Does artificial intelligence necessarily mean a system can somehow "freely" vary its own behaviour?

Question

If an artificial intelligence system existed in a robot and it was able to constantly reprogram and reconfigure itself in any way without disrupting its necessary functions, and without even partly depending on any external data (including any programmers); then such a system could setup its own goals, arguably describable as an AI 'choosing' to reconfigure itself so it can be in a better 'position' to achieve it's certain self-'established' goals. At least it might according to outside observers seem as though such a system possesses the ability to selectively vary its behavior. So could an A.I. system be able to actively 'reprogram' itself and change itself in ways only 'it' could give specific definite 'justifications' for?

By using the term free will here I mean the A.I. system has the 'ability' to change one or more of its 'internal states' relative to the information content and 'information-packages' it has 'access' to in ways that are NOT predetermined by 'previous programming' or information 'set-up' by any 'programmers' or other 'outside' sources of info. or info.-management. The A.I. system is able to change 'itself' information-wise without ANY 'outside' help or interference. This would be 'like' having the 'ability' to 'freely alter' itself regarding what info. to analyse and manipulate next.

Alternatively, instead of using the term free will with all its philosophical baggage one could call this 'non-external-interfering' system self-variation or 'non-outside-influenced' self variation or auto-cybernetic selective variation. Of course any system that can 'vary' itself in some way must be 'doing' some of the important variations 'itself' without any 'externally sourced' instructions; or else the system would have to be 'told' how to do every type of system change and all its important functioning would be run 'by remote'.

Answer 1

Your question raises interesting ideas in the the areas of free will as well as the nature of consciousness. We might begin by assuming that humans have free will, as many philosophers do. It seems, after all, we are able to make choices, to "vary ourselves in some way without any externally sourced instructions". You seem to recognize that free will in a deterministic universe is a complicated issue, so I won't belabor the point here other than to point out some other good questions we have here on that topic:

What are the necessary conditions for an action to be regarded as a free choice?

Is free will reconcilable with a purely physical world?

What is the difference between free-will and randomness and or non-determinism?

The problem I want to highlight to you (which may be why you asked your question in the first place) is: Can we reconcile the notion of freedom while knowing that a robot we programmed — despite making choices we may not have known it would make — is acting in a way we intended (and perhaps even predicted) it to? If you program a robot to reprogram itself to make itself smarter and give itself its own goals, and it does that very thing, are those "choices" it made, or is it just following a set of rules/guidelines we defined? How many successful code-rewrites must occur before we say the robot is self-generating these actions? Can we ever say that?

This issue you run into here, it is no different from the issue of free will in humans. While our brains are made of more mushy materials, we do seem to exist in a very physical and determined universe. Our actions seem very influenced by our physical brain (if you damage it, we start acting differently, and altering the brain in more targeted ways with drugs seems to also have a profound affect). The influence which is perhaps most important to discuss here though (outside of direct external physical brain intervention) is the influence of the past. As we are raised, our brain is shaped not in some vague, ghostly way but physically in the neurons, and these physical changes (what account for our memories and consistency of self, why we see ourselves as the same self over time and not a new person at every moment) greatly influence how we behave and react to things. It is quite literally our own programming. We like to think we are "programmed" in such a way that allows us to "freely" make choices in our life, but do we?

There are many interpretations of free will (compatibilism, incompatibilism, pessimism, etc.) but if you want to hold a hard determinist view (an incompatibilist view, in contrast to libertarianism (philosophy)) view you will find it difficult to reconcile any notion of free will in robots or humans for the same reasons. In my answer to the first free-will-related question above, you will find my own proposal for a new definition of free will and how I reconcile the concept with human choice, and I leave it to others to explain how they do it under other views (as compatibilists, for example).

---

Regarding your comment:

Could a functioning A.I. system that can reprogram 'itself', (without causing any 'internal disfunction); could such a system 'reconfigure' various programs and information it is manipulating to 'come up with' a set of programs and/or info. that actually contradicts certain ideas held by its programmers in such a way that would prove the system was NOT acting even indirectly according to how the programmers initially programmed it? In other words can an A.I. system be 'programmed' and when 'used' it 'tells' the programmers some of their fundamental ideas are wrong?

Can you? Can you act in a manner which does not accord to how you were raised and are genetically/biologically programmed? My gut tells me that nothing is random. A system can act in unpredictable ways, but those ways are not random, they are very caused, they are very determined by the previous inputs. So, sure, a system can act in a manner which contradicts its programmers only in that the programmers themselves had a poor notion of what their code would entail. Just as humans can seemingly act "out of character" too sometimes. The programmers thought that their algorithm would result in behavior somewhat like X, but it turned out like Y. The error is with the programmers, they predicted wrong; the system did not act on its inputs in a way that was random, or inherently unpredictable (just perhaps difficult to predict (obviously so for the programmers if they predicted something else)).

Answer 2

I wonder what the relevance of "constantly re-programming and re-configuring" would be. People don't constantly re-program themselves. And except for some cases that are addicted to plastic surgery, they don't re-configure themselves. (Actually, that isn't quite true. If you eat some food and then get sick, your brain might re-configure itself to dislike that food).

Intelligence and free will are totally separate. Imagine an AI that I could tell "please write 50 symphonies in the style of Mozart, just a lot better", and it would do that and have no choice not to do it, that AI would have an awful lot of intelligence and no free will.

An artificial intelligence is complex. You ask for an AI that is not "pre-determined" by something a programmer programmed. However, even a primitive chess program makes moves that the programmer of the chess program didn't foresee. An AI would constantly do things that the programmers didn't foresee. Do you do anything that isn't pre-determined by your genes? How do you know? It's the same with an AI. We don't know what "free will" is. We don't know if it exists, either in an AI or in a human being.

That aside, electronics seems to be a lot easier than brain surgery. So an AI would probably have a better chance of improving its "brain" than a human has. Actually, we know much better how to improve computers than brains, so obviously an AI that could acquire that knowledge could learn how to improve its brain while being clueless how to improve your brain.

Actually, a sufficiently developed robotic AI with an internet connection could probably find a job as a computing consultant (forging ID and pictures), open a bank account, make money, order parts online to improve itself and so on.

Answer 3

No, there will never be artificial "intelligence" that operates outside of the programming us humans have designed. Computers are nothing but a tool, something that carries out instructions. They don't think for themselves and they aren't capable of doing anything that we have not told them to do. Any adapting and "learning" going on here has to be a product of human-designed programming and it is limited to that programming.

Answer 4

You ask,

” […] could an A.I. system be able to actively 'reprogram' itself and change itself in ways only 'it' could give specific definite 'justifications' for?

Disregarding the like-that-but-not-quite quotes, it appears that you meant to ask

Could a machine intelligence be able to actively reprogram itself and change itself in ways that only it could give specific definite justifications for?

And this is apparently in the context of

” 'ability' to change one or more of its 'internal states' relative to the information content and 'information-packages' it has 'access' to in ways that are NOT predetermined by 'previous programming' or information 'set-up' by any 'programmers'

Well, the idea of a directly programmed intelligence, one where each main function of the mind was implemented by a programmer, was common in the 1970's. But while it's nice as a goal to gain better understanding of how e.g. vision processing works, it's wholly impractical as a way to create an intelligence. As already Alan Turing ⁽¹⁾noted in 1948 or so, the most likely way a machine intelligence is created, is the way that a human intelligence is created, namely by growing up and learning – with some basic instincts and abilities in place, of course.

So a first answer is that the question as posed doesn't make much sense, because it's very unlikely that a machine intelligence will be of the directly programmed variety, that there will be any programming (except of basic functions such as edge detection in vision processing): I'd guesstimate that it's about as likely as a crocodile emerging up through the asphalt in the street, deftly stealing your wallet, only to be hit by a giant iron hippopotamus accidentally dropped from an airplane passing above.

However, running on a digital platform means the possibility of making copies, partly or completely. It means the possibility of trying out things in simulated environments. Not the least it means that explorations of possibilities can be really, really fast. Currently the electronics is some millions times faster than our brain stuff, and that difference increases exponentially, which it's kept on doing since 1965 or so, roughly a doubling every 2 years. So we can expect some really fast evolution as soon as machine intelligences start creating new ones.

---

^{1) Mainly section 9 “The cortex as an unorganized machine” of his 1948 technical report “Intelligent Machinery” (the link is to my transcription of photocopies of the original manuscript).}

Answer 5

Yes, to be intelligent an entity has to be capable of self re-programming. There is no way around it.

You have to look at what comes before intelligence. It is knowledge. Knowledge is about acquiring set of rules. Once you know rules about something you are knowledgeable in it.

It all begins with data. Arrange data in some form, sort it, categorize it, find its limits, average it etc; you now have information.

Analyze information. Generalize it. You find rules that always work from the data. You find rules that never work. You find rules that usually work. You find rules that work only on selected data i.e. you discover ranges of your functions (functions are rules) and/or only on a particular kind of data i.e. you discover domain.

The rules you gather is your information base. As opposed to your database which is just collection of information.

Data -> Information -> Knowledge

The rules are your functions, your formulae, your equations, your theories, your laws. All different names of same general concept.

However, this is not enough. Often complete data is not available. There are big holes in data and we must act before all the data is available.

Intelligence to the rescue. It fill gaps in information. Rules not known are conferred. Given how much data you have and how good you are in finding rules from that data you may be pretty good in guessing rules you dont know.

Now, if an organism is intelligent, it can add rules in its information base that dont come directly from data. This adding of rules necessarily change internal program. The entity is capable of re-program itself.

Behaviour ofcourse is how one is programmed. Re-program definitely means changing of behaviour.

This is as much freewill as we get. Within confines of our initial programming, our instincts we are capable of re-programming ourselves. There is a wide range for us to re-program in.

We can change our habits, learn new skills, even get rid of old habits and forget what we learned.

This is the path any A.I. will follow because its inherent of that last word: intelligence.

Answer 6

For a true artificial intelligence, yes, it must have what you call a free will. For a simulated artificial intelligence, it's not necessary.

I call it the ability to change its cognitive parameters in order to learn, understand or resolve. Also, as said previously by gnasher : Intelligence and free will are totally separate. An AI is defined by its components, hardware and software, and you need to think from this perspective : a free will doesn't mean anything for a set of transistors and chips. It's built to compute. Thinking requires a lot of computations, and what you call free will, I call it engineering and applied (computer) science - in the context of what we know an AI primarily is.

For a given unknown parameter, variable or problem, it needs to think in order to learn how it works, understand why it works, and eventually apply the reasoning to solve what is unknown to "it" yet, and integrate it. One way to integrate it would be to reprogram to optimize itself, without having to go through the process of thinking again, but applying directly the solution in a similar situation.

A free will doesn't mean anything for something made out of the hardware and software we know. It's all about data and results, through processing.

"I think therefore I am".