Does AI rely on determinism?

Question

I don’t believe in free will, but most people do. Although I’m not sure how an act of free will could even be described (let alone replicated), is libertarian freewill something that is considered for AI? Or is AI understood to be deterministic?

Answer 1

I'm going to assume that by free will, you mean something like the philosophical concept of libertarian free will, which is defended by philosophers like Robert Kane. In Libertarian Free Will, individuals have some capability to make choices about their actions. The classic way to argue this is by assuming some kind of spirit-stuff (e.g. a soul) that exists outside the material world, and that this spirit-stuff constitutes the consciousness of a person. Kane tries some mental gymnastics to avoid this, but then concedes something like it in a footnote. I'm not aware of any serious work that doesn't make some kind of non-physical assumption to justify this view. If someone can point at one, I'll update the answer.

By determinism, I'm going to assume you mean the usual notion of philosophical determinism: since people's decisions depend on what happened in the past, and where they are in the present, they don't really have a choice in any meaningful sense. Philosopher's like Dennett adopt a slightly softer view (compatibilism, essentially: you don't get to make big choices, but you do get to make small ones). Appeals to Quantum Mechanics are common to justify that view. In this context, free action means something more like "did something we couldn't predict exactly". An example might be: you are pre-destined to put a can of campbell's brand tomato soup in your shopping cart, but "make a choice" about exactly which of the dozens of cans you will put in. Since small choices can have large impacts (maybe that can will give you food poisoning, and the others wouldn't), this can make all sorts of things impossible to predict exactly.

I think most AI researchers don't worry too much about these issues, but Turing actually addresses them in his paper right at the start of the field, Computing Machinary and Intelligence. The deterministic/compatibilist view point is introduced as Lady Lovelace's objection: Computers only know how to do what we tell them to, so they can't be called intelligence.

Turing's counterargument uses two prongs. First, Turing notes that computers can probably be made to learn (he was right!). If they can learn, they can do things that we don't expect, just like children do. Second, Turing notes that computers already do all sorts of things we don't expect: they have bugs. Anytime a computer exhibits a bug, it did something that was unexpected. Since we cannot generally rule out bugs in programs, computers will always do surprising things. Therefore, computers satisfy the deterministic notion of free will.

Turing also addresses the libertarian notion of free will, which is part of what he calls the "Theological Objection". The objection is that intelligence requires some kind of divine spark (like free will). Turing argues that we can't detect sparks like this right now (he actually thought we would be able to one day, and spent a lot of time looking at supernatural phenomena too). However, there's no reason to suppose that computers with the right programs won't be endowed with them. A divine creator could decide that anytime you build something brain-like, it gets a spark. If we build a program that's brain-like, maybe it gets a spark too. In the absence of some way to detect souls, it seems like we ought to just agree to treat things that seem intelligent as though they had these souls, since otherwise we don't have a really clear way to decide who is and isn't intelligent. The only remaining way is to say "only things made of human meat have souls and are intelligent". While a lot of people do actually say things like this (e.g. animals have no souls), this is a pretty arbitrary view, and I think there's no hope arguing against it. Turing suggests the "Anthropic Principal": we shouldn't assume that we're special, because the earth isn't in a special place in the galaxy, or in the universe, and we have pretty compelling evidence that we're an evolved version of other animals around us, but some groups (e.g. biblical literalists) find this unconvincing.

Answer 2

AI is "deterministic" in the sense that it follows exactly the algorithm. "Deterministic" means different things to a data scientist/programmer, but let's not go into details here.

There is no "freewill" in AI, it's all about mathematics and algorithms. Don't watch too many scientific movies!

Answer 3

Although I’m not sure how an act of freewill could even be described (let alone replicated),

Well, one popular definition goes like this:

[Free will is] the freedom to act according to one's motives without arbitrary hindrance from other individuals or institutions

Source - Wikipedia entry on Compatibilism

Note that this definition is perfectly compatible with determinism (hence the name "Compatibilism"). Actually, proponents usually argue that free will requires determinism, because if your choices were ultimately random, like rolling a dice, how could they be your free choices?

Now, if you assume that an AI can be said to have "motives", then according to this view, it would have free will - if noone hinders it.

The contrary view, Incompatibilism, has been described in another answer by John Doucette. I agree with him that most AI researchers probably don't worry about philosphical questions like that. All the proponents of indeterminism (sometimes called metaphysical free will or libertarianism) I'm aware of assume that there exist "causes" that are neither deterministic / physical, nor purely random. (e.g. Agent causation). Since AIs are more or less by definition based on physical processes, deterministic algorithms and possibly some random number generator, I don't see how they could posess this kind of freedom.

Answer 4

AI is algorithmic not free willed in a sense that humans have free will. So in that sense it is deterministic. Give it the same data each time you would expect the same result. Change something (ie feed it new data to learn from) and then it will give a different result. Hence the determinism.

EDIT: Some algorithms do use some randomising - ie some versions of hill climbing - but if we want to get technical there's no such thing as true random numbers anyway. (Unless you're using one of those supercomputers that use the radiation from the sun as a seeding factor)

Answer 5

The free will in the spiritual world would be for you to have the right to follow or not the main path (God's way). Taking into consideration that whatever path you follow, you will have its consquences (good or bad).

However, an artificial intelligence created to solve problems does not need these distractions. Determinism considers the feeling of freedom a subjective illusion. Trying to apply free will to an AI is to create an inner inner space within the "mind" .. but creating that is already impossible because we can not understand exactly what is going on in our minds. The big question is, how to program something that we can not explain accurately? How to induce an AI to learn something that does not make sense?

I believe that it is not just code and database that will lead the AI ​​to have faith in its existence or to believe in free will. But to construct an AI that reads and understands every human mind, every thought, house illusion, every paranoia, every confusion, every sadness, every lie ... may be a good source of studies, but whoever qualifies for this kind of experiment? What would be the conclusion of an AI by understanding the whole confused and troubled human mind?

It's a very complex question ... let's continue studying and raising this kind of subject. I spent my last years trying to recreate human thoughts and actions in an AI .. it's a study for a lifetime .. my fear is the disappointment at the end of everything :(

Answer 6

Love question. Firstly, as pointed out in other answers Narrow AI today are mostly algorithms following their procedures given some inputs. No need for philosophy here as they are following reproducible steps.

However, if your referring to general AI or AI a kin to human level intelligence or better then maybe the question holds some weight. But again as pointed out it would come back to whether you believe you or I indeed have free will?

For me I believe free will can be modelled as a sort of entropy. If you look on the macro level things are blurry, agents are making decisions and moving around in a unpredicatable way. On the micro level however, given all the data in one state one could predict the next state shattering the idea of free will. I guess it’s up to you to decide whether this fits with your free wil definition or not.

Answer 7

No, because there is no utility in building a "libertarian free AI" as far as I know of.

AI is another tool. What is the purpose in building an AI with such a distinction?

The reason for that question is this. Let's say you want an AI to accomplish some kind of task you want machine assistance in. That's what tools do- assisting with tasks. What exactly would this task be that a "non-libertarian free AI" couldn't achieve but a "libertarian free AI" could?

Answer 8

I was very bored.

The short answer is yes and the long answer is basically yes.

I'm going to gloss over what you could possibly mean by AI and instead focus on what AI programs pretty much are:

A mish-mash of algorithms and methods borrowed from mathematics, perhaps more specifically statistics, inserted into a good old-fashioned program.

Let's start from the very beginning. The very beginning.

For the sake of this argument, everything is a language and there exists an omnipotent alphabet which contains all possible symbols you can come up with to convey a message. Such an alphabet would contain everything you've ever known, all their permutations, and then some. It would contain your clothes as well. Forget countable infinity or all those concepts, those come way later.

It is important to realise that right now I'm communicating with you through not just through the power of a string of characters. You're also experiencing other stimuli that you, or whatever part of "you", is/are capable of interpreting into another language. To be crude, everything you decide to classify as a 'standalone thing' is a compiler that runs in perpetual execution, translating all the 'things' it 'receives' into other 'things' it 'spits out' for other 'things' to 'receive'.

Think about a modern day computer. You write a program in your fancy little language. You hit compile. A compiler goes through your code, and spits out more code. Except this time this code is written in another language, sometimes "closer to the metal", sometimes "just about the same abstract level" and sometimes "even more abstract", and this process repeats itself until somewhere along the line that code you started with, has been interpreted to mere electrical signals, which then themselves are being 'compiled' by an entity we will call 'the universe' and that's up to empirical observation to determine what and what's not going on. (Except "the universe" was always responsible. But we will partition things for the sake of ..partitioning things. You get what I mean, I know you do.)

Now let's jump back to languages. Mathematics is a language in the sense that:

• It is built of symbols contained in an alphabet we will call X
• The several fields specify their own grammatical structure through which we can decide whether or not a statement is well-formed. This encompasses everything from where you can put the + operator in high school algebra, to how you can write a proof in formal logic. There needs to be no justification for how you build a grammar. You can always just make a new grammar and use it instead. Of course, it might not be capable of forming statements which are compatible with other grammars.

What's interesting about X is that its definition is not fixed. Throughout time, we've introduced new symbols into the mathematical alphabet to be able to express more concepts while keeping things separate. (Or rather some people have had the sense to keep it this way.) For example, whenever you see Leibniz's integration symbol, you know you're probably dealing with some kind of integration and not something novel that you've never heard of before.

Now here's where I actually answer your question:

• I assume that by "program" you are referring to the mathematical construct as defined in theoretical computer science: A string of characters from an alphabet.
• This string is then fed to a compiler (lexer|parser|semantic analyser) which spits out another string (mainly the job of the semantic analyser). This string usually is built from characters of a different alphabet. That is to say, the compiler is a function which maps a well formed string of language A to a string of language B
• The end goal of compiling a program is to execute it, which basically means a succession of compilers will take the output of the previous compiler and spit back yet another string, until the string is essentially electrons moving about the circuitry of the computer in your bedroom and producing fancy lights on your monitor

So whenever you write an "AI program", you're just writing a "program" that contains some "AI algorithms" which are really just applications of things we've known in mathematics for 100s of years, which again, are really just a string of characters that are about to be translated by a compiler.

In other words, nothing you can ever write is not deterministic, provided you look at the bigger picture.

A common argument I see is that since AI programs usually "adapt" and "self-optimise" when solving the problem, they're not quite deterministic, in the sense that feeding the program the same input twice will (hopefully) yield better results the second time. Except what really happened is that you had an input string that you partitioned into inputs A and B, and fed them to the algorithm in succession. Had you fed AB initially, you would've obtained the same results.

Answer 9

You guys are missing the point completely.

Everything I read assumes that being is based on science. You are trying to fit the reality of being into a brain, that is, into a machine. You know that you exist. A billion sensors of every imaginable type cannot equate to reality. Even with a billion more neurons than what's in the brain.

It's like the mathematics you use. You believe you can be accurate when the truth is, you cannot be. The position between 2 numbers can never be accurate, they are infinite.

It's like time. You believe it to be organic when the truth is, there is only the now.

Try as you may, you will never fit being into an algorithm. The real illusion is that you believe you can.

Answer 10

"there's no such thing as true random numbers anyway." that's all you really need to deter any idea of AI on any computer. Any software or set of functions on a computer is pretty much (right now at least) all set code, by humans.

Also the execution of actions based on variables not listed is not artificial intelligence, it's just simpler code.

Any REAL Artificial Intelligence will not be made on a board of 1's and 0's, that's defeating the purpose, all of the actions are predetermined (even if they are extended intricately to cover many possibilities) so they have no chance to create something non-deterministic. Real independent intellect is most likely (in my eyes) found, not made.