AI Detectors Catch Who Wrote the Words. A New Study Asks Who Had the Idea.

Quick answer

Every AI detector on the market today, including ours, measures one thing: whether the words in front of it carry the statistical fingerprint of machine-generated prose. A University of Maryland, Google DeepMind, and Simon Fraser University research team published a paper on October 5, 2026, arguing that question misses something policies increasingly care about. They call it "idea provenance": did the argument come from a person or a model, regardless of who typed the sentences? Their prototype detector, IdeaLens, catches AI-originated ideas dressed in completely human prose at rates existing detectors miss almost entirely. It also does the opposite: it correctly clears AI-written text built from a detailed human outline, something word-pattern detectors still flag as machine-made. The gap between "who wrote the words" and "who came up with the idea" isn't theoretical. It's exactly the question Stanley Druckenmiller's editor had to answer in public, two months before this paper came out.

The op-ed that the paper opens with

In August 2026, hedge-fund veteran Stanley Druckenmiller published a Wall Street Journal op-ed criticizing Treasury Secretary Scott Bessent. Economist Claudia Sahm ran it through an AI detector and got a 100% AI-generated score. Druckenmiller didn't deny using AI to write it. He pushed back on one narrower point: that the finished argument wasn't his. "Of course I used AI," he told a reporter. "I write everything using AI now for the same reason I use a calculator when I do math problems." WSJ opinion editor Paul Gigot defended running the piece, arguing the real test was whether the op-ed "reflects an author's original argument," not who typed it.

We've covered what happened next at other newsrooms that started scanning submissions after that story broke — the operational side of the question. The IdeaLens paper takes a different angle entirely. Its authors use the same incident to make a conceptual point: the detector Sahm used, and the ones newsrooms adopted afterward, can only answer "who wrote the words" (what the paper calls prose provenance). Gigot's defense rested on a separate question the paper calls idea provenance: who came up with the argument in the first place. No detector on the market, including ours, measures that second thing. The paper's authors ran Druckenmiller's actual op-ed through their prototype and got a 98.3% probability that its underlying ideas were AI-generated, against a calibrated threshold of 86.3%. They're careful to note that's a model prediction, not proof of what happened during drafting.

How you'd even try to measure "whose idea was it"

Existing detectors, ours included, read the finished text and score its sentence-level patterns: word choice, sentence rhythm, how predictable one word is given the ones before it. That approach is why detectors are good at spotting unedited AI prose and far less reliable once that prose has been paraphrased or rewritten — a pattern we've written about before.

IdeaLens tries something structurally different. Instead of feeding a model the raw text, the researchers reduce each document to an outline: an ordered list of the discourse moves the writer made, stripped of the actual wording used to express them. A 1,319-word opinion piece might reduce to seventeen outline items, starting with something like "an event account of a Treasury announcement" followed by "an evaluation of the market's reaction." Crucially, during training, they paraphrase even those outlines, specifically to stop the model from learning wording-level shortcuts and force it to learn something about the structure of the argument instead.

They trained IdeaLens on a million outlines pulled from web documents, using an existing commercial detector (Pangram) purely as the source of the training labels — a widely used academic shortcut, since nobody has a large dataset of documents with confirmed "whose idea was it" labels. As a control, they trained a second model, ProseLens, on the same documents and labels but fed the raw text instead of outlines. Same backbone, same training data, same labels. The only difference is what each model is allowed to see.

What happened when provenance split

The real test comes when a document's ideas and its prose come from different sources, something that happens constantly now and that ordinary detectors were never built to separate.

In one experiment, the researchers had AI models write essays from increasingly detailed human-supplied outlines, from a bare topic up to a full human-written plan. As the human contribution to the plan grew, IdeaLens's AI-flag rate dropped from 94.9% down to 6.8%. Conventional detectors in the same test kept flagging the resulting prose as AI at rates above 92%, because the words were still machine-written even as the underlying argument became more and more human. That's the expected, useful behavior for a tool answering "who wrote this," and the wrong answer entirely if the real question is "whose idea was this."

The harder test ran the opposite direction. The researchers built a new dataset of 50 stories where a human writer worked entirely from an outline extracted from an AI-generated story, never seeing the original text itself. The researchers call this dataset TwiceTold. IdeaLens flagged 68% of those fully human-written stories as carrying AI-originated ideas. The commercial prose detector in the comparison flagged just 8% of the same stories (the paper notes those stray flags likely came from AI-typical character names the detector spotted directly in the prose), and a prose-only control model trained the same way as IdeaLens flagged none of them. The paper is upfront about IdeaLens's own miss rate too: of the roughly one-third of stories it didn't flag, most were cases where the human writer added speculative or fantasy details beyond the original outline — a real limitation, not a clean result.

On conventional detection benchmarks, where a document's ideas and prose share one source, IdeaLens held its own against the strongest tools in the comparison: it caught about 91% of fully AI-written documents with a 0.6% false-positive rate, and the paper reports its performance held up across 24 languages despite being trained only in English.

What this means for a tool like ours — and for you

Nothing here changes what a prose detector like TheChecker.AI is built to answer, and nothing here makes that answer less useful. If your question is "does this text carry the statistical signature of AI-generated writing," that's precisely what we measure, and IdeaLens's own prose-only control model performed just as well as the commercial baseline on that exact question. What changes is the honest scope of the claim. A detector score has never told you, and still can't tell you, whose argument you're reading. It tells you about the words.

That distinction matters more now because policy is starting to ask the second question directly. A growing number of AI-use guidelines, cited in the paper itself, frame the real issue as whether the submitted work "reflects the author's own thinking," not whether every sentence was typed by a human hand. If your institution's policy is written that way, a word-pattern score, from us or from anyone else, is answering a related but different question than the one the policy actually asks. Treat a flag as a reason to have the conversation Gigot and Druckenmiller had in public: what work did the person actually do, and can they describe it. We've written about why a single score shouldn't be the only evidence in that conversation — this is one more reason that holds.

IdeaLens itself isn't a production tool you'd deploy today. It's a research prototype with a public demo, built to prove the idea-provenance signal exists and can be measured, not a commercial-grade detector. But the underlying finding is durable: prose and ideas can come from different places, and a tool built to catch one will reliably miss the other.

FAQ

What's the difference between "prose provenance" and "idea provenance"? Prose provenance is the question current AI detectors answer: did a machine generate these specific words. Idea provenance is a separate question the IdeaLens researchers define: did a machine generate the underlying argument, structure, or plan, regardless of who wrote the final sentences.

Does this mean current AI detectors don't work? No. The study found that a prose-only detector trained the same way as IdeaLens performed strongly on standard benchmarks, matching commercial-grade accuracy. The finding isn't that word-pattern detection is broken, it's that it only ever answered one of two related questions.

Can IdeaLens replace a text detector like TheChecker.AI? No, and the researchers don't claim that. IdeaLens is a research prototype answering a different question than prose detectors answer. A document can score as fully human on a prose detector and still flag on IdeaLens if its underlying plan came from AI, and vice versa. They measure different things.

Why does the Stanley Druckenmiller op-ed come up in an academic AI detection paper? The researchers use it as their opening example because it's a real, public case where a prose-provenance detector and the newsroom's own idea-provenance standard gave different answers. The detector scored the finished text as 100% AI. The editor's defense relied entirely on a separate claim about whose argument it was.

See how a detector reads your own writing

A detector like ours tells you what the words in front of you look like, not the full story behind them. Run a passage through our free demo to see what a word-pattern score can, and can't, tell you.