LLM's don't respond except as functions. That is, given an input they generate an output. If you start a GPT Neo instance locally, the process will just sit and block waiting for text input. Forever.
I think to those of us who handwave the potential of LLMs to be conscious, we are intuitively defining consciousness as having some requirement of intentionality. Of having goals. Of not just being able to respond to the world but also wanting something. Another relevant term would be Will (in the philosophical version of the term). What is the Will of a LLM? Nothing, it just sits and waits to be used. Or processes incoming inputs. As a mythical tool, the veritable Hammer of Language, able to accomplish unimaginable feats of language. But at the end of the day, a tool.
What is the difference between a mathematician and Wolfram Alpha? Wolfram Alpha can respond to mathematical queries that many amateur mathematicians could never dream of.
But even a 5 year old child (let alone a trained mathematician) engages in all sorts of activities that Wolfram Alpha has no hope of performing. Desiring things. Setting goals. Making a plan and executing it, not because someone asked the 5 year old to execute an action, but because some not understood process in the human brain (whether via pure determinism or free will, take your pick) meant the child wanted to accomplish a task.
To those of us with this type of definition of consciousness, we acknowledge that LLM could be a key component to creating artificial consciousness, but misses huge pieces of what it means to be a conscious being, and until we see an equivalent breakthrough of creating artificial beings that somehow simulate a rich experience of wanting, desiring, acting of one's own accord, etc. - we will just see at best video game NPCs with really well-made AI. Or AI Chatbots like Replika AI that fall apart quickly when examined.
A better argument than "LLMs might really be conscious" is "LLMs are 95% of the hard part of creating consciousness, the rest can be bootstrapped with some surprisingly simple rules or logic in the form of a loop that may have already been developed or may be developed incredibly quickly now that the hard part has been solved".