The Turing-Asimov Dilemma

Question

The Turing Test is a simple test devised by Alan Turing to check for AI. If a machine is able to fool a human into believing it's a human then that machine is AI.

The 3 Laws of Robotics were laid down by sci-fi author Isaac Asimov.

As per The 3 Laws of Robotics an AI (robot) must be able to distinguish human from robot(AI) otherwise it wouldn't be able to follow these laws.

But AI perfectly mimics humans.

The dilemma

1. Either AI can identify humans from AI or AI can't identify humans from AI.

2. If AI can identify humans from AI then AI is more intelligent than humans (AI cognitive capacity exceeds that of humans)

3. If AI can't identify humans from AI then AI can't follow the 3 laws of robotics

Ergo,

4. Either AI is more intelligent than humans or AI can't follow the 3 laws of robotics. [1, 2, 3 constructive dilemma]

Either way we have a problem on our hands (AI is superior or safety issues with AI).

What sayest thou?

Answer 1

The intersection of AI research and deontic logic is nonempty. One preliminary example of the framework for such an intersection is Wieringa and Meyer's "Applications of Deontic Logic in Computer Science: A Concise Overview". Part of the abstract for this paper reads:

Many applications move in the direction of programming a computer in deontic logic to make the computer prohibit, permit or obligate people to do something. We discuss conditions under which this possibility is realistic and conditions under which it would be admissible to do so.

The essay "Toward ethical robots via mechanized deontic logic" more directly pertains to the OP question. "Automated Reasoning in Deontic Logic" has an "AI research" tag on arxiv.org. "Exploratory moral code : formalizing normative decisions using non-modal deontic logic and tiered utility" reflects on how the process of devising a deontic logic for AI can help clarify, or even testify on behalf of, various moral theories.

Regarding the argument of the OP specifically, my counterargument would be that if the AI had been programmed to not harm us, it could be programmed to not harm other AI, too, and in fact would hopefully admit of being programmable to avoid harming any morally salient entities. This, then, even if the AI were unable, in some case, to tell a human apart from a cleverly disguised robotic mule, say.

Answer 2

The problem with this argument is that the three laws are a plot device and not laws at all. Not only has no robot been programmed with them, no robot could be.

You can lay down a general rule that you get an average by dividing by the number of instances and then get a hard error because you neglected to special case zero instances, because numbers have been defined for it, but you can not tell a robot to not harm humans and then be surprised by what it considers harmful or not. You have to define harm for it, and the only way is enumeration in our current languages, or even purely theoretical languages.

Answer 3

There is no dilemma.

Turing Test requires the subject to sit in another room. The subject is not available for observation, only his/her/it's answers are.

That limitation do not exist for Asimov's robots. They can examine the humans directly to know they are humans, and robots to know they are robots. A quick look at face suffices.

Asimov's robots are not cyborgs or in anyway mimic humanity in body appearance. They are made of different material.

Now if you are thinking about robots made of flesh then you are going against definition of robots in all fiction.

If Asimov's robots are forced to analyze a subject on basis of subject's answers only, and they must use Turing Test, and the subject appears to be human in test results then I think the Asimov's robots will be confused. Their circuits will keep on deciding between two equally valid conclusions, getting in a feedback loop and after that just freeze if there is a fuse to burn to stop damage to circuit, or have their circuits burn.

Given equally valid outcomes and forced to choose one there is either a freeze or a burn. Its because in absence of emotions there is nothing superior to logic sitting there solving conflicts.

Answer 4

Your reasoning is faulty. There is no dilemma.

First and foremost, because your understanding of Asimov's laws is flawed. These are the imperatives which govern robot behavior in Asimov's fiction. They are not laws of nature. (Technically, they are incredibly simplified versions of how robots' positronic brains are programmed, but lets assume they are accurate in their simplified form).

As is explored in many of his stories starting with Caves of Steel, most robots in fact can not distinguish between humans and convincing replicas. This does not cause any particular conflicts for commanded robots: their perception is that R. Daniel Olivaw is a human, so they treat him as a human (sometimes). There is no paradox here, the robots are simply mistaken.

This also does not cause any particular safety issues. You have some weird edge cases where a robot might sacrifice itself to save another robot, but that is (mostly) an efficiency issue. You can set up a contrived trolley problem where some convincing AIs are on one track and a robot is manning the switch, but all such problems are described as inevitably driving the participating robots catatonically insane, and the AIs would refuse to participate. So, unsafe only in the sense that you can expend enormous resources to do violence in overly elaborate ways that you might otherwise do cheaply and simply.

Secondly, your arguments are fallacious. "Either they can or they cannot" is a false dichotomy (This is hardly a clean all situations/no situations binary choice). Even assuming that robots are oracularly gifted at spotting humans, you then generalize from that one task to "AIs are superior". It is wildly unclear how you make that leap, what "superior" means in this context, or why this supposed superiority would be bad. (AIs can already do complex calculations better and faster than I can. So could an abacus. Is an abacus "superior" to me? Is this a problem?)

Finally, "AI can't follow the laws of robotics" is a non-starter. AI would have to be perfectly omniscient to follow the second half of the first law, and "harm" is poorly defined. If you want an exploration of how Asimov's robots interact with the world, read Asimov's fiction. He wrote about it more extensively than you would think possible. Start with I, Robot.

Answer 5

I would say that a core issue here is that "intelligence" is not well defined.

Is a system that does a single specific task as well as, or better than, a human "intelligent"? Or does the system need to exhibit human-like performance on multiple tasks at the same time be qualified "intelligent"? Does the system need to show something like creativity to be qualified "intelligent"? Conversely, if a human is very good at one task but very poor at another one, are they "intelligent"? Not "intelligent" when it comes to the task they can't do well?

To me, unless we have a good definition of "intelligence", this kind of discussion is very hand-wavy and not very meaningful.

As for current "AI" - very far from "intelligent". Take ChatGPT - I was testing it last week by feeding it various queries: it routinely makes gross mistakes when it writes code, invoked the thing I asked it to prove in a step of a mathematical proof, and generally revealed what it does which I'm not sure is "intelligent": regurgitate an average of what it has seen in its training set, with not a shred of "understanding" whether that average is accurate or not. It's just an impressionistic patchwork of more or less related tidbits that sometimes passes for reasonable, but usually falls apart when you probe deeper.

Sorry if that's not as thrilling as all the hype you can hear in the media.

Answer 6

A speculation about the Turing Test is that a machine that can pass itself off as a human in text-only communication is 'intelligent' (or 'conscious' or similar - no consensus has been established). This does not mean it is indistinguishable from a human in all circumstances - it doesn't even have to look like a human. And a being that cannot pass the test could still be intelligent. (Someone who cannot type because their arms are broken could fail the Turing Test - this is not good evidence that they lack sentience.)

On the other side, the laws of robotics only make sense if the robots aren't beings with the same feelings as humans - if they were, I think it would be immoral to program them to die rather than harm anyone in any way. Most of Asimov's fictional machines would fail the Turing Test - they couldn't lie if you ordered them to tell the truth, nor could they try to hurt anyone's feelings - but are still 'intelligent' in the sense they they can solve complex problems.

So we have two different hypothetical concepts here. Firstly, machines that can pass for human in conversation, and are therefore presumed to be self-aware, and perhaps deserving of human rights. The other is machines that have no desire other than to serve humans and (less importantly) continue to exist.

In the end, both are flawed concepts, and it doesn't matter if they're compatible or not. Modern GPT-based AIs can now pass the Turing Test to some extent, but most people still don't think that's a sign they're conscious beings, especially those who know a lot about how they work. The Three Laws are designed to make perfectly safe robots, but even if they had a perfect ability to judge harm and recognise humans, we probably wouldn't put those laws into them. Real-life applications for robots would be, for example, to replace soldiers on the battlefield, or to be someone's servant. Would the military want a soldier programmed with the First Law? Would you buy an expensive servant programmed with the Second Law, and therefore equally willing to obey any human order ("Come with me and forget your former master."), not just your own?

Answer 7

We do not have a problem; there is no dilemma and in fact, there is no Question, unless that Question be 'What sayest thou?' Oops…

Asimov was clearly dealing with robots, not androids; certainly not with pretty-much perfect androids that might be mistaken for human beings.

That difference invalidates both the Question and any premise that might have been built upon it.

Without that distinction, I’d agree with your logic but like it not, robots are not androids.

Answer 8

The Laws of Robotics essentially provide their own solution to the dilemma.

Asimov's robots do not mimic human intelligence perfectly. Humans have free will, they can do whatever they want. Robots are bound by the Laws of Robotics, so they can't always do what they want.

Thus, a simple way to make a robot fail the Turing Test is to ask it, "Are you a robot?" and command it to tell the truth. The 2nd Law requires it to obey this command, so unless it can come up with a way that this violates the 1st Law it must confess that it's a robot. Although one out would be that in the Turing Test scenario, the robot cannot tell that the questioner is a human -- both participants are in the dark about the other (the robot may believe that it's giving the Turing Test to the questioner).

As others have answered, the Laws of Robotics are not realistic, they're a plot contrivance that Asimov and John W. Campbell came up with for his stories, many of which have plots that revolve around the difficulty of applying the Laws (much as faster-than-light travel/communication is a fiction that facilitates many stories about space travel, interstellar communities, etc.). The Laws of Robotics are to robots as ethics are to people, and both of these are extremely vague and difficult to implement. Philosophers have spent millenia trying to codify human ethics, but they're still murky.

Answer 9

Your second assumption is wrong.

Ability to distinguish humans from AIs is not based on intelligence (although it might depend on definition of intelligence), but on pure processing power. Given enough (correct) robot and human input AI might be able to identify humans and robots regardless of its lack of understanding the idea of a robot and a human.

It would not even know why it declares something as human or artificial - it just appears a bit more correct after processing amount of data that is impossible for humans to process consciously.

So - machine doesn't have to be more intelligent than human (or intelligent at all) to do better at such tasks.

Answer 10

If a machine is able to fool a human into believing it's a human then that machine is AI.

Just convincing a human that the machine can pass as human is relatively trivial, depending on the human that's doing the test, and how the machine being tested presents itself. I've met plenty of humans that couldn't pass a Turing Test, and plenty of bots on Twitter and other social media platforms fool humans into thinking that they're talking to another human on a daily basis. Not to mention the number of people who don't realise how many articles on click-bait sites are actually written by article generators.

But that's only one specific and narrow window into what AI means. In general we use the term "artificial intelligence" to mean any system that is capable of decision making and/or problem solving in a particular area. Expert systems are a form of AI, neural networks trained to do specific things are a form of AI, etc.

What you're probably thinking of is AGI: Artificial General Intelligence. This entails an ability to generate solutions to novel problems, which thus far is beyond the ability of our machines. While AI researchers are theoretically working towards AGI, it's not currently a primary focus because it's not economically viable at this point.

Or perhaps you're really thinking about Artificial Sentience (machines that experience emotion), Artificial Sapience (machines that can rationalise about what they learn) or Artificial Consciousness (machines that think, feel, etc.) Or some other definition, since what those things actually mean is a hotly-debated topic in philosophical circles.

The 3 Laws of Robotics were laid down by sci-fi author Isaac Asimov.

...who spent most of his Robot stories pointing out how many problems those laws were subject to. One of the recurring characters was Dr Susan Calvin, Robopsychologist, whose main role was trying to figure out how to stop robots from going crazy or exhibiting unexpected behavior due to conflicts between the laws and reality.

4. Either AI is more intelligent than humans or AI can't follow the 3 laws of robotics.

No dilemma here, the 3 Laws of Robotics cannot be followed rigidly even by humans. Adding a 4th law (the Zeroth Law of Robotics) allowed R. Daneel Olivaw to resolve some of the conflicts, and other supplementary laws have been proposed by various people to help out, because the 3 Laws of Robotics are literally impossible to follow.

Answer 11

First of all the "laws of robotics" intrinsically don't work and Asimov himself basically provided a series of short stories called "I, Robot", where he introduces and breaks these very laws in more or less creative ways.

Second of all, afaik the Turing test has the simple premise that humans are intelligent and if machines can pass as humans they would thus also be intelligent.

There are several problems with that, namely that:

• you could fake intelligence by saying the right things without knowing what they mean ("chinese room argument")
• it might be more about the believes, skepticism and phantasy of the the observer than the intelligence of the machine.
• not everything that humans do is intelligent and that not everything that is intelligent is something that humans do. That computers might exceed human intelligence in some domains and lack in others.
• Also it's binary and not really measurable, scalable or otherwise usefully quantified.

Also it's about INTELLIGENCE not ROBOTS. Like a robot, but even more so a human, is a physical entity with a physical, chemical, biological, ..., signature that can be identified as such. "Intelligence" is much more complicated and makes the question "where are you?" and "what are you" much more difficult. That's why this text based interface is possible, because the physical form is not relevant for something to be or not be intelligent. I mean in physicalism intelligence would still need to have a material form in one way or another, but for example the concept of "the mind" is often much harder to localize and pin down than for example a leg.

So no telling machines and humans apart is easy, drawing a demarcation line for intelligence is a much harder task. But for all intents and purposes we are concerned with us fleshy meat sacks not intelligence, ... when it comes to the 3 laws of robotics.

• If AI can identify humans from AI then AI is more intelligent than humans (AI cognitive capacity exceeds that of humans)

As shown, that doesn't have to be the case. Just attach sensors that let it detect human or machine, no intelligence on the machines end.

• If AI can't identify humans from AI then AI can't follow the 3 laws of robotics

Yes. For most cases they'd just treat robots as humans and not harm them either, which would not be a problem. But if you want to bring them to their knees you'd subject them to the trolley problem where every possibly option would violate one or more of these laws, even suicide and contemplating too long.

• Either AI is more intelligent than humans or AI can't follow the 3 laws of robotics.

Even if they are more intelligent they could still be following some arbitrary rules. Like the creation of a limiting mechanism might require intelligence X humans have intelligence X+1 robots have intelligence X+2 and the removal of the mechanism requires X+3 so.