I’m currently working on a long piece about AI embodiment. And in the next week or so, I’m also intending to write a short ‘how to argue against the people who tell you AI is conscious’ piece. I think it’s time I made it very clear why I feel so strongly about that! Increasingly, I think we philosophers have an urgent obligation to spend a great deal of our time thinking hard about AI, and sharing those thoughts in publicly accessible ways.
On topic, I just watched the video of an event I spoke at over the summer — a panel debate also featuring Hollis Robbins and Toby Kurth, on the topic of AI, desire, consciousness, and more. It was organised and moderated by Luke Burgis. So, here are some of the things I said…
1) In response to being played a recording of ‘AI René Girard’ talking about AI and desire, and then being asked: “But is that actually what René Girard thinks about artificial intelligence?”
Look, I'm a philosopher. I spend a lot of time thinking about things like how interesting it is that, as humans, we have desires. And we think. I have my own novel theories about these things, which I can bore you with. But I think I want to say something even more boring than that — something very, very ordinary.
So, I think what I want say is something like: only living things have psychological states. Desire is a psychological state. I think this is a very, very ordinary belief to hold. I actually think pretty much every human being thinks this. They may say other things, like "Machines have beliefs," or "Machines choose". But I just don't think they're picking out this ordinary concept. I think we all know what desire is, and it's a psychological state that only a living thing can have. So, what follows from this is that things like machines, dead people who once existed, fictional characters — to push back a little on René Girard — don't have desires.
Now, that's a very ordinary thing. But I want to draw a conclusion from this, which is: I think it can be quite risky to go around talking as if these other kinds of things can have desires. It can be very conceptually risky. It can lead us to come to some bad conclusions. But I also think it can be risky in a probably more significant way, in that I think it could potentially lead us to do some bad and wrong things.
So, I think most of us think it's special that humans have these capacities: to desire; to think about, reflect on, our desires; to reason on the fact that we have desires; to think about things like what causes my desires, what impact do my desires have? We think this is special, not just ontologically in the sense of picking something out as the kind of creature we are. But also because it gives rise to some moral matters. It gives rise to responsibilities and rights.
And if we start talking about non-living things as having these capacities that they don't have, we might well start losing some of our purchase on the rights and responsibilities we have, as the kinds of creatures that do have these things.
2) On whether fictional characters can have goals
I don't think fictional characters have goals. I think the ordinary sense when we talk about ‘a goal’, when we're picking out the concept, is something that's contingent on having experiences, being able to reason on things, being able to come to conclusions. I don't think the sentence “Sherlock Holmes wants to become more famous” means that there is a thing, there, that has a goal. I think we all can understand the distinction between that and ‘Rebecca has a goal to be a better philosopher’.
Now, I would accept, however, that some of these kinds of concepts, which are contingent on being alive… Words can have more than one meaning. So, it's quite ordinary to say something like “The British government wants to reduce smoking”. Now, we all know — or I'm pretty sure I could persuade you very quickly, if you don't already know this — that the British government can't know [or want] anything! The British government is, like, some collection of people. But what we mean is ‘the people in the British government’. That's an ordinary shorthand. And I think we all understand it's an ordinary shorthand.
If, however, we start going about saying “Mr. Darcy knows something”, or “the AI knows something”, we just may not have in our shared vocabulary, and shared understandings of these things, the neat kind of way of distinguishing there. And it can be particularly costly, for instance, to children to grow up with people saying things like... It's very, very ordinary now for people to say things like “the AI reasons” and “AI reasoning”. I think that can be devastatingly, like, epistemically harmful for children, to not be able to… [someone else offers a suggestion why] But also they just may not be so good at picking out things in the world!
And what's more, if the AI really reasons in the way we reason, we better bloody well start treating the AI differently! Because if the AI reasons, we shouldn't be using it as a tool, for instance. Or, at least, in the way that we do. So, I just think being clear about these things is very important. It doesn't mean that words don't have different kinds of meanings. But there can be real conceptual and moral risk, I think.
3) On the opportunities of AI
You may think that this is incompatible with what I just said about what I see as the limitations of AI. It's not alive, it's not conscious, it doesn't know things, it doesn't have goals. I am incredibly excited about AI! Every day, and I bore my friends by saying this, I'm sure I've said it to Hollis many times […] and the other people who are around me and know me. I just think it's the most amazing time to be alive.
And one of the reasons for that is I just think AI's going to solve so many things. Like these medical problems. I think also it's going to raise the baseline of human knowledge, because we will have much better access to information. I think, for instance, hopefully the CCP will finally be overthrown in China, because they will no longer be able to use restrictions on what people can know.
I think this stuff is astonishing and amazing, and I'm so grateful and excited to be alive. But I do not think that AI is alive, or that it's conscious, or that it knows things. And I think there's serious cost in talking about it in those ways.
4) In response to the idea of using AI to analyse publicly available data to “surface potentially predatory behaviour”
I'd like to make a distinction about the way in which AI can help us learn truths, like facts of the matter about things. So, the AI might look through the evidence in the case, and come to some more reliable conclusion about whether you did a bad thing, or I did a bad thing. What AI cannot do, and neither can a human do, and nothing can do, is tell us truths about the future. There is no truth about the future.
Prediction — we sometimes treat it in this way. We do not know, and it is not the case… We cannot know whether that person will go on and commit a crime. And if we think that predictions, no matter how reliable they are, based upon patterns that have previously been established — if we start to think that AI can tell us truths about the future, then I think we fall into really serious problems, too.
5) In response to a question about AI and Girardian mimetic desire
I'm just going to say the boring, boring thing again, and say it doesn't have experience, it doesn't have psychological states, it doesn't... I think it doesn't make sense to talk about it like this, whether it's mimetic…
I should probably add, though, that there is another way in which people talk about these things, in which they say something like, “The kind of thing that can lead to an outcome that correlates with the kind of thing that the reasoning thing does”. So, sometimes people say things like, "Algorithms choose things", and what they mean by that is that the outcome of the algorithm is something similar to the outcome that the ‘choosing person’ would do.
I think there's real cost to eliding these things, and not talking about them as separate things. But I'm sure there are some interesting questions, at least along that kind of, like, outcome, functionalist, behaviourist, whatever it is, type way of thinking, in which you might want to apply some of Girard's interesting insights about how desires are formed, or something. I just think the idea of talking about an AI forming a desire is incoherent, is all.
6) In response to an audience question about people being criticised for saying that now is a ‘great time to be alive’
As indeed I did [say that]! They might mean something quite general, like the conditions are such that I have more opportunities than I would have done in the past — some kind of relative claim. I think almost everyone in the world, for instance, has access to better medical care. They can live longer, happier, healthier lives.
It doesn't mean, however, that there aren't individual people who are having a horrible time. If the person who just was assaulted, or the person whose kid has gone missing is asked, “Is it a great time to be alive?”, there's also a very smart sense in which they would say, "No, it's not a great time to be alive." And what they mean is, "It's not a great time to be me." So, I think separating out those things is probably quite useful.
We can totally talk on aggregate and say you have better access to schooling, to all of these other things that meet your basic needs, and also acknowledge that for some people, this particular moment, maybe all of the moments in their life, make living hard for them.
7) In response to an audience question about whether people are currently too focused on optimisation
Yeah, if I can just be allowed a little plug — it's for a reason — I have a Substack called ‘the ends don't justify the means’. This is because I'm strongly opposed to consequentialist moral reasoning. And one of the things, one of the many problems I have with it is, just because something is good does not mean that we have the obligation to maximize that. It can be the wrong thing to maximize that.
8) In response to an audience question about learning information and making predictions about the Sherlock Holmes universe — and whether Bayesian theory can help us in this
So, on the Sherlock Holmes thing, I think all I'm saying is something like, let's imagine that Conan Doyle wrote the sentence, "Sherlock Holmes knows that Dr. Watson is awesome," right? Like, that's... I like reading Sherlock Holmes, they're great. I don't think that just because that sentence is truly in the book, that means that Sherlock Holmes is the kind of thing that can know anything.
The interesting thing, of course, about the Sherlock Holmes universe, qua the writing that Conan Doyle did, is there are “facts” about the “future” — in that page ninety-nine says Dr. Watson dies, and on page three, it is the case that on page ninety-nine Dr. Watson dies.
But, of course, if we apply this kind of pattern thing, and we go through the first one hundred Sherlock Holmes stories, and we try to do some kind of “psychologising” of the character. Sorry, I'm using all the scare quotes! [I was doing a lot of little ‘scare quotes’ hand signals] And we think, look, I can reliably infer from the actions that Conan Doyle has attributed to Sherlock Holmes that if Sherlock Holmes ever meets a woman with blonde hair, who is over five foot seven, and loves singing opera, and is from New Zealand, and all of these things, then Sherlock Holmes will fall in love with her. That sounds like an interesting thing to do. But of course, the hilarious thing is, in the story one hundred, Conan Doyle can just do something entirely different! Such that he punches her in the face. And you're like, "Oh my God, I didn't know Sherlock Holmes would ever punch a woman."
So, I mean, this is just... I mean, this stuff is amusing, but I worry that we kinda lose grip on the important things. Like, what is it to be a human who reasons? And how might we use, for instance, inferential reasoning within our justice system? And we get on to stupid analytic philosophers like Rebecca talking about Sherlock Holmes.
You can watch the whole event here — thanks, Luke Burgis!