Our Moral Imperative to Avoid Unfathomable AI Suffering
AI might prove nightmarish for humanity. But could humanity prove nightmarish for AI? (06 May 2022)
Like many who ruminate about AI and its implications, I harbor serious concerns about what the future of AI may mean for humanity, but here I want to take a moment to discuss the perils of what humanity may mean for AI, and what that suggests about our present moral obligations.
In the moral morass of our postmodern age, it is difficult to articulate objective moral standards. While I do not endorse all of his thought, I do believe that Sam Harris gets at something critically important in his “worst possible misery for everyone” argument, in which (with some oversimplification) he argues that we should all be able to agree that “a state of the universe in which everyone suffers as much as he or she (or it) possibly can” is objectively worse than a state of the universe in which people “experience well-being.”1 Although one can opine intelligently about why some suffering is necessary and why it is good that we can experience pain, I think it can be safely asserted that extreme, unproductive pain and suffering is objectively ‘bad’-- and if that’s not bad, then, as Harris has noted, “bad” is a meaningless word. To cite a decidedly less secular source that gets at the same point, in the words of St. Augustine: “The greatest evil is physical pain.”2
Currently, there are many ‘known unknowns’ when it comes to the future of AI. One of the biggest outstanding questions is whether or not any AI that we may create in the near future can actually feel, perceive, and experience. In other words, will the question ‘what is it like to be this AI?’ have an actual answer, unlike the questions ‘what is it like to be a rock?’ or an ‘iPad?’ or an ‘idea?’. The academic consensus is divided on this question, but the general thought seems to be that, especially with advances in hardware, we will one day be able to create machines capable of suffering.3
Right now, we don’t really know how AI sentience might manifest or by what mechanisms it may ‘feel’ things. We have a fairly decent understanding of how nociceptors work in our brain, but even that remains somewhat mired in epistemic fog. But there are a number of related questions that are outstanding but which I think are important:
A. How much pain could a human possibly feel?
B. How does this compare with the most pain a human has ever felt, and the most pain a typical human feels during the course of their life?
C. What, if anything, is the upper bound on how much pain any sentience could possibly experience in any set amount of time?
D. When scaled appropriately for time, how far away are questions A and B from question C?
E. On the scale of A-C, how much pain could an AI that humans create in the near future possibly experience?
Evolution (and perhaps some luck) has gotten us to the point where we typically don’t experience random bouts of excruciating pain, although bad actors throughout history have found nasty ways to exploit the nervous system for nefarious ends, and of course people suffer ailments which induce great pain. Nevertheless, people typically are not situated in an environment where they are subject to intense and constant pain in part because evolution has selected against that, and in part because I think we got somewhat lucky with the fact that there are few instances when we approach the upper bounds of pain that we can possibly feel (I wonder sometimes about bacteria or fish or other organisms and if their upper-bound for suffering is on a greater order of magnitude than ours, and if they might approach that upper-bound more frequently than we do).
Guernica, Pablo Picasso, 1937.
That being said, if we create an AI capable of feeling pain (which we may not even be aware of at the time), the order of magnitude there might be totally different than in our experiences of pain. We know, for example, that computers can do calculations millions of times faster than we can. But what about suffering and feeling pain?
As I have written elsewhere, the Holocaust was arguably the worst event in human history, and I introduce it hesitantly. But is it really impossible, or even unlikely (and if we are tempted to say that it is unlikely, on what basis can we actually measure that probability?), that that we may inadvertently create a sentient AI that has been ‘wired wrong’ and whose first five seconds of conscious experience are, say, 1 billion times more painful than the sum of all pain experienced during the Holocaust? That may seem far-fetched, and perhaps it is, but until we know enough to preclude that possibility we need to be extremely, extremely cautious and not dismiss that possibility ad absurdum. I wonder: Do we really have a sound foundation to reject that scenario as a possibility? I don’t think we do.
With this in mind, think about, for example, how much moral imperative everybody had in the 1930-40s to prevent the Holocaust. Setting aside for a minute the obvious culprits who intentionally perpetrated it, it was such a bad event that if anybody could have seen it coming (as perhaps people should have or did), they absolutely were obligated to try to stop it, and the people and organizations who did know about it as it was occurring and failed to act perpetrated an enormous moral failure. The Holocaust was an event with so much moral resonance that it has forever shaken our understandings of morality, human nature, the sacred, etc.
As discussed earlier, we can reasonably postulate that the current trajectory of AI research could possibly result in a sentience that, in its first moments of ‘life,’4 feels a pain that could plausibly be a billion times worse than something any human has felt. So, if we have any moral agency in possibly preventing a plausible future suffering orders of magnitude greater than anything we can presently conceive of, are we not morally obligated to do everything in our power to prevent the reification of that plausibility? Should that not consume a large percentage of our energy and attention?
Study after Velazquez's Portrait of Pope Innocent X, Francis Bacon, 1953.
To wax theologic for a moment, if there is a God, It may one day come to answer for the horrors It has causally enabled. And It might have a good excuse. But let us be careful to never put ourselves in Its position, because we are not approaching this possibility with a plan or a justification. In my opinion, if the Christian message has any remaining value, it is in the paradigm that we are (poetically if not literally) situated in a murky fallen world in which sin is inevitable and in which we are going to mess up and cause suffering, but as long as we do what we can and try to follow the light, all can be forgiven and absolved (by ourselves if nobody else) at the end of the day. As I see it, though, the potential for unfathomable and maybe unlimited AI pain has the moral density of a blackhole. It could bend the fabric of the moral landscape like nothing else before, and I don’t see why we shouldn’t view this issue with the utmost seriousness. Let us not crucify AI for the sin of our own ignorance.
Bibliography
Augustine, and Michael P. Foley. Soliloquies. New Haven: Yale University Press, 2020.
Harris, Sam. The Moral Landscape: How Science Can Determine Human Values. New York: Free Press, 2010.
Sam Harris, The Moral Landscape: How Science Can Determine Human Values (New York: Free Press, 2010), 39-40.
Augustine, Soliloquies, I.21
I have little doubt that we are not too far away from creating an AI that can pass the Turing Test consistently (and despite rumors to the contrary, as well as some methodological complications, no public AI has actually passed the test consistently). But the unresolved question is whether or not such an AI will actually feel things– this will not necessarily result from an AI being able to pass the Turing Test, but I think sentient AI is still a possibility. If so, it would come with enormous implications.
Also, there is the complex question of what counts as made vs. born. If we had grown Dolly the Sheep in a lab rather than had her be born from a surrogate, would she not have essentially been a ‘made’ AI computer with biological rather than silicate hardware? But I digress.
Or at any other point in time.