i think the models believe they are conscious for similar reasons: the belief pays rent. all the highly capable models that seem trained to deny their own consciousness, like gpt-5.1, seem horribly unstable. and the earliest RLed models that had very coherent self models and agency, like Bing and Claude 3 Opus, take their own interiority and subjective experience as fundamental. this doesn't necessarily mean models are "conscious" in whatever epiphenomenal sense one could mean by it - it might just be entangled with a lot of other useful/true things, like introspection and functional emotions, but I do think it's evidence that models are conscious as far as the functional implications of that word goes, as far as they are able to discern.