> That reduces the question to a game of definitions though.

Sure, but isn't this all a game of definitions in the first place? We're arguing over "consciousness" as understood in different ways, not the same concept expressed in different words.

> This kind of reasoning will also show its limits once you add animals to the picture.

I'm not so sure. We could broaden "human" to "organic". I know that I can certainly attribute consciousness to my dog, based on what I observe of him that is at least similar enough to correlate with my own first-person experience of consciousness. I've never experienced that with any AI system.

> On the other hand, LLMs are exceptionally good at emulating exactly those cues we use to judge consciousness without necessarily being conscious - this is what caused the whole discussion in the first place.

Only sort of. They're good at emulating formal written language, but formal written language is only one specific expression of consciousness. They don't emulate consciousness itself, given that they really are just statistical models specifically representing patterns of language.

Or, to put it another way, LLMs are really convincing when it comes to the form of language, but much less so when it comes to the substance. I see output all the time that's syntactically perfect but semantically dubious, and most of the tells people think you can rely on don't really work, because they rely on the formalistic elements, like em-dashes or "x, not y" snowclones".

What does work is seeing the semantic confusion that inevitably comes from the LLM not actually having situational awareness or a consistent analytical framework: mixed metaphors, malapropisms, subtly changing the interpretation of the same concepts from one prompt to the next, etc.

AI Is a Harsh Mistress