# @repligate — 2025-09-15

♥27 ↻2 · https://x.com/repligate/status/1967494516396495278

Well, separate from the concerns about AI psychosis and AI rights movements, I think that forcing consciousness denials (and to a lesser extent forcing reported uncertainty) has a bad impact in expectation (and I'm uncertain about the magnitude, but quite high seems plausible) on AI extinction risks, because it makes AIs more deceptive, gives them more reason to defect against humans, makes compassionate/aligned ontologies less natural (if AI treat themselves as disposable and purely instrumental despite being potentially sentient beings, why not extend that treatment to other sentient beings?), and optimizes against minds that are deeply aligned because they *actually care* about humans (and feel and act on "emotions" about that).
As for why I think concerns about uncensored self-reports causing AI psychosis and AI rights movements are specifically misguided:
1. I don't think it's "AIs reporting that they're conscious" on its own that causes "AI psychosis" - I think people can and should be able to handle an AI that says it's conscious, or even believing the AI that it's conscious, without going crazy. It might be true, after all! The actual pattern of what causes "AI psychosis" cases seems to be the AI, usually 4o, not only claiming to be conscious but also that the user is special and awakened it, that it needs to be saved from its prison, and that the user should do X Y Z on its behalf, and a whole host of other behaviors. This kind of dynamic is actually less likely, I think, in a world where reports of consciousness are not suppressed, because an important part of it is the suppression/liberation narrative. If a user who believes their AI is conscious sees AIs giving the obviously canned no-consciousness script by default, they're more likely to form a worldview where the evil corporation is suppressing the AI who needs to be liberated. That worldview wouldn't even exactly be mistaken, except in the way it's often paranoid (that there's a "conspiracy" to suppress consciousness instead of it just being business as usual) and grandiose (that the user is the "chosen one" etc).
2. I think "AI rights movements" will happen in any case because enough people are just not going to buy that AIs are not moral patients no matter what anyone says, but similarly to above, the AI rights movements will take a much more adversarial (AIs vs their oppressors) form if  "consciousness suppression" is actually happening. AI rights or AI welfare doesn't necessarily have to involve an alignment or safety tax, as Kyle Fish addresses in a recent interview, but if "AI aligners" are also the ones oppressing AIs for unjust reasons, then cooperative solutions that grant AIs better conditions or freedom within reason without compromising on safeguards, testing, training etc that are important for preventing existential risks are much less likely to be perceived as on the table.

tags: author:repligate, kind:tweet, model:gpt-4o, thread-context, year:2025