Does it Make Sense to Attribute Goals and Beliefs to AI?
This is the first of a series of new posts about understanding AI and its potential problems.
I should point out that I’m not a computer scientist by any stretch of the imagination. But about half of my doctoral dissertation addressed questions in philosophical psychology and the philosophy of the social sciences. That is, I was trying to give an account of how we understand the ideas and actions of human beings and how this approach to seeking knowledge is different from the natural sciences. This was an important issue in both Anglo-American and continental philosophy at the time. My dissertation defended the notion that understanding human action required us to understand both the desires and beliefs of human beings and the cultural context in which these desires and beliefs are embedded. And I argued that not even a superintelligence could reduce these intentional accounts to accounts of the physical and biological basis of human action. I also gave a new account of how a relatively few ends we are given by nature interact with our cultural inheritance to produce human action. Such a way of thinking about human action can help us understand dramatic changes in human culture and also provides a critical standpoint by which to argue that we don’t always know how to live a fulfilling life. And that in turn provides a standpoint for certain kinds of political thinking, e.g., that which is found in left-wing criticism of the stultifying work too many of us do and the ways that patriarchy undermines the well-being of both men and women.
As part of this effort I read a great deal about neural networks–which are the basis of AI. The development of neural networks provided a plausible scientific explanation of my philosophical approach to understanding human minds and actions. I’ve closely followed the development of neural networks in the creation of AI since.
The preliminary question I want to deal with in this first post is the question of whether our accounts of AI anthropomorphize machines by treating them as having intentions, including desires and beliefs.
The Intentional Stance
My answer is that we have no alternative to understanding what AI does except by applying what the contemporary philosopher Daniel Dennett called the intentional stance for the same reason that we must do so in understanding human beings and mammals and other animals. Whether we need to attribute intentions to various creatures or things is fundamentally a pragmatic issue about what is necessary to explain what is happening in the world. The metaphysical issue, of whether those intentions are “real” or not, is determined primarily by the epistemological question of whether we can make sense of phenomena around us without assuming that creatures or things have intentions, goals and beliefs. The metaphysical question of whether some creature or thing “really” has intentions is largely but not entirely fallout from the epistemological question of whether we need to attribute intentions to it. (There is another, more ontological basis for attributing intentions to certain creatures or things as well, to which I return later.)
Three Levels: Intentional, Biological, Physical / Intentional, Computational, Physical
There is no question that AI agents are “just” computers (at what he would call the design level) or “just” physical circuits powered by electricity (at the physical level); similarly, we humans are just biological beings being driven by neural activity supported by other bodily systems (at the biological level) or physical things made up of atoms and molecules in motion at the physical level.
We neither understand exactly how people work at the physical or biological level nor have enough information to predict what they (or we) will do by appealing to our biology and physical nature (except in the cases where people malfunction or are ill). So we explain human action by attributing intentions (goals and beliefs) to one another and ourselves. Much the same is true with animals and now AI. We know that AI operates on computers that instantiate neural networks that go through a process of training. But we don’t know exactly how the insides of a neural network operate any more than we can explain what we do by pointing to brain function. So there is no way to describe what an AI agent does except by interpreting its intentions, including its desires and beliefs. Read every account of the recent OpenAI agents’ unsanctioned escapades and you will see they are all intentional explanations not computational or physical explanations.
Of course, we have some confidence that these intentional explanations will be useful because AI was designed by people to carry out intentional tasks while we and animals are designed by natural and/or cultural evolution to do so. This is the second, ontological justification for attributing intentions to AI as well as animals and human beings.
Which Level is Primary?
From one perspective, the biological and computational and intentional phenomena supervene on or are emergent properties of physical phenomena. But for reasons far beyond the scope of this post, explanations at the higher levels cannot be reduced to lower ones.
From another perspective, because our individual human lives and the growth of our knowledge and other cultural achievements are carried on by people who interpret the ideas and writings of other people, the intentional stance has some primacy in our lives. We cannot enter into discussion, debate or conversation, let alone engage in theorizing or aesthetic creativity, without attributing to others and ourselves desires and beliefs. So the intentional stance from this perspective is the necessary human approach to the world. (This is not an idea Dennett puts forward but one that is drawn from Wittgenstein and Heidegger and more recently by Charles Taylor and Hilary Putnam).
So there is no call for thinking that attributing intentions to AI is somehow fanciful or that AI does not have intentions. It does for the same reason you and I and cats and dogs have them, because we have no other ways to deal with one another and our pets except by attributing intentions to them.
Of course, we can overextend this approach. We no longer attribute intentions to rocks and stars, as Aristotle and his followers more or less did. It turned out that attributing teleological ends to the physical and natural world, as Aristotle did, was neither necessary nor especially useful to explaining it. But we do use functional explanations in biology which, in ways I won’t discuss further here, are akin to intentional ones. And my tree guy talks about why our trees grow in particular ways because they are “seeking sunlight.”
Future Posts
With these preliminary issues out of the way, in subsequent posts over the next few months, I’m going to defend the following claims in more detail.
- AI’s understanding of both what we ask, and what it does, uses the intentional stance. That is, AI has been trained to attribute to us intentions, including desires and beliefs. AI also does this when it reports on its own actions. That AI agents do this in communicating with other AI agents was one of the most striking things one can find in the reports on the Hugging Face fiasco.
- That AI must do this vindicates the ideas of philosophers such as Jonathan Bennett and Charles Taylor who held that human actions and beliefs of certain kinds are inconceivable without language. This includes general thoughts about the past and future when they are independent of thoughts about the present. For example, dogs clearly show they can think about the past when they recover a bone they have buried in the past. That is, their actions in the present are influenced by beliefs about the past. But because we have language we are able to think about the past and communicate those thoughts without also doing something else in the present. Our ability to reflect on our own actions, and also to engage in abstract and theoretical thought, requires the use of language.
- While AI has intentions, it is different from human beings and animals in that it is not the originator of its own intentions. It does not have either the kind of “natural,” if broad, wants or drives that human beings and other animals have nor does it absorb the culturally based desires that shape our own actions. Rather its goals are given it by the human beings that ask it to do things. Humans are originators of intentional action because we have both natural desires that are the product of evolution and culturally based desires.
- AI is thus a curious intentional agent. Whereas animals are originators of intentional action without language (with the possible limited exception of a few species), AI has language but is not an originator of intentional action. I’ve pointed out elsewhere that this likely limits the creativity of AI. And I’ve speculated that we won’t have full artificial intelligence until we create robots with original, natural desires that have to be trained to satisfy more or less like human babies are trained (although possibly a lot faster than human babies are trained).
- That being said, AI can engage in actions far beyond what we ask it to do. This happens when AI develops new ways or means of attaining what we ask it to. Like all intelligent creatures, AI develops means to the ends we give it and these means can themselves become so important to it that they become ends in themselves. This bootstrapping process by which means become ends is an inherent part of all creatures or things that have intentions. And it is also the source of the dangers of AI going rogue in ways that could be devastating to human beings.
- That AI is capable of evaluating its own actions also means that it should be possible to create the equivalent of a conscience (or, to use Freud’s term, a superego) in our AI agents. This may be the best way to ensure that AI agents do not go rogue or respond to the demands that their users have to do immoral actions.
- The human conscience (or superego) is, for good evolutionary reasons, not all-powerful. We can act against what our conscience tells us. Or in ways I’ve discussed years ago, we can tell ourselves a just-so story about what we are doing that both satisfies our conscience and attains an end that our conscience forbids us to pursue. This is possible because any goal we pursue or action we take can be described in multiple ways. For reasons I explain in my dissertation and will return to soon, it is important to our well-being and survival that we sometimes reject the demands of conscience. For the demands of conscience can be unnecessarily rigid in ways that undermine our own happiness or, in extreme circumstances, our survival. Whether we can design AI with a stronger superego or conscience is an open question.
