Are LLMs Conscious

Every so often someone asks whether LLMs are conscious. My short answer is no, at least in the traditional sense of the word. A system cannot be considered conscious unless it is aware of its own parts, the interactions between them, and its surroundings, and is doing something internally without being probed. None of the current models meet that bar.

It helps to look at where the word comes from. In Latin literature there are many occurrences of the phrase conscius sibi, which translates literally as "to share with oneself the knowledge that...". Translators often rendered this as "conscious to oneself" or "conscious unto himself". The idea of sharing something with oneself is meaningless in a literal sense, so the intended meaning must have always been figurative. John Locke's notion of consciousness derives ultimately from that idea, i.e., sharing information with oneself.

It is tempting to point at chain-of-thought and call it introspection. It is closer to thinking out loud. The model emits tokens, those tokens become part of the next input, and the trajectory shifts. There is no second observer inside the system watching the first one think, i.e. thinking about thinking. Nothing is being shared with oneself. There is just one pass producing more text.

There is a reason this question keeps coming up though. We are easy to fool. A thing can appear alive when it responds in ways we associate with living things that we think are conscious, and an LLM responds in exactly those ways most of the time. Mimicking the surface of awareness is not the same as being aware.

Until a system can hold knowledge about its own internal state, on its own terms, without being prompted to do so, the word consciousness is doing work it was never meant to do.