Can chatbots actually assert things or just mimic assertion?

The paper ("Chatting with Bots: AI, Speech Acts, and the Edge of Assertion") asks whether chatbots can assert — "one of the many things that we do with language... a speech act" — framing the pro-view as the Thesis of Chatbot Assertion (TCA): "at least some of the output produced by current-generation chatbots qualifies as assertion." It finds both the motivations for TCA (chatbot outputs' "proper function" is to be informative; fine-tuned LLMs acquire "descriptive functions") and the objections to it "weighty," producing "something of a dilemma." Two standard ways of splitting the difference — treating chatbot output as proxy-assertion, or as fiction — are tried and rejected. Instead the authors propose a third category, proto-assertion, placing chatbots "somewhere in the space between beings that are fully-fledged asserters and those that are non-asserters," as "'edge-cases' when it comes to the capacity for assertion."

The argument runs through an ontogenetic analogy. Before full assertion, a child passes through a stage resembling a "proto-walker" learning to walk: a toddler who answers "zoo!" to "what did you do today?" seems to assert something, yet it's doubtful she meets the full conditions for assertion — understanding the content, holding the right mental attitudes, mentalizing, or being fully "sanctionable" — even though she has a "rudimentary grasp" of assertoric norms and can be corrected ("that's not a dinosaur, that's a rhinoceros"). Because each condition on assertion "admits of gradations" rather than switching on at one threshold, a system can satisfy some conditions partially without satisfying all fully. The authors distinguish proto-assertion from Nickel's proxy-assertion (responsibility displaced wholly onto a principal, as when a father sends his daughter to buy flour) and from Freiman & Miller's "quasi-testimony" (mere phenomenological resemblance plus a truth-tracking norm), judging both inadequate to the chatbot case.

This complicates the sharper line drawn in Are language models and human speakers doing the same thing?, which treats human communication and LLM string-production as "structurally different operations" — the proto-assertion account instead treats the human side of that line as itself gradient, since children only gradually acquire the capacities that note treats as constitutive of communication. Its developmental argument runs alongside Does human language use ever exist outside communication?, which also roots language use in communicative acquisition — but where that note draws a hard species line excluding LLMs, this paper uses the same developmental lens to place chatbots partway along the continuum children travel. It is a philosophical cousin to What attitudes hide behind identical claims that chatbots are conscious? and to How do chatbots enable distributed delusion differently than passive tools?: all three resist a binary verdict on surface-identical chatbot behavior by inserting a distinctive middle category.

The excerpt does not show that any specific chatbot clears the proto-assertion bar in practice; the argument is conceptual — that a middle category must exist and chatbots plausibly occupy it — not an empirical test of particular outputs. It also concedes, without resolving, that the child case might instead have a "sharp, clearly-defined" boundary merely hidden from observers, in which case the analogy's premise fails. The implication, held only as strongly as that premise holds, is that "can chatbots really assert" may be better posed as a question of degree than of kind.