Most AI writing detectors attempt to determine whether a person or an artificial intelligence model produced a piece of text. They typically analyze linguistic patterns, word choices and sentence structures to estimate whether the writing resembles machine-generated content. However, these systems face a fundamental problem: the person who writes the words is not always the person—or system—that originally developed the ideas.
A new research project called IdeaLens explores a different approach. Instead of concentrating primarily on the final wording, the detector attempts to identify whether the underlying ideas in a document originated with a human or an AI model.
The research, published as an arXiv preprint on October 5, 2026, was conducted by researchers affiliated with the University of Maryland, Google DeepMind and Simon Fraser University. A report published by Tech Xplore on October 13 brought the work to wider attention.
The distinction matters in academic publishing, education and professional content creation. A writer might develop an original argument and use AI to improve the wording. Another writer might produce every sentence manually while following a detailed outline generated by an AI system. Conventional detectors can struggle to distinguish these situations because they primarily evaluate the text itself.
IdeaLens aims to address that gap by analyzing the ideas behind a document. Its early results are promising, but the technology remains a research development rather than definitive proof of who created a piece of work.
Table of Contents
What Is the IdeaLens AI Detector?
How Does IdeaLens Work?
AI Idea Detector Results: What the Research Found
Key Features of IdeaLens
Benefits for Education, Publishing and Businesses
Limitations and Ethical Concerns
How to Try IdeaLens
The Future of AI Idea Detection
Frequently Asked Questions
Conclusion
What Is the IdeaLens AI Detector?
IdeaLens is an experimental AI detection system designed to estimate the origin of a document’s ideas, regardless of who wrote the final sentences.
The research team calls this concept idea provenance. It focuses on where a document’s claims, arguments, events and overall structure appear to originate, rather than simply identifying the source of its wording.
This is an important distinction because AI involvement can take several forms.
A student might independently develop an argument and ask an AI assistant to improve grammar and clarity. A researcher might use an AI system to brainstorm possible explanations before writing a paper independently. A content creator might request an entire outline from an AI model and then write the article from scratch.
These examples involve different relationships between human work and AI assistance. A detector that examines only the final prose may not distinguish them reliably.
IdeaLens attempts to evaluate the underlying ideas instead. Its analysis looks for patterns associated with human-generated and AI-generated concepts, including the way arguments are supported, how specific details are presented and whether a document acknowledges unresolved questions or limitations.
The system is not simply searching for AI-related words. It converts a document into a structured outline designed to preserve the content and organization of its ideas while removing much of the original wording.
The research paper reports that this approach can help distinguish between AI-generated text based on human ideas and human-written text based on AI-generated ideas.
Why Is Idea Provenance Important?
The difference between writing and ideation is becoming increasingly relevant as AI tools become part of everyday work.
A person may contribute the central argument, research direction and evidence while using AI to help express those thoughts. Conversely, a person may type an entire document while relying heavily on an AI model to determine what the document should say.
Neither scenario can be understood fully by looking at writing style alone.
For universities, publishers and organizations, the relevant question may be whether a person developed the intellectual contribution, not merely whether an AI system helped produce the sentences.
IdeaLens represents an attempt to measure that distinction. However, an automated prediction about the source of ideas is not direct evidence of a person’s intentions, intellectual contribution or compliance with an institution’s AI policy.
How Does IdeaLens Work?
IdeaLens uses a different input representation from traditional prose-level detectors. Rather than relying exclusively on the original document, it processes a structured outline intended to preserve the document’s ideas while minimizing overlap with its wording.
The research team trained the system using approximately one million documents drawn from the FineWeb collection. These documents received initial human-written or AI-written labels from an existing detector, which the researchers then used as training labels for the new approach.
This detail is important: the training labels were estimates of prose origin, not independently verified records of who originally created every idea. The research team designed the outline-based method to encourage the model to learn signals associated with idea origin, but the training process still involves assumptions and potential noise.
1. Convert the Document Into an Outline
The first stage converts a long-form document into a structured outline.
Instead of preserving every sentence, the process identifies elements such as the document’s main claims, supporting arguments, narrative events and other relevant concepts. Each element receives a role and a brief description.
An outline might capture the main argument of an article, the evidence supporting it and the limitations acknowledged by the writer without copying the original sentences.
This transformation is intended to remove surface-level writing signals, making the analysis less dependent on vocabulary and sentence construction.
2. Reduce Direct Wording Overlap
The outline descriptions are paraphrased to reduce overlap with the original text.
This is central to IdeaLens’s design. If the detector received the entire original document, it could potentially rely on the same linguistic patterns used by conventional AI writing detectors.
By using a simplified representation of the document, IdeaLens attempts to focus its classification on what the text communicates rather than exactly how it communicates it.
The method cannot guarantee that all stylistic information disappears. The outline is itself produced through a language-model-based process, so its structure and content may retain signals influenced by the source document and the outline-generation method.
3. Analyze Patterns in the Ideas
The system evaluates the resulting outline and estimates whether its ideas appear to originate from a human or an AI model.
In the research team’s analysis of approximately 90,000 item-level predictions, some patterns were associated more strongly with human ideas. These included specific, verifiable references and meaningful discussion of limitations or unresolved problems.
Other patterns were associated more strongly with AI-generated ideas, such as generic details, sweeping claims and numbers or methods presented without enough context to verify them.
These are statistical observations across the evaluated data, not universal rules for judging individual pieces of writing. Human authors can produce generic arguments, and AI systems can generate detailed, well-supported work.
4. Produce a Classification
After processing the outline, IdeaLens produces a score that can be converted into a human-or-AI classification using a selected decision threshold.
That threshold affects the balance between detecting AI-generated material and incorrectly flagging human work. A setting designed to minimize false positives will generally involve a trade-off in detection sensitivity.
The system is therefore best understood as a probabilistic detector, not a definitive authorship verification tool.
AI Idea Detector Results: What the Research Found
The early results suggest that examining ideas can provide useful information that conventional text-focused systems may miss. The most important finding is not simply that IdeaLens detects AI-generated writing, but that it responds differently when the origin of a document’s ideas differs from the origin of its words.
The researchers evaluated the system on conventional AI detection benchmarks as well as experiments involving mixed human-AI workflows.
Results on Conventional AI Detection Benchmarks
Across 19 evaluations designed to distinguish fully human-written documents from fully AI-generated documents, the researchers reported that IdeaLens identified approximately 91.1% of AI-generated documents, with a false-positive rate of about 0.6% at the selected operating threshold.
| Evaluation measure | Reported result |
|---|---|
| AI-generated documents identified | 91.1% |
| False-positive rate on human documents | 0.6% |
| External benchmark evaluations | 19 |
These figures indicate promising performance within the study’s evaluation settings. They do not mean that the tool will be 91.1% accurate for every submitted document, subject area, institution or AI model.
A false-positive rate also differs from the probability that a particular flagged document is human-written. The interpretation of a detector’s result depends on the quality of the evaluation, the prevalence of AI-generated material in the documents being assessed and the threshold chosen.
When AI Writes From a Human-Developed Plan
One experiment examined AI-generated writing produced from increasingly detailed human plans.
This scenario is common in assisted writing. A human may establish the topic, argument, structure and key supporting points, then ask an AI model to turn that plan into polished prose.
As the human plans became more detailed, IdeaLens’s AI flag rate fell from approximately 95% to around 7%. In comparison, the conventional detector Pangram 4 still flagged approximately 92% of the documents as AI-generated.
This does not mean the AI stopped generating the words. It means IdeaLens increasingly treated the ideas as human-originated, while the prose detector continued responding to the machine-generated writing.
The difference illustrates the separate questions the two approaches are designed to answer.
When Humans Write From AI-Generated Ideas
The researchers also tested the opposite scenario: human authors writing stories based on outlines created from AI-generated stories.
In a dataset of 50 stories, IdeaLens flagged 68% of the human-written stories as having AI-originated ideas. Pangram 4 flagged 8%.
This result suggests that IdeaLens can detect potential AI involvement even when the final wording was produced by a person. However, the experiment used a relatively small dataset, and its result should not be treated as a general error rate across every form of human-AI collaboration.
The finding also raises an important policy question. If a person writes every sentence but uses an AI-generated outline, should the resulting document be classified as AI-generated, human-written or collaborative? The answer depends on the applicable rules and what the classification is intended to measure.
What the Results Really Mean
The research supports a broader way of examining AI-assisted writing: the origin of words and the origin of ideas are separate dimensions.
IdeaLens may offer useful evidence about the second dimension, particularly when AI and human contributions are mixed. But its prediction cannot, on its own, establish misconduct or determine whether AI use violated a specific policy.
The researchers also reported weaknesses in detecting content from certain newer AI models, indicating that continued testing and refinement are necessary.
Key Features of IdeaLens
1. Idea-Level Analysis
IdeaLens is designed to evaluate the underlying concepts and arguments in a document rather than relying solely on surface writing patterns.
That makes it conceptually different from detectors that primarily look for the statistical characteristics of AI-generated prose.
2. Support for Mixed Human-AI Workflows
Many documents are not entirely human-created or entirely machine-generated. IdeaLens’s experiments explicitly examine cases where humans supply ideas but AI writes the text, and cases where people write text based on ideas generated by AI.
This makes its approach relevant to discussions about AI-assisted writing, authorship and intellectual contribution.
3. Structured Outline Processing
The detector uses role-labelled outlines that capture a document’s content in a more compact form. The outlines are intended to preserve ideas while reducing dependence on the original wording.
This is the central technical feature that separates IdeaLens from a conventional prose-level classifier.
4. Research Benchmarks and Evaluation Data
The project includes evaluations across multiple benchmarks, alongside purpose-built datasets designed to examine idea provenance.
The researchers have also made the code, models and evaluation resources publicly available, allowing other researchers to study the approach and assess its limitations.
5. Public Demo and Open Research Resources
People interested in the technology can explore the IdeaLens demo and inspect the project’s public GitHub repository.
The repository describes a pipeline that classifies a document’s format, extracts a structured outline and scores that outline with the IdeaLens model. The availability of these resources supports further investigation, but it does not mean the detector has been validated for every high-stakes decision.
Benefits for Education, Publishing and Businesses
More Detailed Academic Integrity Reviews
Educational institutions increasingly need to distinguish between permitted AI assistance and the unauthorized use of AI in assessed work.
An idea-focused detector could add a new source of information to an academic integrity review, especially where an institution’s policy considers the origin of arguments or intellectual contributions.
However, its results should not replace examination of drafts, notes, research records, disclosure statements or direct communication with the author.
A student who is incorrectly flagged could face serious consequences. For this reason, a detector’s score should never be treated as conclusive proof of misconduct.
Better Context for Scientific Publishing
Researchers and journal editors need to evaluate whether a paper presents original thinking, reliable evidence and appropriate attribution.
IdeaLens could potentially help identify patterns worth reviewing when a document appears to reproduce generic AI-generated arguments or when the intellectual contribution is unclear.
Still, detecting the likely origin of ideas is not the same as assessing the scientific validity of those ideas. Original ideas can be incorrect, and AI-assisted research can contain valid contributions when it is conducted transparently and responsibly.
Human editorial judgment remains essential.
Clearer AI Use Policies
Organizations often use AI tools for brainstorming, planning, drafting, editing and summarization. These activities can involve different degrees of human contribution.
An idea-level perspective may help policy makers distinguish among forms of assistance instead of applying one broad definition of AI-generated work to every situation.
Organizations should define what counts as acceptable AI use before relying on detection tools. A system cannot resolve a policy ambiguity simply by producing a classification score.
Support for Content Quality Reviews
Businesses producing reports, research summaries, marketing material and technical documentation may find the idea-provenance concept useful when reviewing the originality of intellectual contributions.
It could serve as one signal in a wider review process, alongside source verification, editorial evaluation and the examination of document history.
It should not be used as a substitute for checking facts, citations or whether the content genuinely meets the intended brief.
Limitations and Ethical Concerns
Despite its promising results, IdeaLens has important limitations that researchers, educators and businesses should understand before using it in real-world decisions.
A Research Preprint, Not Definitive Proof
The work was published on arXiv as a research preprint. It provides detailed experiments and a foundation for further investigation, but its reported performance should not be confused with universal real-world validation.
Independent evaluation across broader populations, document types and institutional settings will be important before the detector is used in high-stakes contexts.
The Training Labels Have Limitations
The researchers trained the model on approximately one million documents labelled by an existing prose detector. Those labels were not direct, independently verified records of the source of every idea in each document.
Converting prose-level labels into an idea-level detector is a central part of the research contribution, but it also creates a methodological challenge: the model learns to infer idea provenance from patterns in data whose original labels concerned the wording.
The approach is promising, but the assumptions behind the training process deserve ongoing scrutiny.
False Positives Can Harm Human Writers
No AI detector should be assumed infallible. The reported false-positive rate was low in the study’s selected conventional evaluations, but actual results may differ when the detector encounters unfamiliar topics, writing styles or models.
The issue is especially sensitive when the detector flags human-written text because the ideas may have originated from an AI system.
A classification should be treated as a reason for careful review, not an automatic verdict about authorship or academic misconduct.
The Origin of an Idea Is Not Always Clear
Ideas are often developed collaboratively. A writer may combine conversations with colleagues, research papers, brainstorming tools, AI suggestions and personal experience.
In such cases, there may be no simple, objective division between human-generated and AI-generated ideas. A detector’s binary label can oversimplify a process that involves multiple contributors and stages.
AI Models Continue to Evolve
New models may produce ideas that differ from the patterns learned by the detector. The researchers themselves found that some newer AI models were harder to identify, highlighting the need for continued evaluation.
Detection systems must adapt as generation methods improve, and their performance should be reassessed as the technology changes.
Detection Should Not Replace Transparency
For many organizations, a clear disclosure of AI use is more useful than attempting to infer every contribution after a document has been completed.
Maintaining drafts, research notes, source records and appropriate declarations can help establish how a document was created. These records provide context that a detector cannot reliably reconstruct from the final text alone.
How to Try IdeaLens
The research team has published a public demo and an open-source repository for people who want to examine the technology.
Try the detector: IdeaLens demo
Explore the code: IdeaLens on GitHub
Read the research: IdeaLens: Detecting AI Ideas in Long-form Writing
The repository explains that IdeaLens classifies a document’s format, extracts a role-labelled outline of its ideas and scores that outline to estimate idea provenance. It also provides access to the companion ProseLens detector and the project’s evaluation resources.
Anyone trying the demo should interpret its output cautiously. A detector’s classification cannot independently establish who wrote a document, whether a particular idea was original or whether a writer broke an institutional policy.
For research, teaching and professional use, it is best to treat the result as one data point in a wider evaluation process.
The Future of AI Idea Detection
AI detection is moving toward a more nuanced discussion about what it means for a document to be human-created or AI-generated.
Early detection approaches concentrated mainly on whether the wording resembled output from a language model. IdeaLens explores a different dimension by estimating whether the underlying concepts and arguments originated with a person or an AI system.
That distinction may become increasingly important as AI tools are used for more than text generation. They can assist with outlining, research planning, argument development, summarization and creative brainstorming before a final document is ever written.
Future detection systems may therefore need to examine several dimensions of authorship, including idea provenance, prose provenance, source reliability and the history of human-AI collaboration.
For that to work responsibly, developers will need to explain what their systems measure, publish clear evaluation results and acknowledge the limits of their predictions. Educational institutions and publishers will also need transparent rules about how detection results can be used.
IdeaLens is a meaningful research step in that direction, but it is not a complete solution to AI authorship. Its biggest contribution may be showing why the source of ideas and the source of words should not be treated as the same question.
Frequently Asked Questions
What is IdeaLens?
IdeaLens is an AI detector designed to estimate whether a document’s ideas originated from a human or an AI model, regardless of who wrote the final text. It uses a structured outline of the document rather than relying solely on the original wording.
How is IdeaLens different from traditional AI detectors?
Traditional detectors often focus on linguistic patterns that may indicate AI-generated prose. IdeaLens attempts to identify the origin of a document’s ideas, which can differ from the source of its words in mixed human-AI workflows.
How accurate is IdeaLens?
The researchers reported that IdeaLens identified 91.1% of AI-generated documents with a 0.6% false-positive rate across 19 conventional evaluations at the selected threshold. These figures apply to the study’s test settings and should not be interpreted as a guarantee of accuracy for every document.
Can IdeaLens detect AI ideas in human-written text?
In a research experiment involving 50 stories written by humans from AI-generated outlines, IdeaLens flagged 68% as AI-originated in terms of ideas. This result suggests potential for detecting AI involvement beyond the final wording, but the experiment was limited in size and does not establish a universal detection rate.
Does IdeaLens prove that someone used AI improperly?
No. A detector estimates the likely origin of ideas or text; it does not establish misconduct, intent or whether a particular use of AI violated a policy. Any high-stakes decision should consider additional evidence and human review.
Can I try IdeaLens online?
Yes. The research project provides a public demo at ideadetector.ai, alongside its GitHub repository and research paper. Users should review the tool’s information-handling practices before submitting sensitive documents.
Is IdeaLens ready for professional use?
It is available for exploration and research, but users should be cautious about relying on it for high-stakes decisions. The research is an early development, and further independent evaluation is important, particularly for newer AI models and mixed human-AI workflows.
Conclusion
IdeaLens introduces a different way to approach AI detection by looking beyond the words in a document and examining the ideas those words communicate. Its research results suggest that this approach can reveal forms of human-AI collaboration that conventional prose detectors may overlook.
The distinction could matter in education, scientific publishing and business content creation, where the origin of an argument may be as important as the wording used to express it.
However, IdeaLens remains a research-stage detector whose predictions require careful interpretation. Its results should support transparent review rather than replace evidence, human judgment or clear AI-use policies.
As AI becomes more involved in brainstorming, planning and research, understanding the relationship between ideas and words will become increasingly important. IdeaLens offers an early look at how future detection systems might approach that challenge.
Need Expert IT Support for Your Business?
AI tools can support research, content workflows and decision-making, but getting reliable results requires a thoughtful approach to technology, data protection and digital operations.
Progressive Robot provides software development, application development, cybersecurity, IT consulting and digital solutions to help businesses adopt technology effectively. Whether you are assessing AI tools or planning a broader digital transformation, our team can help you build secure and practical technology workflows.
