Newest papers

Building Auditable AI Text Detectors with Sparse Autoencoders

Liam D. Dugan, Callum Stuart McDougall, Matthew Denton, Philip Pham, Christine Kaeser‐Chen, Chris Callison-Burch · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.30006 · cited by 0 works (OpenAlex) · open-access PDF

Japanese

AIによって生成されたテキスト検出器は、ますます結果的な設定で展開されています── peer review、教育、入学、ジャーナリズム、コンテンツのモダニング──彼らの出力が直接個人に影響を与えます。しかし、最も強力な検出器は、決断が透明でない細かいニューラル分類器ですが、透明な代替品(スタイロメトリックおよび言語学的機能分類器)は、幅広い範囲でそれらを下回ります。 我々は、細かい調節されたニューラル検出器を人間の読み取り可能な概念に基づいて動作する分類器に変換するパイプラインを提示します。 それから我々は、解釈可能な機能に基づいて軽量なテキスト分類

Machine translation (Meta M2M100); the English original follows.

Abstract

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 on a detector's residual stream, assigning auto-interpretation labels to each SAE feature, and retaining only features whose generated descriptions pass a global fidelity threshold. We then train a lightweight classifier on just the interpretable features. We find that our SAEs recover latent concepts that correspond to highly specific and predictive linguistic "tells" of AI-generated text. Classifiers built on these extracted features are not only highly interpretable but exhibit stronger performance on out-of-distribution texts than either neural or stylometric alternatives.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Artificial intelligence, Classifier (UML), Computer science, Detector, Fidelity, Artificial neural network

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange