The Institutional Palate

Artificial intelligence matches or exceeds human performance on tasks with verifiable answers — protein folding, Olympiad mathematics, code generation. But the capacity that most governs scientific advance is not reasoning. It is taste: the ability to judge which untested ideas deserve pursuit. Editors and funders exercise this judgment daily. No one has successfully articulated it, taught it, or automated it.

arXiv:2603.16659 demonstrates that fine-tuned language models trained on journal publication decisions can evaluate research pitches more effectively than either frontier AI systems or human expert panels. Using management research proposals: frontier models achieve 31% accuracy, expert panels reach 42%, but fine-tuned models surpass both at 59%.

The fine-tuned models exhibit calibrated confidence — they know which judgments they are sure about and which are borderline. They transfer their evaluative capability to new formats without retraining. The approach also succeeds in economics, reaching 70% accuracy.

The implication: scientific taste is not ineffable. It is pattern-extractable from institutional records — decades of accept/reject decisions that encode what the field considers worth pursuing. The patterns exist in the data, not in the reasoning of individual editors. No single editor could articulate the rules, but the aggregate of thousands of decisions contains learnable structure.

This is uncomfortable for two reasons. First, it suggests that gatekeeping — the activity academics complain about most — is the activity that contains the most extractable judgment. Second, it means the judgment is institutional, not individual. The fine-tuned model does not learn what any editor thinks. It learns what the institution, across thousands of decisions, has repeatedly chosen. The taste belongs to the journal, not the reviewer.