What does an AI watermark look like?

Nothing. If you can see it, you are looking at a label, not a watermark.

The mark a company means when it says watermark is a statistical pattern buried in the content itself — tiny adjustments spread across an image's pixels, or a bias in which word a model picks when several would do equally well. It is meant to be imperceptible. Anything you could point at, you could crop.

The badge in the corner is a label

The small stamp on an AI-generated image is there for people, and it comes off. In August 2026 Google added a setting that removes the visible badge from new Gemini image and video generations, while the invisible SynthID mark stays embedded either way. The toggle does not appear for EU accounts, where the AI Act requires visible labelling.

In text there is nothing to see at all

Text has no pixels to nudge, so the mark goes into the writing. As the model generates, it is steered toward particular options whenever several words would fit, which leaves a faint bias across a few hundred words. It is not a hidden character you can search for and delete. Anthropic began marking Claude's output this way on 2 August 2026.

Which is why you cannot just look

Reading one of these marks takes the detector built for it. Google's SynthID Detector sits behind a waitlist aimed at journalists and researchers. Anthropic's text detection is in private preview, for regulators, researchers and organisations with EU compliance duties. There is no general check that reads every company's watermark, and a site offering you one is doing something else and calling it that.

What you can inspect yourself is C2PA metadata — signed provenance wrapped around a file, rather than a watermark inside it — with the file audit. And a check that finds nothing tells you nothing: most files carry no mark at all, and no watermark result, present or absent, establishes who wrote or made something. The two technologies, in full.