Large language models and the freedom of expression

This paper defends the view that the speech-like outputs of large language models (LLMs) ought pro tanto to be extended the same protection from state interference as human-generated speech. I argue that influential audience-based and speaker-based justifications for free speech rights ground the protection for LLM outputs to the same or nearly the same extent that they ground human beings’ rights to free speech. Then I consider four potential defeaters: deception (people are likely to be deceived into thinking they’re interacting with other humans, not LLMs); responsibility (no one can be held responsible for LLM outputs); discursive pollution (persuasion by LLMs is likely to covertly favor the interests of the powerful); and scale (swarms of LLM-powered chatbots may distort our understanding of political reality). I find that they do not in fact defeat PT at present, though some concern for the future is warranted.