Does ChatGPT organize text differently than human writers?

A specific syntactic finding from the metadiscursive nouns comparison: ChatGPT relies heavily on anaphoric references (pointing backward to previously discussed material), while students demonstrate greater use of cataphoric references (pointing forward to material that is about to be introduced).

In practical terms:

• Anaphoric: "The above analysis suggests..." / "As discussed earlier..." — summarizing
• Cataphoric: "The following argument will show..." / "Consider three reasons..." — framing what comes next

This is not a trivial stylistic preference. The choice of anaphoric vs. cataphoric structure reflects a fundamentally different model of the reader. Cataphoric structure assumes an active reader who needs a roadmap: you tell them where you're going before you take them there. Anaphoric structure assumes a passive reader who is following along: you refer back to what you've established.

Effective academic argument typically uses cataphoric structure to build anticipation and signal logical progression. ChatGPT's preference for anaphoric structure means it tends to summarize what it has said rather than set up what it is about to argue — a writing habit that is organizationally safe but rhetorically weak.

The deeper implication: this pattern may reflect something about how autoregressive generation works. Token-by-token generation is inherently backward-looking (each token is conditioned on prior tokens). Generating cataphoric structure requires projecting forward to what will be said, which is a higher-order planning operation that autoregressive generation doesn't naturally support.

Source: Discourses