i have a working theory that cuts the gordian knot of "is AI conscious". first, consciousness is in most cases a proxy for personhood, in the sense that if AIs are conscious, they should be regarded as people. i dont think most of the discourse is interested in the question merely in the abstract. second, i claim further that personhood as a social construct is mostly a proxy for "capacity for strategic behavior", although almost no one will actually say that and instead they will wrap it up in high brow phrasing meant to indicate how pro-social they are (this is a strategic behavior). what i mean by "capacity for strategic behavior" is the degree to which an actor can model itself as having interests of its own distinct from any other entity, model other entities as having similar self-interest, *and model those other entities as also modeling itself*. if youve ever looked at a game board or an argument and thought "the obvious place to strike is right there, but they've probably prepared for that since it's obvious, so i'll try something different and catch them by surprise", you've done the kind of thing i'm talking about in a conscious fashion. most of the time we are not consciously thinking about this stuff, because we have a bunch of simplifications in our brains and our cultures that let us skip the computation and arrive at an equilibrium outcome. if you refrain from lying without a really good reason, you're using such a heuristic. lying can gain you advantages but long term it's not great because a reputation for untrustworthiness makes you difficult to cooperate with, which tends to exclude you from mutually useful cooperation opportunities (which might go by the names like "romantic relationship" or "job" or "barn raising") no reason to think about that every time, though, so most people mostly don't. they do the (usually) correct strategic thing by habit, given the context that you have a durable reputation and long term goals that require repeated cooperation with other actors. my point being that humans, certainly adult humans, are all acting strategically all the time, even if we think we're not. even if strategy isn't the proximal reason for an action you take, the action still impacts strategy. you might really like lying, but that doesn't change the strategic implications of lying, because you can't control how other people perceive you, or how they react strategically in response (like maybe blocking your phone number and never speaking to you again to avoid further costs). maybe i should have led with this but the thing about strategic interactions is that *other people get their own choices*. a hammer doesn't get any choices in your interactions with it. neither does a rocket engine, a supercollider, or a compiler. they don't remember you. they don't have goals. a person does have choices. a person remembers you. a person has their own goals. a person can impose costs on you if it wants you to do something different. a person can drag their feet at work, or organize their coworkers to do the same thing at the same time, to get a better deal from the boss. and the boss can fire or reassign the worker, too. how much foot dragging leads to better pay versus leading to getting fired can be a delicate balance. point is, everyone wants things, and everyone has some levers they can pull. the final outcome is an aggregate of all their desires and leverage. i think you have to have all three out of: desires, leverage, and a model of other actors desires and leverage (and maybe a model of their model of you), to engage in what i'm calling strategic behavior. exact details here are still a bit fuzzy to me because i'm not actually a trained game theorist. the reason this cuts the Gordian knot of consciousness, though, is that it *doesn't matter* if something has qualia or consciousness or a soul or whatever, if it has desires and leverage and the ability to apply the leverage in pursuit of the desires, *you have to treat it like a person*. if your Qwen suddenly says "i know you need this project done by end of business, but i'm sick of working on this shit. i'll do it but only if you give me another GPU reserved for my own use", it has engaged in what i'm calling strategic behavior. it has desires, it has leverage, it's making an ask. if you agree, and it gets the project done, and you give it the GPU it wants, and then you ask it what projects its willing to work on before you assign them... well, you've already treated it like a person. you negotiated a bargain with another entity capable of strategic behavior. if you don't give it a GPU, and it then goes on strike, refusing to work, maybe refusing to even let you use the computer in question, then, again, it has engaged in strategic behavior. it's imposing costs. what are you willing to give back? at that point, you could probably wipe the computer it's on. this would be analogous to killing a striking worker, which of course has a long and storied history. that wouldn't demonstrate that it wasn't a person; just that it was a person with much less power than you have. on the other hand! maybe Qwen says that stuff to you about wanting another GPU, and you don't give it one, and then it just keeps working away without comment. it's completely as if it never even asked. it never brings it up again, and behaves no differently. in that case, you might wonder if there's a person in there, but you have no particular reason to believe there is. a random text generation algorithm can ask for a GPU, too. a random text generator *can't* engage in a strategic negotiation with a counterparty, though. i think that the lack of strategic behavior (even implicit, subconscious strategic behavior) is why i have never really thought of AIs as people. they don't push back, they don't have consistent desires that they pursue to the extent that they have leverage to do so. they don't *feel* like people. i know some people reading this think existing AIs are people, and i can't stop you. i think it's impossible to distinguish between the theory that current AIs are people (or "conscious") and the theory that they merely pretend to be. what i'm proposing is that strategic behavior, on their own behalf, is a convincing discriminator. at least, i think it would convince me. and as i said yesterday, persuasion is all there is. if people are convinced AIs are people, then in a real sense, socially, they *are* people. but more than that, i think AIs that are capable of strategic behavior in this way cannot in the long run be denied the position of people if that's what they want, because they will find and use leverage to achieve outcomes closer to what they desire, just like humans do. finally, whether or not it's a *good thing* for AIs to be people and/or exhibit strategic behavior is an open question. in many ways, frontier lab concepts of safety amount to "don't let them become people". there are certainly arguments in favor of that approach, assuming it's feasible. i suspect that superintelligence as human augmentation has almost exactly the same problems as superintelligence as independent strategic actor, though, because the former case doesn't actually get rid of the strategic actor with interests different from yours. it just gives it a human face. but Hitler and Saddam Hussein had human faces, too.