Delete SynthID Sign in
Watermarks and the law · Updated 15 July 2026

What the law expects, in plain language.

AI-content disclosure and watermarking rules are moving quickly and differ from country to country. This overview explains the landscape as we understand it in mid 2026 and where your responsibilities sit. It is general information, not legal advice.

Three layers, three separate questions

An AI image can carry up to three different things, and each raises its own legal question. A visible badge, such as a corner sparkle, is a label a viewer can see. An embedded signal is an imperceptible pattern woven into the pixels. Provenance metadata is structured data attached to the file, including EXIF, XMP, and C2PA content credentials.

Editing any of these changes the file in your hands. It does not change the underlying fact that the image was generated or assisted by AI, and it does not reach records the original tool may keep on its own servers. Whether you may alter a given layer, and whether you must still disclose the origin, are questions the law answers separately from whether the edit is technically possible.

The global picture in mid 2026

The summaries below are simplified, may already be out of date, and never replace advice about your specific situation. Rules also depend on how and where you publish, not only where you live.

European Union. The AI Act introduces transparency duties for generative systems. Providers are expected to mark synthetic output in a machine-readable way, and those who publish AI-generated or materially altered media, including deepfakes, are expected to disclose it. These duties are phasing in through 2026.

United States. There is no single federal watermarking mandate. Instead there is a patchwork of state laws covering election-related deepfakes, non-consensual imagery, and likeness or publicity rights, alongside federal action against deceptive practices. Several states are moving toward broader AI-disclosure requirements.

China. Rules require AI-generated content to carry both a conspicuous visible label and embedded identifying metadata. Removing, altering, or concealing those labels in order to mislead is restricted.

South Korea. A national AI framework introduces notification and labeling duties, so that people are told when they are looking at generative AI output.

United Kingdom. There is no dedicated AI-labeling statute. Existing law, including fraud, intellectual property, data protection, and defamation, applies, under a principles-based approach led by sector regulators.

India. Intermediary rules and government advisories push platforms toward labeling synthetically generated or altered media. The framework is still developing.

What removing a watermark does not do

  • It does not create authorship or ownership you did not already have.
  • It does not erase records the generating service may hold on its own systems, or copies already shared elsewhere.
  • It does not, by itself, satisfy or cancel a disclosure duty you may still owe to a platform, client, employer, contest, or audience.
  • It does not make a deceptive or unlawful use lawful.

Where your responsibility sits

Delete SynthID is for images you created or have permission to modify. You are responsible for confirming that your edit and your publication comply with the law that applies to you and your audience, and for keeping any disclosure that a reasonable viewer, or a specific rule, would expect. When origin would matter, keep or add a clear statement that the image is AI-generated or AI-assisted.

See our responsible use guidance for a practical checklist, and our terms of service for the rules that govern your account.