Local AI Detector 作者： Fred

Local AI Detector analyzes the text and images on the page you're reading
for signs of AI generation — entirely on your device, with no servers, no
API keys, no accounts, and no data collection.

FEATURES

• Fusion mode (default): mix up to six on-device detectors (Mozilla's Fakespot
RoBERTa, TMR, ModernBERT, a lite model, perplexity, experimental Binoculars),
pick how they're combined, and see how many agree. No cloud API.
• Shows a calibrated probability ("AI 91%") rather than a made-up confidence, and
"—" when a snippet is too short to judge.
• Three highlight styles for flagged sentences: heatmap, flagged-only, or
underline.
• A hidden/invisible-Unicode-character report, always shown separately from
the AI score.
• Image provenance and watermark checks: C2PA Content Credentials (verified
against a bundled, attributed trust list), unsigned generator metadata,
and open-source Stable Diffusion/SDXL/FLUX invisible watermarks — with an
honest list of schemes that CAN'T be checked locally (Google SynthID,
Anthropic's and Gemini's text watermarks, Meta Content Seal, and others)
rather than pretending they don't exist.
• Page-aware scoring: one score for an article, a score per comment or reply
on Reddit/Hacker News/forums/reviews/chat sites, and small markers on
flagged search-result snippets. An optional slop filter dims flagged
items; an optional local-only site memory tracks a per-domain tally.
• On YouTube, reads the transcript and, experimentally, samples the audio
itself to flag likely AI narration — nothing about the video leaves the
device.
• A fast Quick check runs automatically (confirmed against the full
detector set before showing a high score); a one-click Deep check runs
everything.
• Model updates you control: check for newer model revisions, see their
licence before updating, and roll back if needed. Bring your own custom
Hugging Face model per detector slot.

PRIVACY

Every analysis runs locally, in your browser. The only network requests are:
downloading/updating AI models directly from huggingface.co (once you
consent, or when you ask to check for updates), and — only for a website
you've explicitly granted permission to, one site at a time — fetching a
single image's bytes to check its provenance. See our privacy policy (linked
from the extension's Options page and its source repository) for the
complete, exact accounting. This add-on declares no data collection to
Mozilla, because it collects none.

ACCURACY, HONESTLY

This is a small, transparently-documented project, not a commercial
forensic tool. On about 1,900 held-out web texts (Reddit posts, reviews, news,
how-tos, stories; AI side from 2024–26 models), the default Fusion tells AI
filler from human writing with an AUROC of 0.91. Its slop filter hides almost only AI
text (99% precision), catching about half of it. Unedited assistant-voice filler is
caught well. Paraphrased or edited text mostly isn't. Full methodology, numbers, and caveats are published in
the source repository rather than a vague marketing accuracy claim.

OPEN SOURCE

MIT-licensed. Every bundled library and model is under an open licence
(MIT/Apache-2.0/MPL-2.0) — nothing gated or non-commercial. Source code,
architecture notes, and the full calibration methodology are linked from the
extension's Options page.