The difficulty of self-modeling explains why humans feel like they have free will.
Cross-posted to Substack.
What exactly is free will, and do humans have it? This is one of the main metaphysical problems in Western philosophy, which has been debated for at least two millennia. There are many answers, which have changed in popularity over time. Many hold that God granted us the fundamental freedom to choose our actions, even knowing that some of us might choose badly. Others argue that the probabilities of all events, including human actions, are determined by the laws of Nature; and therefore humans cannot be fundamentally free.
None of these explain why free will is such a persistent question. But it is weirdly persistent! Many people (materialists) believe that only physically measurable Nature exists, and therefore do not believe in souls or an afterlife. In a Christian worldview, the world follows regular laws, so free will only makes sense as a property of the soul. Yet, materialists typically still believe in free will. Why is that?
Because from the inside, it feels like we are free. It is difficult to discard belief in something that you directly and clearly observe, even if it contradicts the rest of your worldview. So we get positions like compatibilism which asserts that it’s technically free will so long as it is caused by the human’s mental state, even if that is deterministic.
Why does it feel like we are free? In this essay I take no position on whether free will is ontologically real[1]. I contend that what people actually mean when they talk about free will is the feeling of unpredictability that the behavior of other humans, and their own behavior, generates. They feel like they have free will because they can’t accurately predict their own behavior in advance. If they could predict their behavior almost perfectly in advance, they would not believe they have free will. Other humans have free will for the same reason. This reframes free will as a testable concept, which aligns with intuitions, and nevertheless exists under materialism.
Models of self and of others
Intelligences are pattern-matching and prediction machines. We attempt to explain the world around us, and form expectations about how objects will behave. Most objects are rather simple: for example, a metal fork mostly does not change shape, it matches the temperature of the things it touches, and falls to the ground if you let it go. Air blows on your face and typically makes you cold. Cloth drapes. Most animals follow their behaviors.[2] All these things have fairly simple behavioral properties, and our brain can model them very well.
One type of thing is difficult to predict: other human beings. The evolutionary arms race that produced intelligence was in part about modeling other humans’ behavior, and in part about making your own hard to predict. Humans have long memories and act according to subtle cues they saw years ago. Humans use language, and may say anything at any point in time. It is not possible to study human behavior like we study rocks or plants or most animals, unless you study traits that are temporally stable or averaged across many people.
Just like we have models for object behavior, we have models for other humans: we may know someone is brave, charismatic, eager to help, shy. We know what catchphrases a friend employs (“this is true”, “wild”). Yet, most human behavior remains inexplicable (especially that of strangers). In many cases even our own behavior is inexplicable. (If not, why do I keep leaving work until the last minute?)
Under this much uncertainty, if you expect there is some underlying regularity but don’t expect to be able to reduce it even with lots of information, the only sensible approach is to mostly give up on predicting the action, and plan for a wide range of possible responses.[3] Hence the feeling, and from there the concept, of free will.
It is possible to somewhat predict your own behavior in a new situation, if you pretend to be in the same situation and reason about what you should do. This still feels like free will because you’re determining a decision, and also in many cases these determinations are wrong.
Implications of reframed free will
Suppose this is true: the feeling of free will comes from the unpredictability in the model of self, or of others. Then, one can say that free will is when this unpredictability is actually there (along with there being something interesting to predict, to exclude randomness). What are the implications?
Free will is a two-place, relative concept
Instead of saying without qualification that some agent ‘has free will’, we have to say they have free will with respect to another agent’s model. It’s also a matter of degree. For example, my dog might not have much free will with respect to me: I know it likes to eat certain foods, it will be happy to play catch, it will enthusiastically want to go on a walk but then actually be tired, etc. However, if it has a model of self (which I think is likely), it is likely not so transparent to itself.[4] Likewise people that know each other very well have reduced free will with respect to each other; but this comes with advantages such as easier cooperation.
Natural phenomena that you don’t understand (e.g. magnetism, to use a previously mysterious example) don’t have free will in the same way, because they turn out to have relatively simple laws that you can learn to understand. That said, people routinely use the presence of entities with ‘free will’ to explain confusing phenomena (ghosts, animist spirits).[5]
Implications for the free will of AIs
Under this new model, it is also an empirical question whether AIs have free will. Do they have a model of themselves and other agents? And does this model sometimes make a mistake, and therefore surprises its own model?
I’m pretty sure the answer to both of these questions is yes. LLMs learn to model human text, which means they learn to model the humans that write it, so they’re good at modeling other entities. It is practically inevitable that they turn their modeling faculties on themselves, therefore having a model of self. However, this self-model is contained within their generating process, and it is thus necessarily less complex than that process. This means that there should be a gap between the self-model and the actual behavior of LLMs, thus generating free will.
Conclusion
Free will is the feeling of unpredictability that opens up between your model of yourself, and your actual actions.
Though the existence of this essay is kind of silly, if you believe that free will is God-given. ↩︎
Lower animal behaviors (e.g. insects) can be very intricate, and scientists study them constantly. But they also have to study elementary particles, because everything is complicated, but this does not give the particles free will. It also pains me to say this, because I don’t want to imply that nonhuman animals therefore have the same moral worth as rocks (they don’t). However, animals are simple enough that, if they were in the ancestral environment, we mostly know how they behave and how to manipulate them to do what we want. (To disastrous moral consequences, in many cases, but that is beside the point.) ↩︎
Interestingly, this would mean true uniform random draws don’t have free will (there’s no regularity that generates them), but pseudo-random numbers do have free will (you can predict them if you know the current hidden state). I think we can posit another rule like “you could not perfectly predict it even in principle, your brain is not big enough” but there might be other objections to this one too. ↩︎
I find it quite likely that dogs self-model in this way. After all, they have to predict the behavior of other dogs in the pack, and of prey. Therefore they probably have a model of the prey and other dogs, and from there it’s easy enough to generalize to the model of I. ↩︎
Nowadays, they might invoke the Efficient Market Hypothesis, since financial markets are unpredictable, and they’re also self-aware (they have derivatives markets about their own prices). ↩︎