AI ranked 6,617 ICML papers for impact; humans favored novelty

AI and human reviewers picked different kinds of machine-learning research when ranking the same papers. The AI favored perceived impact and performance across different settings, while humans gave more weight to novelty.

Different AI systems also agreed strongly with one another, but their rankings were sensitive to online hype. That creates a live concern about handing scientific direction to models that reward visibility and broad applicability over originality. The comparison covered all 6,617 ICML 2026 papers.