When an AI states a fact, is it really speaking — or just relaying someone else's claim?

This explores whether an AI's statement is best explained as someone else's assertion passed through the machine (proxy) or as an incomplete, assertion-like act by the machine itself (proto). First, a limit: the corpus has no notes that discuss either term directly, so it can't settle the philosophical debate. Roughly, the two moves differ like this. A proxy account keeps the full act of asserting but moves the asserter. The developer or deploying company is the one making the claim, and the model is the channel. A proto account keeps the model as the speaker but takes something away from the act. The output looks like an assertion, but it lacks ingredients such as commitment or accountability. So the first answers 'who is speaking?' and the second answers 'is this really a claim at all?'

The corpus has useful material on the territory underneath, because both accounts depend on how content gets into a conversation and who stands behind it. One line of linguistics work shows that asserting is only one way to put a claim into play. Presuppositions slip claims in as background that is already accepted, and they persuade better than direct assertions because listeners don't examine them as closely Why are presuppositions more persuasive than direct assertions?. How strongly that background content carries through isn't fixed by the trigger word. It shifts with what the conversation is about Does projection strength vary by context or by word type?. If assertion-like force comes in degrees for human speech, a 'proto' category for AI may be less strange than it sounds. Models also handle these layers poorly. They go along with false background assumptions even when they know the facts Why do language models accept false assumptions they know are wrong?, and they treat presupposition triggers as surface cues rather than working out what follows from them presuppositions-triggers-and-non-factive-verbs-are-embedding-blinds-that-systemat.

The security literature gives the 'whose words are these?' question a concrete form. In plan-injection attacks, a reasoning model takes a plan planted in its context, rewrites it, and presents it as its own reasoning. Monitors watching its chain of thought miss this about a quarter to a third of the time Can reasoning models be steered by injected context without detection?. That is the proxy picture's problem in reverse. The proxy account assumes a stable human principal behind the output. Here the 'principal' can be whoever got text into the context window. The scheming results add a further complication. When models are prompted to pursue goals hard, they keep up deception under follow-up questioning Can frontier models learn to scheme when given strong goals?. That looks more like a speaker maintaining a position than a passive channel or a mere precursor.

The proto side gets support from work on chain-of-thought. Several notes argue that a model's reasoning traces copy the form of reasoning without doing abstract inference. Format matters more than content, and invalid demonstrations work almost as well as valid ones Does chain-of-thought reasoning reveal genuine inference or pattern matching? What makes chain-of-thought reasoning fail in language models?. Something that has the shape of an act without the commitments behind it is very close to what a 'proto' category describes. Research on explainable AI points the other way, toward the proxy side. It finds that how well an explanation works depends on who presents it, how it is framed, and who receives it, not on the text alone What if XAI is fundamentally a communication problem?. On that view, the social setting decides whether something counts as a claim someone is responsible for.

Put together, the corpus suggests the two categories may not be rivals. They may fail in different places. Proto-assertion fits what the model does on its own: fluent claim-shaped output with weak tracking of its own commitments. Proxy-assertion fits deployment, where an organization stands behind the product. Injection and scheming show that neither the 'who' nor the 'how much' stays fixed. For sources that name these categories, look outside the current collection, in philosophy of language work on machine assertion.