When billionaire investor Stanley Druckenmiller was asked whether he used artificial intelligence to help write his Wall Street Journal op-ed, he did not deny it. “Of course I used AI,” he said, adding that he did not see a problem with it.
The Journal’s editorial page editor, Paul Gigot, defended the piece on the grounds that its opinion was actually Druckenmiller’s. The defense raises a question most AI detectors aren’t capable of answering.
Detectors such as Pangram, which The Washington Sun said was used to flag the piece, judge who wrote the words, not who came up with the ideas.
IdeaLens, a new tool from researchers at the University of Maryland, Google DeepMind and Simon Fraser University, now judges the ideas themselves.
IdeaLens never scores the finished text. First, an AI model reduces the document to an outline, a list of points tagged by the job each one does, such as presenting evidence or making a recommendation.
Existing tools such as Pangram only look at patterns in a document’s words and phrases. The detector was trained using outlines from roughly 1 million web documents that Pangram labeled either human-written or AI-generated. Documents that came back as mixed were discarded. The outlines were paraphrased using Google’s Gemini first to ensure that the model learned from what was said, not how it was phrased.
A twin model named ProseLens, trained using the same documents but fed the full text, served as the control. The two models disagreed sharply whenever a document’s ideas were conceived by a human but the text was AI-generated, or vice versa. Training on outlines pushed the model to track where the idea originated, rather than who wrote the final words.
When AI wrote documents from increasingly detailed human outlines, IdeaLens’ AI flag rate decreased from 94.9% to 6.8%, while Pangram still flagged 92.4%. When 17 computer science graduate students or faculty hand-wrote 50 stories from AI-generated outlines, IdeaLens flagged 68% of them, compared with 8% for Pangram.
The results also give insight into what makes an idea read as machine-made. Ideas that used vague details that could describe almost anything, or claimed more than the evidence supported, scored as more AI-like. Ideas that cited verifiable sources or discussed downsides and open problems scored as more human.
The researchers spell out the limits. Detection dropped to between 51% and 79% on stories from four newer AI models. When documents near the tool’s cutoff were run five times, only 52.5% got the same verdict every time.
This matters because an AI flag can become an accusation. The authors write that the tool’s predictions “must not be used by themselves as conclusive evidence” of AI use.
Schools face a difficult question with new technology like this. A student who brainstorms with an AI chatbot and then writes every sentence alone could pass today’s detectors and still be caught by this one.
Baruch College’s AI guidance allows generative AI only with an instructor’s permission and requires that its use be disclosed. Beyond that, the guidance does not say whether brainstorming with AI counts the same as using it to write, leaving that line to instructors, if they draw it at all.
As detectors start to examine ideas instead of words, schools and other institutions may need to reevaluate guidance on the use of AI.
