SynthID: what can and cannot be checked

SynthID is Google's family of invisible watermarks for images and text. Slop Squash is often described as a SynthID checker; the honest answer is "partly".

What is checked

Every ChatGPT, gpt-image, DALL-E and Codex image since May 2026 carries a SynthID watermark and a Content Credentials manifest, and OpenAI runs a public verification endpoint. Our API sends the image file, and nothing else, to that endpoint and reports its SynthID and Content Credentials findings in the card as Image watermark. See The OpenAI provenance check.

What cannot be checked

• Google's SynthID for images is an undisclosed learned watermark with no verifier outside Google. Google's Content Detection API stopped taking new applicants in September 2026, and the consumer surfaces (the Gemini app, Search, the Chrome right-click) have no API. Our adapter for Google's API exists but stays dormant until Google grants access, and the card says the watermark could not be checked. Community attempts to template-match SynthID only recognise the exact images used to build the template and fail on anything else, so we do not ship one.
• SynthID for text is keyed to the generator. Detecting it requires the generator's secret key, which Google does not release. No extension and no third-party service can read it; anyone claiming otherwise is running a classifier. For text the SynthID row is always listed as not available.

What replaces it for Google images

Google signs Gemini, Imagen and Nano Banana output with Content Credentials, which we validate cryptographically when the platform preserved the metadata. When it did not, a Pro-triggered generator classifier gives a probability and a best guess at the generator, and that result is cached for everyone. Both are strictly more informative than a yes/no watermark check when they fire.