What reader assumptions underlie anaphoric versus cataphoric discourse patterns?

This explores the implicit picture of the reader baked into backward-pointing (anaphoric) prose that recaps what was just said, versus forward-pointing (cataphoric) prose that sets up what's coming. The corpus has a direct anchor here: ChatGPT defaults to anaphoric organization — summarizing the ground already covered — while human student writers lean cataphoric, flagging arguments before they arrive Does ChatGPT organize text differently than human writers?. The interesting part isn't the stylistic difference; it's that each pattern encodes a different bet about who's reading and what they need.

Cataphora assumes a reader you are steering through time. To say 'I'll argue three things, and the third is the surprising one' is to model a reader whose understanding is still being built, who needs orientation toward a shared destination you both haven't reached yet. That's a forward-looking, jointly-constructed relationship. Anaphora assumes much less: a reader who mainly needs to be reminded of what was already established, so the text consolidates rather than projects. One treats meaning as something being assembled with the reader; the other treats it as something to be tidied up after the fact.

Why would a language model gravitate to the second? Two threads in the corpus converge. First, autoregressive generation is mechanically backward-looking: each token is a continuation of the tokens already present, so the natural gravity is toward summarizing and extending the existing context rather than committing to a not-yet-written future Does LLM generation explore competing claims while producing text?. Second, and deeper, the model can't actually hold the kind of reader-relationship cataphora presupposes. It treats the prompt as a fixed frame and never jointly updates the shared 'scoreboard' of a conversation — the user is the sole maintainer of common ground Can LLMs truly update shared conversational common ground?. Cataphoric writing is a promissory note to a reader whose evolving state you're tracking; if you can't track that state, you default to recapping the state you can see.

This lines up with a broader finding that the model misses the *communicative stakes* a real reader brings. It fails to adjust scalar implicature to context — it doesn't ask 'what does this listener need to infer here?' Can language models adapt implicature to conversational context? — which is the same competence cataphora demands: anticipating what a reader will need before they need it. It connects to the claim that we talk *at* language models rather than *to* them, because genuine address presupposes a partner capable of mutual orientation toward a shared future Are we really communicating with language models?.

The thing you might not have expected to learn: a dry-sounding distinction between two ways of ordering sentences turns out to be a tell. Forward-pointing structure quietly assumes a reader whose mind you are co-building toward something; backward-pointing structure assumes a reader you only ever catch up. Which pattern a writer reaches for reveals whether they think anyone is actually traveling with them — and that's exactly the capacity the corpus argues current models lack.