Can AI systems genuinely assert claims about the world?

Butlin and Viebahn set out two requirements an entity must meet "to be capable of assertion": its outputs must have "a descriptive function," and it must be "capable of being sanctioned by agents with which it interacts." Testing current systems against both, they conclude neither pre-trained nor fine-tuned LLMs clear the bar. "Pre-trained large language models that have not been subject to finetuning fail to meet the first requirement" — their outputs take the surface form of assertions without the function. "Language models that have been fine-tuned for 'groundedness' or 'correctness' may meet the first requirement, but fail the second." Using LaMDA and Sparrow as their primary examples, the paper's verdict is blunt: "optimism about AI assertion in current systems is misplaced."

The mechanism runs through what the authors call descriptive representation: an output has a descriptive function when it exists "to convey information to an observer or consumer system, so as to cause the consumer to behave as though some condition holds" — illustrated with a firefly's flash (fitness via natural selection), a thermometer's mercury level (designed for users), and a sentence spoken with communicative intent. Sanctionability is a separate, "much more demanding" bar, grounded in assertion as "a norm-governed social practice": to assert, a system must be the kind of thing that can be held accountable for getting it wrong. Fine-tuned LLMs can be "sensitive to credibility scores" and those scores can shape outputs, but a low score is not "bad or undesirable for them" in any way that matters to the system itself — the sensitivity is present without the stake. The paper also closes off proxy assertion (modeled on, e.g., a father sending his daughter to buy flour) as an escape route, arguing it does not make AI assertion easier to establish.

That flat rejection puts this paper in direct tension with Can chatbots actually assert things or just mimic assertion?, which takes up the same dilemma — chatbot outputs look assertion-like, but full assertion seems too strong a verdict — and resolves it by inserting a graded middle category, proto-assertion, rather than denying assertion outright. Butlin and Viebahn consider and reject proxy-assertion as that kind of middle ground; the proto-assertion paper treats the gradedness of the underlying conditions, not proxy status, as the resolution. The two papers agree on the diagnosis (existing binary categories don't fit) and disagree on the remedy (deny assertion vs. find a third category). The conclusion also converges, by a different route, with Are language models and human speakers doing the same thing?: both end up denying that current LLM output counts as the full human speech act, one via a sharp operation-level distinction, this one via unmet functional and social requirements.

The excerpt does not test today's frontier chatbots directly — its primary examples, LaMDA and Sparrow, are earlier fine-tuned systems, and the argument is conceptual rather than an empirical audit of any specific current deployment. It also leaves open, rather than resolves, what would need to change: the authors name "the feature which grounds interests, which may be consciousness, emotions, desires or higher-order preferences" as the missing ingredient, without claiming to know whether or how AI systems could acquire it. The implication they draw, held at just this strength, is practical rather than metaphysical: since the systems themselves cannot be sanctioned, "it is the humans who employ such systems who are responsible and sanctionable for the outputs and their effects" — a conclusion that follows from the argument even if sanctionability itself later turns out to admit the same kind of gradations proto-assertion proposes.