Is it legal to remove a watermark?

It depends entirely on whose work it is. Removing a watermark from something you made, or hold a licence to, is ordinarily fine. Removing someone else’s credit so you can reuse their work is not — and in the US that is its own offence, separate from the infringement.

It turns entirely on whose work it is. Removing a watermark from footage you shot, from an image you hold a licence to, or from something you generated yourself in an AI tool is ordinarily fine. Removing a photographer’s credit from their photograph so you can reuse it is not, and it can be expensive.

Almost every page on this topic stops there. The useful detail is in which law applies, because two very different ones get confused constantly.

People reach for section 1201 of the US Digital Millennium Copyright Act — the anti-circumvention rule that covers breaking technical protection measures. It generally does not apply here. A visible watermark does not control access to anything; you can already see the image. It is a mark printed over the top, not a lock.

The provision that does apply is section 1202, which covers copyright management information — the author’s name, the copyright notice, the terms of use, and identifying marks that carry them. Intentionally removing or altering that information, knowing it will conceal infringement, is its own violation, separate from the infringement itself.

A section 1202 claim carries statutory damages of $2,500 to $25,000 per violation, and does not require proving any actual loss.

A section 1202 claim carries statutory damages in the range of $2,500 to $25,000 per violation, and it does not require proving actual loss. That is why stock agencies and photographers reach for it: the watermark removal is a second, independently actionable act on top of using the image without a licence.

The practical consequence is that “I removed the watermark from a stock preview” is a worse position than “I used an image without a licence”, not a lesser one.

When Sora, Veo, Kling or Meta AI stamp a badge on something you generated, no third party’s credit is being removed. Those are provenance markers rather than authorship claims, and the terms of the generator you used govern what you may do with the output — worth actually reading, since they differ.

The part that matters more is the distinction between the visible badge and the invisible marker. C2PA provenance records and Google’s SynthID are embedded in the file or the pixels themselves and survive the removal of the visible logo. Stripping a visible badge does not make AI content undetectable, and any tool claiming otherwise is either misleading you or doing something you should not want.

There is a widely repeated claim that California now bans tools whose purpose is removing AI provenance disclosures. We went and read the bills, because the claim kept appearing without a citation.

The California AI Transparency Act (SB 942), as amended by AB 853, became operative on 2 August 2026. It requires large generative AI providers to attach latent and manifest disclosures, obliges large online platforms not to knowingly strip provenance data, and requires hosting platforms to ensure the systems they offer carry disclosures — with further platform duties from 1 January 2027. Penalties run to $5,000 per violation, enforced by the Attorney General, a county counsel or a city attorney.

In the EU, Article 50 of the AI Act puts transparency duties on providers and deployers of generative systems rather than on tool makers. The direction of travel in both jurisdictions is obvious enough, though, and anyone building in this area should expect the ground to keep moving.

These are ordinarily fine:

• Footage or photographs you created
• Output you generated yourself in an AI tool
• Media you hold a licence to, where that licence permits alteration
• A client’s material you have been engaged to work on

These are not:

• Stock previews, where the mark exists precisely because the licence has not been bought
• Removing a creator’s credit so their work can be reused as if unattributed
• Republishing another person’s post as your own
• Stripping a provenance badge to pass AI video off as real footage

We built FrameCleaner for the first list. Our acceptable use policy says so plainly, and we remove visible marks only — embedded provenance records are left intact, deliberately.

We also cannot police it, and would rather say that than imply otherwise. Your files never reach our servers, which is what protects your privacy and equally means we have nothing to inspect. What you remove, and from whose work, is on you.