AI Detectors Were Checking the Wrong Thing — This One Reads Your Ideas, Not Your Words

You know the problem: paste AI-polished text into a detector and it flags you even though the thinking was yours. Let an AI brainstorm the whole plan and rewrite it yourself, and the same detector calls it human. Today's detectors judge how you wrote, not what you thought — and a new paper finally fixes that mismatch.

IdeaLens, by Rishanth Rajendhran and 7 co-authors (submitted October 5, 2026, arXiv:2610.06778), is the first AI detector that asks: who came up with the ideas? — regardless of who wrote the words. In a controlled study, its AI flag rate collapsed from 95% to 7% as models wrote from increasingly detailed human plans, while the prose detector Pangram 4 still flagged 92% of those human-planned documents. Flip it — humans writing from AI-generated plans — and IdeaLens caught 68% where Pangram 4 caught just 8%.

What the researchers did

The core trick: strip the prose away before you judge. IdeaLens converts every document into an outline — each item pairs a discourse role (claim, evidence, example, caveat) with a brief paraphrased summary of the content. The paraphrase deliberately minimizes word-level overlap with the original, so the detector cannot cheat on writing style. What's left is the skeleton of the thinking itself.

Training data was the hard part — no "idea provenance" labels exist. So the team trained on 1 million FineWeb documents with silver labels from Pangram, an existing prose detector. Because the outlines carry almost no surface-level information, those labels could only be learned through the ideas.

Stress tests: (1) AI flag rate 95% → 7% as models wrote from increasingly detailed human plans — it correctly credited the human ideas; (2) above 96% detection on AI-derived plans; (3) on 50 human-written-from-AI-plan stories, IdeaLens flagged 68% as AI vs 8% for Pangram 4; (4) strong detection at low false-positive rates across 19 existing detection benchmarks, holding across domains, formats, and languages; (5) a 90,000-prediction analysis mapping human vs AI ideation fingerprints — human ideas cite verifiable sources, carry meaningful details, and acknowledge limitations; AI ideas lean on generic details, unverifiable numbers and methods, sweeping claims, and suspiciously perfect solutions. Caveat: IdeaLens is less effective on some newer models — the team is upfront about that. Models and datasets are released for further research.

Why it matters

Universities, publishers, and employers are quietly rewriting AI rules around ideas ("AI may polish, not think for you") — but until now no tool could enforce that distinction. A student with a human outline plus an AI draft got flagged; AI-brainstormed work with light human editing sailed through.

For developers, the bigger lesson is the method: strip surface features before you classify. Paraphrase away style, keep structure, and the signal gets cleaner. Expect "idea-level representation" in plagiarism tools, peer-review assistants, and content-quality scoring within a year.

How YOU can use it

1. Try the tool itself

Public demo at ideadetector.ai, code on GitHub (github.com/RishanthRajendhran/IdeaLens). Feed it a draft to sanity-check whether the thinking in a document is really yours.

2. Run the 90K-prediction self-audit on your writing

Human-idea fingerprints: verifiable sources (specific names, studies, dates); meaningful concrete details; acknowledges limitations and trade-offs. AI-idea fingerprints: generic details that sound specific but check out as nothing; numbers and methods that can't be verified; sweeping claims; suspiciously perfect solutions. Before publishing AI-assisted work, ask: could every claim survive a fact-check, and does it admit what it doesn't know?

3. Prompt for human-style ideation, not human-style prose

IdeaLens proves detectors are moving to the ideas layer. Use this:

Before writing, produce an outline of your ideas. For each point:
- CITE a specific, verifiable source (name, study, date)
- INCLUDE one concrete detail that can be fact-checked
- STATE one limitation, trade-off, or counterargument
Then write in plain language. Do not invent statistics,
do not make sweeping claims, do not present anything as perfect.
If you lack a verifiable source for a claim, flag it as uncertain.

4. Steal the outline-strip technique for your own evaluations

Building any classifier or reviewer? Copy the core move: paraphrase the input into (discourse role, content summary) pairs first, then judge. Minimal sketch:

# 1. Convert document to outline (style stripped)
outline_prompt = """Convert this document into an outline.
Each item: [DISCOURSE ROLE] brief paraphrased summary.
Roles: claim, evidence, example, caveat, method, conclusion.
Paraphrase - do not copy the original wording."""
# 2. Judge the outline, not the prose
judge_prompt = """Does the thinking show verifiable sources,
concrete details, and acknowledged limitations? Flag unverifiable
numbers, sweeping claims, perfect solutions."""

Key takeaways

• Ideas are the new provenance: 95% → 7% flag collapse on human plans, 96%+ on AI plans.
• Current detectors punish the wrong people on both sides of the line.
• The fingerprint checklist is free: verifiable sources + concrete details + admitted limitations = human thinking.
• Strip style before judging — a reusable pattern for any content-evaluation task.
• Watch the arms race: weaker on some newer models; this opens a front, it doesn't end the war.

Read the original paper: "IdeaLens: Detecting AI Ideas in Long-form Writing" — Rishanth Rajendhran, Minjoon Choi, Jenna Russell, Ramya Namuduri, Deniz Bölöni-Turgut, Marzena Karpinska, John Wieting, Mohit Iyyer — arXiv:2610.06778 (cs.CL, submitted 5 Oct 2026). Models and datasets released; demo at ideadetector.ai.