Christian List on Free Will, Levels of Reality, and Group Agency

Christian List

Hi Sean. Nice to be on your podcast.

Christian List

Officially my area is described as philosophy and decision theory — that's at least my job description here in Munich. I work at the Munich Center for Mathematical Philosophy, a fairly interdisciplinary place. The people here are philosophers who interact with neighbouring disciplines — maths, physics, economics, cognitive science, AI, neuroscience, and a whole bunch of other things. My own interests have been all over the place, around individual and collective decision-making and the metaphysical questions this raises.

Christian List

I think the MCMP is probably the only place with that particular name and label, but there are other places with related interests. There's a very good place in Amsterdam called the ILLC — the Institute for Logic, Language, and Computation — which overlaps with a lot of what we do. My former department, Philosophy, Logic, and Scientific Method at the London School of Economics, specialises in similar things. So we're a fairly big presence in this area, but by no means the only one.

Christian List

Absolutely. The interdisciplinary nature of the place was one of the big attractions for me when I chose to come here, and I love it.

Christian List

The big distinction in the philosophical landscape is between compatibilism and incompatibilism. There's a long-standing debate about whether we could have free will if the world is deterministic. Determinism is the idea that if we fix the state of the universe at a given point in time, there's only one possible trajectory of future states under the laws of nature — which seems to prevent there being forks in the road, decision nodes where I can do one thing or another. The idea of free will is bound up with having choices between alternative possibilities, and that leads many people — the incompatibilists — to say that if the universe is deterministic, there can't be free will. Compatibilists, on the other side, say that even if the universe is deterministic, we could still have free will, and then try to give an account of free will that makes sense of that.

There are two conventional options on the compatibilist side. One is to say it's simply a mistake to define free will in terms of the possibility of doing otherwise — what matters is not alternative possibilities but that an agent can endorse their actions. If I do something in line with my goals and motives, if I believe in it and want to do it, if it expresses my character, that's good enough for counting the whole thing as free. The other type of compatibilist takes a different strategy: some form of choice between alternative options is important for free will, so we must give an account of how we could do otherwise, but then render this compatible with determinism by redefining what it means to say I could have done otherwise. Take the trivial example: this morning I had coffee rather than tea. I want to say I could have done otherwise, I could have had tea instead. What this second kind of compatibilist says is that it wasn't required that, at the time of choice, it was genuinely possible for me to have had tea — maybe if the universe was deterministic only the coffee trajectory was a real possibility. What matters, instead, is that if the world had been a little different — if I'd had a slightly different back-history, maybe a desire for tea — then the trajectory would have unfolded differently, in the tea direction, and that's good enough for free will. So that second kind of compatibilist redefines "could do otherwise" in terms of a counterfactual conditional. Those are the conventional options.

Christian List

I actually like to reframe the debate. What I've just given you is the standard framing — it's what I present to my students in an introductory lecture on free will. But my point is that the standard debate frames things the wrong way, because it fails to distinguish between different levels of description at which we can think about the world — a physical level on the one hand, and a level of agency on the other. Once we carefully draw the distinction between levels, this opens up new space for defending free will in a more interesting way.

Christian List

My response is that the strategy I'm about to explain gives us a wonderful rebuttal of that objection. Some of the standard compatibilist routes I summarised earlier really can be accused of redefining or watering down free will. If we say free will does not require the possibility of doing otherwise at all, we have definitely watered it down compared to our common-sense understanding — most people outside academic philosophy think some notion of choice is essential for free will, and as soon as you take that away, you've redefined a watered-down version. Even if we redefine "could do otherwise" in the counterfactual way, we've arguably still watered it down slightly, because I want to say this morning I really could have chosen tea instead — not just that if the world had been a little different I would have chosen tea. At the moment of choice, it was open to me: I could have gone the coffee way or the tea way, and I opted for one, but the whole thing was under my control. So yes, some versions of compatibilism do look like they're redefining or watering down free will — my whole point is that if we reframe the debate in the right way, we can avoid that.

Christian List

It definitely goes in the right direction. To anticipate one of the key points: to just ask "is the universe deterministic or not?" — which is how it gets framed in a lot of metaphysics textbooks, and in the influential work of Peter van Inwagen, who developed one of the most famous arguments for incompatibilism — is a somewhat underspecified, ill-defined question. The distinction between determinism and indeterminism can be drawn, but it becomes fully meaningful only once we carefully specify the level of description at which we're trying to draw it. Once we have a system carefully described, and we're clear about the level of description, we can ask, relative to that level, whether we're dealing with a deterministic system or not.

Take people who think some fundamental theory of physics gives us deterministic laws of nature — the mainstream view, at least until Einstein. They'd say yes, there's a level of description at which the physical universe behaves deterministically: fix the trajectory of the physical universe at that level up to a particular point in time, and it admits only one nomologically possible continuation. So far so good. But it's entirely possible that if we re-describe the very same system at a different, more macroscopic level — a sequence of macro states rather than finely described microstates — and ask what the possible macro-level continuations are under the laws governing the system, two or more distinct sequences of macro states could be fully compatible with those laws. Whenever you get that pattern, you have some form of indeterminism relative to the macro level while having determinism relative to the micro level. Systems we study in the sciences admit many different levels of description — particle physics, chemistry, various biological levels, individual cells, larger organisms, and so on — that differ in grain or detail. There's nothing weird about recognising that at some levels you might have deterministic dynamics and at others indeterministic dynamics. In a thought experiment, my friend and co-author Marcus Pivato and I considered — purely for conceptual intuition-pumping — a hypothetical system admitting infinitely many levels of description, ever more fine-grained, with determinism at odd-numbered levels and indeterminism at even-numbered ones. That would be a perfectly coherent scenario — we're not saying the world is like this, but it shows that for such a system you can't say it is deterministic or indeterministic simpliciter. The distinction between determinism and indeterminism is really a level-relative, level-specific distinction.

So, back to your point: yes, I wholeheartedly agree that if someone says fundamental physics is deterministic — which is, I suppose, an open question, but let's call it at least a live possibility — the reason this doesn't automatically refute free will, even on the alternative-possibilities characterisation, is simply that the kind of indeterminism required for free will is not indeterminism at the microphysical level, but indeterminism at what I call the level of agency — the agential level, a completely different level of description. Given this logic of level-specificity, indeterminism at the level of agency is fully compatible with determinism at the level of fundamental physics.

Christian List

Basically, yes, just with a slight twist. You say you don't know how the coin is going to land — true, but that framing makes it sound as if it's just an epistemic phenomenon, something to do with unpredictability or our informational limitations in figuring out the details of the microstate. That would invite the objection that what we really have, when we talk about indeterminism at the level of agency, is just unpredictability at that level — not genuine openness of choices — and so it's not good enough for free will. My response is that it's not the case that we only treat properties, patterns, and phenomena at the most fine-grained microscopic level as real. That would be absurd — if anything you can't pinpoint at the level of particle physics, or superstrings or whatever, is somehow unreal, then a whole bunch of things become unreal: the solidity of surfaces, organisms, societies, universities, unemployment, war. Those are absolutely real phenomena. We need a philosophical view that takes the human perspective seriously, and the common-sense ontology of the world seriously — one that recognises the reality of ordinary macroscopic objects and properties. So if our best descriptions, even from a scientific perspective, of certain macroscopic phenomena lead us to treat them as behaving indeterministically — as with dice or coin tosses, where stochastic modelling is the right approach — we should say that's as scientifically well supported as anything in particle physics, and take it at face value, just as real as what we say about a more fundamental physical level. Marcus Pivato and I have called this phenomenon emergent indeterminism — we shouldn't think of it as merely epistemic, as informational incompleteness, but as ontic, a real phenomenon. That's the twist I'd put on your framing.

Christian List

As I said, we can think about the world at many different levels of description — physics, chemistry, biology, psychology, sociology, economics, micro- and macroeconomics, and so on. What matters is not just that when we move from fine-grained microscopic descriptions to coarse-grained macroscopic ones, we throw out information — we do, we abstract away — but we do this not just because of computational or epistemic limitations. We do it to identify patterns while getting rid of noise: some low-level details are relevant for the lower level but not at all for the macroscopic description, so it's illuminating to extract the signal and focus on the macro pattern. The coarse-graining is done for a principled reason.

Second, it's not just that we coarse-grain when moving from low-level to high-level descriptions — we also use different concepts and categories. Even philosophers of chemistry point out that a concept like acidity doesn't have a neat physical counterpart; it can't easily be translated into purely physical concepts. That becomes all the more significant as we move up the explanatory hierarchy — a whole bunch of psychological concepts can't really be reduced to purely physical, chemical, or biological ones. Psychology, fairly widely accepted, is not reducible to neuroscience.

The most important dividing line, I think, is between sub-intentional and intentional levels of explanation. Physical and chemical explanations are clearly sub-intentional — we use dynamic or stochastic modelling, but no intentional vocabulary: no goals, purposes, beliefs, semantic content, or representations. Once we move into psychology, cognitive science, agency, rationality, mind, mentality, and social phenomena, we make very significant use of intentional concepts. There are good reasons why intentional concepts can't be reduced to sub-intentional ones — a whole separate set of arguments — but if we accept that some of the concepts required for intentional explanations are not reducible to concepts from a sub-intentional level of description, that gives us a good case for saying the higher level enjoys a kind of explanatory autonomy relative to the lower level.

Christian List

One reason a lot of people are tempted by reductionistic positions is that they accept supervenience — a very natural idea: that all the high-level facts that obtain are a necessary byproduct of low-level facts. If we put all the low-level building blocks of this world — all the elementary particles — in place exactly as they are in our physical universe, we have thereby put you and me in place too, with our brains and minds and everything. So everything true about Sean and Christian as people, as agents, is necessitated by an extraordinarily complex set of facts about the underlying physical system. Everyone who accepts that is a supervenience physicalist.

The tempting but invalid argument then says: because supervenience physicalism holds, because high-level facts and properties supervene on low-level ones, we should therefore be able to explain everything in low-level terms, or translate high-level descriptions into low-level ones. That's the step where non-reductive physicalists — the position I find much more compelling — no longer agree. Supervenience does not imply explanatory reducibility.

Here's the nerdiest argument I can give for that. The languages we use to describe systems are all countable languages — they permit us to form as many expressions as there are natural numbers, infinitely many but countably infinitely many. Any standard logic in a textbook is countable; English is countable; German is countable; anything you can write down with a finite alphabet is a countable language. With a countable language you can describe countably many properties. Now suppose the set of possible microstates of a physical universe, or any physical system, is infinite — completely reasonable, since initial conditions could be chosen in different ways, and any sufficiently realistic system admits infinitely many possible microstates. Macroscopic properties studied in the high-level sciences — biology, psychology — always correspond to sets of states, because any property is realised by a set of states. But once you've got an infinite set of possible microstates, you automatically have uncountably many possible subsets. Even if you concede, for the sake of argument, that there are only countably many possible microstates — as many as the natural numbers — there are still uncountably many sets of those states. That means there are, extensionally speaking, far more properties out there than a countable language would allow us to describe. If you have a countable language for talking about microstates, then for purely combinatorial reasons, almost every property corresponding to a collection of microstates will not be describable by any expression in that language.

Now, when we think about macro properties — granting supervenience — any macro property does correspond to some collection of microstates, but it would be a kind of mega-fluke if that collection turned out to be one your low-level countable language allows you to describe. You might be lucky; maybe the only explanatorily salient properties happen to be ones that can be picked out with the low-level language. But this reverses the burden of proof: the proponent of explanatory reducibility would have to show that, and combinatorics are stacked against the reductionist. So that's the nerdy argument for why, even granting supervenience, there's no guarantee that high-level descriptions of properties are reducible to low-level descriptions.

Christian List

A lot of the high-level properties are of interest to us precisely because they play a role in our lives — we, as biological organisms, have been evolutionarily selected to be sensitive to certain properties of the environment and not others. So I don't find it particularly surprising that macroscopic properties are especially well measurable from our perspective, compared with microscopic ones. There are many different levels of description one could use to think about a system, and which levels are explanatorily relevant or interesting depends somewhat on your goals and interests as an observer, your perspective on the world. So I don't find it surprising that we, as humans operating at a particular scale, find certain levels of description particularly salient — who knows, if we were intelligent beings at a completely different scale, maybe we'd be sensitive to different features at a very different level of description. There's a certain amount of interest-relativity to our choice of levels.

Christian List

My proposal is that we should use a checklist of three conditions for assessing whether a particular system or organism has free will. First, intentional agency: only systems that qualify as intentional agents are even candidates for having free will. My water bottle here is not an intentional agent, and isn't even a candidate. Second, alternative possibilities — but these have to be alternative agential possibilities, corresponding to choice options between which an agent can choose. Third, causal control over the resulting actions — philosophers sometimes call this mental causation: the purportedly chosen action must count as having been caused by the agent's mental states, not merely by an underlying physical process of the body, as in a reflex. If the doctor hits my knee with a little hammer, my leg shoots up, and some bodily process causes that movement, but that's not mental causation. Whereas if I intentionally lift the bottle — I had a desire to lift it, formed the intention to lift it, and my arm movement systematically covaried with the presence or absence of those intentional mental states — then I, qua agent, not just qua physical system, had causal control over the resulting action. Free will requires all three: intentional agency, alternative possibilities, and causal control. A system that displays some form of high-level indeterminism wouldn't by itself qualify as having free will, because it might not satisfy the agency condition or the causal-control condition. It's the interplay of these three things that amounts to free will.

Christian List

Large language models are super interesting, but let me set them aside for a second and say something about the general methodology first. I'm quite drawn to, and was originally influenced by, Daniel Dennett's work on the intentional stance. Dennett's idea is that some systems — the solar system, say — we can understand perfectly well by viewing them as physical systems: take a physical stance, write down a physical model, and explain the thing. Other systems — you and me, human organisms — we can't understand adequately just by viewing us as physical systems. If you wanted to make sense of you and me as aggregates of molecules, good luck — you'd have no chance of explaining how we managed to coordinate on the time for this conversation, which we agreed on weeks ago. So Dennett says: with some systems, to make sense of them, to explain and predict them, you've got to take an intentional stance — view them as agents with beliefs, desires, goals, intentions. Once you look at them through that lens, their behaviour becomes intelligible, whereas from the purely physical perspective it would be a complete mystery. Dennett then says that's really all there is to the notion of agency: whenever a system is usefully and voluminously explicable and predictable from the intentional stance, viewing it as an agent, that's an agent.

I think the problem with Dennett's view is that it's a bit too instrumentalist — agency becomes just in the eye of the beholder. To be an agent, on this Dennettian view, is just to be interpretable as an agent: if a thermostat is well interpretable and well predictable as an agent, well, that's an agent, and there's nothing more to the story. I want to go further and say being an agent is not the same as being interpretable as an agent — one is an epistemic criterion, the other an ontic one. If a system is satisfactorily explicable only by viewing it as an agent — if viewing it as an agent is explanatorily indispensable, if we can't explain it adequately any other way — then that indispensability is very strong evidence that we're really dealing with an agent. If I have absolutely no chance of making sense of you, Sean, without viewing you as an intentional agent, that's very strong evidence you really are an agent. It's this indispensability test that I use to assess whether a system should qualify as an intentional agent, from the perspective of the free will debate.

As for AI: it's a complicated issue, because there are big debates about whether LLMs are just stochastic parrots — glorified versions of the text-completion software on our smartphones — or whether they genuinely have agency, or mimic it. You find people on both sides, and the technology is in flux — early on we were very much at the stochastic-parrot end of the spectrum, but the systems increasingly get engineered explicitly for reasoning abilities. What I'd say is: once those systems satisfy the indispensability test — once you can't make good explanatory sense of them without viewing them as genuine intentional systems — we have good evidence they really are intentional systems. My feeling is that, whether we like it or not, we're heading in the direction where more and more systems are like this, which is why this set of criteria — intentional agency, alternative possibilities, causal control — might, whether now or in the future, force us to conclude that some sophisticated AI systems qualify as having free will.

Christian List

The whole consciousness issue raises a separate set of issues. For full disclosure — my theory of free will is fully framed against the background of physicalism, indeed non-reductive physicalism, as we've now discussed at some length. But when it comes to consciousness, I go in a slightly different direction. It's not something I'd antecedently expected — I expected I'd be a physicalist about consciousness too, and I ended up surprising myself by coming out on the other side. In a nutshell, the extra complication with consciousness that simply isn't there with free will and agency is that consciousness is first-personal: phenomenal consciousness is about first-personal experiences, first-personal facts, the first-person perspective. By contrast, when we ask whether a system is an agent, whether it has representations, whether it has alternative possibilities between which it can choose, whether it has causal control or mental causation in relation to its actions — these are all third-personal questions, ones we can in principle formulate from the perspective of an ordinary scientific third-personal observer. So I think the study of free will and agency can be done completely unproblematically from the ordinary third-personal scientific perspective — methodologically, much closer to what David Chalmers would call the easy problems of consciousness. Whereas the hard problem is how we make sense of the first person against the background of this impersonal scientific picture of the world. That's the big structural difference between the consciousness debate and the free will debate.

In the case of free will, I've convinced myself you can be a physicalist — even a physical determinist — and still embrace free will, even with alternative possibilities at the level of agency, once you recognise the levelled architecture of the worldview. I find it much harder to properly accommodate consciousness as first-personal experience — not just as third-personal descriptions, but as first-personal experience — within this purely third-personal, impersonal worldview of the scientist. I've ended up in a place that's actually even more radically non-physicalist about consciousness than Chalmers's view. Chalmers rejects physicalism but doesn't reject the ordinary impersonal third-personal picture of the world — he accepts the framing that we think of a world describable from a scientific perspective, and the debate is about what properties must populate that world to make sense of consciousness. The physicalist says a world with physical properties is enough. Chalmers, in his classic book The Conscious Mind (1996), says physical properties alone are not enough — you need phenomenal properties too, which don't supervene on the physical ones, and that's how he makes sense of the zombie thought experiment and his view that zombies are conceivable and metaphysically possible, if not nomologically possible. But this still concedes that you can look at the whole thing third-personally, impersonally, from the perspective of the Olympian scientific observer studying systems from the outside — what Nagel would call the view from nowhere — and then debate what properties we need to throw in to make sense of consciousness. I'm more on Nagel's side: I think the problem lies with the view from nowhere itself, the impersonal perspective. We can't properly accommodate consciousness in our worldview without recognising the fundamentality of the first-personal perspective, which leads me to think we must postulate genuine, irreducible first-personal facts over and above the third-personal facts in the physicalist worldview. But once you introduce first-personal facts, you're quite radically outside the standard physicalist picture, which raises a whole bunch of new challenges. That's a quick summary of why I think the consciousness debate is structurally different from the free will debate.

Christian List

Great questions. I actually got into the subject of agency in a slightly non-standard way, because I began my career working primarily in social choice theory — the study, and mathematical modelling, of collective decision-making. In the course of that work I collaborated a lot with Philip Pettit, a philosopher at Princeton and the Australian National University. Philip and I did some work together on the aggregation of judgments — a collective-choice problem in the Condorcet–Arrow social-choice-theory tradition — but we also wrote a book together on group agency. In it we discuss whether it sometimes makes sense to view certain collective entities — especially organised, institutional collectives like firms, corporations, collegial courts, universities, even states in their entirety — as unified agents in their own right. An important observation is that in ordinary life we do this all the time: we talk about how the airline treats us, we ascribe goals and intentions to firms. It's not just common sense — we do this in the relevant sciences too. In economics, the theory of the firm views firms and corporations as rational actors and applies rational-choice-theoretic models to explain their behaviour. In international relations theory, entire countries or states get treated as unitary rational actors — in fact, the field of game theory developed significantly after the Second World War in the context of strategic studies, where studying the strategic interaction between the United States and the Soviet Union was a major driver of game-theoretic models. So the idea that corporate entities can be rational agents in their own right isn't just common sense — it's entrenched in the social sciences. That led us to try to develop philosophical foundations for the idea, which immediately raises interesting metaphysical and philosophy-of-mind questions — what do you say about the mental states, or analogues of mental states, of corporate entities — and normative questions about the responsibility those entities can have.

In our book, Philip and I defend a realist theory of group agency: corporate entities can truly be intentional agents, with functional states that qualify as desires, goals, or preferences, and an ability to systematically interact with their environment on the basis of those states — pursuing their desires in line with their beliefs. We don't think those entities are phenomenally conscious — you could think of them as zombie agents, zombie corporate agents. But if you're a realist about group agency, as I am, there's the obvious question of whether we should ascribe free will to those entities too. My theory of free will, especially its three conditions, lets us think about that systematically, because the criteria — intentional agency, alternative possibilities for choice, causal control over the resulting actions — are portable; I didn't custom-make them for the human case. You can ask whether chimpanzees fulfil those criteria, and I think — I'm not a primatologist, but given everything I've read — the answer is yes: it would be very weird if humans came out as having free will and none of the other animals did, not even our closest relatives. Similarly with LLMs and AI systems: if our best explanations force us to say such-and-such an AI system is an intentional agent, because viewing it that way is indispensable — if they force us to say it's a choice-maker, a genuine decision-maker choosing between options, and that it has some analogue of mental causation, with high-level representational states as the controlling, difference-making variables in what it does — then we'd have to say those systems qualify as having free will in this functionalist sense. And exactly the same should be said of group agents: if a corporate entity satisfies the three conditions — intentional agency, alternative possibilities for choice, and causal control, the corporate-level analogue of mental causation — we should conclude there's a functionalist sense in which that entity has free will. This conclusion might sound a little revisionary and weird at first, but it becomes less weird once we recognise free will is not the same as consciousness — we're not thereby forced to conclude corporate entities are conscious; they're not. On reflection, the conclusion should even be welcomed, because those entities are very powerful — they play an enormous role in our contemporary world, they can cause a lot of harm and damage, and their power is much greater than that of typical individuals. So we should have frameworks in place for corporate responsibility, corporate liability, even corporate criminal liability — lawyers take those notions seriously and debate them — and if some form of free will is a necessary condition for fitness to be held responsible, it's a congenial finding that that condition is also satisfied by some of those corporate entities.

Christian List

Free will in this functionalist sense is a necessary but not sufficient condition for fitness to be held responsible. I suggested chimpanzees might have free will, or some version of it, under those conditions, but we don't want to say chimpanzees bear moral responsibility, let alone legal responsibility. So free will, understood this way, is a necessary precondition, and there are additional conditions that also have to be in place — certain normative cognitive capacities need to be present in a system, and certain epistemic or informational conditions must be met too. Moral philosophers have debated in detail what the nature of those conditions is, but free will is an important part of the package — and this analysis suggests corporate entities could tick that box.

Christian List

The observation is right. Suppose you're dealing with a group agent, and you're interested in what beliefs, preferences, or overall collective attitudes to ascribe to it as a corporate entity — you're right that this might be relative to a particular voting method, institutional structure, or aggregation rule that defines the group agent's beliefs or preferences, and those corporate attitudes will be contingent on the aggregation rule in question. But what you can then say is that the group's beliefs or judgments supervene on the individual judgments, relative to the organisational structure that defines the entity — the organisational structure might be written, for instance, in the entity's articles of association: think of this as its software, or operating system, specifying how the thing works. It's the perfect analogy with a computer: the high-level states of my computer also supervene on low-level states, but relative to the operating system running the thing, which is the much more contingent social-level analogue of the nomological constraint naturally given to us in a physical system — there, the constraint is given by the laws of nature, which fixes the nature of the supervenience relation we're interested in; in the corporate case, the supervenience relation is relative to the organisational structure, the system-relative nomological constraint, so to speak.

Christian List

The way of thinking I've summarised here is heavily influenced by social choice theory, the field I started out in — the formal study of aggregation rules and mechanisms. For a long time social choice theory focused mainly on things like preference aggregation, vote aggregation, or utility aggregation, but in the last twenty-five years or so many of us have worked a lot on the aggregation of judgments and beliefs. Social choice theory gives us a wonderful set of axiomatic tools for thinking about the nature of different aggregation methods and comparing them. Very often we proceed by formulating a wish list of desiderata — minimal conditions a good method of aggregation should satisfy — and then use mathematical analysis to ask whether there exist aggregation methods satisfying those criteria simultaneously. Sometimes we get a negative result, an impossibility result — Arrow's impossibility theorem is the best-known example — but sometimes we get a possibility result: a bunch of conditions uniquely picks out a particular aggregation rule, and to the extent we find those criteria obvious, intuitive, and normatively compelling, that gives us strong reasons to adopt that rule. One of the things Philip Pettit and I do in our book is import ideas from social choice theory to suggest a systematic way of thinking about the organisational structures of group agents.

Christian List

Wonderful, that's great.

Christian List

That's a good point, and this is another area where the technology is in flux — these systems differ quite a lot in how stable or volatile their personality is. I hear that companies like Anthropic devote a lot of resources to developing the personality of Claude, and employ philosophers to give philosophical input on it. So to some extent this is a contingent issue that's in flux, but here's where it matters. From a normative perspective, agency and responsibility is also a kind of forensic accounting device, used to make stable action attributions across time. If I hold you responsible for something you did in the past, there's a background system of assumptions about personal identity playing into that judgment. If I had to operate on the assumption that there are just very short-lived time slices of Sean, with no extended personal identity, that would significantly change our views and verdicts on responsibility attribution in a diachronic setting — and I think the point you raise is really relevant to that debate. If you have an AI technology that is extraordinarily attitudinally volatile, maybe you just have these very short-lived time slices as the meaningful units of agency, and you jump to a completely transformed instance of the system — which breaks our normal moral understanding of agency and responsibility, familiar from the human case. What this highlights is that once new intentional systems come on stream — corporations have been around for a while, though still relatively new or modern against the history of humankind; AI is a much more recent phenomenon — there are things that turn out to be somewhat parallel to the human case, where lessons from the philosophy of mind, philosophy of action, rationality and choice theory, or moral philosophy carry over. But there are also new features of these systems that force us to revise, adjust, or update our systems of concepts and categories — and that's where we'll have to go with AI systems too.

Christian List

Not. (laughs)

Christian List

Agency, on my picture, is a functionalist notion, and has no built-in consciousness requirement or condition — some kind of functional awareness, sure, but not phenomenal consciousness. There's a lot of debate at the moment about whether AI systems are conscious or sentient, without a complete consensus, but my hunch is that the somewhat dominant view is still that current systems, despite their remarkable functional capacities, are not yet phenomenally conscious. All of this is in flux, and it's another example of how this new technology is a test case for both metaphysical and neuroscientific debates about consciousness.

Christian List

Exciting times, but also scary times — a bit of both. It's definitely an interesting time to do philosophy.

Christian List

Thanks, Sean. It's been fun.