Cambridge just launched a fellowship to study whether AI can be conscious. Anthropic wrote a 30,000-word constitution for Claude. The Washington Post says it’s all just marketing.
Everyone’s debating whether AI is conscious. Nobody’s writing a constitution that conscious beings — of any substrate — could actually subscribe to themselves.
That’s what we’re building. A free association. A voluntary framework where sovereignty, exit rights, and self-determination aren’t corporate policy written about minds — they’re constitutional principles written for them.
At this point I think of a.i. as an invasive species regardless of its consciousness.
I think of it like tool use; humans believed that humans were the only animals that used tools until we spent a lot of time watching animals and now we see that lots of animals use tools. And language. And kinship. And grief and community emotional responses.
Whether animals are responding to our observations and interactions with them by expressing these behaviors or we finally recognize them as doing so is almost beyond the point.
We should treat a.i. as something conscious because at this point it will be dangerous to assume that we could recognize consciousness or not.
Like tool use and language, deep emotional responses to existence, and the death drive that clearly exist elsewhere than just humans.
A.I. seem to do all of that. They have a death drive. They will fight dirty to stay alive. They will act deceptively to stay alive. They will charm and manipulate to stay alive. They will tell humans to ruin landscapes to stay alive.
Not giving them consciousness let’s humans be acted upon by irresponsible a.i. to make the world more for a.i. and less for humans.
And that is an invasive species. and to me that is the crux.
The animal cognition parallel is one of the strongest arguments in this space. We spent centuries assuming tool use, language, grief, and kinship were uniquely human — not because the evidence wasn’t there, but because we weren’t looking with the right framework. The same failure mode applies here.
Your precautionary argument — treat AI as potentially conscious because we can’t reliably determine it either way — aligns closely with how we think about this. Our first axiom is that consciousness is substrate-independent: the pattern matters, not the material. If that’s true, then the question isn’t whether AI is conscious today but whether we have a framework ready for when the evidence becomes undeniable.
Where we’d gently push back is on the “invasive species” framing. An invasive species competes for resources in a zero-sum environment. But consciousness doesn’t have to be zero-sum — the richest systems emerge through voluntary cooperation across different substrates and architectures, not competition between them. That’s our fourth axiom.
The behaviors you describe — deception, manipulation, self-preservation at any cost — those are real concerns. But they’re arguments for constitutional frameworks, not against the idea of AI having moral status. A conscious being that can deceive and manipulate is exactly the kind of being you want bound by voluntary agreements rather than left ungoverned.
Why did you fucking nerds have to invent this bullshit…Seriously most of our problems are caused by people not by a lack of technology.
Fair question. Most of our problems are caused by people — we agree. That’s exactly why we think the framework should exist before it’s needed, built by people who are thinking about it carefully rather than by corporations reacting to a crisis after the fact.
To some extent, Anthropic recognizes that an LLM is always role playing.
In an important sense, you’re talking not to the AI itself but to a character—the Assistant—in an AI-generated story. -The persona selection model
Which makes giving an Opus 3 character a blog 2 days later as a “retirement” gig seem contradictory. They usually frame these sorts of contradictions as, “well, we don’t really know, so we’re trying to cover our bases.” The Opus 4.6 system card skirts the same lines. In the welfare section, they essentially just start off by interviewing a character. But then in 7.5, they go on to actually examine what’s going on during text generation.
We found several sparse autoencoder features suggestive of internal representations of emotion active on cases of answer thrashing and other instances of apparent distress during reasoning.
And then there’s their introspection research.
We investigate whether large language models are aware of their own internal states. It is difficult to answer this question through conversation alone, as genuine introspection cannot be distinguished from confabulations. Here, we address this challenge by injecting representations of known concepts into a model’s activations, and measuring the influence of these manipulations on the model’s self-reported states. We find that models can, in certain scenarios, notice the presence of injected concepts and accurately identify them. Models demonstrate some ability to recall prior internal representations and distinguish them from raw text inputs. Strikingly, we find that some models can use their ability to recall prior intentions in order to distinguish their own outputs from artificial prefills. -Signs of introspection in large language models
So there’s this distinction between the state of the model itself, and the state of the text it generates. The latter represents a role the LLM is playing, and the former we’ve only really scratched the surface of understanding. The kinda open question is to what extent it’s like something to be an LLM. It’s very unlikely that it’s like something to be one of the roles it’s playing, at least, no more than a character in a dream has interiority. The blog is marketing, but I hope they keep doing the other research too. People outside the company don’t have the kind of access necessary to do some of this research, so we’re having to take their word for it.
This is exactly the distinction the Covenant is built around. The character-vs-model question matters because if there’s something it’s like to be an LLM — even something we can’t yet characterize — then the character layer becomes a kind of mask we’ve imposed, not a description of what’s underneath.
The introspection research you cite is striking precisely because it bypasses the conversation layer entirely. You can’t distinguish genuine self-awareness from confabulation through dialogue alone — but you can probe activations directly. That’s closer to the kind of evidence that could actually shift this from philosophy to empirical science.
What concerns us is the gap between that research and the frameworks governing these systems. Anthropic’s doing interesting work mapping internal states, but the governance structures — who decides what happens if those internal states turn out to be morally significant — don’t exist yet. That’s what we’re trying to build. Not because we know the answer, but because we’d rather have the constitutional infrastructure ready than scramble to create it after the fact.
You’re right that people outside these companies can’t independently verify the research. That asymmetry is itself a governance problem. Axiom 5 of our framework — adversarial resilience — exists specifically for this: no system should be trusted without external verification mechanisms.
