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

Chinese

人工智能生成的文本探测器越来越多地部署在相应的环境中──对等审查、教育、录取、新闻和内容调控──他们的输出直接影响个人。然而,最强大的探测器是细节化的神经分类器,其决定是透明的,而透明的替代品(风格学和语言学特征分类器)在广泛的范围内表现不佳。我们提出了一个管道,将精确调节的神经探测器转换成一个基于人类可读的概念运作的分类器。我们通过直接在探测器的残余流中训练稀缺的自动编码器(SAEs)来实现这一点,将自动解释标签分配给每个SAE特征,并只保留其生成的描述超过全球忠诚度门槛的特征。

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