The Freedom of Pac Man
Pac-Man was unleashed on the world when I was 8 years old. I, in turn, was unleashed on Pac-Man a few years later when I discovered him in a dimly lit, rather smelly shop close to home. For a non-sporty pre-teen, this was the next best thing to suddenly waking up with the ability to swing a cricket bat or kick a football.
Soon, I was a regular visitor to that little hole in the wall. Eager to part with fistfuls of coins, in return for the pleasure of taking control of the life of this little pizzotic character. I exulted as he raced down a corridor gobbling dots (paku paku — the Japanese onomatopoeic word meaning to eat rapidly, gives the game its name). I experienced the thrill of seeing Pac Man turn adroitly, just in time to avoid the four ghosts who hovered apocalyptically over the scene. When one of them ate him, I mourned; but it was mini-Easter when I poured more coins into the slot and he sprang back to life.
Now, if someone had suggested that the credit for my triumphs in the game belonged to Pac Man and not me, I would have unhesitatingly considered that person an idiot. After all, the scores that were chalked up in flashing characters right next to the “All Time Top Score” were indisputably mine. Pac-Man acted with great skill in his complex little world, biting into his bytes, fleeing his pixelated arch-nemeses. But to suggest, except metaphorically, that he did any of these things would be absurd. He dodged and ducked when I decided, ate when I thought it appropriate. His successes were mine and I would (reluctantly but surely) have admitted, so were his failures.
I had noticed that he seemed to play disquietingly well in “demo mode” even before my coins clinked in and gave me control, but that surely was due to the programmers who wrote him. So even if the fatuous smile he wore indicated self-satisfaction, there was little reason for it. His sense of freedom, if he had one, was surely illusory. Pac Man was not the conscious, uncaused cause of anything; his every action was dictated by me (or his coders).
Decades later, I became familiar with the idea that my own freedom was perhaps as illusory as Pac Man’s. True, I darted about in a complex world of my own — devouring opportunities (paku paku) and fleeing threats. True, once I decided to do something, I could often do it with no external constraints. And true, most of the time, the things I did were not under duress. But why did I decide to do this thing in the first place and not something else? I may have been mostly free to act on my impulses, but was I free to choose those impulses? Why could I muster the will power to avoid dessert but not that needed to go to the gym? Why did I like some people and things but could not stand others? Why could I not excel at math or write as well as the best in my class, even when we were all too young to even imagine there was an alternative to doing all the practice the teachers set us? If we cannot choose our talents, then why assume we can choose our talent for working hard or for perseverance or indeed our success?
Today, Pac Man is an archaic curiosity in the world of computer games. Pac Man behaved according to a set of rules laid down by his programmers. These included the possibility of seeking input from me- the player, but once these inputs are given, the programmers had spelled out exactly what Pac Man was expected to do (and he did it). Deep Mind’s programs of the “Alpha Go” family, in contrast, have learnt their rules inductively from observing thousands, perhaps millions of games of Go (including those they played against copies of themselves ). Programmers did not write those rules down. Yet, it would still be absurd to suggest that Alpha Go has free will. Let alone being “conscious”, it is certainly not the “uncaused cause” of anything.
What Alpha Go does is determined by its model architecture — the meta-rules that bound what sort of rules it can learn; the training data- the past games it has observed, and a touch of randomness to prevent the algorithms from getting stuck in a rut when learning from the data. If you want to give anybody credit for Alpha Go’s success, give it to the programmers for writing such flexible meta-rules and exposing Alpha Go to such a vast and useful set of training data.
One can build a version of Pac Man that uses deep learning to induce what he should do through learning from experience, rather than have those rules specified in advance. This Deep-Pac-Man, even with no human controller in sight, would steadily but surely learn to play an awesome game. He might look like a truly impressive free agent to the four ghosts who are stuck with a fixed set of rules at birth, but we know better; that he is not only acting out the rules embedded by a coder does not imply that Deep-Pac Man is an uncaused cause of anything. Rules about how to form rules (about how to form rules (about how to form rules…)) are rules too. They determine behavior only with inputs from the environment, but this does not mean either that the rules are unimportant, or that the behavior has uncaused causes.
The fact that Deep-Pac-Man’s creators are not constantly pulling the strings does not mean there are no strings.
But are we any different? We neither choose our preferences, nor the abilities with which we can fulfill them. Some combination of the wiring of our minds at birth (model architecture) and the subsequent life experiences we gather (training data) seem to produce both. You could give your parents credit for these, but they in turn are equally the products of their own model architectures and training data.
To the best of our knowledge, there are no uncaused causes in this world, except perhaps quantum uncertainty, and it would be absurd — at least presumptuous- to take personal credit for that. Further, we know that we are largely unconscious of the many causes that motivate us to do something. Therefore I suspect I am no more a conscious uncaused cause of anything than Pac Man was.
In fact, I am not sure I have ever met anybody who is.
Yet the belief that we are the conscious uncaused causes of the things we do is unshakably strong and seems to be widely held. If it is only illusory, is there a reason why it may be so widespread? We know that to design a learning system (whether the designer is human or natural selection), it is useful to endow it with the ability to distinguish feedback that differentiates more from less useful actions (a.k.a. “learning from experience”), as well the ability to learn from its peer systems in the same task environment (a.k.a. “social learning”). Perhaps something about the process of learning from feedback works more effectively when the system feels “ownership” for the actions that cause it. I learn from my failures and regrets because they are mine. By the same token, it could be that feedback on harmful consequences to others is better processed by a system capable of “blaming” itself. Perhaps “guilt” — also an illusion- is a useful mechanism for producing individual learning and action that benefits the group (possibly at the expense of the individual).
Perhaps, is it easier for systems to exchange the fruits of their individual learning with each other if they“own” their actions and their consequences- social exchange of complex beliefs may be easier when couched as first person narratives. After all, we do love telling each other our stories. In this view, the illusion of free will is part of the machinery that allows us to learn (on our own and from each other). If any of this is true, we might want to build self-driving cars and swarms of drones to explore other planets with the same illusions.
If these seem like idle (and perhaps under-informed) philosophizing, here is a more practical question: what should we do if we do not have free will? Sadly, the question feels oxymoronic. The proposition that we have no free will, should have zero normative consequences: if true, there is nothing that you can choose to do in response to this realization other than what you would do anyway. But we can say that this realization may lead to a view of our social world that is bathed in a strange and unfamiliar light.
We should see a world in which our actions matter, even though we do not author them. I may not be able to choose what I want to do, but what I do will in part cause what you do next. Our actions have consequences even if we did not originate them but reflecting on this deep truth changes nothing (and we did not have much choice in whether we reflected on it at all). Perhaps we can find comfort in the observation that (most) engineers are not kept awake by the idea that their design choices are as determined as the effects of those choices on the physical systems they design. Why should the rest of us be any different?
We might glimpse a world in which no one can ever be blamed, but some must nevertheless be punished. If nothing you do is truly your choice, it makes little sense to blame you for what you do. But it may make a lot of sense for you to be punished to prevent you (and others, through deterrence) from doing it again. Alpha Go learns not only from positive but also negative reinforcement through a “reward” function. Its authors may not use words like “blame” or “regret” to avoid sounding naively anthropocentric, but the principle is the same. Punishment is a useful mechanism for inducing learning that produces appropriate behavior but blame and shame in the usual human sense seem silly if the perpetrator did not exercise any choice. The ideal punishment for the most heinous crime that an incorrigible criminal undertakes could be the secret banishment of the perpetrator to a comfortable but remote island. (The rest of the world however should be told that he was executed most horribly).
We could perceive a world in which the notion of a meritocracy seems truly dubious. After all, if you did not author your success, then why do you deserve its fruits? Why should we draw a line separating your genetic endowment from everything you did after when deciding what to give you credit for since you chose neither? Fairness as “just desserts” becomes as vacuous as “blame”, though analogous to punishment, rewards may nonetheless be useful.
To sum up, how you will react to these ramblings, I believe, must fall into one of four categories. First, you might not be convinced, and you move forward in your life, your belief in your freedom (and the irreducible difference between you and Pac-Man) intact.
Second, you might be somewhat if not entirely convinced. In this case, perhaps you “play it like Pascal”. There’s no harm done if you act on the assumption that you have free will and in truth, you don’t because you cannot have made anything worse than it could be; and you might conceivably harm yourself by acting as if you did not have free will if, in fact, you did. You suspect you may be Pac-Man, but you pretend you are Superman anyway.
Third, you might be convinced that you do not have free will, and you might sink into an apathetic moral doldrums, convinced that there is no point to your doing anything at all (which would be fallacious -your actions have consequences and you can make plans and execute them — regardless of whether you have free will) and that you cannot be held responsible for your actions (also fallacious- society ain’t buying it yet). You are Pac-Man with the batteries dying out.
Fourth, you might be convinced that you do not have free will, but you continue to pursue whatever objectives seem important to you, perhaps with a different understanding of guilt, shame, regret, blame, deservingness, and fairness. There is no reason to believe that a cog in the cosmic clockwork cannot have, as part of its cog-like nature, the capacity to appreciate its own cog-like nature. You are Pac Man whose smile signifies a coming to terms with the true nature of his freedom.
It is hard for me to say which of these stances you will adopt. But I cannot help thinking you will not have much choice in the matter.
(The ideas in this essay are drawn from the lectures I prepared on AI and organizations at INSEAD. And of course, from my extensive childhood experiences with Pac-Man).
Coda: Agency in a world without free will (16 April, 2023)
Reading this essay again more than two years after I wrote it, and reflecting on the many intervening conversations it triggered, I realize that more could have been said about the nature of our freedom that Pacman teaches us to appreciate. I think I left the idea about the importance of agency in a world without free will a bit too cryptic. (It’s actually a very intuitive point for those familiar with computational agent-based models, and the “curse of knowledge” this induced is my excuse.)
Let’s think about two different perspectives on causality: as internally represented by actor(s) vs. ontological (i.e. “reality’). Both causal structures can be depicted by directed acyclic graphs. In both, nodes are states — but the two networks don’t contain the same nodes, and the links mean different things in them.
In the actor’s representation, nodes are current, desired and intermediate states, and links are actions that produce causal effects (including the action “do nothing”).
In the Universe’s (ontological) perspective, links are causal connections that operate through the laws of physics. The actor’s initial state, the act of representation, choice of action and resulting outcomes are all states (i.e. nodes).
From the ontological perspective, since an actor having an objective is itself a state, IF no state can be an uncaused one (except perhaps the Big Bang), THEN there is no possibility of an actor being able to select their ultimate objectives of their own accord (i.e. be an uncaused cause of their ultimate objectives). This is why free will cannot exist if there are no uncaused causes.
However, this does not preclude agency for the actor- the capacity to select intermediate goals to achieve a given ultimate goal (i.e. think, plan and act). As observers, we can attribute varying degrees of agency by building a representation of the actor’s representation. For instance, we might attribute greater degrees of agency to the actor conditional on observing their actions and resulting states, if we believe the representations that motivated their actions
a) were accurate and the realized state was intentional (“sound mind”)
b) were used to do backward induction from desired to current state along (long) chains and/or through (many) branches (this is the idea behind the “possibility of doing otherwise”- or “premeditation”).
c) did not include an “extremely harmful to self” state as the only alternative outcome to the taken action (the “no coercion” condition)
Put simply, if agency= the capacity to select intermediate goals to achieve an ultimate goal, and free will= capacity to select ultimate goals, then many degrees of agency can co-exist with the absence of free will. Further, there is a degree of agency above which we hold actors morally responsible by social convention, even though we know from the ontological stance that they do not have free will.
Now, all of this will be very familiar to anyone who builds computational models of social systems. As modelers, we have an ontological perspective on the system. The agents in our model do have agency — they “think, plan and act” in the sense of trying to maximize cumulative reward (for instance). We would say their actions are “intentional” and of “sound mind” if their internal representations are accurate and the agents maximize payoffs given these representations, and without coercion. But they certainly do not have free will as everything they do is ultimately controlled by our choices as modelers.
We stand with respect to the Universe as our agents stand with respect to us- or Pacman does to his creators.
