Google DeepMind released an AI text detector that explains each decision

Google DeepMind and the NLP group at the University of Pennsylvania released DetectorScope on 24 October. The repository is open source under Apache 2.0 and has 260 stars in six days. It is a different kind of AI text detector. Most say yes or no. This one shows the reader why.

The tool takes a passage of text. It returns a label of "ai" or "human", a probability, and a list of named features that drove the decision. Each feature carries a plain-English description and a weight. Each has a location in the text so a user can see which words it picked up.

What the demo shows

The repository README includes a worked example. The authors fed the detector a one-sentence passage from a chat. It returned a label of "ai" with a probability of 0.88. It also returned the three features that pushed the score up or down.

The strongest feature was "the word about used as a preposition" with a weight of +0.23. The second was "repeated phrases or sequences in text" with +0.14. The third was "conversational filler phrase like i said" with -0.05.

A live demo runs on Hugging Face Spaces at huggingface.co/spaces/liamdugan/detector-scope-demo. The repository links it from the top of the README. The project also publishes to PyPI as detectorscope.

How it works

The paper describes a layered pipeline. The authors train sparse autoencoders on the hidden layers of a classifier. Each autoencoder feature gets a plain-English description, generated from where in the text that feature fires. A separate check scores how well each description predicts the feature's own activations.

A sparse head over the scored features makes the final decision. The output is a sum of per-feature contributions. Every contribution has a description, a weight and a location.

A user controls a trade-off through a parameter called tau. A higher tau keeps fewer features, each with a well-matched description. A lower tau keeps more features and gets higher accuracy at some cost to readability. The default is 0.85.

What the paper reports

The authors tested the detector on two public sets. The first is the COLING 2025 GenAI Detection shared task, which covers news, essays and reviews with 40 generators. The second is the AI Peer Review Detection Benchmark. It covers machine-learning peer reviews written by humans and by GPT-4o or Llama.

The authors report that the auditable detector matches a fine-tuned black-box classifier on both sets. Full tables are in the paper at arxiv.org/abs/2610.30006. The paper will appear at EMNLP 2026 in Budapest.

Who it is for

An editor or teacher who flags writing as AI can now show which words made the detector decide. A buyer of AI-detection services can audit a vendor by inspecting its features. A researcher can build on the same pipeline to study which signals the detectors use.

The tool is for text in domains close to its training sets: news, essays, reviews and peer review. Writing far from these is not covered. The repository holds the three model collections and the pipeline scripts.

Source

• google-deepmind/detectorscope, GitHub
• Paper: DetectorScope, arXiv 2610.30006