New AI Tool IdeaLens Detects Machine Ideas, Not Just Machine Writing
Writers, publishers and academic editors may soon be able to judge whether a text’s ideas — not just its wording — came from a machine: researchers at the University of Maryland and Google DeepMind introduced IdeaLens, an “idea detector” that examines the argumentative structure of a text rather than its prose, according to a preprint that has not yet been independently peer-reviewed.
Schools, universities, science and media are all caught in a cat-and-mouse game with artificial intelligence: which homework, studies and reports come from a human, and which were written by a large language model? Texts can be examined, more or less reliably, for AI phrasing — but the question of where the ideas originated has remained difficult to answer.
Conventional detectors such as Pangram scan for surface features like word choice, sentence structure and statistical patterns. IdeaLens instead breaks a text — whether a study, an email or a 1000-page novel — into segments, describing each passage’s content and function, for example whether it is evidence, a recommendation or a description, according to Heise. Each entry in this plan receives a role, such as claim, evidence or context, along with a short description of its content.
To eliminate the influence of writing style, a separate language model rewrites the text, removing the author's characteristic vocabulary and phrasing; for training, these descriptions were also revised to minimize the information revealed by the original text's typical word choices. The tool can be tried out on a website, Heise and T3N report.

The system was trained on roughly 800,000 texts — a collection of one million longer English-language web texts, of which about 842,000 were used for training, Scinexx reports — labeled by the AI detector Pangram as machine-generated or human-written.
The team told Heise it views Pangram’s imperfect analyses as an advantage, adding deliberate “noise” to the training data. Testing used a corpus of texts partly from the pre-ChatGPT era, where human authorship is relatively certain.
IdeaLens caught an average of 91.1 percent of AI documents across 19 standard detection tests, while falsely flagging 0.6 percent of human texts, Scinexx reports. Its decisive advantage showed with mixed authorship: in an experiment with 50 stories written by humans from an AI-generated outline, IdeaLens classified 68 percent as based on AI ideas, while Pangram 4 flagged only eight percent as AI-generated, according to scinexx | Das Wissensmagazin.
In the reverse case — language models elaborating increasingly detailed human plans — IdeaLens raised a false alarm in only 6.8 percent of fully human-planned texts, compared with 92.4 percent for Pangram 4. On texts written with AI assistance but built on human ideas, conventional detectors like Pangram 4 classified up to 92.4 percent as AI-generated, while IdeaLens, depending on the discursive structure of the text, brought the rate down to as low as 7 percent, Heise reports.

In an analysis of roughly 90,000 individual assessments, the system linked human ideas to concrete, verifiable sourcing, and rated vague claims, barely checkable figures and solutions that gloss over difficulties as more AI-typical — though Scinexx notes this is not a universal rule of human or machine creativity. According to scinexx | Das Wissensmagazin, the figures apply only to the datasets examined and are no guaranteed hit rate for arbitrary manuscripts.
The limits remain considerable: the training labels came from another detector that can itself err, the mixed-authorship test rested on just 50 stories, and genuinely collaborative idea development between humans and models is poorly captured.
IdeaLens also cannot say whether an idea is novel, correct or scientifically valuable. It remains a research tool, not proof of a single text’s authorship — and independent tests have yet to show how reliably it performs in everyday scientific practice.