Detection
What each check looks for, how the score is built, and where it can be wrong.
16 articles
Can it detect AI-written text reliably?
No tool can, and Slop Squash does not claim to. Text is scored on writing patterns, every triggered pattern is listed, and a text flag is a best guess.
Updated 2 September 2026Community reports and the verdict cache
A shared, anonymous database of verdicts keyed by a hash of the content. It holds earlier rechecks, classifier results and report tallies, never who produced them.
Updated 2 September 2026Content Credentials (C2PA)
Full signature validation against the C2PA trust list, not just a presence check. A trusted signature that declares AI generation is the strongest signal we have.
Updated 2 September 2026Can it detect images made by Gemini, Nano Banana or ChatGPT?
Often, yes. OpenAI images are verified through OpenAI's endpoint; Google images through their Content Credentials signature when it survives, otherwise through the classifier.
Updated 2 September 2026EXIF and IPTC metadata
Generator names in the EXIF Software and XMP CreatorTool fields, and the IPTC digital source type, are read on every image our API fetches.
Updated 2 September 2026The generator classifier
A pixel-level classifier that works when metadata was stripped and names the likely generator. Pro triggers it; the cached result is shown to everyone.
Updated 2 September 2026How accurate is it?
Accurate enough to be useful, not enough to be a verdict. Signatures and watermarks are near-certain; classifiers and writing patterns are probabilities the card always attributes.
Updated 2 September 2026How verdicts are scored
One combined score from 0 to 1, with provenance signals on top, then classifiers, then the recheck, then writing patterns blended with community reports.
Updated 2 September 2026The model recheck, explained
A vision-language model on Cloudflare Workers AI answers a fixed rubric about a post and returns a probability with notes. Pro triggers it on request; the answer is cached for everyone.
Updated 2 September 2026The OpenAI provenance check
Our API sends the image file alone to OpenAI's content provenance endpoint, which reports SynthID and Content Credentials signals for OpenAI-made images.
Updated 2 September 2026Recheck versus classifier: which Pro check to use
The classifier is fast, numeric and names a generator; the recheck reasons in words. Use the classifier first and the recheck for borderline cases.
Updated 2 September 2026SynthID: what can and cannot be checked
OpenAI's images are verified through OpenAI's public endpoint. Google's SynthID can only be read by Google, and text watermarks by nobody else, so the card says so instead of guessing.
Updated 2 September 2026Videos, GIFs and thumbnails: what is actually checked
Video frames are never analysed. Containers are scanned for Content Credentials, posters and YouTube thumbnails are checked as images, and animated images get provenance only.
Updated 2 September 2026What "uncertain" means
A combined score between 40% and 70%. The evidence leans one way without settling it, and the card says so.
Updated 2 September 2026Why false positives happen
Classifiers and writing patterns are probabilities, not proof. Here is what trips them, what a flag does and does not mean, and what to do about a wrong one.
Updated 2 September 2026Writing-pattern analysis (stylometry)
The on-device text check scores sentence rhythm, overused vocabulary, lists of three, em dashes, markdown in comments and more. Every triggered pattern is listed in the card.
Updated 2 September 2026