Can human review keep pace with AI-accelerated research generation?

The argument behind PAT is structural, not incremental: once AI drives generation — hypotheses, code, proofs — human peer review becomes the load-bearing constraint, because the cognitive labor of line-by-line verification does not scale to match automated output. The authors' resolution is uncomfortable but consistent: if you accept AI-accelerated generation, you are logically committed to AI-accelerated verification, or the whole pipeline stalls at review.

To govern that transition they propose a taxonomy of four progressive levels of AI-human collaboration in evaluation. PAT today sits at Role 1 (a tool for authors to self-check before submission) and Role 2 (a tool for reviewers to augment their reading), with higher levels of autonomy explicitly deferred. This staging is the intellectually honest move — it refuses the binary of "AI reviews everything" vs. "humans review everything" and instead names intermediate contracts where the human stays in the loop but is relieved of the exhaustive parts.

This complicates the "just detect AI content" reflex. Since Does more automation actually hide rather than eliminate errors?, a taxonomy of collaboration roles is exactly the governance scaffolding that detection alone cannot provide — it specifies who is accountable at each level rather than assuming a classifier can police the boundary. And since Can automated review loops handle AI-generated research at scale?, the taxonomy is the missing rung: it describes the human-in-the-loop levels that a fully closed automated venue skips over.