SV-Detect: AI-generated Text Detection with Steering Vectors

A research team has proposed a new method for detecting AI-generated text by extracting steering vectors from the hidden layers of a frozen language model, an approach designed to remain robust even when the writing is polished or rewritten. The technique, detailed in a paper submitted to the arXiv preprint server on 5 June 2026, constructs a direction at each layer of a language model that separates human-written from machine-generated text [1][2]. Each input is then represented by its layer-wise alignment with these directions, and a lightweight classifier trained on these projection features produces the final detection score [1][2]. The authors report that the method achieves strong performance both in-distribution and under distribution shift, including across domains, source models, and machine-editing transformations such as polishing and rewriting [1][2]. Interpretation analyses indicate that the learned directions align with recognizable stylistic cues while capturing substantial additional signal beyond surface features [1][2]. The work positions fake-text detection as a representation-space probing problem and argues that steering vectors provide a simple and effective solution [1][2]. The paper was authored by Tatiana Gaintseva and submitted as a 3,181 KB PDF at 14:34:37 UTC [1]. Large language models, which are typically based on transformer architectures and pre-trained to predict the next word, have made machine-generated text increasingly difficult to distinguish from human writing [6]. The arXiv repository, where the paper appears, hosts non-peer-reviewed preprints and has grown to receive about 24,000 submissions per month as of late 2024 [4]. The detection of AI-generated content has become a parallel challenge to the models' own development, especially as editing tools can obscure the original machine fingerprints [2][6].