The problem of AI alignment is usually defined roughly as the problem of making powerful artificial intelligence do what we humans want it to do. My aim in this essay is to argue that this problem is less well-defined than many people seem to think, and to argue that it is indeed impossible to “solve” with any precision, not merely in practice but in principle.
There are two basic problems for AI alignment as commonly conceived. The first is that human values are non-unique. Indeed, in many respects, there is more disagreement about values than people tend to realize. The second problem is that even if we were to zoom in on the preferences of a single human, there is, I will argue, no way to instantiate a person’s preferences in a machine so as to make it act as this person would have preferred.
Problem I: Human Values Are Non-Unique
The common conception of the AI alignment problem is something like the following: we have a set of human preferences, X, which we must, somehow (and this is usually considered the really hard part), map onto some machine’s goal function, Y, via a map f, let’s say, such that X and Y are in some sense isomorphic. At least, this is a way of thinking about it that roughly tracks what people are trying to do.
Speaking in these terms, much attention is being devoted to Y and f compared to X. My argument in this essay is that we are deeply confused about the nature of X, and hence confused about AI alignment.
The first point of confusion is about the values of humanity as a whole. It is usually acknowledged that human values are fuzzy, and that there are some disagreements over values among humans. Yet it is rarely acknowledged just how strong this disagreement in fact is.
For example, concerning the ideal size of the future population of sentient beings, the disagreement is near-total, as some (e.g. some defenders of the so-called Asymmetry in population ethics, as well as anti-natalists such as David Benatar) argue that the future population should ideally be zero, while others, including many classical utilitarians, argue that the future population should ideally be very large. Many similar examples could be given of strong disagreements concerning the most fundamental and consequential of ethical issues, including whether any positive good can ever outweigh extreme suffering. And on many of these crucial disagreements, a very large number of people will be found on both sides.
Different answers to ethical questions of this sort do not merely give rise to small practical disagreements. In many cases, they imply completely opposite practical implications. This is not a matter of human values being fuzzy, but a matter of them being sharply, irreconcilably inconsistent. And hence there is no way to map the totality of human preferences, “X”, onto a single, well-defined goal-function in a way that does not conflict strongly with the values of a significant fraction of humanity. This is a trivial point, and yet most talk of human-aligned AI seems to skirt this fact.
Problem II: Present Human Preferences Are Underdetermined Relative to Future Actions
The second problem and point of confusion with respect to the nature of human preferences is that, even if we focus only on the present preferences of a single human, then these in fact do not, and indeed could not, determine with much precision what kind of world this person would prefer to bring about in the future.
One way to see this point is to think in terms of the information required to represent the world around us. A perfectly precise such representation would require an enormous amount of information, indeed far more information than what can be contained in our brain. This holds true even if we only consider morally relevant entities around us — on the planet, say. There are just too many of them for us to have a precise representation of them. By extension, there are also too many of them for us to be able to have precise preferences about their individual states. Given that we have very limited information at our disposal, all we can do is express extremely coarse-grained and compressed preferences about what state the world around us should ideally have. In other words, any given human’s preferences are bound to be extremely vague about the exact ideal state of the world right now, and there will be countless moral dilemmas occurring across the world right now to which our preferences, in their present state, do not specify a unique solution.
And yet this is just considering the present state of the world. When we consider future states, the problem of specifying ideal states and resolutions to hitherto unknown moral dilemmas only explodes in complexity, and indeed explodes exponentially as time progresses. It is simply a fact, and indeed quite an obvious one at that, that no single brain could possibly contain enough information to specify unique, or indeed just qualified, solutions to all moral dilemmas that will arrive in the future. So what, then, could AI alignment relative to even a single brain possibly mean? How can we specify Y with respect to these future dilemmas when X itself does not specify solutions?
We can, of course, try to guess what a given human, or we ourselves, might say if confronted with a particular future moral dilemma and given knowledge about it, yet the problem is that our extrapolated guess is bound to be just that: a highly imperfect guess. For even a tiny bit of extra knowledge or experience can readily change a person’s view of a given moral dilemma to be the opposite of what it was prior to acquiring that knowledge (for instance, I myself switched from being a classical to a negative utilitarian based on a modest amount of information in the form of arguments I had not considered before). This high sensitivity to small changes in our brain implies that even a system with near-perfect information about some person’s present brain state would be forced to make a highly uncertain guess about what that person would actually prefer in a given moral dilemma. And the further ahead in time we go, and thus further away from our familiar circumstance and context, the greater the uncertainty will be.
By analogy, consider the task of AI alignment with respect to our ancestors ten million years ago. What would their preferences have been with respect to, say, the future of space colonization? One may object that this is underdetermined because our ancestors could not conceive of this possibility, yet the same applies to us and things we cannot presently conceive of, such as alien states of consciousness. Our current preferences say about as little about the (dis)value of such states as the preferences of our ancestors ten million years ago said about space colonization.
A more tangible analogy might be to consider the level of confidence with which we, based on knowledge of your current brain state, can determine your dinner preferences twenty years from now with respect to dishes made from ingredients not yet invented — a preference that will likely be influenced by contingent, environmental factors found between now and then. Not with great confidence, it seems safe to say. And this point pertains not only to dinner preferences but also to the most consequential of choices. Our present preferences cannot realistically determine, with any considerable precision, what we would deem ideal in as yet unknown, realistic future scenarios. Thus, by extension, there can be no such thing as value extrapolation or preservation in anything but the vaguest sense. No human mind has ever contained, or indeed ever could contain, a set of preferences that evaluatively orders more than but the tiniest sliver of (highly compressed versions of) real-world states and choices an agent in our world is likely to face in the future. To think otherwise amounts to a strange Platonization of human preferences. We just do not have enough information in our heads to possess such fine-grained values.
The truth is that our preferences are not some fixed entity that determine future actions uniquely; they simply could not be that. Rather, our preferences are themselves interactive and adjustive in nature, changing in response to new experiences and new information we encounter. Thus, to say that we can “idealize” our present preferences so as to obtain answers to all realistic future moral dilemmas is rather like calling the evolution of our ancestors’ DNA toward human DNA a “DNA idealization”. In both cases, we find no hidden Deep Essences waiting to be purified; no information that points uniquely toward one particular solution in the face of all realistic future “problems”. All we find are physical systems that evolve contingently based on the inputs they receive.*
The bottom line of all this is not that it makes no sense to devote resources toward ensuring the safety of future machines. We can still meaningfully and cooperatively seek to instill rules and mechanisms in our machines and institutions that seem optimal in expectation given our respective, coarse-grained values. The conclusion here is just that 1) the rules instantiated cannot be the result of a universally shared human will or anything close; the closest thing possible would be rules that embody some compromise between people with strongly disagreeing values. And 2) such an instantiation of coarse-grained rules in fact comprises the upper bound of what we can expect to accomplish in this regard. Indeed, this is all we can expect with respect to future influence in general: rough and imprecise influence and guidance with the limited information we can possess and transmit. The idea of a future machine that will do exactly what we would want, and whose design therefore constitutes a lever for precise future control, is a pipe dream.
* Note that this account of our preferences is not inconsistent with value or moral realism. By analogy, consider human preferences and truth-seeking: humans are able to discover many truths about the universe, yet most of these truths are not hidden in, nor extrapolated from, our DNA or our preferences. Indeed, in many cases, we only discover these truths by actively transcending rather than “extrapolating” our immediate preferences (for comfortable and intuitive beliefs, say). The same could apply to the realm of value and morality.
Magnus Vinding 