The Annalist

Draft

OpenAI adds Google's invisible marks to AI images so fakes can be spotted

In May 2026, OpenAI began embedding SynthID, invisible marks invented at Google, in images made with AI. Kakao and ElevenLabs are adopting them too, and you can now check whether an image is AI-generated right in Google Search.

The better AI gets at drawing, filming and speaking in other people's voices, the harder it is to tell a real photo from a fake. One answer is a digital "watermark": a mark inside the file that the eye cannot see but special software can find.

Google has such a technology, SynthID, developed by its Google DeepMind lab. On 19 May 2026, the day of the Google I/O conference, OpenAI announced it would also add SynthID marks to images created in ChatGPT, Codex and through its API (application programming interface, programmatic access to its models). OpenAI also writes provenance information into files under the C2PA standard (Coalition for Content Provenance and Authenticity). But that information is easily lost when a file is re-saved, while the invisible mark stays.

According to Google, SynthID is also being adopted by South Korean internet company Kakao and by ElevenLabs, which makes voice synthesis. Together with Nvidia, Google will mark video created with Nvidia's Cosmos models. By this date SynthID had marked more than 100 billion images and videos.

Checking is getting easier. You can ask "is this AI-generated?" right in Google Search, through Lens, AI Mode and Circle to Search, and in the Chrome browser in the coming weeks. In the Gemini app, the check had already been used 50 million times.

Why it matters

Competitors began using one system of marks, and OpenAI adopted a rival's technology rather than pushing its own. Decisions like this add up to a shared "record-keeping system" for AI content that social networks, media and regulators can rely on.

For ordinary people, it means the "is this AI?" check is appearing in services they already use. But there is an important caveat: SynthID finds only its own marks. If there is no mark, that does not mean a person made the image: open models run on a home computer mark nothing.

For the curious

Besides 100 billion images and videos, SynthID had marked the equivalent of 60,000 years of audio. On the same day, OpenAI previewed its own image verification tool: it looks for both SynthID marks and C2PA metadata, but for now recognizes only OpenAI's own content. Google also announced C2PA verification in the Gemini app, which shows whether a shot is an unaltered camera original or has been edited.

Sources

  1. Making it easier to understand how content was created and edited, Google, 19 May 2026 (primary source)
  2. Advancing content provenance for a safer, more transparent AI ecosystem, OpenAI, 19 May 2026 (primary source)