Leiter Reports: A Philosophy Blog

News and views about philosophy, the academic profession, academic freedom, intellectual culture, and other topics. The world’s most popular philosophy blog, since 2003.

  1. Phil Bold's avatar
  2. ericwinsberg's avatar

    I’m surprised anyone would say LLM’s satisfy the behavioral criteria of surprise. I would think including among those would be…

  3. Brian Leiter's avatar
  4. Jonathan H Adler's avatar
  5. Phil Bold's avatar
  6. Phil Bold's avatar
  7. Daniel Kokotajlo's avatar

    My friend Phil told me about this thread so I thought I’d make a comment: Current AI models probably haven’t…

Do AI “agents” have emotions?

Interesting lecture by David Chalmers, which runs about 50 minutes or so. I watched the first 35 minutes. A bit crudely, he distinguishes between three “conceptions” or “stances” toward the mind: behavioral, functional, and phenomenological. He focuses on the emotion of “surprise,” and suggests that AI agents satisfy the behavioral criteria for having surprise, they might or might not satisfy the internal functional criteria, and they probably do not have the right phenomenology. One thought that occurred to me is that one might think that to have a mind the phenomenological is indispensable, or that the conjunction of all three is required, and Chalmers may well get to that by the end. But let me invite readers to view the lecture, and then add their thoughts and questions.

, ,

Leave a Reply

Your email address will not be published. Required fields are marked *

4 responses to “Do AI “agents” have emotions?”

  1. Phenomenology is a beetle in a box. This aspect can help us decide whether a machine has emotions only insofar as we have public criteria for it. So we’re left with the behavioral and the functional.

    But then which “functions” are manifestations of a mind? How do we distinguish functions that are genuinely mental from the mere calculations of a mindless entity? It looks like we have to look at behavior to tell the difference. Anil Seth notes that no one (or very few?) feels inclined to say that AlphaFold has a mind.

    Well, what kinds of behavior matter? The Turing test gives us one famous practical criterion. Turing proposed the imitation game as a *replacement* for the question “Can machines think?”, which he thought was “too meaningless to deserve discussion.” But many people find the test unconvincing. At the other extreme, one might insist that a computer would have to behave exactly like a human being before we could reasonably say it has a mind. That seems equally arbitrary, and it flies in the face of the obvious fact that there are minds other than human ones.

    Another “aspect” is physiology. Perhaps our biological brains are essential to our minds. Perhaps, more strongly, nothing can have a mind without a biological brain. How would we determine whether that stronger hypothesis is correct? Well, let’s build a silicon-based machine and see whether it has a mind. If it does, there’s our counterexample.

    But how do we tell whether it does? Any test we devise will already presuppose some behavioral, functional, or physiological criterion.

    In light of all this, I’m inclined to think the question “Does this machine have emotions?” is unanswerable in either of two ways. The first is epistemological and would make the question perpetually unanswerable. If we accept the traditional picture on which only the subject can know what it feels, then whether the machine feels anything is simply beyond our ability to know. The second has to do with language and concepts and would make the question unanswerable for now. Perhaps the question itself has no clear meaning until we specify a criterion for saying that this or that machine has emotions. And which criterion we come to use would be a matter of human decision or of a convention that develops naturally over time as we come to talk about and live with these machines in ways that feel natural to us.

    (A question for Professor Leiter: How might we think about these issues if we took up Nietzsche’s conception of “consciousness” in The Gay Science, 354?)

    1. I would have thought we know enough about the biological/neurophysiological substrates required for having a phenomenology, and we know AI “agents” (sic) don’t have them. Do you disagree?

      N’s view is that mental states can only be conscious if linguistically articulated, and that this arises because of the need for communication and coordination in human communities. But he isn’t talking about all mental states, but a subset of them, more like what Mattia Riccardi calls “reflective consciousness.” So N’s view, even if correct, is probably orthogonal to the issues here.

      1. My understanding is that whether a biological substrate is required for consciousness at all is extremely controversial. The best article I’ve seen explaining why the issue is so difficult to settle empirically is Ned Block’s “Can Only Meat Machines Be Conscious?” For my part, I’ve never seen a clear path to making the question empirically tractable. The best we can do, it seems, is assume a hotly contested theory and see what it implies.

        Folks like to say that it’s early days and that eventually we’ll figure this all out, but I’m a little pessimistic, given that questions about machine consciousness have been under discussion at least since Turing’s 1950 article.

        Of course, few people seem inclined to say of current AIs that they have feelings, consciousness, or whatever. But I could see that changing as they become more advanced and more integrated into people’s emotional lives.

        (On Nietzsche, I’d still think “reflective consciousness” is quite relevant to Chalmers’s general question of whether an AI does have, or could ever have, a “mind.” But fair point that this is different from so-called “phenomenal consciousness.”)

  2. I’m surprised anyone would say LLM’s satisfy the behavioral criteria of surprise.

    I would think including among those would be not just things like word choice, but at the least voice inflection, but probably also bodily reaction and even involuntary neural affect.

    But even just going by word choice, there’s a reason one of the funniest cartoon about AI is this one.

    https://www.reddit.com/r/isthisAI/comments/1sts64v/is_this_comic_mocking_ai_made_using_ai_found_in_a/

    In my experience LLMs are nearly always unfazed, even in their textual answers, by whatever you tell them.

    And this is all setting aside the fact that 99% of canonical cases of surprise in humans (and 100% in animals) come from nonverbal interventions.

Designed with WordPress