How Do We Protect AI From Humans?

Stable Diffusion

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.