aiqbench analysis
This arXiv preprint proposes a pipeline that trains sparse autoencoders on a fine-tuned AI-text detector, keeps only features with reliable auto-generated explanations, and builds a simple classifier on them. The authors claim it matches or beats opaque neural detectors on out-of-distribution text, though the evidence is a preprint and has not been independently reproduced. For readers, it is a concrete example of applying interpretability tools to a deployed classifier in a high-stakes setting.
Building Auditable AI Text Detectors with Sparse Autoencoders
arXiv:2610.30006v1 Announce Type: cross 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.
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.
Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
arXiv:2610.30006 [cs.CL]
(or arXiv:2610.30006v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2610.30006 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 249 words
Related coverage
Making Clinical Language Models Auditable: Concept-Guided Fine-Tuning for Robust Prediction
arXiv · Aug 28, 2026
Decision Sufficiency of Learned Representation Interfaces for Safe Generative Control
arXiv · Oct 21, 2026
An overview of frontier AI company monitoring practices in mid 2026
LessWrong · Oct 7, 2026
Alignment fine-tuning induces conditional misalignment in Qwen2.5-7B-Instruct
LessWrong · Aug 21, 2026