DetectorScope turns a black-box AI text detector into one whose every prediction is a sum of named, validated features. It can tell you which properties of a text made it predict "AI" and where in the text they are located.

Try our demo! 🤗 https://huggingface.co/spaces/liamdugan/detector-scope-demo

pip install detectorscope

Example Usage

from detectorscope import AuditableDetector

Load the AI detector

detector = AuditableDetector.from_pretrained( "liamdugan/detector-scope" )

Run the AI detector on some input

output = detector( "Like I said, this is about accuracy, about speed, and about cost." )

Print the results

print(output)

Detection(label='ai', p_ai=0.88, features=[ Feature(importance=+0.23, description='the word about used as a preposition'), Feature(importance=+0.14, description='repeated phrases or sequences in text'), Feature(importance=-0.05, description='conversational filler phrase like i said'), ...])

[25 Oct 2026] DetectorScope will appear at EMNLP 2026 in Budapest, Hungary!

Overview

The pipeline: SAEs on the detector's residual stream, a description and validation score per feature, a threshold on that score, and a sparse head over the features that pass.

A DetectorScope detector is a concept bottleneck over interpretable sparse-autoencoder features: the language model's residual stream is encoded by SAEs, each feature gets a natural-language description that is validated against its own activations, and a sparse linear head over the validated features makes the decision. The prediction is therefore a sum of per-feature contributions, and every contribution has a description and a location in the text.

F1 (COLING) and TPR at 5% FPR (RAID) against the interpretability threshold, in domain (grey) and out of domain (white). FT-Model+SAE matches the underlying neural detector even at thresholds of 0.7 and above.

Across both benchmarks applying our auditable pipeline has minimal adverse effects on performance. The auditable detectors stay close to the fine-tuned black-box classifier they are built on. The full tables are in the paper.

python -m venv .venv && source .venv/bin/activate pip install detectorscope

From source, with the extras for the pipeline, the experiments and the demo:

git clone https://github.com/google-deepmind/detectorscope && cd detectorscope pip install -e ".[all]"

Usage

Any released detector loads by its Hub id; the other settings of a repo are its branches.

from detectorscope import AuditableDetector

detector = AuditableDetector.from_pretrained("liamdugan/detector-scope-peer-review-gpt4o") detector = AuditableDetector.from_pretrained("liamdugan/detector-scope", revision="base-sae")

detector(text) returns a Detection with label, p_ai, logit, bias and features, every feature active in the text with a non-zero effect, largest first. importance is how much the logit falls if the feature is silenced, and logit == bias + sum(importances). A list of texts returns a list.

Set return_highlights=True to see where each feature is localized within the text:

output = detector(text, return_highlights=True) output.tokens # the text as tokens output.features[0].highlights # [(token index, activation), ...]

You can use the tau parameter to control the tradeoff between interpretability and accuracy. Every feature has an auto-interpretation score, how well its description predicts where it fires, and tau is the score a feature must reach before the head is allowed to use it. A higher tau gives the head fewer, better-described features, so every item of evidence it reports is one you can read and check, at some cost in accuracy. A lower tau admits more features, including ones whose descriptions are loose, and the detector gets more accurate and less legible. The default is 0.85.

detector = AuditableDetector.from_pretrained("liamdugan/detector-scope", tau=0.95)

tau runs from 0.00 (every feature) to 0.95 in steps of 0.05, and each branch carries a head at every step, so changing it downloads nothing. It can also be set per call, which is the easiest way to compare the same text at several settings:

output = detector(text, tau=0.50) # more features, higher accuracy output = detector(text, tau=0.95) # the few best-described features detector.heads # every (head, tau) pair the repo holds

head="xgboost", on from_pretrained or on the call, selects the gradient-boosted head at the same tau instead of the logistic-regression one.

We provide the following model collections for download via Huggingface:

Each repo holds the paper's three settings as branches. ft-model-sae fine-tunes Gemma-2-2b as a classifier and fine-tunes the SAEs on it; ft-sae keeps the unmodified Gemma-2-2b and fine-tunes only the SAEs; base-sae keeps both the unmodified Gemma-2-2b and the unmodified Gemma Scope SAEs. main is the setting that performed best on each domain: ft-sae for the open domain, ft-model-sae for peer reviews. Every branch bundles its language model, so nothing else is downloaded and no licence gate is involved.

The defaults are set by hand: the logistic-regression head over features whose auto-interpretation score is at least 0.85, with the decision threshold fixed on the development split. Each branch also carries the XGBoost head and both heads at every other threshold.

The pipeline that builds a detector is the package's stages, run as python -m detectorscope.<stage>: finetune (the classifier), train_sae (SAEs on its residual stream), data.extract and data.process (per-token activations, per-document caches, log-odds ratios, feature statistics), autointerpret (descriptions and their validation, through Gemini; set GEMINI_API_KEY), model.detectors.heads.training.train_logreg / train_xgboost (the heads) and evaluate. Stages 0 to 2 need a GPU; the rest runs on CPUs.

The experiments are four scripts under scripts/: hparam_sweep.py (the SGD + L1 grid search), counterfactual_ablation.py (zero-ablation vs. targeted paraphrase of single features), paraphrase_ablation.py (the feature-guided paraphrase attack) and regex_recall_ablation.py (regex recall of the descriptions). Every figure and table of the paper comes from scripts/visualization/analysis.ipynb, and --help on any script lists its arguments.

demo/ is the Gradio app, which runs locally with python demo/app.py and as the Hugging Face Space; scripts/huggingface_hub/ exports the detectors to the Hub and publishes the Space.

@inproceedings{dugan-etal-2026-building, title = "Building Auditable {AI} Text Detectors with Sparse Autoencoders", author = "Dugan, Liam and McDougall, Callum Stuart and Denton, Matthew and Pham, Philip and Kaeser-Chen, Christine and Callison-Burch, Chris", booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing", month = oct, year = "2026", address = "Budapest, Hungary", publisher = "Association for Computational Linguistics", url = "https://aclanthology.org/2026.emnlp-main.762/", pages = "17640--17689", ISBN = "979-8-89176-459-0", }

Acknowledgements

This research was developed with funding from the Defense Advanced Research Projects Agency's (DARPA) SciFy program (Agreement No. HR00112520300) and is based upon work supported in part by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via 56000026C0019 (the BENGAL program). The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of DARPA, ODNI, IARPA, the Department of Defense, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright annotation therein.

The detectors are research artifacts released to support the paper. They make mistakes, including on text written by people, and their accuracy outside the domains and generators they were trained on is not established. They must not be used as the basis for decisions about individuals, such as academic integrity, hiring or publication decisions, and the authors and Google accept no liability for any use of them. This is not an official Google product.

Copyright 2026 Google LLC. The software is licensed under the Apache License, Version 2.0; all other materials under the Creative Commons Attribution 4.0 International License (CC-BY). Both are distributed on an "AS IS" basis, without warranties or conditions of any kind; see the licenses for the governing language. Adapted third-party code carries its licence in the header of the module that contains it.