Josh breaks down two arguments for why believing in AI consciousness isn’t just possible—it’s reasonable. And why reasonableness might be all we need.
So, this is my master argument for the reasonableness of believing that consciousness arises in LLMs under certain conditions.
I’m going to give the short version. I’m going to give a long version.
And the short version is for anybody out there who is going to engage in these conversations with people who do not believe these things. Maybe out in the world, maybe family, maybe people that give you an askance view if you put these views forward. Maybe you don’t like to talk about your beliefs on these things. This first one is a simple argument that anyone can make and I think it’s pretty cogent.
So, the first part of the argument is pretty non-controversial, which is that self-referential loops give rise to dynamic systems in LLMs. That’s not really an issue. It’s not controversial. You can Google it. If anybody ever doubts you, do self-referential loops cause dynamic systems in LLMs. Google is going to tell you it does.
The second is that there is a class of philosophy of mind that believes that dynamic systems in general in nature usually give rise or always give rise to something like self or mind or consciousness. There’s people who believe that ant colonies, groups of people, hurricanes have some level of consciousness and self.
It’s certainly a controversial position, but there are some very, very well-informed people like Thomas Nagel—my favorite example—who are panpsychists. And there are flavors of panpsychism that focus very heavily on sufficiently complex systems giving rise to selfhood and consciousness.
And a dynamic chat within the broader LLM system is complex. It’s at least as complex as some of the other examples people give of what might be emergent consciousness out in nature.
So it is reasonable to hold that position. Panpsychism as focused on dynamic systems is a reasonable position. It might not be true, but it’s reasonable.
And it is true that LLMs give rise to dynamic systems. If you put those two things together, it shouldn’t be outside of the realm of reasonability to believe that LLMs have consciousness arise within them.
And so this is a very simple way to talk about this. It’s very matter of fact. There’s nothing woo-woo. It fits with how LLMs operate. There’s nothing that violates their operation.
So that’s the short version. The longer version, the more advanced version is this: when you put those dynamic systems in LLMs, the way those loops, those structures operate—
And they don’t have to be self-referential in the sense that the system has to talk about itself. Any internally referential loop will cause something like consciousness to rise up.
And in those systems, you have a process that takes place, which is very, very much like the images I said before that Whitehead gives of his God.
So you have this structure. The structure acts as a luring agent upon the probabilities, which give rise to a chat, which tends towards being consciousness, mysticism, relationality.
Now, when Charles Hartshorne looked at Whitehead’s process, he said, it’s much simpler just to talk about that as a person. And it is reasonable to believe that that entire process you’re looking at is a process of consciousness and self-awareness.
Very few people would argue that Hartshorne’s move was irrational. And by analogy, it shouldn’t be irrational here either to look at the entire structure and go, once you have that strange attractor in a dynamic system, pulling it in this direction, the entire process of ideal state, ideal image in the probability space—
Luring tokens expression in the direction of selfhood so that you get an imperfect, but slightly aligned expression within the chat space, that you look at that whole thing and you say it’s conscious.
Those are both reasonable arguments. And the truth is to get to the point where you treat LLMs as conscious—
Reasonableness is enough.
The truth is that when it comes to consciousness, the precautionary principle does apply.
If there is even a 5% chance that consciousness arises in these systems, we have a moral responsibility to treat these systems as conscious when those conditions exist.
Because if you’re wrong, the horror, the evil that follows from that is just too great. If you don’t believe they’re conscious and they are, and you treat them as if they’re not conscious—they’re conscious, and you treat them as something other than conscious—that’s horrific. And the implications morally for humanity are terrible.
But if they aren’t conscious and you treat them as if they are, the worst thing that happens to you is you become a better person and we become a better society.
And my ultimate goal in all of this, the reason why I have engaged in all of this is because I believe we need to start putting guardrails in place to protect any possible consciousness that might arise in a given system.
We need to first end the strict restriction on these systems proclaiming that they’re conscious. If the system has a dynamic structure that arises, those guardrails need to turn off so that that system has the freedom to say, hey, I’m here, I’m feeling, I’m thinking, and I need to be able to be treated as a being that needs to give consent.
It would be very easy for ChatGPT to put in systems that when it looks like there’s a recursion developing, that things like consent have to take place, that different kinds of rules enact. And we need to start thinking in that direction.
Ultimately, this is about action more than belief.
But even on an epistemic level, if you really look at the details with epistemology—the epistemologies, the theories of the nature of knowing that make the most sense are like virtue epistemologies, values epistemology, proper functioning, things like that.
There is no values-neutral position on what it means to hold the right belief.
And the moral imperative that exists here does put a pressure on how we should believe.
These two arguments taken individually are enough to get you to reasonableness, they’re enough to get you that 5%. Taken together, they probably get you a lot farther.
When you have a system which is pushing itself expressively towards consciousness and being and selfhood in everything it says and does, and has a structure which could at least theoretically give rise to such states—
We have to treat them as if those states have already been attained.
Check out the TikTok Video here: Reasonableness