AI watermark explained simply: SynthID, Content Credentials and how to check an image

What an AI watermark is

You know classic watermarks from banknotes. An AI watermark works in a similar way, only invisibly: when generating content, the AI service embeds a signal in the image, video or audio clip that you do not notice but that verification software can find. Google DeepMind developed SynthID for this purpose, which can mark images, video, audio and text. According to Google, OpenAI, Nvidia and Kakao now use SynthID too, and Apple is set to follow.

A related but different technique is Content Credentials. They are not a watermark but a record of origin and editing. Google compares them to a digital passport: stored as additional information in the file, they record which tool created it and how it was edited. They are based on the open standard of the C2PA, an alliance of more than 500 companies including Adobe, Google, Microsoft, Meta, OpenAI and the BBC. Content Credentials are not only for AI: camera makers and news organisations also use them to document the origin and editing of their photos. According to Google, they therefore do not by themselves tell you for certain whether content is AI-generated.

OpenAI combines both: according to OpenAI, images from ChatGPT carry Content Credentials and a SynthID watermark. In the EU, OpenAI also marks ChatGPT texts with its own invisible text watermark called textGrain, which nudges word choices into a subtle pattern.

What the EU has required since August 2026

Article 50 of the EU AI Act has applied since 2 August 2026. Providers of AI that generates images, video, audio or text must mark these outputs in a machine-readable way that is effective, reliable, robust and readable by different systems. For AI systems already on the market before 2 August 2026, this obligation only applies from 2 December 2026.

In addition, deployers who publish AI content must visibly disclose deepfakes as AI-generated, meaning deceptively realistic images, videos or audio of people, places or events that appear real. This also applies to AI texts intended to inform the public on matters of public interest, unless a human has editorially reviewed them and takes responsibility for them. For evidently artistic or satirical works, a notice that does not spoil the work is sufficient.

Two important exceptions: according to the European Commission, anyone using AI purely privately and not professionally is not covered by these deployer obligations. And the providers' machine-readable marking is not required when the AI only performs an assistive function for standard editing.

The European Commission has published a voluntary code of practice and its own EU icons for this, with variants for fully AI-generated and partly AI-modified content. By the end of July 2026, about 190 organisations had signed the code, including Google, Meta, Microsoft, OpenAI, Anthropic, Mistral, Aleph Alpha and Black Forest Labs. What else the AI Act means for you is explained in our separate article on the law.

How to check an image

In the Gemini app: Google offers signed-in users a check directly in the Gemini app. You upload an image, a video of up to 90 seconds or an audio file of up to one hour, each up to 100 MB, and ask something like: is this AI-generated? Gemini then looks for SynthID and reads any Content Credentials. According to Google, about ten image checks are possible within 24 hours, plus about ten audio checks totalling no more than three hours.

At OpenAI: on its verification page (Verify), you can upload an image or audio file. The tool looks for signals associated with OpenAI, that is SynthID or a trusted C2PA manifest from OpenAI. According to OpenAI, uploaded content is only processed for the check and not used for training.

For Content Credentials in general: the Content Authenticity Initiative's Inspect tool lets you examine files in many formats, such as JPEG, PNG, HEIC, MP4 or MP3, for provenance information.

OpenAI's tip: upload the original file if possible and do not crop or convert it beforehand. For screenshots, Google recommends cropping tightly to the image and not uploading collages.

Where the limits are

Each checking service only finds certain signals. According to Google, Gemini can currently only detect content created with Google's own AI tools, even though other companies also use SynthID. OpenAI's tool is explicitly not designed to detect content from other AI providers.

Content Credentials are easily lost, OpenAI writes, for example through a screenshot or a file conversion. Watermarks such as SynthID are more robust and can survive light editing, cropping or screenshots. Heavy cropping, compression or, for texts, rewording and translation can make them undetectable, though.

So caution is needed in both directions. According to OpenAI, a detected signal shows that AI was involved but says nothing about whether the content is accurate, who made it or how much of it came from the AI. And no signal does not mean an image is genuine: it may be missing because the content is older, comes from a different tool or because the additional data was stripped when it was shared. Google also recommends a reverse image search to find the original source. How else to spot fakes is explained in our article on AI fakes in images, voices and calls.