From a functionalist point of view: when is an algorithm an A.I., and when is it just software?

Question

Recently, The Atlantic published an article claiming that "Google Taught an AI That Sorts Cat Photos to Analyze DNA". When you look at the original paper published by the Google team, what they really did was take a neural network model normally used for image classification and apply it to classifying DNA data instead (see also this debunking of The Atlantic's article). From one point of view you can claim that they used "an AI" to analyze DNA data, from another they just applied a more advanced mathematical model than the previously used ones (universal approximators vs. parametric models).

This begs the question: Assuming that the computational theory of mind and functionalism are true, is there a way of discriminating algorithms that are just software and mathematical models from algorithms that qualify as AI?

One straight up answer to this is obvious: It's an AI if it passes the Turing test (or some modern more advanced variation of it), otherwise it is just software.

But beyond the Turing test there seems to be another issue with The Atlantic's article: To say that "an AI" did this or that implies that the neural network they used had some level of agency or autonomy, when in fact it was a completely inert program that starts and stops based on a set of pre-defined parameters and instructions.

The same applies to Amazon's Alexa: It doesn't respond to any communication unless someone prefaces their sentence with the word "Alexa". It never starts communicating on its own or spontaneously inserts itself into a conversation.

My questions:

1. Does this autonomy/agency aspect of a program really make it possible to distinguish intelligent AI from "just software"?
2. Has the question of agency been studied in philosophy of mind and philosophy of artificial intelligence?
3. How is agency (as described above) related to intentionality and to Daniel Dennett's intentional stance?

Dennett says in "The Intentional Stance" that: "first you decide to treat the object whose behavior is to be predicted as a rational agent; then you figure out what beliefs that agent ought to have, given its place in the world and its purpose." but can we treat an object as an agent with purpose if it doesn't have the autonomy that I described above?

4. Similarly how does this tie into Chalmers notion of Philosophical Zombie? Why might imagine that such autonomy can serve to discriminate between philosophical zombies and humans, since zombies act like they are autonomous, but aren't being driven by an autonomous agent on the "inside" (See my previous post about "freewill zombies")?

Answer 1

The position of functionalists on AI is similar to the position of compatibilists on free will in two important respects. First, they distance themselves from the Cartesian idea that there is some extra special "substance" or "essence" there, and dissociate what is so bundled into effects that can be modeled piecemeal causally and/or computationally. Second, as a consequence of the first, to them there is no bright line between intelligence and non-intelligence just as there is no bright line between freedom and non-freedom, there is only a continuum with extremes and pragmatically placed thresholds. Roughly, the more useful agency talk and other intentional vocabulary is for describing and anticipating behavior of a system the more likely it is to be classified as AI. The Turing test, in its classical form relying only on conversations, is insufficient, it has to be extended to the totality of behavior, including non-verbal actions and responses displaying certain levels of sensitivity to changes in the environment and complexity in adjusting to them, learning from experience as it were.

This is not, in principle, different from inferring agency and intentionality in other people. Functionalism is officially agnostic as to how the function is implemented, in this sense Kant was a functionalist about the mental. Although the computational theory of mind strongly suggests some form of material implementation and may even go hand in hand with eliminative physicalism, technically it is not married to either. But whatever one's belief about the underlying basis of agency, in practice we go by observable patterns of behavior and adopt intentional language when it works. This is behind late Quine's acquiescence to "mentalistic language" in Pursuit of Truth:

"Still the mentalistic predicates, for all their vagueness, have long interacted with one another, engendering age-old strategies for predicting and explaining human action. They complement natural science in their incommensurable way, and are indispensable both to the social sciences and to our everyday dealings. Read Dennett and Davidson."

Davidson is, of course, a non-eliminative materialist who believes in irreducibly mental descriptions, and Dennett is a champion of as if intentionality. As for the rest of the OP questions AI in Context chapter from Alison Adam's book Artificial Knowing addresses them directly:

"Dennett's intentional stance offers a get-out clause, a way of acting as if certain objects have intentionality without worrying about whether they actually do have. Taking an intentional stance towards something is a way of granting it some level of agency. I believe that Dennett's position strikes a significant chord with both Collins's ideas and recent research in actor network theory. I am surprised to find that neither body of work refers to Dennett's writings even though he seems to lend such obvious support. Under this view we no longer have to worry about whether a machine or other object can really think; we worry instead about the appropriateness of designating it an intentional system, or in the language of ANT we worry about granting it agency. The job which ANT sets itself is to enfranchise the world of objects, although the process of delegation is done by humans - to see that knowledge and sociability are not properties just of humans but of humans accompanied by those objects that are delegated humanity.

[...] There are some interesting points of contrast here as it is as if the social sciences and AI have separately discovered the idea of agency at about the same time. Both in the popular sub-domain of distributed AI (DAI), where knowledge is distributed through several knowledge bases, or where intelligent agents act in concert to solve a problem, and also in robotics, it is curious to see that the language of agency and intentionality abounds, possibly much more so than in other areas of symbolic Al. Why should this be so? It seems to stem, not so much from a will to enfranchise objects in the way of some Continental sociologists, but rather a will to import sociological models into the design of their systems.

Leaving aside the question of whether the sociological models employed would be seen as oversimplified to the point of naivety by the sociologically sophisticated, I argue that using the language of agency permits a use of intentional language, almost by sleight of hand. If you call something an 'agent' then you can use intentional terms without examining them, without justifying them and indeed without grounding them. Such terms can be used in a purely operational way and then the metaphor of their functionalism can be allowed to slip into a reality. Now this might not be a problem to the ANT sociologists, but others, such as Forsythe, in her criticism of the positivist stance towards knowledge adopted by the knowledge engineers in her study, would find the same fault at work here. And of course, Searle, the arch-intentionalist, would have no truck with such sleight-of-hand intentionality. He would not allow it in an all-singing, all-dancing robot so I doubt whether he would let it in by the back door here."

Answer 2

On Agency, Autonomy and AI

Nice question. I believe there is no answer to it since it is a matter of definition. . Those who have hopes for AI tend to use the word to describe complicated software, and if they're a bit fussy they restrict the term to software that can 'learn', not that a computer ever knows it's learned anything.

1. Does this autonomy/agency aspect of a program really make it possible to distinguish intelligent AI from "just software"?

I think if some piece of software exhibited agency and autonomy I'd be tempted to award it AI. But this is a deep issue. Many philosophers question whether human beings are agents and this would include all in the the perennial tradition. Agency is illusory for this view.

2. Has the question of agency been studied in philosophy of mind and philosophy of artificial intelligence?

Not to much effect afaik. Agency is simply assumed by most Western philosophers and they rarely delve so deep into consciousness as to start questioning the whole idea. But there will be a lot of work of which I'm unaware.

3. How is agency (as described above) related to intentionality and to Daniel Dennett's intentional stance?

It seems to me that agency would be impossible without intentionality.

4. Dennett says in "The Intentional Stance" that: "first you decide to treat the object whose behavior is to be predicted as a rational agent; then you figure out what beliefs that agent ought to have, given its place in the world and its purpose." but can we treat an object as an agent with purpose if it doesn't have the autonomy that I described above?

This may be your choice. Dennett seems to advocate making assumptions about the agency of objects and developing a fantasy about their beliefs, purpose and place in the world. I would doubt that an object can have beliefs and find the whole approach highly unscientific. I'd agree with you that unless we see some autonomy, even if it's ultimately illusory, then it will look like commonplace software.

5. Similarly how does this tie into Chalmers notion of Philosophical Zombie? Why might imagine that such autonomy can serve to discriminate between philosophical zombies and humans, since zombies act like they are autonomous, but aren't being driven by an autonomous agent on the "inside" (See my previous post about "freewill zombies")?

There is no way to make this discrimination since zombies are defined as being indistinguishable from humans. If we believe that zombies are possible and are a coherent concept then we must also believe that it would be impossible to tell if a computer or a human being is conscious or has autonomy and agency. If zombies are possible then it's possible that human beings are zombies and there would be no way to rule this possibility out.

I feel that Chalmers' zombies are best consigned to the realm of unicorns and spaghetti monsters. As for how to distinguish between software and AI this is a matter of definition and definitions vary. I would say there is no distinction to be made between software and what most folks are calling AI. Most forms of AI could be reproduced on an abacus if its was big enough and had a motor. A seasonal 'Bah humbug' would be my response.

Answer 3

Definitions

Learning: the acquisition of knowledge or skills through experience, study, or by being taught.

Primary Argument

Artificial intelligence is a field which attempts to mimic animal or human intelligence in software. One thing which differentiates AI from other types of algorithms is that AI learns. Other forms of algorithms do not. Because of this difference, AI is useful when we might not necessarily know all the details of the problem we are trying to address.

There are two types of AI, basic machine learning and general artificial intelligence. When the field of AI started, AGI was essentially the kind of AI. These days, the focus is on basic machine learning, using methods like artificial neural networks. There are philosophical questions about whether or not we can produce true AGI, and there is a basic test, the Turing Test, which attempts to see whether or not we succeed.

Regardless, the key difference between a general algorithm and an AI is that an AI is designed to learn.

Rebuttal

Basic linear regression techniques are not AI because they do not learn. You do not set a linear regression technique off on its own and have it learn as it goes. It is a pre-modeled system. Neural networks can do linear regression and do so only the fly and therefore do learn.

TimB argues that "neural networks et al don't 'learn'; they collect data. They are then capable of using that data in the next 'iteration' of their algorithm." However, that is exactly what learning is: taking data and using it in the next iteration of the decision making process.

TimB and others also seem to only accept AGI, but this question is at least partially a Computer Science question so the definition is largely based on the field of AI in computer science. The field distinguishes between weak and strong AI, or AGI vs AI as I described it above (Computer World). An AI does not have to act like a human. It just needs to be able to accomplish some kind of learning task that a human could.

Animal Intelligence

Moving on from the Computer Science argument, consider the following. It is true that we generally measure intelligence using our understanding of human intelligence as the paradigm, but we admit that many other creatures are intelligent, just in different ways. It is important to look at this point if we want to see what constitutes intelligence. If we using the Turing test to determine whether something is "AI" then even our close relatives would fail the Turing test. Does that mean that primates, other than humans, are not intelligent?

Answer 4

Let me start by stating that all algorithms are "just software."

There are two types of software: 1) Software that performs the same function over time, and 2) one who's function changes over time.

For the first type, the programmer provides the "intelligence" for the function. For the second type, the programmer provides some basic functions and the means for the functions to change themselves (self modifying code).

Self modifying code is the "critical" requirement to be able to implement some level/version of A.I. Therefore, if after some interval of run time, you examine the computer's code/function and it has changed, then it has AI capability, if it is the same, then it does not.

Answer 5

Consider the specific questions.

Question 2: Has the question of agency been studied in philosophy of mind and philosophy of artificial intelligence?

Regardless of what has been done, one can start to question it now by defining the choice of a freewill agent and the state of a deterministic machine. The following definitions are provisional especially the use of "something".

Define the choice of a freewill agent as something for which there does not exist a complete explanation.

Define the state of a deterministic machine as something for which there exists a complete explanation.

An answer to the first question follows from these definitions.

Question 1: Does this autonomy/agency aspect of a program really make it possible to distinguish intelligent AI from "just software"?

Since there does not exist a complete explanation for choices made by a freewill agent but there do exist complete explanations for states of a deterministic machine, deterministic machines are not freewill agents. That implies there is no way, based on agency, to distinguish intelligent AI from just software. Both “intelligent AI” and “just software” are deterministic machines and not freewill agents.

Question 3: How is agency (as described above) related to intentionality and to Daniel Dennett's intentional stance?

Dennett, based on the provided quote, apparently needs a “rational agent” to get intentionality and perhaps agency.

Are agents in general “rational”? Do agents in general even have a brain? These questions are not merely rhetorical. Note that John Conway and Simon Kochen’s “The Strong Free Will Theorem” proved that if humans have free will then so do quantum particles. Quantum particles do not have brains. They are not rational. A brain is not necessary for agency.

When Dennett restricts agents to “rational agents” I suspect he is simplifying the problem so he can intelligibly explain it with his theory. Such simplifications may produce results that do not adequately reflect reality. That is, Dennett’s explanations may be neither correct explanations nor complete explanations although they may be useful for AI. The correctness and completeness of his explanations for human reality require justification.

In the next references to Alan Turing and Jonathan Haidt, I will argue why such a justification may not be possible.

Turning’s “Computing Machinery and Intelligence” expresses a concern about extrasensory perceptions:

Once one has accepted them it does not seem a very big step to believe in ghosts and bogies. The idea that our bodies move simply according to the known laws of physics, together with some others not yet discovered but somewhat similar, would be one of the first to go.

He does not deny extrasensory perception, but he offers the following as one solution.

One can say in reply that many scientific theories seem to remain workable in practice, in spite of clashing with ESP; that in fact one can get along very nicely if one forgets about it. This is rather cold comfort, and one fears that thinking is just the kind of phenomenon where ESP may be especially relevant.

It might not only be “thinking”, but choosing as well. Notice that Turing does not deny the existence of extrasensory perception although he does not find it intelligible or meaningful in terms of his theory. He faces the possibility that his approach may not be complete or even accurate.

Haidt’s “The Righteous Mind” raises questions about our human rationality. His lecture, “The Rationalist Delusion” summarizes his position.

If Dennett is trying to explain human beings as “rational agents”, he may not be explaining human reality as it really is although his explanations may work well for AI.

Question 4: Similarly how does this tie into Chalmers notion of Philosophical Zombie? Why might imagine that such autonomy can serve to discriminate between philosophical zombies and humans, since zombies act like they are autonomous, but aren't being driven by an autonomous agent on the "inside" (See my previous post about "freewill zombies")

Given the Conway-Kochen theorem, it is worthwhile separating freewill agency from other forms of subjectivity since there are results that pertain to free will alone.

Consider a quantum particle that has free will. Such particles may be close approximations to what a freewill zombie might be. There are at least two objections to saying they are freewill zombies. (1) A freewill zombie is a p-zombie. A p-zombie is supposed to be a human being. A quantum particle is not a human being. (2) One needs to verify that there is nothing more to the subjectivity of quantum particles than freewill agency.