IdeaLens Targets AI Ideas; Pangram Misses 79.8% of Muse-Glimmer Rewrites

A Tokyo Metropolitan University study found Pangram missed 79.8% of scientific abstracts rewritten by Meta's Muse-Glimmer, while catching 93.5% of those rewritten by GPT-5 and flagging only 1 of 5,000 human-written abstracts. The broader pattern from the same research: detectors trained on a vendor's older models can drop from catching over 99% of AI-rewritten abstracts to catching only 3.8% after a new model generation is introduced, leading researchers to conclude detectors should be re-tested with every new model release. Separately, Google DeepMind and other labs proposed IdeaLens, a detector that analyzes a document's outline rather than its wording to determine whether ideas originated from AI or a human. On documents where AI wrote from a human's full outline, IdeaLens flagged only 7% as AI while Pangram 4 flagged 92%; on 50 stories written by humans from AI-generated outlines, IdeaLens flagged 68% while Pangram 4 flagged only 8%. A New York Times op-ed by philosopher Simon Critchley, whose piece argued AI cannot do philosophy, was detected as roughly 76% AI-written using Pangram, a result multiple observers independently replicated.