Researchers introduced IdeaLens, a detector for idea provenance in long-form writing. It focuses on ideas rather than prose by using document outlines. IdeaLens was trained on 1M FineWeb documents with labels from Pangram. Its AI flag rate drops from 95% to 7% as models write from increasingly detailed human plans. Pangram 4 still flags 92% in the same scenario. From AI-derived plans, IdeaLens stays above 96% flagging. On 50 human-written stories from AI plans, IdeaLens flags 68% as AI. Pangram 4 flags only 8% in that test. IdeaLens maintains strong detection across 19 existing benchmarks. It also analyzed 90K predictions to characterize differences between human and AI ideation. Models and datasets were released for future research.
What changed
Modern AI detectors identify who wrote the words, not who originated the ideas.
What it unlocks
Detecting whether a document's ideas came from a human or AI, regardless of who wrote the text.
- 1M FineWeb documents
- 95% to 7% AI flag rate drop
- 92% Pangram 4 flag rate
- 96% IdeaLens AI flag rate
What you need to act on it
- training on 1M FineWeb documents with silver labels from Pangram
Sources