Does AI generate diverse claims or diverse perspectives?

The mechanical output of an LLM is claims — well-formed propositions, arranged in plausible sequence. What LLMs do not produce is points of view. A point of view is a position a speaker occupies relative to other positions, which means it requires knowledge of the field of positions, investment in one of them, and responsiveness to counterpositions. Claims can be generated without any of that.

The mechanism matters. Does LLM generation explore competing claims while producing text?: the model does not canvas the rhetorical neighborhood of a claim before producing the next token. It produces the most probable continuation given the prompt, which means the generation path tracks the contour of the training distribution, not the contour of argumentative space. Best-of-N and beam search rank output by scoring functions that are not rhetorical — they do not know which counterposition this claim is answering. RLHF and alignment tune further against exploration, because exploration surfaces friction, and friction reduces user satisfaction.

So the output grows in volume without growing in perspectival diversity. A thousand AI-generated articles on a topic contain a thousand claims and approximately one point of view — the one the training distribution and alignment regime jointly privilege. This is why AI text often feels diverse at the token level and monotonous at the argumentative level. The proliferation is real; the perspectival proliferation is an illusion.

This bears directly on discourse quality. Discourse is not a collection of claims; it is a distribution of positions in tension with each other. AI increases the claim count while compressing the position count. How does AI writing escape the conversations that govern knowledge? is the consequence at the system level; this is the consequence at the output level.

Source: Epistemic Inflation