Most AI detectors ask who appears to have written the words. IdeaLens asks a different question: who appears to have supplied the plan behind them? The distinction matters when a person writes polished prose from an AI-generated outline, or when a model turns a detailed human plan into fluent text.[1][2]
Those mixed cases are increasingly common in education and publishing. A prose detector may identify the visible writer while missing the origin of the argument, sequence of claims or proposed solution.
Strip away the wording, keep the discourse
IdeaLens converts a document into an outline. Each item pairs a discourse role—such as a claim, limitation, example or proposed method—with a short paraphrase of the content. The representation is intended to minimise surface overlap with the original wording, forcing the classifier to rely more heavily on conceptual organisation.[1]
Four authorship combinations that ordinary detection can confuse
| Ideas | Final prose | What a prose detector sees | What IdeaLens is designed to test |
|---|---|---|---|
| Human | Human | Human language and human plan | Whether the plan resembles human ideation |
| Human | AI | AI wording over a human plan | Whether detailed human planning remains visible beneath the proseSource 1 |
| AI | Human | Human wording over an AI plan | Whether AI-derived structure survives a human rewriteSource 1 |
| AI | AI | AI wording and AI plan | Both language and conceptual structure may signal AI |
The controlled results expose the gap
What changed when prose and plans were separated
| Measure | Value | Context |
|---|---|---|
| AI writes from increasingly detailed human plans | 94.9% to 6.8% IdeaLens flag rate | As the human plan became more detailed, the ideas looked increasingly human even though AI wrote the prose.Source 1 |
| AI writes from AI-derived plans | Above 96% IdeaLens flag rate | The detector continued to identify AI-like ideation.Source 1 |
| Humans write 50 stories from AI plans | 68% IdeaLens; 8% Pangram 4 | The idea detector found much more AI involvement than the prose detector.Source 1 |
| Existing benchmark suite | 19 benchmarks | The authors evaluated performance across domains, formats and languages.Source 1 |
The asymmetry is the contribution. When AI writes from a detailed human plan, IdeaLens mostly stops flagging the work even though a prose detector still recognises AI wording. When humans write from AI plans, the reverse occurs: the language looks human, but the conceptual outline retains signals the model associates with AI ideation.
What the model thinks human and AI ideas look like
The researchers analysed about 90,000 predictions. Their account suggests that human plans were more likely to include verifiable sources, meaningful details, unresolved problems and limitations. AI-derived plans more often contained generic details, sweeping claims or apparently complete solutions. Those are statistical tendencies in the training and evaluation data, not universal rules for human creativity.[1]
A detector is not a provenance record
IdeaLens was trained on one million FineWeb documents using “silver” labels supplied by another detector. Its predictions therefore inherit assumptions and errors from the labelling pipeline. The authors also report weaker performance for some newer models and explicitly state that a prediction is not direct evidence of the writing process.[1][2]
The broader insight is more durable than any one classifier. Authorship is no longer a single variable. A document can have a human problem, an AI outline, human evidence, AI prose and a human final edit. Policies that ask only whether “AI wrote this” are becoming too crude for the workflows they are trying to govern.
Sources & further reading
- IdeaLens: Detecting AI Ideas in Long-form Writing arXiv · 5 Oct 2026 · Accessed 15 Oct 2026
- Is this idea from AI or humans? A detector claims it can tell Nature · 13 Oct 2026 · Accessed 15 Oct 2026