Many MindedA detector that reads a paper's argument structure to spot AI authorship
A team at the University of Maryland, College Park has built IdeaLens, an AI-text detector that ignores phrasing and looks at the structure of a paper's ideas, arguments and narrative. It is still in development.
To train it, the researchers used the Pangram detector to label about one million texts, including news articles, blog posts and academic papers, as human- or AI-written. A large language model then rewrote each text as an outline, stripping the original wording and leaving background, claims, evidence and arguments. A classifier learned to tell human-origin outlines from AI-origin ones. Human outlines tended to hold more study-specific details, while AI outlines held more generic points that could fit other studies.
The reported numbers: IdeaLens wrongly labeled text from ChatGPT and other large language models as human about 1 percent of the time. It correctly identified human origin in about 95 percent of lightly AI-edited texts. Across evaluation datasets it missed roughly 10 percent of AI-generated work. Author Rishanth Rajendhran told Nature the team prefers missing some AI documents over wrongly flagging humans.
Why it matters: trustworthy knowledge gets cheaper when readers and reviewers can filter out generic machine output. The pressure is real. One estimate says about 12.5 percent of biomedical articles show signs of substantial AI involvement, and 100 or more hallucinated citations got past reviewers at NeurIPS 2025.
The caveats are significant. IdeaLens is a black box, and the team cannot fully explain its verdicts. They did not individually verify the origin of every training document, and the tool was not tested on hybrid articles where humans and AI co-develop ideas. It will not stop the flood of AI content alone. The source text was truncated, so other limitations may be missing.
Source: Singularity Hub