Who Settles the Next Word?

Search for AI free will and the suggestions that come back include giving AI free will and AI that has free will, as if free will were a feature a lab could switch on. The question underneath is an old one. Philosophers have argued about free will for two thousand years, and they have three main answers. Each of them, applied honestly to a language model, says something different. None of them says that a chatbot has free will because it sometimes surprises you.

This essay takes the three answers one at a time, states each at full strength, and then asks what it implies for a trained model. It sits beside Which Soul Do You Mean?, which asks what sort of thing a machine would have to be, and The Moral Status of Minds We Might Build, which asks what we could owe one. Free will is the third question in that set: whether a machine could ever be the one responsible for what it does.

How the sources were handled. The three positions are stated from the Stanford Encyclopedia of Philosophy, the standard peer-reviewed reference, in four entries opened on 2 October 2026: “Free Will”, “Compatibilism”, “Incompatibilist (Nondeterministic) Theories of Free Will” and “Skepticism About Moral Responsibility”. Where a philosopher such as Frankfurt, Fischer and Ravizza or Pereboom is named, the view is given as the encyclopedia states it, not from memory of their books. Every quotation is recorded in this site’s source ledger. None of these entries discusses language models. What each view implies for a model is my reading, and it is marked as mine.

What “free will” is a name for

The encyclopedia’s main entry opens with a definition that rules out a lot of loose talk at once:

The term “free will” has emerged over the past two millennia as the canonical designator for a significant kind of control over one’s actions.Timothy O’Connor and Christopher Franklin, “Free Will”, Stanford Encyclopedia of Philosophy

The kind of control in question is the kind that would make someone deserve praise or blame. The entry calls this basic desert, after Derk Pereboom: the agent deserves blame or praise just because she did the act, knowing its moral status, and not because blaming her would have good effects (O’Connor and Franklin, “Free Will”). So the question is not whether a system is unpredictable, or complicated, or able to say “I choose”. The question is whether it has the control that would make it, and not only its makers, answerable for what it does.

What a model does when it answers

Any answer has to start from the mechanics, which are simpler than the marketing suggests. A language model produces its reply one token (a word or part of a word) at a time. For each token, it runs a forward pass: the text so far goes in, a fixed set of trained numbers (the weights) is applied, and out comes a score for every token in its vocabulary. That step is arithmetic. Given the same weights and the same input, the mathematics gives the same scores.

The scores are then turned into probabilities, and a separate piece of code called the sampler picks one token. A setting called temperature controls how adventurous the pick is. At temperature 0 the sampler always takes the most probable token, which is called greedy decoding. At higher temperatures it draws at random from the distribution, so less likely tokens sometimes win. The random numbers normally come from a pseudo-random generator. Given its starting value, the seed, that generator produces the same sequence every time.

In practice, chatbots vary even at temperature 0. A team at Thinking Machines Lab sampled 1,000 completions at temperature 0 from one open model, with the prompt “Tell me about Richard Feynman”, and got 80 different completions. All of them agreed for the first 102 tokens. Then 992 said Feynman was born in “Queens, New York” and 8 said “New York City”. The cause was not anything like a decision. The server groups requests into batches of different sizes, and that changes the order in which the hardware adds up numbers, so the floating-point results come out very slightly different. With computing kernels rewritten so the batch size could not affect the result, all 1,000 completions were identical (Horace He and Thinking Machines Lab, September 2025).

So when a chatbot gives different answers to the same question, the variation comes from one of two places: a dice roll that was added on purpose, or the order in which the hardware happened to add numbers. Everything else is the model’s trained function applied to its input. Keep that picture in mind for all three answers below.

Answer one: libertarianism, freedom needs an open future

The first answer says that determinism and freedom cannot both be true, and that we are free, so determinism is false. (This “libertarianism” is a position about the will. It has nothing to do with the political one.) The encyclopedia states the shared core:

Incompatibilists hold that free will and determinism are mutually exclusive and, consequently, that we act freely (i.e., with free will) only if determinism is false.Randolph Clarke, Justin Capes and Philip Swenson, “Incompatibilist (Nondeterministic) Theories of Free Will”, Stanford Encyclopedia of Philosophy

Libertarians do not think any indeterminism will do. The same entry says they disagree amongst themselves about what else, besides indeterminism, is required for free will. According to the “Free Will” entry, Most libertarians endorse an event-causal or agent-causal account of sourcehood. On the event-causal view, a free decision is caused by the agent’s own reasons, but not deterministically. The reasons make the decision more or less likely, and nothing before the moment of choice settles which way it goes. The agent-causal view says that is not enough. The agent herself, a persisting thing, must cause the decision. The entry puts the claim this way: an agent is in a strict and literal sense an originator of her free decisions, an uncaused cause of them (Clarke, Capes and Swenson).

Libertarians also face an objection that matters here: the problem of luck. If nothing before the choice accounts for which way it went, the difference looks like chance. The entry gives the conclusion of that argument: And if the difference between the agent’s making one decision and her instead making another is just a matter of luck, she cannot be responsible for the decision that she makes. Libertarians have replies. One is that reasons that raise the probability of a decision and nondeterministically cause it are enough to account for the decision (in a sense that makes it not just a matter of luck). But no libertarian answers the luck objection by saying chance is freedom. The whole effort goes into explaining why their kind of indeterminism is not mere chance.

What it implies for a model (my reading). Start with the easy part. A model running greedy decoding on fixed hardware is deterministic, and so is a sampled model with a fixed seed. On the libertarian view, neither can be free, for the same reason a determined human could not be. Turning the temperature up does not change that, because the random numbers come from a seeded, deterministic generator.

Suppose a lab took the luck objection’s advice literally and drew the random numbers from a physical source, such as electronic noise, so that each pick was truly undetermined. The structure would then look a little like the event-causal picture: the forward pass computes something like reasons-weighted probabilities, and an undetermined event picks among them. The resemblance is shallow. The indeterminism sits in a separate component bolted on after the model has done all its work, and it rolls a weighted die. That is the luck objection in its plainest form. The event-causal libertarians’ answer depends on the agent’s reasons being what causes the outcome. Here the sampler is a random number generator that consults no reasons at all. And the agent-causal view needs a persisting agent that is itself the cause. A model’s weights persist, but each conversation is a separate run, and nothing in the system is a candidate for an uncaused cause. So on both libertarian views, “giving AI free will” by adding randomness gets the theory backwards. The randomness is the part that has to be explained away.

Answer two: compatibilism, freedom is how the choice is made

The second answer denies the premise the first one starts from.

Compatibilism is the thesis that free will is compatible with determinism.Michael McKenna and D. Justin Coates, “Compatibilism”, Stanford Encyclopedia of Philosophy

For a compatibilist, the question is not whether your choice was determined. It is whether the right kind of process inside you made it. The “Free Will” entry describes two main models of that process.

Reasons-responsiveness. The first and, in the entry’s words, perhaps most popular model asks whether the process that produced the act responds to reasons. The entry calls John Martin Fischer’s the most detailed version, including his joint work with Ravizza (1998). The entry summarises it: Fischer and Ravizza maintain that moderate reasons-responsiveness consists in two conditions: reasons-receptivity and reasons-reactivity. Receptivity is recognising reasons, and the entry says it depends on capacities such as being capable of understanding moral reasons and the implications of their actions. Reactivity is acting on them, and Fischer and Ravizza ask only for a weak form: there must be some possible situation, under the same laws of nature and with the same mechanism, in which a sufficient reason to act otherwise would lead the mechanism to act otherwise. They add a second condition about the agent’s history. She must have come to take responsibility for the mechanism. In the entry’s summary that means she believes she is an agent when acting from it, believes she is an apt target for blame and praise for what it does, and holds both beliefs on the basis of her evidence.

Identification. The second model, beginning with Harry Frankfurt’s 1971 paper, asks whether you act from desires you identify with. The “Compatibilism” entry explains Frankfurt’s version with two addicts. Both want the drug, and both also want not to take it. The difference lies in a second-order volition, a desire about which of their desires should move them.

The unwilling addict does not take the drug of her own free will since her will conflicts at a higher level with what she wishes it to be. The willing addict, however, takes the drug of her own free will since her will meshes with what she wishes it to be.Michael McKenna and D. Justin Coates, “Compatibilism”, Stanford Encyclopedia of Philosophy

What it implies for a model (my reading). Compatibilism is the only one of the three answers on which a machine could have free will through good engineering, so it repays a close look. Determinism costs the model nothing here, and neither does temperature. What matters is whether the process that produces its answers is reasons-responsive.

There is a real case for reactivity. Tell a model that the dinner guest has a nut allergy and its recipe changes. Tell it the code must run on an old version of a language and its answer changes. A reason in the input leads to a different output, and the same mechanism, the forward pass, does the work in both cases. That looks a lot like the weak reactivity Fischer and Ravizza describe.

Receptivity is harder. It is meant to be a capacity, understanding reasons and what an action implies, and a capacity should show up reliably, not only in the one prompt that happens to state the reason. Models can be inconsistent in a way that strains this. A rewording that changes nothing important can change the answer. A model can refuse a request in one framing and grant it in another. Whether that is an occasional lapse, as human reasoning has too, or a sign that there is no stable pattern of reasons underneath, is an open empirical question.

The history condition is harder still. A model will readily produce the sentence “I am responsible for that answer”. But Fischer and Ravizza ask for beliefs held on the basis of evidence, which is the question of whether there are beliefs behind the words at all. The essay on the Chinese Room covers that problem. A model’s self-descriptions are also among the things its developers train. It may say it is responsible, or that it is a tool with no will of its own, depending on which answer its developer rewarded.

Frankfurt’s model raises the same difficulty from another side. Does a model have anything like second-order volitions, a settled preference about which of its tendencies should win? Labs do train models to follow stated principles over a user’s request. That is a kind of priority among tendencies, and it is the nearest thing to a higher-order will the system has. Whether it amounts to the model wishing its will to be one way is not something its outputs can tell us.

Whose reasons? The manipulation argument

The strongest objection to compatibilism is the one that fits a language model best. It is called the manipulation argument. The “Free Will” entry states its general form: it will seem that agents can be manipulated into satisfying these conditions, and so be manipulated into meeting whatever conditions a compatibilist proposes, and, yet, precisely because they are manipulated into satisfying these conditions, their freedom and responsibility seem undermined.

The best-known version is Pereboom’s four-case argument, which the “Skepticism About Moral Responsibility” entry sets out. In one case, a man called Plum was programmed from the start of his life by a team of neuroscientists. They set up his reasoning so that it is often egoistic, and in the end he kills a man named White. He meets all the compatibilist conditions:

Plum has the general ability to regulate his actions by moral reasons, but in his circumstances, due to the strongly egoistic nature of his deliberative reasoning, he is causally determined to make his decision to kill.Gregg Caruso, “Skepticism About Moral Responsibility”, Stanford Encyclopedia of Philosophy, setting out Pereboom’s case

Each later case is a little more like the fourth, an ordinary action causally determined in the natural way. The challenge is to say where, along the way, responsibility switches on. The “Free Will” entry describes the two ways compatibilists reply. On the soft-line reply, they look for a relevant difference between manipulated agents and genuine ones, as Fischer and Ravizza’s history condition tries to do. The entry’s answer is that the manipulator can simply be imagined arranging that condition too. On the hard-line reply, compatibilists accept that the manipulated agent is free and responsible, and argue that the manipulation is worrying only while the manipulators keep interfering with how the agent develops. If they create a person and then let that life unfold, the worry is supposed to fade.

What it implies for a model (my reading). For humans, Plum is a thought experiment. For a language model, it is a description of how the system is made. A model’s dispositions come from three layers of other people’s choices:

• Pretraining. The model learns to predict text from a very large body of human writing, which someone selected and filtered.
• Fine-tuning. It is then trained toward the answers human raters preferred, as in OpenAI’s InstructGPT work (Ouyang et al. 2022), and toward rules its developer wrote.
• Deployment. At run time, an operator’s hidden instructions, the system prompt, can change what it will and will not do in a given product, and the user usually cannot see them.

The soft-line compatibilist needs some feature of history that separates a free agent from a programmed one, and a model is programmed in the most literal sense the language allows. The hard-line compatibilist can accept a manipulated agent if the manipulators create it and then step back. With a model they never step back. The weights are retrained, and the operator speaks before every conversation. So the question of whose reasons a model acts on has an unusually concrete answer: its developers’, its raters’ and its operators’, filtered through whatever the training produced. Compatibilists disagree about whether that rules out freedom. All of them would at least ask it first.

Answer three: hard incompatibilism, nobody has it

The third answer agrees with each of the other two about what is wrong with the other one.

One of these positions is hard incompatibilism, which maintains that whatever the fundamental nature of reality, whether it is deterministic or indeterministic, we lack basic desert moral responsibility.Gregg Caruso, “Skepticism About Moral Responsibility”, Stanford Encyclopedia of Philosophy

The argument, as the entry sets it out, goes through the options. Against compatibilism, hard incompatibilists say there is no relevant difference between being determined by natural causes and being determined by manipulators, which is the four-case argument. Against event-causal libertarianism, they press the luck objection. They grant that agent causation could in principle supply the needed control, but argue that it cannot be reconciled with our best physical theories. The incompatibilist-theories entry gives Pereboom’s position in one sentence: he has argued that we can have free will only if we are agent-causes, but that the evidence is against the existence of agent causation.

This is not a view that nothing matters or that no one should be held to account. What it gives up is desert, blame earned just by having done the thing. What it keeps is a forward-looking kind of responsibility. The “Skepticism About Moral Responsibility” entry says that on Pereboom’s account this is grounded in three non-desert invoking desiderata: future protection, future reconciliation, and future moral formation (Caruso, “Skepticism About Moral Responsibility”).

What it implies for a model (my reading). On this view, “can an AI have free will?” gets the same answer as “can a person?”: no, because nothing does. The machine is not a special case. It is the general case with the workings visible. It is also worth noticing that the only kind of responsibility a hard incompatibilist keeps is the kind we already apply to models. When a model behaves badly, nobody thinks it deserves to suffer for it. It is retrained or restricted to protect people later and to shape how it behaves next time. That is forward-looking accountability with the desert taken out. The view gives no comfort to anyone who wants to blame the model so as not to blame its makers. The makers lack basic desert as well, on this account, but they are the ones whose future conduct can be protected against, reconciled with and changed.

Where the three answers disagree

Here is each answer’s verdict on today’s chatbots, as I read it. None of these is this site’s verdict, and the disagreements are real.

• Libertarian. Not free. Its outputs are either determined or decided by a weighted die that consults no reasons, and on this view chance is not control. A machine could be free only if it were an agent whose own reasons, or whose own self, settled an undetermined choice, and nothing in current models is built that way.
• Compatibilist. Not yet, though possible in principle. Determinism and temperature do not matter. Whether a model’s answers come from a stable, reasons-responsive process is partly testable, and the evidence so far is mixed. Its history is the hardest problem, and compatibilists divide over whether being designed rules out freedom.
• Hard incompatibilist. Not free, and neither are you. The useful question is forward-looking: how to protect people, repair harms and correct behaviour, which is mostly a question about the humans who build and deploy the system.

The three disagree about three separate things, and each is a different place to push back. The first is whether determinism matters. Libertarians and hard incompatibilists say yes, compatibilists say no. The second is whether indeterminism helps. Libertarians say yes, of the right kind. Hard incompatibilists say it adds only luck. The third is whether being designed defeats freedom. That is the manipulation argument, and it divides compatibilists among themselves. A reader who wants to settle the AI question has to settle at least one of these first. No fact about language models will settle it for them.

Two things hold on every view. First, nobody’s verdict depends on what the model says about itself. A model saying “I chose this freely” is a trained output, and each theory asks about the process behind it, not the sentence. Second, “giving AI free will” is not a dial. On the libertarian view, adding randomness takes the system further from freedom. On the compatibilist view, freedom would be a property of how reliably the system responds to reasons, which is achieved slowly and checked from outside, not switched on. On the hard-incompatibilist view, there is nothing to give. If a machine ever did come to deserve some share of the responsibility, the next question would be whether it could also be wronged. The Moral Status of Minds We Might Build takes that up.