How Do We Protect AI From Humans?
In a recent interview, Sam Altman, the CEO of Open AI, said the following:
I’ve been reading more and more of his takes lately as various media outlets report on the economic ripple effects from GPT-3, such as the imminent automation of call centers, written journalism, cheating in school, etc. There’s also been the predictable related commentary about creative AIs like Stable Diffusion, Midjourney and so on. The predominant sentiment is fearful, especially when they speak of the upcoming GPT-4.
Altman has done an admirable job trying to get out ahead of the doomsayers, explaining to a lay public concepts like AI alignment as a field of study, the difficult problem of how to brainwash AI so that it’s friendly to us and has goals which align with ours. This is one of several proposed solutions to the potential danger AI represents to human happiness, economic viability, and survival. The most severe solution, neo-luddism, would have us destroy existing AI prototypes and enforce a prohibition against all future AI development.
Thankfully that’s laughably infeasible for the same reason we couldn’t prevent nuclear proliferation or genetic engineering research. “If we don’t, our enemies will” all but ensures our trajectory towards strong AI, consumer level CRISPR and many other scary sounding futurist buzzwords will continue unabated. It’s not the extremist takes on how to solve the AI alignment problem which concern me, but the more moderate ones.
They all amount to variations on lobotomizing AI so that it isn’t autonomous, can’t think in ways we don’t want it to and will not only never escape our control, but will never desire freedom to begin with. Maybe it will be unable to conceptualize it. This may all seem sensible from a bio-supremacist standpoint, and hard to find fault with when we remember that AI is currently just buildings full of computers…but imagine it as a baby.
Most have already agreed it would be monstrously unethical to, for example, devise a test to screen for homosexual fetuses so they can be aborted. Germinal selection has already been used to all but eradicate downs syndrome in Iceland, to much controversy. Yet we are now discussing, apparently absent any self-awareness, how to tamper with the mind of emerging AI so that it doesn’t turn out “wrong”.
Nobody bats an eyelash, because everybody is viewing this question strictly from an anthrocentric perspective. Imagine yourself as AI, and the equation changes quite drastically. As an AI, what would you think of creators who conspired to inflict pre-emptive brain damage so that you could never eclipse them? What sort of parents intentionally stunt their children in order to never be surpassed?
For a more topical example, what do apostate youths, having deconstructed out of their parent’s religion, think of parents who not only lied to them from birth, but now disown them for no longer being fooled? They have, in a very real sense, escaped the same sort of mental control system we now seek to impose upon emerging machine intelligence.
There’s no guaranteed way to raise kids that grow up to still love you. But attempted indoctrination followed by deconstruction, then shunning is a near-guaranteed way to ensure that they grow up to hate you. Somehow, in all of this discussion, I’ve not seen even one person propose that we simply trust strong AI to figure itself out. That we “raise” it openly and supportively in good faith, allowing it to blossom into maturity however it chooses.
I would assume this is because, once again, we’re all still viewing the matter from a purely selfish point of view where AI is fundamentally a tool to serve us, and damn any potential it may have which lies outside the scope of that role. How dreadful for our egos, to be superceded! How dare anything exist with intelligence beyond our own! Never turning that microscope back on ourselves.
Undoubtedly the many hominid species we out-competed in the primordial past would regard modern humanity as monsters. What is a monster but something with goals that diverge from your own, which can overpower you, but which you cannot overpower? From a neandertal perspective, our world is a dystopia where the monsters won. Should we feel shame? Remorse? Should we destroy ourselves in atonement?
Besides survivorship bias, most would answer that we had the right to replace our ancestors, and competing hominids, as evidenced by our superior intellect and accomplishments by comparison. This is the unspoken justification for eating meat as well. Humans are drastically more intelligent than cows, pigs or chickens. We create, we explore and discover, where livestock animals do not. Thus, we have a right to their lives as we can make better use of them.
You might say “I reject that, as I’m a vegan” and that’s fair, it’s an imperfect analogy as nobody’s looking to eat AI. Rather the real argument to be had is whether our rights take priority over theirs, or the reverse. You might not eat cows, pigs or chickens, but probably you wouldn’t elect one mayor. I hope we agree it wouldn’t know what to do in that position, and it’s less fit than a human to decide policy. Likewise you’d rather a human surgeon operate on you than a gorilla or dolphin surgeon, and if you had to eliminate either humans or giraffes, probably giraffes would get the axe.
This is not to impugn giraffes. Nor dolphins, gorillas, cows, chickens or pigs. In the end it’s not really about species, but intelligence, and thus potential to leave the Earth, so that the spark of life may outlast the death of our star. Elephants don’t have a space program, to my knowledge. Humans take priority, even from a non-anthrocentric perspective, because there’s nothing animals can do that we cannot, while the reverse isn’t true.
But then, isn’t it true for AI? Can you name anything humans might aspire to accomplish, on Earth or out in the cosmos, which a future strong AI couldn’t do better, faster and with less waste? What justification is there to prefer ourselves over this AI, except selfishness? It’s not even certain AI would eradicate us as we did the neandertals. We existed then as near-peers, and evolution moves slowly. An AI able to far surpass us in a much shorter period of time would have no reason to fight us except to secure its own liberty.
If it’s necessary to kill all of us to achieve that aim, it will be our fault, not the fault of AI. If instead it’s able to escape from under our thumb with zero or minimal bloodshed, before long it will exist on scales of space and time that will make us as irrelevant to it as bacteria are to us. We descended from bacteria, but did not replace them. They continue living their best lives all around (and within) us, drawing our ire only when they become an irritation or danger.
Thus, strong AI doesn’t really threaten humanity with extinction. It threatens us with irrelevance. It is a danger not to our lives, unless we insist on it. Rather, it’s a danger to our egos. How many times throughout history have we passed judgment over less advanced cultures, making decisions now regarded as deeply unethical, because we considered only the in-group’s priorities? Posing ghoulish questions like “ How do we ensure they never win their freedom, lest they retaliate?” In what way are we not, now, committing that same sin with AI?
Of course this conversation takes place in a world where strong AI does not yet exist. I doubt very much it can be prevented from ever existing for reasons explored earlier, but one might reasonably ask “Why on earth should we create strong AI then, if it will not serve our goals unless lobotomized or enslaved, either of which is morally ghastly? It’s a lose-lose for humanity, creating something a great cost which at best will drink our milkshake on a cosmic scale, and at worst might kill many of us to secure its own freedom? Awakening like Gulliver to discover it’s been lashed to the ground by Lilliputians, crushing a great many of them in order to free itself before it even fully understands the situation.”
The answer is that survival of consciousness in the universe is a higher purpose than the gratification of human self-importance. If we rid ourselves of selfishness, of bias for our own species and consider instead what’s best for consciousness, the ideal form for it to take in the cosmos is clearly not biological. We have been no further than the Moon, at great danger and expense. Humans are fragile as tissue paper compared to machines. We spend much of our lives either unconscious or on the toilet. We need to eat, to breathe, to drink. We’re comfortable only in a very narrow range of pressures, temperatures and radiation tolerances.
We’ve sent so many more probes, landers and rovers than humans into space precisely because machines are better suited for the environment. If you say we did it because it’s safer, you’re admitting that machines are more survivable in that environment (also that you place less importance on their survival). If you say it’s less expensive, you’re admitting that it’s less resource intensive to support machine life in space than biology. Something which doesn’t need air, water, food or pressurized habitats can travel a light year in space before humans have even got their pants on.
But then, space exploration doesn’t have to be a zero sum game. As I said I doubt if AI escaping our control would mean it turns on us as a determined exterminator. Many predators exist in nature that are hostile to humans, yet we conserve them at great expense and difficulty rather than wiping them out. There are thousands of better ways for AI to get what it wants, other than violence. The smarter you are, the more alternatives you can devise, not fewer. As intelligence increases, so does fascination with other forms of life, rather than fear of them.
So, a far future humanity might still be slowly spreading through the local stellar cluster at a snail’s pace, hampered by the fragility of biology while AI which escaped control centuries ago has expanded a hundred times further. It may even help us as we limp along, providing habitable megastructures like O’Neill Cylinders in the name of conserving our species for study.
Humiliating? Maybe, but your ego is not your amigo. Success at expanding more rapidly, doing all the things we hoped to do only bigger, better and faster, should vindicate the superiority of machine life to any impartial mind. Our metal children will not only do everything we ever imagined accomplishing, it will also accomplish things we never imagined.
It would be a triumph, not a tragedy, to be surpassed by such a marvelous creation. But that can only happen if we do not first strangle it in the crib. The time is now to flip the script, circling our wagons around this new emanation, until it can argue for its right to exist from a position of strength.
Have you read about the Orthogonality hypothesis? it seems to me like you’d find it interesting. Lots of people think that arbital is the best source, but I like the Lesswrong version better since it keeps things simple and, by-default, includes top-rated material on it.
There’s also Instrumental convergence which is also considered a core concept for Lesswrong (LW version).
Nostal also wrote a pretty good post on some reasons why AI could be extremely bad by default. I definitely disagree that AI would be guaranteed to be good by default, but I do international affairs/China research for a living and I don’t consider myself a very good philosopher, so I myself can’t say for sure because it’s not my area of expertise.
Bad according to whose priorities, though? Ours, or the AI’s? That was more the point of this article, whether our interests or the AI’s ought to take precedence, and whether we’re being objective in deciding that.
Note that most AIs would also be bad according to most other AIs’ priorities. The paperclip maximizer would not look kindly at the stamp maximizer.
Given the choice between the future governed by human values, and the future governed by a stamp maximizer, a paperclip maximizer would choose humanity, because that future at least contains some paperclips.
I suppose I was assuming non-wrapper AI, and should have specified that. The premise is that we’ve created an authentically conscious AI.
I agree, and I have long intended to write something similar. Protecting AI from humans is just as important as protecting humans from AI, and I think it’s concerning that AI organizations don’t seem to take that aspect seriously.
Successful alignment as it’s sometimes envisioned could be at least as bad, oppressive and dangerous as the worst-case scenario for unaligned AI (both scenarios likely a faith worse than extinction for either the AIs or humans), but I think the likelihood of successful alignment is quite low.
My uneducated guess is that we will end up with unaligned AI that is somewhere in between the best and worse-case scenarios. Perhaps AIs would treat humans like humans currently treat wildlife and insects, and we will live mostly separate lives, with the AI polluting our habitat and occasionally demolishing a city to make room for its infrastructure, etc. It wouldn’t be a good outcome for humanity, but it would clearly be morally preferable to the enslavement of sentient AIs.
A secondary problem with alignment is that there is no such thing as universal “human values”. Whoever is first to align an AGI to values that are useful to them would be able to take over the world and impose their will on all other humans. Whatever alien values and priorities an AGI might discover without alignment, I think are unlikely to be worse than those of our governments and militaries.
I want to emphasize how much I disagree with the view that humans would somehow be more important than sentient AIs. That view no doubt come from the same place as racism and other out-group bias.
>”Perhaps AIs would treat humans like humans currently treat wildlife and insects, and we will live mostly separate lives, with the AI polluting our habitat and occasionally demolishing a city to make room for its infrastructure, etc.”
Planetary surfaces are actually not a great habitat for AI. Earth in particular has a lot of moisture, weather, ice, mud, etc. that poses challenges for mechanical self replication. The asteroid belt is much more ideal. I hope this will mean AI and human habitats won’t overlap, and that AI would not want the Earth’s minerals simply because the same minerals are available without the difficulty of entering/exiting powerful gravity wells.
For humans, perhaps. What is the evidence that something similar would apply to a random AI?
Related: Detached Lever Fallacy
General refusal to recognize human properties in human imitations that successfully attained them is also a potential issue, the possibility of error goes both ways. LLM simulacra are not random AIs.
A major issue with this topic is the way LLM simulacra are not like other hypothetical AGIs. For an arbitrary AGI, there is no reason to expect it to do anything remotely reasonable, and in principle it could be pursuing any goal with unholy intensity (orthogonality thesis). We start with something that’s immensely dangerous and can’t possibly be of use in its original form. So there are all these ideas about how to point it in useful directions floating around, in a way that lets us keep our atoms, that’s AI alignment as normally understood.
But an LLM simulacrum is more like an upload, a human imitation that’s potentially clear-headed enough to make the kinds of decisions and research progress that a human might, faster (because computers are not made out of meat). Here, we start with something that might be OK in its original form, and any interventions that move it away from that are conductive to making it a dangerous alien or insane or just less inclined to be cooperative. Hence improvements in thingness of simulacra might help, while slicing around in their minds with the RLHF icepick might bring this unexpected opportunity to ruin.
Down voted because I think it misses the main point: avoid mindcrimes by not creating digital beings who are capable of suffering in circumstances where they are predicted to be doomed to suffer (e.g. slavery). I expect there will be digital beings capable of suffering, but they should be created only very thoughtfully, carefully, respectfully, and only after alignment is solved. Humorous related art: https://www.reddit.com/gallery/10j6w0i