AI watermarking
AI watermarking is embedding a hidden, machine-detectable signal in text, images, audio or video at the moment an AI model generates it, so software can later identify the content as AI-generated.
Also called: content credentials, C2PA, SynthID, AI content watermark
By Divya Sharma · Checked Nov 15, 2026 · How we review
As generated images, voices and text get harder to tell apart from human work, platforms, newsrooms and regulators want a way to check what was made by a model. Watermarking is one approach. It is different from the visible logo watermark that photo tools add, and from detectors that guess after the fact by analysing style.
How it works
For images, audio and video, the generator makes small changes to pixels or samples that people cannot see or hear but a matching detector can find. Google DeepMind's SynthID is an example: it embeds watermarks directly into AI-generated content and offers a way to check for them.
For text, Kirchenbauer et al. proposed a scheme that, before each word is generated, picks a random "green" subset of tokens and nudges the model towards them. Readers notice nothing, but a statistical test can detect the bias from a short span of text, without access to the model.
Watermarks vs content credentials
C2PA takes a different route. Instead of hiding a signal in the content, it is an open standard, called Content Credentials, for recording where a file came from and how it was edited. A watermark lives in the content itself; a credential travels with the file as a record of its history.
Limits
Watermarks only exist if the model maker adds them, so open models run locally usually produce unmarked output. Cropping, compression, paraphrasing or translation can weaken or remove a mark. Each scheme needs its own detector. And the absence of a watermark proves nothing about whether a human made the content. For builders working with text-to-image or voice models, check whether your provider marks output, and how your users can verify it, especially where deepfakes are a risk.
Related terms
- C2PA content credentials
- C2PA content credentials are signed provenance records, defined by an open standard from the Coalition for Content Provenance and Authenticity, that travel with an image, video, audio file or document and record its origin and edits.
- Text-to-image
- Text-to-image is the task of generating a new picture from a written description, called a prompt. Most current text-to-image models are diffusion models conditioned on the text.
- Deepfake
- A deepfake is image, audio or video content made or altered with AI so that it convincingly shows a real person, place or event in a way that did not happen and could pass as authentic.
- Synthetic data
- Synthetic data is training or test data generated by software, often by another AI model, rather than collected from real-world sources. It is used to fill gaps, add examples cheaply, or avoid exposing real personal data.