TextGrain vs SynthID

TextGrain is OpenAI\'s statistical text-watermarking method. SynthID is Google DeepMind\'s watermarking technology spanning text, images, audio and video. SynthID for text exists, so this is not simply a comparison of text versus images. The applicable implementation and verification access matter more than the brand name alone.

Official announcement, access guide and technical report reviewed October 6, 2026. Availability below distinguishes announced rollout from verified access. This site is independent and does not offer a TextGrain detector.

Compare the content and signal first

Question	TextGrain	SynthID
Origin	OpenAI\'s text provenance work.	Google DeepMind\'s watermarking technology.
Content	Language-model text.	Text, images, audio and video, with media-specific implementations.
Text mechanism	Keyed block optimal transport calibrated to an entropy-loss budget.	A text-watermarking implementation; not an interchangeable TextGrain test.
Our site	No TextGrain detection or verified removal.	No SynthID detection or verified removal.

Which one does OpenAI use?

OpenAI\'s October 5 announcement describes TextGrain for eligible text, while its provenance documentation lists SynthID for supported images and audio. Supported images can additionally have C2PA Content Credentials. A Google-origin technology can be used in another provider\'s product; a detected watermark should be interpreted in the context of the actual verifier. Read the OpenAI watermark guide.

Detector access is specific to the service

OpenAI text verification requires approved organizational access. Google\'s SynthID overview describes media checks in Gemini and a detector portal for images, video and audio. Those media interfaces do not establish public verification of every text watermark. Neither service\'s access rules grant this independent site detector access.

Does the comparison establish a winner?

OpenAI reports favorable TextGrain results against approaches it evaluated, including SynthID for text. That is a provider\'s evaluation, not an independent test performed here. We do not turn it into a universal ranking: passage length, domain, transformations and detector thresholds affect interpretation. Consult the OpenAI technical report and the originating evaluations before comparing performance claims.

Character cleanup is a different task

Removing zero-width characters or normalizing spaces is not a matched statistical test of either watermark. Our browser-local cleaner reports specific character changes, not provenance. A negative watermark result also does not prove human authorship, factual accuracy or successful removal. For the TextGrain construction, continue to the technical report guide.

Google sources for this comparison

Google DeepMind\'s official SynthID overview was reviewed October 6, 2026. It explains supported media and current media-verification interfaces. The OpenAI sources below support the TextGrain and OpenAI-specific access statements.