Does ChatGPT have free will? And do humans?
Have you ever asked yourself if ChatGPT has free will? You might think that the answer is obvious. After all, ChatGPT is a machine who follows rules. It might look like it can choose, but ultimately this is just an illusion produced by code.
For a long time, I assumed the question ended there. Humans have free will but obviously AIs don’t. But when I reflected more on what human choice actually involves, I was less and less convinced by the contrast.
Intuitively, I think most of us would agree that AI is programmed while humans are free. But is this really a case? After all, we humans are also shaped by our circumstances: our upbringing, language, culture, trauma, privilege and also luck play a crucial role in making us who we are. Before we are able to reflect on our decisions we already inherit ways of seeing the world – our ‘training data’ you might say. Even the possibility of being a good or a bad person seems to be subject to circumstance.
So perhaps instead of asking “Is ChatGPT programmed?” we should ask what kind of autonomy is needed for moral responsibility, if nobody starts unshaped?
Bernard Williams’s idea of moral luck is interesting to consider in the discussion of free will and moral responsibility. Moral judgment, Williams argues, often responds to factors the agent cannot control, even though morality often pretends it should not. We praise and blame people for actions that are shaped by circumstance, temperament, and coincidence because we cannot help responding to what actually happened, even if it wasn’t what someone intended in a frictionless world.
If Williams is right, we are never fully autonomous. Our concept of responsibility already involves contingency because humans are always shaped by forces they did not choose. AI systems too are shaped by training data they did not choose. If neither begins from scratch, where exactly does responsibility enter? Where can we draw the line between genuine moral responsibility and “programmed response”?
The tempting answer: neither of us is free
One response to this question is to accept determinism, which claims that free will is just an illusion. Our choices are just the effects of previous physical causes stretching back beyond our awareness. So no matter what, I was going to cook a steak for myself for lunch today because of my biology, my past experiences, my environment, and the state of my brain moments before the decision. Once you zoom in on this causal chain, there is never a clear point where an uncaused “self” steps in to will an action.
If this is right, then ChatGPT looks less alien. It responds to prompts because it has been trained on patterns. Humans respond to stimuli because we have been trained by experience and have evolved into our current states. Some critics argue that there is a degree of randomness in the world, drawing on quantum physics, however this doesn’t necessarily entail freedom. An LLM uses probabilities whilst a human brain is noisy and occasionally unpredictable. Unpredictability is not autonomy though as no choice is still being made by our own capacities.
Determinism also appeals to ethics because if nobody ultimately deserves blame due to their actions being out of their control, perhaps punishment should focus on protection and rehabilitation rather than retribution. However the ethical cost is ultimately very high. The more we discredit human agency, the harder it becomes to justify why anyone should answer for anything at all if everything was simply pre-determined.
P.F. Strawson offers an alternative to the determinist picture by focusing on ‘reactive attitudes’ like gratitude, resentment, forgiveness or indignation. For Strawson such attitudes are a crucial aspect of what it means to be a moral being amongst others.
Strawson does not deny that causal explanations exist. His point is that demanding a God’s-eye form of ultimate freedom misunderstands what responsibility is for. Responsibility is not about escaping causation. It is about participation in a social practice where we hold one another accountable, offer reasons, and expect responses in return which are not completely ridiculous.
This doesn’t settle the philosophical worries just yet. But it changes what needs to be shown. We do not need to prove that humans are uncaused authors of themselves. We need to identify what capacities make someone a suitable participant in these practices.
ChatGPT does not have free will, but not because of determinism
All LLMs, including ChatGPT, take large sets of data, identify statistical regularities and then generate their answers according to the most likely continuation of the input. ‘Stochastic parroting’ is the term some experts use for this: LLMs are very good at mimicking fluently without actually understanding what they are saying. Nonetheless, it is also true that LLMs consistently impress us with their ability to do reasoning-like task.
But the mechanism is important if we want to assess ChatGPT’s responsibility. ChatGPT’s answers are not shaped by a character, a past or evaluative commitments. They are simply the product of completing patterns in the most probable way.
OpenAI is very aware of this and does not ascribe autonomy to its current models. ChatGPT cannot really plan, improve itself or pursue projects in the way a human agent does. When ChatGPT seems to be aiming at a goal, this goal has been specified from the outside, by designers and users, and the way ChatGPT can reason is constrained by human programmers.
So we should not ascribe free will to ChatGPT, not because it is determined – we are too after all – but because ChatGPT cannot be considered morally answerable: it cannot affectively recognise obligations to others, it cannot feel guilty or hurt when blamed, we cannot ‘make up’ with ChatGPT when it fails us. ChatGPT never takes a moral risk, because it has nothing to lose. It cannot participate in our moral community, because it does not have a biography that would make confessions, apologies or improvements matter. It can only simulate, but simulation just is not the same as participation.
But humans are also shaped, so what is the difference?
We have already seen that our actions are shaped by luck and circumstance, but surely this is not the only factor. We do not just generate outputs (i.e. behaviour), we also actively develop a character, we can practise self-control and we feel empathy. Unlike ChatGPT we can ask ourselves why we acted in the way we did and we can revise our reasons. We have the ability to see that there is a gap between who we are and who we should be, even if we did not get to choose the factors that shaped us.
In Giving an Account of Oneself, Judith Butler states that what it means to be a person is not a total mastery of oneself or a perfect knowledge of one’s motives. Instead, Butler argues that what makes someone a person is answerability: the possibility of being held accountable by others and to take that accountability seriously even while recognising that the way we understand ourselves is also incomplete and shaped by society.
A similar move is made by compatibilist philosophers of free will: they argue that being responsible does not have to mean that we could have acted otherwise in a possible world with the same conditions. Instead, compatibilists focus on practical control: we are responsible if we act by responding to reasons accompanied by understanding, and thus if we are able to justify our actions.
Still, neither Butler nor the compatibilists can fully alleviate the worries about moral luck. Ultimately it seems like what we deserve is always shaped by our circumstances. We simply cannot neatly separate actions and character from the contingent factors that – at least partly – bring them about.
Still, as demonstrated by philosophers like Strawson, human can participate in a moral community in a way that is impossible for current LLMs.
The real danger is under-crediting ourselves
It is very important that we don’t fall into the pitfalls of what philosophers of technology call “moral deskilling”: where we offload judgement to systems that can generate plausible reasons until we lose the habit of reasoning together. There is a danger that even if we do accept that CHatGPT lacks free will, we may start treating it like an agent, which can slowly erode human responsibility.
As AI systems become increasingly complex, a “responsibility gap” can form where harms occur through systems whose behaviour is hard to predict, and responsibility dissolves into a blame game without clear answers. The model did it. The user did it. The developer did it. Nobody did it.
This is where a firm philosophical stance matters. If we treat AI as responsible, we risk laundering human choices through machine outputs. If we treat humans as not responsible because of luck, we risk the same laundering from the other direction. Therefore it is crucial that we accept that we have a degree of free will, and accept moral responsibilities without blaming it on luck all the time whilst also taking responsibility for the creation and use of AI.
Where I land
Ultimately I think that ChatGPT does not have free will. ChatGPT cannot be morally responsible, not because it is determined, but because it is not a person who can reflect on herself, understand and be bound by moral obligations, or interact with others in a way makes responsibility intelligible in the first place.
We humans are responsible even when our options are highly constrained by contingent factors. While our autonomy may not be perfect and always subject to luck, we are able to recognise reasons, be held accountable for our actions by others, to reflect on ourselves and feel that what we do is morally significant.
That is what makes the difference between genuine responsibility and ChatGPT’s programmed responses: not the presence or absence of causes, but a life that can be answered for.