AI text detectors are already deciding who gets into school, whose papers pass peer review, and which posts stay online. But the vast majority of detectors cannot even explain a single judgment they make.
A research team from Google DeepMind and the University of Pennsylvania wants to change that. Their new method turns an opaque detector into one that reasons with human-readable concepts, and the cost in accuracy is surprisingly small.
The paper is titled "Building Auditable AI Text Detectors with Sparse Autoencoders," submitted on October 21, 2026, and accepted to EMNLP 2026—one of the major academic conferences in natural language processing.
The problem the researchers describe is a familiar trade-off. The best-performing detectors are fine-tuned neural classifiers whose internal logic is essentially closed. The alternative is transparent tools built on stylometric and linguistic features (such as word choice and sentence structure patterns), which you can open up and inspect, but which, according to the paper, lag far behind neural models in accuracy.
The team's answer: keep the neural model's knowledge but "translate" it out. The pipeline starts with a detector fine-tuned on Gemma-2-2B, after which the researchers apply sparse autoencoders (SAEs) to the model's residual stream activations.
As the model processes text, it passes information through an internal channel called the residual stream. These signals are dense and entangled, with many concepts mixed into every single value. Sparse autoencoders disentangle them, learning to represent the same information with a large set of independent features, of which only a few are activated for any given input.
Next, the pipeline automatically labels these features, assigning each one a human-readable description. Only features that pass a unified fidelity threshold are kept—that is, the label must genuinely match the feature's behavior. The features that pass the filter are fed into a lightweight logistic regression classifier, where each input has a weight, so you can clearly see how much each item influences the final judgment.
On in-distribution data (i.e., text similar to the model's training data), this interpretable classifier reaches an F1 of 96.8, compared with 98.0 for the original black-box detector.
Even more striking are the results on out-of-distribution data, i.e., text that differs in form from the training set. Here, the new model achieves an F1 of 92.2, while the black-box baseline plunges to 75.2.
The code and trained models have been made public in the google-deepmind/detectorscope repository on GitHub.
Auditability is the real highlight
The paper grounds its motivation in settings where detectors are already in practical use: peer review, education, admissions, journalism, and content moderation. In each of these settings, a detector's output can directly affect a person's fate.
An auditable detector changes the nature of the conversation. A reviewer can no longer only say "the system judged this as AI-written"; they can point to the specific concepts driving that judgment and assess whether they hold up. This also gives the person being judged a foothold to argue their case.
The method relies on automatic feature labeling, so the quality of the labels sets an upper bound on how meaningful these explanations can ultimately be. The fidelity threshold is designed to filter out weak labels, but only independent testing on the public code can show how robust it is across domains. In addition, whether this approach can scale beyond the Gemma-2-2B detector used in this paper remains to be seen.
