IdeaLens Looks Beyond the Words to Detect AI-Originated Ideas
A research team has introduced IdeaLens, a detector designed to estimate whether a document’s underlying ideas came from a person or an AI system—even when the prose itself was written by someone else. The approach reframes a fast-moving debate in research and publishing, but its authors stress that the tool is experimental and its findings should not be treated as proof of who originated a particular idea. [1] [2]
Table of Contents
- Why idea provenance is different
- How IdeaLens works
- What the tests found
- Important limitations
- What this could mean for publishing
- FAQs
Why idea provenance is different
Most AI-writing detectors analyze the language on the page. They look for statistical or stylistic patterns associated with machine-generated text. That can be useful when the question is whether a model produced the words. It does not necessarily answer a different question: who came up with the argument, plan, or sequence of ideas?
The distinction matters because human-AI collaboration can happen in either direction. A person might develop an argument and ask a model to turn it into polished prose. Or a person might write every sentence while following a structure or set of ideas proposed by a model. A detector focused only on wording can struggle to distinguish these cases. IdeaLens is intended to estimate the source of the ideas rather than the source of the final wording. [1]
The question has practical significance for universities, journals, conference organizers, and workplaces as they set rules for AI assistance. Policies may allow spelling or language editing while asking authors to disclose more substantial AI contributions. The difficulty is that idea development is often collaborative and gradual, rather than a simple handoff between a person and a machine. [2]
How IdeaLens works
Instead of feeding a full document directly into its classifier, IdeaLens first converts it into an outline. Each outline item pairs a rhetorical role—such as a claim, event, or supporting argument—with a short description of what the document says in that section. The descriptions are paraphrased to reduce overlap with the original wording. The model then evaluates this representation, which is meant to emphasize the content and structure of the ideas rather than sentence-level style. [1]
The researchers trained the system using a one-million-document subset of the FineWeb collection. Because there is no large, independently verified dataset labeling who originated every idea in long-form writing, the team used existing authorship labels from the commercial detector Pangram as “silver” labels. The researchers assumed that, for documents classified as wholly human- or AI-generated, the ideas and the prose shared the same source. They excluded documents labeled as mixed. [1]
That training choice is central to interpreting the results. IdeaLens is not trained on direct records of brainstorming or authorship decisions. It learns from prose-authorship labels after the documents have been transformed into outlines. The outline technique is an attempt to make those labels useful for idea provenance, but it does not turn the labels into ground truth about how each idea was formed.
What the tests found
In one controlled test, language models wrote documents from plans that varied in how much human input they contained. As the plans became more detailed and human-authored, IdeaLens’s rate of labeling the resulting documents as AI fell from about 95% to about 7%. Pangram 4 still labeled about 92% of documents generated from the most detailed human plans as AI. In a separate condition using AI-derived plans, IdeaLens continued to flag more than 96% of the documents as AI-generated. These results suggest the outline-based approach may be more sensitive to the origin of the plan than a detector that mainly evaluates the finished prose. [1]
The paper also tested a small set of stories written by people who followed AI-generated plans without seeing the original AI-written stories. IdeaLens flagged 68% of those human-written stories as AI-derived, while Pangram 4 flagged 8%. This is a notable result because the text was written by people, but the underlying plans came from AI. It also shows that the system did not identify every AI-influenced case: roughly one-third of these stories were not flagged. [1]
For conventional benchmarks that distinguish wholly AI-written documents from wholly human-written ones, the authors report that IdeaLens flagged 91.1% of AI documents on average at a 0.6% false-positive rate across 19 evaluations. The paper says its model also transferred to 24 languages despite being trained on English, using English outlines as an intermediate representation. These figures describe the researchers’ evaluation setup; they are not a guarantee of performance on every discipline, language community, or real-world submission process. [1]
The team’s analysis of roughly 90,000 item-level predictions offers a possible clue to what the model notices. Outlines were more likely to receive AI-associated scores when they made claims unsupported by their evidence, used numbers or methods without enough context to check them, or relied on details that could fit many topics. Specific sources and discussion of limitations or open problems tended to receive more human-associated scores. The authors present these as patterns in model predictions, not as a definitive description of how people and AI systems think. [1]
Important limitations
IdeaLens is a research prototype, and its predictions should not be treated as proof of misconduct or as a substitute for evidence about a writing process. The training labels came from another detector rather than an independently audited record of idea origins. If those labels contain errors or systematically reflect properties other than provenance, those weaknesses can carry into the new model.
The researchers’ own discussion and independent reporting highlight a deeper challenge: many documents mix human and AI contributions. A researcher might use a chatbot to brainstorm, revise the ideas, and then write the paper independently. Another might supply the core argument and use AI for drafting. A binary “AI or human” result cannot describe the full range of these workflows. The 50-story evaluation is informative but small, and it cannot represent every discipline, genre, or kind of collaboration. [2] [1]
Experts quoted by Science also questioned whether the system might rely on cues that are only indirect proxies for idea origin, such as proper nouns or other features of the outline. One expert said the method needs more stress testing and validation before it becomes part of AI policy. A model can produce a probability or classification, but that output is not a record of who first proposed an idea. [2]
What this could mean for publishing
The contribution of IdeaLens is less a ready-made verdict machine than a new way to frame detection. It asks whether a system can examine the structure and substance of a document after stripping away much of its surface language. If further research validates the approach, it could help editors investigate cases where prose detectors produce misleading results—for example, when a human’s ideas are expressed in AI-polished language.
But the method also makes a careful appeals process more important, not less. Institutions would need to explain what a score means, disclose its error rates for the relevant context, and give authors a meaningful way to provide evidence about their process. Draft histories, notes, citations, and transparent disclosure may offer context that a classifier cannot observe. The tool should not be used alone to make high-stakes judgments about authorship, academic integrity, or employment.
The researchers have released models, data, and code to support further work. The next steps include testing on independently verified examples, evaluating more realistic mixtures of human and machine contributions, and examining whether performance holds across fields and communities. Until that evidence exists, IdeaLens is best understood as an experimental research direction—one that highlights the difference between writing the words and originating the ideas. [1]
FAQs
What is IdeaLens?
IdeaLens is a research model that estimates whether the ideas in a long-form document were generated by a person or an AI system. It analyzes an outline representation rather than relying only on the original prose. [1]
How is it different from an AI-writing detector?
Traditional text detectors generally focus on who wrote the words. IdeaLens aims to focus on the ideas represented in the document, so it may distinguish human-planned ideas written by AI from AI-planned ideas written by a human. [1]
Can IdeaLens prove that a person used AI?
No. It produces a model prediction based on patterns in an outline. The training labels are indirect, mixed workflows remain difficult to classify, and researchers have called for further validation. A score alone cannot establish who originated an idea. [2]
Is the system ready for journal or university policy decisions?
The evidence is not yet sufficient to treat it as a standalone decision-maker. The paper is a preprint, and independent experts quoted in Science called for more stress testing before such tools are used in policy enforcement. [1] [2]
References
- Rajendhran et al., “IdeaLens: Detecting AI Ideas in Long-form Writing,” arXiv, 5 October 2026. https://arxiv.org/abs/2610.06778
- Jeffrey Brainard, “Did an AI or a human write that paper? New detector decides by examining its ideas,” Science, 12 October 2026. https://www.science.org/content/article/did-ai-or-human-write-paper-new-detector-decides-examining-its-ideas