Answers are grounded in this paper and cite their sources.
AI-text detectors are marketed as judges of "AI-ness," yet what they learn is rarely audited. We present a Chinese detector built entirely from structural, CPU-only features (160 dimensions) with a 14-class (7 genres × {human, AI}) head, and we evaluate it under an FP-controlled protocol with a GAN-like adversarial loop in which a language-model attacker plays the generator and a human selects the direction. The 160-dimensional detector reaches 17/18 on a held-out exam at 5% FP calibrated on ≥15-line humans (the 94-dimensional specialist head reaches 15/18 at a single global threshold), runs at 173 documents/s without a GPU, and holds false positives at 0.3% on the human calibration set, 0–1.7% on independent non-author corpora, and 9.4% on the author's unpublished poems under the no-private model. Its central result is negative but informative: a leave-one-paradigm-out matrix shows that the detector is a fingerprint of the generation paradigm, not of "AI-ness" - 43–86% detection when only the model or prompt changes inside direct generation, but 2–13% when the paradigm changes. We map the failure modes that any claim of "AI detection" must confront: line-level paraphrase (4%), short texts (30.8% at 0–50 characters), and targeted pastiche (46.2% on 7-line imitations at its default operating point, where human FP is 7.5%; at a fixed 5% FP: 35% for the baseline, 48–57% for the repaired cross-poet models, and 84% on decontaminated 6–14-line pastiches with a dedicated short head). The repair is instructive: one hole closes with 135 poems from four poets, while paraphrase and cross-paradigm transfer do not close at all.
No takes yet. Share an insight, caveat, or question.
Haowei Luo (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: