A Detector That Separates AI Ideas From AI Phrasing, and Says Where It Fails

A student writes her own argument and asks a chatbot to tidy the sentences. Another asks the chatbot for the argument, then rewrites every line in his own voice. Hand both essays to an AI-text detector and it returns much the same thing: one score for how machine-like the writing looks. The two documents differ in exactly the way a teacher cares about, and a single number cannot say which is which.

That gap is where a preprint posted to arXiv on Aug. 27, 2026, begins. Xin Chen, Fuwei Zhang and colleagues at the Institute of Artificial Intelligence at Beihang University in Beijing propose dropping the human-or-machine question and asking two narrower ones: where did the ideas come from, and where did the wording come from. Each answer is either human or AI, so the pair sorts a document into one of four kinds: human ideas in human words, human ideas in AI words, AI ideas in human words, or AI throughout.

That four-way scheme is not theirs. It comes from HART, a benchmark published last year, and the study notes outright that it does not claim to introduce the taxonomy. What the Beihang group adds is the machinery between the text and the label. Their system, D2C-Routing, sends two families of evidence down two separate paths. The content path is built from how information is organized: which entities recur, how the argument is structured. The expression path is built from word choice, rhythm and the regularity of the surface. Each path trains its own yes-or-no head, and a learned gate combines the two into the final four-way call.

On their own reworked split of the HART files, which they call MixD2C, the best result is 0.8603. That figure belongs to the full detector system, a probability fusion of three models; the single D2C-Routing model scores 0.8440, and the researchers keep the two apart carefully. Neither is an accuracy. The measure is the share of texts in each of the four kinds that the system flags correctly, once it has been tuned to wrongly flag only one in a hundred of the texts it should leave alone, then averaged. It is a deliberately harsh setting, and the reason for it is that a detector producing false accusations at any real rate is worthless.

The 6.5-point gain in the abstract is measured against one specific thing: the authors' own rerun of RACE, the closest existing method, on the identical files. They ran it themselves because the released version does not ship the sample identifiers or checkpoints needed to reproduce the published result, and they note twice that their rerun is not that reproduction. Against RACE's published number, the gap would be nearer three points, and the study marks even that comparison as context rather than a result.

Then comes the part a summary would drop. That result holds inside the benchmark and does not travel. The authors retrain on three of HART's four writing domains and test on the fourth, and the average falls to somewhere between 0.31 and 0.68 depending on which domain is held out. On three of them, a plain text-only model (no routing, no two paths) scores better than theirs. Transferred cold to a different mixed-origin benchmark, the system lands below chance. Their limitations section opens with the sentence: "The main positive result is in-domain on MixD2C."

The hardest single case is the one with the sharpest consequences. AI-written content that a person has rewritten in their own words remains, on this evidence, very hard to separate from writing a model produced end to end. On those texts the single model gets the content question right about 95% of the time and the expression question about 64% percent, and on that class it sits below the RACE rerun rather than above it. The breakdown is on the expression side, which is the half of the split meant to notice a human rewrite.

The routing idea itself is supported less firmly than its name suggests. Removing the two supervised heads, or removing the learned gate, does measurably hurt performance. But when the authors deliberately swap the evidence between the paths, or assign it at random and freeze it there, the headline number comes out statistically similar. What the experiments support is supervising the two dimensions separately and learning how to combine them, not the particular decision about which linguistic cue belongs on which side.

The work is an arXiv preprint scheduled to appear at the EMNLP 2026 conference. The authors have released their code.

They also wrote an ethics statement that anyone tempted by the headline number should read first. Detectors like this "should not be used as sole evidence for punitive decisions about authorship, academic integrity, employment, or access to services." The labels in their data "describe controlled construction protocols rather than a person's intent or honesty," and evaluating at a strict false-positive rate "does not remove the risk of false accusations." The use they propose is decision support, auditing and further research, with a person reading the output. Nothing in their own tables argues for more.