Who Can Read the AI Mark? SynthID Detector, OpenAI's Text Watermark and C2PA
AI watermark detectors mostly read only the marks their own company, or its partners, put in. On October 7, 2026, Google opened SynthID Detector to everyone for images, video and audio. Two days earlier, OpenAI said its new text watermark detector would go first only to approved researchers and expert organizations.
What is an AI watermark?
An AI watermark is an imperceptible signal embedded in AI-generated output. In images, video and audio it sits in the pixels or the sound; in text it sits in the model's word choices. Finding it usually requires a key held by the company that embedded it, which is why each company tends to ship its own detector.
What does Google's SynthID Detector check?
SynthID Detector checks whether an image, video or audio file was made with AI from Google or its partners. Google's announcement names OpenAI, NVIDIA and Kakao as partners, with Apple coming soon. The tool is available globally in English.
According to Ars Technica, users must sign in, and each user gets a daily quota of approximately 10 checks. A file without a SynthID watermark, such as one made with another company's tool, gives the detector little to work with.
Why did OpenAI limit access to its text watermark detector?
OpenAI said its text watermark, called textGrain, has significant limitations, and that these contributed to giving detector access first to approved researchers and expert organizations. In OpenAI's own evaluations:
In 400-token passages, replacing 10% of words with synonyms reduced detection from about 92% to 66%.
Replacing 25% of words reduced it to 17%.
Shorter text, and text with less flexibility in word choice such as mathematics, was harder to detect.
These are OpenAI's own figures, not independently verified. OpenAI also states that the absence of a detected watermark does not prove human authorship.
How is a watermark different from C2PA Content Credentials?
A watermark asks whether a specific company's AI was involved. C2PA Content Credentials record who signed a file and what happened to it afterwards. C2PA is an open standard from the Coalition for Content Provenance and Authenticity.
AI watermark (e.g. SynthID) | C2PA Content Credentials | |
|---|---|---|
What it carries | An invisible signal | A signed record of origin and edits |
Who reads it | Usually the embedding company's detector | Tools that follow the open standard |
Weak point | Can weaken with editing | Can be lost through screenshots or re-saving |
When it is missing | Not proof a human made it | Not proof the content is fake |
The two are layered rather than interchangeable. OpenAI's help documentation describes both C2PA metadata and invisible watermarks as provenance signals.
What does the EU AI Act require?
Article 50 of the EU AI Act requires providers of generative AI to mark outputs in a machine-readable way. OpenAI said its EU text watermarking for ChatGPT and Codex is a response to that requirement. The duty to mark exists; the tools to read those marks are still separate for each watermarking technology.
Who certifies C2PA implementations?
C2PA publishes a Conforming Products List of implementations that meet its conformance requirements. MuseBlossom's Contents Defence is listed for image (June 11, 2026), audio (August 11, 2026) and video (September 21, 2026). A listing confirms conformance to the standard; it is not a guarantee of individual product features.
FAQ
If a detector finds no watermark, did a human make it?
No. OpenAI says text may go undetected if it is short, edited, translated, or made with another company's tools.
Can anyone use SynthID Detector?
Yes, with sign-in. Ars Technica reports a daily quota of approximately 10 checks per user.
Do Content Credentials prove content is true?
No. They show who signed a file and what was recorded about its history, not whether what it shows is accurate.
Learn how Contents Defence works with C2PA Content Credentials: MuseBlossom blog