What is an AI watermark?

A visible watermark is text or a logo drawn onto the image itself: a brand name, a “made with AI” note, a copyright line, or a label repeated across the whole picture. Anyone can see it without special software, although it can be cropped or painted over.

An invisible watermark alters pixels, audio samples, or word choices in a way people do not notice but a matching detector can recognise. It is embedded by the generator, so it only exists for content produced by systems that support it.

Provenance metadata, often called a content credential, is a signed record attached to the file that lists who made it, with which tool, and what edits followed. It is not drawn into the picture at all, which is why it is sometimes confusingly described as a watermark.

SynthID is Google DeepMind’s watermarking tool for AI-generated content. According to DeepMind, it embeds digital watermarks directly into AI-generated images, audio, text or video. They are imperceptible to humans but detectable by SynthID’s technology.

For images and video, DeepMind says the watermark is designed to withstand cropping, filters, frame-rate changes and lossy compression. Text written in the Gemini app and audio from Lyria and NotebookLM are marked as well. A public SynthID Detector lets anyone check whether an image, video or audio file was made with AI from Google or from partners such as OpenAI, NVIDIA and Kakao. What SynthID does not do is label the picture for a viewer.

The Coalition for Content Provenance and Authenticity, or C2PA, maintains an open technical standard for establishing the origin and edits of digital content. Its steering committee includes Adobe, Amazon, BBC, Google, Meta, Microsoft, OpenAI, Sony and TikTok. The consumer-facing name for the standard is Content Credentials.

A C2PA manifest is a set of statements about an asset, digitally signed by the software or hardware that performed each step. It can record that an AI system generated or altered the content. The manifest is usually embedded in the file as metadata, and inspection tools such as the Content Credentials Verify site display it. The C2PA explainer is candid that this metadata can be removed and is not always complete, so “durable” credentials pair it with an invisible watermark or fingerprint.

Invisible watermarks and credentials are built for verification: a platform or a journalist can check them, but a casual viewer cannot. A visible watermark is built for communication. It tells the person looking at the picture that it is a draft, a concept, an AI render, or the property of a studio.

Many marketplaces, publishers and clients ask for AI imagery to be labelled, and a concept visual sent for approval should not leak as a finished product. A label repeated across the frame also makes casual reposting more work than a small corner mark, and a line with a © symbol and a year is easy to add with the photo copyright watermark tool.

Any AI-generated JPG, PNG or WebP can be watermarked here without an account. The text watermark editor repeats a short line across one image; a “made with AI” note or your name works well. To place a brand mark instead, use the logo watermark tool. For a whole set of renders, the batch watermark tool applies the same label to up to 50 files and downloads them as a ZIP.

Choose a font and colour that stay readable over light and dark areas, keep the opacity moderate, and check the preview at full size. Exported files are re-encoded by the browser, so any credential or invisible watermark the generator attached may not survive, so keep the original. The image is processed on your device and is not uploaded to our servers.