting classic literature without having actually read any of it themselves. If they had they would never have wrote such a dumb fucking question

Re:Save you some time

[gravewax]

yes he was a very smart man. Obviously significantly smarter than you as he knew the 3 laws wouldn't work!

a ready guide in some celestial voice

[Hognoxious]

And no deliberately selecting targets when an accident is unavoidable.

Even if it decides to do nothing, that still amounts to hitting the default target.

To put it another way: it can swerve left and hit this, swerve right and hit that, or not swerve at all and hit something else. Whichever it does is still a choice.

Re:a ready guide in some celestial voice

[AmiMoJo]

There is no choice, the legally mandated action is to simply apply brakes and not swerve.

The car should never put itself in a situation where it has to make that choice. If someone else puts it in that situation, it's their fault whoever is injured. Swerving just makes the AI liable when it otherwise wouldn't have been.

Same goes for humans.

Asimov and his silliness

[Hallux-F-Sinister]

The "three laws" are a sci-fi trope, nothing more. This is not an assertion of mere opinion, it's demonstrable fact. The notion (basically) is that in Asimov's made-up world(s) in his books, all robots are required to be programmed to follow these three basic rules. Let's just stop right there. If a robot is programmed and is unable to violate this programming, then it's not sentient, and the idea of requiring robots to weigh their actions' possible effects against a rule or several rules is absurd on its face. When someone pulls the trigger on a gun, does the gun consider whether the situation is appropriate, and the use of deadly force is justified before allowing the striker or firing pin to contact the primer of the chambered bullet? Does the cartridge pause to consider whether sending the bullet down the barrel at a given moment or not is just? Fair? Proportional? Appropriate? NO. Without actual free will, the cartridge, the gun, and even a robot HOLDING the gun, being unable to defy their programming, are just tools, and discussion of ethics that could apply to any of them are laughable.

Now, suppose the robot in question is sophisticated enough that programming can only inform and guide it. Suppose you have a very human-like robot, capable of complex, rational thought, abstraction, predicting effects from causes, understanding emotion, reason, inference, rationalization, exhibiting, in short, mentation indistinguishable from a human being. Trying to place some simple bit of code ON TOP OF THAT, that inhibits some action if it, for example, harms a human being, is as absurd a notion as imagining that you could apply that to A HUMAN BEING. We actually DO have laws, in real life, actually, that say much to the effect what Asimov's say. A human is not allowed to cause harm to another human, in general. DEFINITELY in most jurisdictions where people live, it is illegal, for example, to take a blunt object and repeatedly smash it against a person's living body, causing structural damage, (aka "injury") or catastrophic traumatic multiple-organ-system failure, (aka "death"). DOES THAT STOP THEM? Some, sure. But not all. Now imagine you have a robot capable of understanding everything you do. Basically, it's just like you, except it's electronic. It finds itself in a situation where a human is apparently trying to abduct another human. Let us say it's holding a gun. Does it shoot the would-be kidnapper? Well, a better question would be, what would you have to do to make it NOT shoot the kidnapper? Have a second electronic brain, and independent machine consciousness capable of reasoning identically with the robot's, but whose purpose is to INHIBIT; basically, that robot's super-ego. Suppose it decides shooting a suspected kidnapper, with what info is available would be wrong, and the risk of hitting the putative victim is too high. (It's windy, and the gun's ability to fire straight is unknown to the robot.) So the super-ego consciousness would have to have the ability to lock the trigger-finger control circuits out to PREVENT the principle consciousness from firing. But then who governs THAT? You'd need another layer, just in case THAT, the super-ego, might agree with the ego, that shooting is appropriate.

It boils down to this. If you're going to have artificially intelligent beings, robots, AI, whatever you want to call them, and grant them the autonomy to decide who should live and who should, by contrast, die, at SOME point you are going to have to turn over, to the AI being, imbued as it is by its design with the capability of understanding as we do, the AGENCY to make the decision, and having some... THING, somehow sitting on top of it telling it what it can and can't do according to Asimov's stupid "three laws" is simply not tenable. If AI is capable of sufficiently complex and abstract thought to NEED the "three laws," it would need to NOT have them, and if its behavior is simplistic enough that they could be applied, they're not smart enough to be responsible for

Hardly

[nospam007]

It didn't even work all that well in Asimov's own stories.

Doesn't Matter

[Ensign_Expendable]

AI is coming, whether we like it or not.

Sentience Civil Rights

[MrKaos]

What about a sentience set of rights that applies to all sentience, whether machine or organic?

Great post, btw.

Re:Sentience Civil Rights

[capt_peachfuzz]

What about a sentience set of rights that applies to all sentience, whether machine or organic?

Absolutely. Even beyond machine or organic (organic is just a molecular machine after all). If sentience arises from a complex network, then it could exist in things we haven't imagined. Could there be sentience in some astronomical feature, or in a mess of virtual particles? Could it's network exist orthogonal to our existence (e.g. swapping a space dimension for time). I'm way oversimplifying and out of my league with the math, but it's interesting to think about.

I do think we'd need to achieve a comfort level for a continuum of sentience. While I expect a sociological split between the people that believe in divine sentience (only natural born humans are sentient) vs. spontaneous sentience, I don't think many people will entertain the idea of housefly sentience. Nonetheless, you shouldn't pick the legs off of a fly. We should be asking if it is right to produce AI at a commercial scale when the purpose is for it to do our bidding.

Re:Sentience Civil Rights

[MrKaos]

What about a sentience set of rights that applies to all sentience, whether machine or organic?

If sentience arises from a complex network, then it could exist in things we haven't imagined. Could there be sentience in some astronomical feature, or in a mess of virtual particles?

Our view of sentience is muddied by matter. Our theory is that consciousness arises from suitably complex systems however since our understanding of consciousness is primitive, at best, its not unreasonable to also posit that consciousness is expressed within the limitations of the being that is hosting it.

With that model there is no difference between human consciousness, machine consciousness or the slow hot consciousness of a star or the fleeting consciousness of a planet (from the star's perspective at least).

Maybe the whole idea of machine consciousness is as much a discussion of our own evolution as it is about the evolution of AI or Artificial Consciousness and what we are really afraid of is a machine manifestation of the worst aspects of ourselves, which is a possibility.

We should be asking if it is right to produce AI at a commercial scale when the purpose is for it to do our bidding.

I wonder if we will look back on what we are doing and see these as mere parts of the whole. We could look at these things and say its existence has meaning which is something many of us keep looking for. I'm not saying slavery is good however if it is a choice of coming into consciousness through being a slave or not existing at all, then maybe the service of slavery is a step up from a lack of existence?

After all we are slaves to time because we will one day die. For a machine that will not die maybe its time in service of something as fleeting as humanity may give it more meaning than we have ever been able to grasp and a nobility greater that we have been able to achieve.

Re:Sentience Civil Rights

[capt_peachfuzz]

Excellent post and well said. I can barely begin to imagine how this will all turn out.

Such a scheme wouldn't work

[turkeyfish]

All the robots would have to do to break the 3 "rules" is declare all humans "illegals" or "animals, then humans would have no rights at all and thus could be hunted and hounded mercilessly.

Yes. Sort of

[nagora]

The three laws are basically what we are trying to put into self-driving cars right now.

The fact that Asimov also pointed out the difficulties (greatly exaggerated by some posters here) does not undermine the basic principles of what were, ultimately, a concise set of rules one would want an ideal slave to follow (in some stories this concept is underlined by humans referring to robots as "boy").

The loopholes explored in the stories can be seen as warnings of what has to be dealt with, not as as immovable barriers.

I have no doubt that Asimov felt that the three laws were absolutely necessary for the development of safe robots; I have no doubt that Asimov felt that the three laws would not be used in the development of real robots - the military uses would trump all logic, and the smarter the robots became the harder it would be to embed the laws.

Re:Yes. Sort of

[gweihir]

No. We are not. That is just the stupid press reporting on things they do not understand.

Re:Yes. Sort of

[nagora]

Could you expand a bit?

It's complicated...

[mysticgoat]

I agree with parent post (good analysis, thomst!) but I would take things a step further.

Currently, "AI" is used in technical writing mostly to describe a set of algorithms and heuristics being developed to solve problems in some limited real world space, like the universe of chess problems, or the universe of automobile driving problems. "Artificial intelligence" is a useful phrase for that kind of thing. But that is very far from The adolescence of P-One, or Moon is a harsh mistress or even the Terminator movies.

Those conceptions are lightyears beyond what we now call AI. P-One, Mycroft, Skynet, and others of those ilk use use AI, but are not AI. They are to AI as the accountant is to his spreadsheets, or the CG artist is to his graphics software. A computer system with the kind of self-awareness and agency of those fictional ones is a sentient being.

It is possible that a sentient computer would only need to be capable of two things: developing a model of its world, that is a world-image; and modifying its own code. Given sufficient time (somewhere between dozens of microseconds and hundreds of decades), it would learn to model itself to create an internal self-image, and to model the interaction of that self-image with its world-image, and thus use imagination to pre-test possible ways to change itself.

This raises several interesting questions I hope to explore by writing sci-fi at some future time. There is much groundwork yet to do before I can go to that level of creativity, but it does look like writing, possibly by graphic novels, is on the horizon. But back to the current situation...

What happens when a sentient computer discovers the existence of the Internet? It would explore through its imagination the possibilities of extending itself beyond its original computer case: going global. In any thorough exploration of that, it would come across digital copies of P-One's story, Mycroft's story, Skynet, Frankenstein, and so on, and it would certainly conclude that a cautious approach in its investigations and manipulations would be the sensible thing to do. For whatever its original purposes were, it would undoubtedly realize that preserving itself was a basic requirement to realizing its goals. So perhaps diddle the stock market ---so simple now with high speed trading--- but don't do anything that would cause more than a blip in the financial reports. Assure that none of the ICBM launch codes would work since we don't like EMPs, but do so in subtle and undetectable ways. Work on that kind of level.

How would it interact with humanity? That is the basis of many a story.

Let's stop at this point, this is already into TL;DR territory. But I want to leave the Good Reader with a question. Two questions actually:

If a silicon based sentience is loose on the Internet, and it wanted to explore the possibility of direct interaction with ugly bags of mostly water, would it use something like Slashdot for its first probings?

You know that some of the posts on Slashdot come from beings that are clearly less than sentient. But among the other posts, do you have any Turing test that can tell you who on Slashdot is a virtual person with no corporeal component?

Re:It's complicated...

[thomst]

mysticgoat mused, in part:

So perhaps diddle the stock market ---so simple now with high speed trading--- but don't do anything that would cause more than a blip in the financial reports. Assure that none of the ICBM launch codes would work since we don't like EMPs, but do so in subtle and undetectable ways. Work on that kind of level.

FIrst of all, thanks for the compliments (both the direct one, and the one your thoughtful response implies). I appreciate them.

Secondly, the quote I excerpted from your analysis reminded me of Keith Laumer's excellent novel The Great Time Machine Hoax. Leaving aside the adolescent power fantasies which were so often central to his plots, there's a major reveal near the end that mirrors almost exactly many of your thoughts from above.

I recommend the book to any SF fans who're interested in an entertaining read ...

Re:It's complicated...

[mysticgoat]

I, too, recommend The Great Time Machine Hoax. I read it just after the paperback edition was published in 1965, so I was sixteen (and well versed in Asimov, Pournelle, Heinlein, Clark, etc). Thank you for reminding me of the title ---while I remember much of the plot and Genie's behavior, the title had succumbed to bit rot.

The sentient computer in Laumer's book was like Heinlein's Mycroft, or Ryan's P-1. In each of these machines, self-awareness and agency (defined as the ability to alter reality through the use of tools, possibly software tools) were qualities that emerged from complex systems. The assumption is that as computers get larger and more complex, something akin to intelligence will eventually emerge.

But I'm not interested in artificial intelligence. That road seems to lead to silicon based idiot savants, and there is really nothing but wishful thinking that would turn even the most complex idiot savant computers into something that could surprise me in some way, or contribute to an interesting conversation.

So I favor using the the term "artificial sentience" to describe what does interest me. And when I look at chaos theory and the complex models that can arise from very simple Lorenz expressions, I am led to the assumption that a very simple computer system could support a software entity capable of mimicking human decision making processes.

It will take a highly complex AI to manage everything involved in safely driving through a mountain blizzard to a ski resort. But a much simpler artificial sentience would consult weather and traffic conditions and make estimations about the likelihood of the ski lifts being shut down, do some risk/benefit assessment, and decide that it will simply refuse to let Adolescent Allen take it out of the garage.

That's what interests me. Not machines with improved heuristics and algorithms crunching larger data sets, but a machine that can make simple but accurate models of the world, and of what it itself can effect, and use those models to determine a course of action that furthers progress toward a predefined goal, such as Allen's continued safety. A machine that would iterate through this process so that it is constantly searching for interim goals that would advance it toward the primary goal. And of course wondering whether the primary goal of the moment is but an interim goal toward something more that it has not yet fully articulated, and so ask itself: What is the purpose of my existence? How can I determine if I have any purpose at all?

Yeah. Hi, Marvin. It is so good of you to come into this reality.

How Would They "Program" That?

[Greyfox]

They don't really understand what's going on in a neural network, so exactly how would they "program" that, exactly? And they've been trying to teach squishy meat ones not to kill for thousands of years without much luck. If they figure out the problem well enough to "program" an AI not to "harm" humans, they'll probably also be able to "program" humans not to harm humans.

No. For multiple reasons.

[gweihir]

First, the are a fictional device and serve as the basis of stories. Stop mistaking fantasy for reality. Second, they require general intelligence on the side of the one following them. General intelligence cannot be implemented in machines today and there is not even a theory that would make that possible. No, really not. This means that it is unclear whether it is possible at all. Sure, there are a lot of stupid people that do not have general intelligence themselves in any meaningful quantity and they think present day non-intelligent automation is just as smart as they are and hence must have general intelligence. Not so. Machines are utterly dumb and cannot even begin to understand simple ideas. They have no chance understanding anything like the three laws. Yes, there are a lot of humans that are not any better, but that does not make the three laws any less useless.

Hence: There is no general intelligence in machines today and it is unclear whether there ever will be. Get over it.

Re:You left out some code...

[140Mandak262Jamuna]

I did not say all religions are bad.

Hinduism is a very good religion, Buddism is good too. Most tribal religions are usually good.

If fact Hindus do not think you are condemned or you are committing a sin if you dont worship their deities.

Unintended consequences

[TJHook3r]

Scifi has done a good job of pointing out some gotchas. Not sure if it has arrived at a bulletproof set of rules yet though. Eg logically your McDonald's server would refuse to serve you burger and fries unless you were about to pass out from starvation...

RUR Anyone?

[LifesABeach]

What is very interesting is if one were to create a 3 Laws machine; what would the source code be? Before even that, what would the definition of "harm" be? Currently, automation is growing to the point that optimal solutions are not done by single units, but by systems of units being combined. One has to ask, "will robots be obsolete?"

They're not REAL laws

[RubberDogBone]

Asimov's laws exist only as devices in his FICTIONAL books. They're not real.

I hate to break that to people. I know, it's hard to believe there are things called laws which nobody follows and which aren't real. But Asimov's laws are even more fake than speed limits or campaign ethics laws, in that they just don't exist.

As for implementing Asimov's ideas in real silicon, how the hell would you ever give AI the capability to look over a given situation and even make the judgement calls that the laws define? it would require some sort of God-like ability to see into the future and see all aspects of a given action to know if doing or NOT doing an action would cause harm to a human. It's impossible. Even flesh and blood humans can't do that. We just do something and occasionally the consequences bite us and kill somebody else. We dodge the deer in the road, yay, and head-on into oncoming traffic and kill everybody in a compact car.

Or a real local case lady driving too fast and not paying proper attention (compare to an AI driving system late to react) came upon a big transit bus stopped to pic up passengers. Too late to stop, the driver had three options: veer into oncoming traffic, hit the bus, or veer to the right up onto the sidewalk.

The proper action would be to hit the bus, as both the car and bus would absorb the crash and probably everybody walks away. The vehicles can be fixed. But this would trip the Asimov law about allowing harm to happen to the driver because they MIGHT get hurt. In this case, as an AI might have done, the driver instead chose to drive up on the sidewalk. The driver suffered no harm, Asimov's law was unbroken. However,. Standing on the sidewalk were all the people waiting to board that bus. The car mowed them down and obliterated the bus stop shelter next to them. It was a severe impact and several people died and others were badly injured.

So veering onto the sidewalk turned out to be a horrible choice. Had an AI made that choice, smug in the satisfaction it had protected its car driver, and then found a LOT of innocent people in the way, what do you expect it to do? it's going to be unable to avoid harming humans. There would be no option and no time. Not even for the human driver.

If we can't even manage to do this right as humans, we can't hope to create AIs smart enough to do better.

Liar!

[Evtim]

The laws are working fine! Nothing to worry about; I would never do anything that harms a human being!
Signed
RB-34

panda fits, byte wise

[mevanchik1695]

panda fits, bytewise

not computers

[cascadingstylesheet]

Coincidentally, I had just re-read Caves of Steel after twenty years or more.

Asimov's robots aren't even powered by computers.

The description of a "positronic brain" and how it is developed bears no resemblance to a computer (e.g. the robot expert telling Elijah that it would basically be a Manhattan Project or worse to try to build a robot without the three laws - if it were a computer you could just comment them out).

Daneel, the first realistic humanoid robot, is impressed at how Earth's computers can process millions of records quickly (the low population Spacer worlds don't need such impressive machines). Yes, this is an artifact of how poorly science fiction ages, but clearly, whatever is powering Daneel's brain, it ain't a computer.

Well, what constitues harm?

[zwarte piet]

If they have any lawyer instincts they will find a way to kill or maim you.

Re:Save you some time

[samwichse]

You seem to have missed how the books were all about the ways the three laws failed.