⬢github Python · 10 ★ +5 since we first saw it · pushed 1 d ago · Apache-2.0
google-deepmind/detectorscope
Build AI text detectors whose decisions are mechanistically grounded and human-auditable (EMNLP 2026)
DetectorScope is a Python library from Google DeepMind and Penn NLP that converts black-box AI-text detectors into auditable ones. It uses sparse autoencoder features with natural-language descriptions, so each prediction is a sum of named, validated features showing which text properties triggered an 'AI' verdict and where they occur. You can also convert your own detectors and analyze LLM output tendencies.
Why now: It was just released and will appear at EMNLP 2026, with a Hacker News mention highlighting 'AI text detectors that show their work' plus a live Hugging Face demo.
Who it is for: Researchers and practitioners who need to inspect, trust, or analyze AI-text detection decisions, such as NLP researchers, moderators, and educators.
ai-text-detectionexplainable-aiinterpretabilitymechanistic-interpretabilitysparse-autoencoders
Stars over our 138 snapshots: 5 to 10, since 1 d ago.
Where people talked about it
API: https://socialmediatrends-api.osmike.com/v1/repos/google-deepmind/detectorscope