Google DeepMind and UPenn build AI text detectors that show their work
A new method makes AI text detectors easier for people to understand while keeping most of their accuracy.
Photo: Tima Miroshnichenko / Pexels
AI text detectors already help decide who gets into a school, whose research paper passes review, and which posts are allowed to stay online. But most of these tools cannot explain why they made a particular decision.
A research team from Google DeepMind and the University of Pennsylvania wants to solve this problem. They have developed a method that makes an AI detector explain its decisions using ideas that people can understand. It loses surprisingly little accuracy in the process.
The research paper, called “Building Auditable AI Text Detectors with Sparse Autoencoders,” was submitted on October 21, 2026. It has been accepted at EMNLP 2026, a major conference for research on how computers understand and work with human language.
Opening the black box
The researchers describe a common problem: the most accurate detectors use neural networks that are specially trained to classify text. But it is very hard to see how these models reach their decisions.
Other tools are easier to understand because they look at features such as word choice and sentence structure. People can inspect these features to see how the tools work. However, according to the paper, these simpler tools are much less accurate than neural network models.
The team developed a series of steps that keeps what the neural network has learned but makes its decisions easier to understand. First, they use a detector that has been specially trained for this task and is based on Gemma-2-2B. Then they apply sparse autoencoders, known as SAEs, to the signals inside the model's residual stream.
As a model reads text, it passes information through an internal pathway called the residual stream. The information in this pathway is packed together, so many different ideas can be mixed into the same numbers. A sparse autoencoder separates this information into many individual features. For any given piece of text, only a small number of these features become active.
The next step is to give each feature an automatic label written in words that people can understand. The researchers keep only features that pass a single, fixed test of how well their labels match what the features actually do. These selected features are then used in a simple classifier called logistic regression. It gives each feature a weight, making it possible to see exactly how much each feature contributed to the final decision.
What the numbers show
When tested on data similar to the text used to train the models, the easier-to-understand classifier achieved an F1 score of 96.8. The original detector, whose decisions are harder to explain, scored 98.0.
The results were even more interesting on data that differed from the training examples. On this unfamiliar type of text, the new model achieved an F1 score of 92.2, while the original detector's score fell to 75.2.
The code and trained models are available to the public on GitHub in the google-deepmind/detectorscope repository.
Why being able to check the results matters most
The paper explains that these tools are already used in areas such as research paper reviews, education, school admissions, journalism, and content moderation. In all these settings, a detector's decision can directly affect someone's life or work.
A detector that people can check makes a big difference. Instead of simply saying, “the system says this was written by AI,” a reviewer could point to the specific ideas that led to the decision and check whether they make sense. The person affected by the decision would also have something specific they could challenge.
The method relies on automatically generated labels for the features, so the quality of those labels limits how useful the explanations can be. The researchers use a minimum quality threshold to remove weak labels. However, independent tests using the publicly available code will help show how well this approach works on different kinds of text. It is also not yet clear whether the method can work at a larger scale or with models other than the Gemma-2-2B detector used in this study.