Well it much more unpredictable than a human. Liability would require AGI and the AGI agents to be predictable. But the LLM models are more like semi intelligent toddlers with an encyclopedic knowledge base. They can reason somewhat but understand little. With the right prompt, you can get wildly weird behavior.

You could demand that corporations are responsible to mitigate if something goes wrong - if it can be mitigated. Similar to a forum admin who needs to take down an illegal post after it's become known to them.

You can say for something like an application for taking medical notes there is liability. Because you have a choice to not use that, or to double check their ai assistant. Or with the company who sells the software if they make false claims or misleading marketing.

But in general? Who knows what they'll say lol.

LLMs generate language, but not a single human on earth understands how they reason. How they actually "think". We do not have a theory for that, not for human and not for LLMs. We know how it starts (text prediction) and figured out how to increase this massively and we understand the basic atoms just like we understand neurons - but we do not know how the actual "thinking" works. This is what makes LLMs actually miraculous (at least to me).

We can understand human minds through psychoanalysis. Obviously LLMs aren't sentient but they are also much less "bound by nature". They are fever dream machines that you can sort of put in a harness, or let rerun until some test criteria is fulfilled. But a user could take some AI model and retrain it or do all sorts of things. How could the LLM take responsibility? It would have to predict what happens outside of it's "black box"? It would literally have to understand what the question means and what it will be used for, where they are in what country with what legislation etc. That requires general knowledge / intelligence.