A research team has introduced IdeaLens, a model designed to estimate whether a document’s ideas came from a person or an AI, even when someone else wrote the final words. In one small experiment, it flagged 68% of 50 stories written by people from AI-generated outlines.
The distinction matters because an AI can write sentences from a person’s plan, or a person can turn an AI’s plan into prose. A text detector asks who wrote the words. IdeaLens instead tries to classify the source of the document’s claims, events and structure. Nature highlighted the project in an Oct. 14 news report; the research paper itself was posted to arXiv on Oct. 5.
IdeaLens reads an outline instead of the original prose
The system first turns a document into an outline. Each entry pairs a role, such as a claim or event, with a short description of its content. Language models create these outlines; during training, another model paraphrases them to reduce clues copied from the original wording. IdeaLens then classifies the outline rather than the full text.
To train it, the researchers used one million web documents with authorship labels supplied by Pangram, a detector of AI-written prose. They assumed the ideas and words in each training document came from the same source, making those labels rough stand-ins for idea provenance. The paper calls them “silver” labels: large-scale but noisy, rather than verified records of who first proposed each idea.
Human-written stories from AI plans were the harder test
In a controlled experiment, models wrote from plans that contained increasingly more detail from human authors. IdeaLens’s AI-flag rate fell from about 95% to 7% as the human plans became more complete; the prose detector Pangram 4 still flagged 92% of the resulting documents as AI-written. In a separate set, 17 computer-science researchers wrote 50 stories from scratch using only outlines generated by an AI. IdeaLens flagged 68% of those stories as AI-idea work, compared with 8% for Pangram 4.
The 50-story test was small and staged, so the preprint does not establish whether IdeaLens can reliably resolve authorship disputes in real publishing or academic settings.