Squint

Checks whether text was likely written by AI, and shows why.

Try it at squint.nickymolina.com. The site runs the classifier only (verdict, % AI, flagged passages; up to 500 words). The full version with the word-level heat map runs on your own Mac: see Run it on your Mac.

Squint uses two methods, both running locally:

• A trained classifier (ModernBERT-base fine-tuned on human, ChatGPT and Claude writing, built in classifier/) gives the verdict.
• Binoculars gives the word-level view: two related language models (Falcon-7B and Falcon-7B-Instruct, 4-bit) read the text, and the app compares how predictable each word was against how predictable the models expected it to be.

What you get:

• Verdict: Likely human / Uncertain / Likely AI, from 50 words; "Too short to tell" under 50; low confidence under 150 words.
• % of text that looks AI-written: from 150 words, by running the classifier on overlapping ~150-word windows.
• Show me why: word-level heat map, hover any word for the model's top 5 guesses and where the actual word ranked, per-sentence and per-paragraph leanings, plain-English reasons.
• Tested on held-out human writing (including 512 non-native English essays) and ChatGPT and Claude output. At 50+ words it wrongly flagged 3 of 1,340 human texts (0 of 512 non-native), and caught 747 of 751 ChatGPT and 699 of 733 Claude samples.

Status: live as a public preview. The accuracy target in docs/roadmap.md (GPTZero-level: at least 97% of AI text caught, at most 1% of any kind of human writing flagged) is not met yet; college essays are the main gap below.

Run it on your Mac

Download Squint for Mac (292 MB). Needs an Apple Silicon Mac with macOS 14 or later. The word heat map needs 16 GB of memory and about 11 GB of disk; with less memory Squint runs the classifier only. Models download on first launch (10.5 GB, or 557 MB under 16 GB).

The app isn't signed by Apple, so macOS blocks it the first time:

1. Unzip it and drag Squint into Applications.
2. Open Squint. When macOS says it can't check it, click Done.
3. Open System Settings → Privacy & Security, scroll down, click Open Anyway and confirm.

To build the app yourself, see mac/README.md. To run the server from source, see server/README.md.

Known gaps

• College essays are sometimes flagged. About 4–6% of a set of 70 real college essays were wrongly flagged as AI, against about 0.4% for other human writing. Don't use Squint alone to judge a personal statement. See R1 in docs/roadmap.md.
• Some Claude writing gets through, mostly under 150 words (about 90% caught there, 95% at 150+). Binoculars alone missed nearly all current AI text, which is why the trained classifier now gives the verdict. See docs/known-gaps.md.
• Edited or paraphrased AI text has not been tested.
• Older or other AI models are mostly missed: 2022-era ChatGPT answers got through almost every time. A fresh re-check on new sources matched the main results for current ChatGPT and Claude, with 0 of 1,000 new human texts flagged.
• Gemini is untested. Its samples have not been generated yet.

Layout

Path	What
docs/api-contract.md	The API between the web UI and the server. Change it first.
server/	Python service: Binoculars scoring, verdicts, queue, rate limit, preview mode for the public site.
web/	The web UI.
mac/	Squint for Mac: the Swift window, build script, and the classifier manifest the app checks for updates.
calibration/	Building the tuning set, fitting thresholds, the accuracy report, and classifier.json (the promoted classifier's cutoffs). Data is not committed.
classifier/	Overnight classifier training, evaluation, promote.py and publish.py (the release the app downloads). Weights go to server/models/ (not committed).
ops/	The public site: edge Worker (ops/edge/), cloud image (ops/cloud/), and the Mac mini service and tunnel.

Nothing users submit is stored. Model files and datasets stay out of Git.

License and credits

Code: MIT (LICENSE). Classifier weights: CC BY-NC-SA 4.0, following the most restrictive training source. Every model and dataset is credited in MODEL_LICENSE.md.

Built by Nicky Molina.