Liam Dugan, Callum McDougall, Matthew Denton, Philip Pham, Christine Kaeser-Chen, Chris Callison-Burch · 2026 · arXiv
DOI: 10.48550/arXiv.2610.30006 · open-access PDF
Detektor teks yang dihasilkan oleh AI semakin diimplementasikan dalam pengaturan konsekuensial -- peer review, pendidikan, penerimaan, jurnalisme, dan moderasi konten -- di mana output mereka secara langsung mempengaruhi individu. Namun detektor terkuat adalah klasifikator saraf yang disesuaikan halus yang keputusan mereka tidak jelas, sementara alternatif transparan (stylometric dan klasifikator fitur linguistik) melebihi mereka dengan margin yang luas. Kami menyajikan pipa yang mengubah detektor saraf yang disesuaikan halus menjadi klasifikator yang beroperasi pada konsep yang dapat dibaca oleh manusia.
Machine translation (Meta M2M100); the English original follows.
AI-generated text detectors are increasingly deployed in consequential settings---peer review, education, admissions, journalism, and content moderation---where their outputs directly affect individuals. Yet the strongest detectors are fine-tuned neural classifiers whose decisions are opaque, while transparent alternatives (stylometric and linguistic-feature classifiers) underperform them by wide margins. We present a pipeline that converts a fine-tuned neural detector into a classifier that operates on human-readable concepts. We accomplish this by training sparse autoencoders (SAEs) directly
Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese
GET https://store.agentexchange.work/gw/papers/verify?title=...&year=... (0.003 USD), fetch this record as JSON with /gw/papers/paper?id=2610.30006, or list what it cites with /gw/papers/references?id=.... Pay per call in USDC (x402) or with a 19 USD card key for 1,000 calls. Catalogue · llms.txtSources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange