Science & SpaceAI Novelty Judges Are Too Unreliable to Trust, New Benchmark Shows
A study from the Allen Institute for AI tested six leading models on whether they could spot a genuinely original research idea. A single prompt tweak sent one model's accuracy below random chance.