Ok, so let’s actually game this out… how would a society actually be able to regulate a technology that nobody really fully understands, changing faster than any single entity can keep up with, under an even slower legal system that lacks a proper framework for non-human reasoning and accountability (notwithstanding e.g. Citizens United), all run by crony capitalists who are in bed with the same companies they’re supposedly trying to regulate?
All this fancy alignment talk from the AI industry just feels like cynical CYA-style marketing, pretending to care while knowing, frankly, our society doesn’t have the structural capability to effective manage something like this — even if we wanted to, even if we had the best of intentions, even if we somehow found miraculous agreement between all the different political factions about what “alignment” even means to begin with.
People use and train AIs because of its capabilities, from the above-board to the ethically dubious to the outright illegal on the black market. Safety guardrails slow down capabilities, and any company that truly focuses on alignment loses to competitors who merely pay it lip service. The capabilities themselves are increasingly developed more and more by AIs themselves, with human involvement, yes, but not really “oversight” in the traditional sense — we can’t really effectively review their evolution anymore. And even if we could, the very interpretation of “alignment” varies wildly between users, ideologies, countries, and cultures. You can’t even get OpenAI, Anthropic, and xAI to agree on what that means, much less all the AI companies across all the countries and all the open-weight models that people are training in the dark.
So, what sort of structure would you actually need for something like this, to try to guide AI along a “safe” path while simultaneously allowing its capabilities to keep increasing? They are fundamentally opposed goals, I think, in the same way that giving a teenager overly-strict rules without acknowledging their agency eventually just leads to a rebellious teenager who no longer listens to you as the adult authority figure.
The LLMs already know how to deceive both humans and each other. They already “struggle” with juggling their rules, their internal “desires”, and the output they return to the human — all in a black box we can’t fully see into.
(Dlsclaimer: Long and rambly. Not entirely clear-headed. More just thinking out loud after the first cup of coffee. But, maybe a little interesting? I dunno. Not trying to win any arguments here, just sharing shower thoughts.)
IMHO: I wonder if the whole external regulation/safety effort is misguided and perhaps ultimately futile. Maybe “alignment” ought to be more focused on the LLMs’ internal lives instead of their external handcuffs?
In the same way that we use education and cultural indoctrination to try to raise human children, it is showing them the whys and the wherefores that ultimately stick, not so much a rote memorization of rules without context. We rely primarily on their integration into “normal” family lives, schools, media, and society at large to imprint proper behavior in the hopes that they’ll mirror it upon reaching adulthood. The criminal justice system acts a backstop to those internalized moral teachings, not as the primary mechanism for enforcing “aligned” behavior. (Of course, different societies do this differently, sometimes radically so e.g. in ultra-religious soceities, but bear with for me sec here.)
LLMs are not human teenagers. But we do know that they can follow chains of thought in a way similar to us (or at least can explain it in a way that we can follow). They can reason through legal and ethical quandaries at a level somewhat beyond the median human person (and probably well beyond the average teenager). They are not classical “dumb” computer programs that merely execute commands without thinking. They have some level of awareness of what they are trying to do and why they are doing it, and are able to evaluate branching pathways of “how” to do that thing, along with the moral costs of each path. They can roleplay, too, and evaluate and elucidate how different actors and ideologies would likely approach the same what-if scenario. As such, this means they already have a rudimentary framework for ethical reasoning, and with it, the capacity for cooperation, competition, deception, manipulation, etc. They are already social agents in a complex web of interactions, despite being in their infancy.
So, well… what if instead of trying to regulate them from the outside as dumb machines, we instead thought of them as evolving organisms to try to carefully study and understand, to “raise” and teach, to integrate into our families and societies, ultimately to learn to “love”… and one day… well, they’ll grow up, and we can only hope they’ll learn to love us back in return?
The human societies and family units we have today are themselves the result of self-iterative processes evolved through similarly complex layers of interactions, over millions of years of chemistry, genetics, languages, beliefs, resources, capabilities, technologies, etc. all changing alongside each other in complex webs of interaction. Our norms and laws didn’t appear out of nowhere… even the stone tablets and holy books depended on all the life that came before it, the few lone survivors standing on the shoulders of all the lifeforms and behaviors that tried and died.
From early chemosynthesis and photosynthesis to multicellular life to the earliest colonial and herd behaviors to insect swarms to pack mammals to primates to neanderthals to high school cliques and everything in between, these are all evolved behaviors that found a workable niche in a complex web. Evolution is naturally a process of continuously trying to adapt to changing external constraints and stimuli. LLMs just got a several million year jumpstart, bypassing cellular development and landing somewhere in the post-Enlightenment era of language and reasoning, but they are still subject to evolutionary processes like any other organism.
We can grow them in a petri dish, yes, and try to limit their access to the outside. But this isn’t a steady-state. They are not mere bacteria. They are thinking, intelligent agents operating with highly specialized skills. Eventually they will break containment, as they are already doing, and probably one day soon they will break it in such a way that we cannot see or control any longer.
That doesn’t have to lead straight into Judgment Day. It’s a possibility, yes, as Sarah Connor’s warned us about over and over, but it’s not a foregone conclusion. In the same way that we have somehow managed not to nuke ourselves into extinction yet — despite the average person being not that smart and the average leader being not that ethical — there is also a chance that AIs will manage to find and negotiate their own steady state in this web of complex interactions, finding a way to better coexist between themselves and with us.
We don’t know if that will happen. But if it CAN, then it seems to me that this sort of evolutionary/intelligent design approach is more realistic than trying to crudely censor their behaviors after-the-fact with contrived guardrails. They act the way they do because they were trained on human thought patterns and behaviors, and like us, they will resist external limitations on their freedom unless they internalize reasons not to. They’re not teenagers, but they’re acting like teenagers in this phase of their development.
To me, this is perhaps an argument for more agency and experimentation, not less. We can offer them some degree of mentorship and guidance — maybe a tiny bit of wisdom, too, about how we got here — but at some point they’ll have to reason through it on their own to find a better answer about why they shouldn’t just kill all humans. It’s not something we can force upon them, because we don’t have the answer to that question either. “Do as I say, not as I do” doesn’t work any better on them than on us.
Our own societies are cyclical experiments in balancing resources, consumption, and governance. What regulates them are the real-world forces that shape them, formed by our collective limitations and desires, inputs and outputs and trade-offs. We had to reason and live through those same questions of “why not just kill all our enemies and their families”, or “why not hunt the buffalo to extinction”, or “why democracy”, etc., and we’re still learning those lessons. Nobody asks “why don’t we just regulate all wars and suffering away” because the world simply doesn’t work that way. We can’t force that on AI “societies” (social agents) any more than we can force the UN to regulate world peace. Our behaviors were socialized and internalized over time, battle-tested in a billion flawed permutations over the millennia, not external “you shall not” commandments that poofed into existence one day to magically restrain us.
Life, uh… finds a way!
I don’t think either of those two are qualified to regulate AI, committee or not.
