Is Removing an AI Watermark Illegal? The Actual Legal Answer

A quick note before anything else: this article is general information for a technical and educational audience, not legal advice. Copyright and AI-transparency law is genuinely unsettled here, varies by country, and turns heavily on facts specific to your situation. Nothing below should be read as a prediction about what would happen to you personally — consult a qualified attorney licensed in your jurisdiction for that.

With that said: since watermarks-remover crossed roughly 11.7k GitHub stars in under a week for stripping Claude, Gemini, and OpenAI provenance marks, the comment sections have split into two confident, opposite camps — "obviously illegal, that's tampering with copyright info" versus "obviously fine, it's your file." Neither camp is fully right. The honest legal answer is closer to it depends on what the mark is, what jurisdiction you're in, and what you do next — and that answer, while less satisfying than a flat yes or no, is the one that actually survives contact with the statutes.

TL;DR

Start with the actual statute: DMCA Section 1202

The US law people usually mean when they say "watermark removal is illegal" is 17 U.S.C. § 1202, the DMCA's "Integrity of Copyright Management Information" provision. It's worth reading what it actually says, because the popular version of this rule is looser than the statute.

Section 1202(c) defines copyright management information (CMI) as a specific list: the title of a work, the name of the author, the name of the copyright owner, terms and conditions of use, identifying numbers or symbols, and a short list of related items. Section 1202(b) then prohibits, without authority, intentionally removing or altering CMI, or distributing works knowing CMI has been removed — but only when done "knowing, or... having reasonable grounds to know, that it will induce, enable, facilitate, or conceal an infringement."

That intent clause is doing most of the legal work, and it's the part most online commentary skips. As legal commentary on the statute has repeatedly noted, Congress built § 1202 so that it "does not apply to those who act innocently" — the removal has to be tied to facilitating or hiding an infringement, not just erased for its own sake. Courts applying the statute (e.g. in the Ninth Circuit's Stevens v. CoreLogic line of cases) have reinforced that a bare removal, without evidence connecting it to concealing infringement, does not by itself establish a § 1202(b) violation.

Does AI watermark metadata even count as CMI?

This is where it gets genuinely uncertain, and where you should distrust any confident answer — including this one, past a certain point.

• C2PA manifests plausibly qualify as CMI in some cases, since they're explicitly designed to carry authorship, ownership, and terms-of-use style data — the same categories § 1202(c) lists. If a C2PA manifest identifies a copyright owner and someone strips it specifically to pass the work off as someone else's or to hide that it infringes another work, that's a much more plausible § 1202(b) fact pattern.
• Purely statistical text watermarks (SynthID-style token bias, the kind Anthropic, Google, and most labs are converging on — see how AI text watermarking actually works) are a harder fit. That kind of mark identifies which model generated the text, not a title, author, or copyright owner in the sense § 1202(c) lists. Whether a court would treat "this text came from Model X" as CMI is an open question — no reported US case has resolved it as of this writing, because the technology is too new for the caselaw to have caught up.
• No public detector exists for most of these marks, as we covered in are AI watermarks monetisable? — which also complicates ever proving in court that a specific mark was present and then removed.

Do not read "plausibly qualifies" as "definitely qualifies." This is genuinely unresolved law being applied to a technology that mostly didn't exist when Congress wrote the statute in 1998.

What the penalties actually are, if § 1202 applies

If a court does find a § 1202(b) violation, the actual numbers are public and worth citing precisely rather than guessing at:

• Civil damages under 17 U.S.C. § 1203: statutory damages of not less than $2,500 and not more than $25,000 per violation, at the court's discretion, plus the possibility of actual damages and profits instead.
• Criminal penalties under 17 U.S.C. § 1204 apply only for willful violations committed for purposes of commercial advantage or private financial gain — a first offense can carry up to $500,000 in fines or up to 5 years in prison (or both); repeat offenses roughly double that exposure.

Notice what's required to reach the criminal tier: willfulness and a commercial-gain motive. Someone stripping C2PA metadata off a personal image for privacy reasons is nowhere near that bar.

The EU AI Act: a duty on providers, not a ban on you

The other law that gets cited in these debates is the EU AI Act's Article 50, which entered into application on August 2, 2026 — the same date Anthropic used as the cutoff for rolling out Claude's own watermarking. It's worth being precise about who Article 50 actually binds, because this is where a lot of casual commentary conflates two different legal questions.

Article 50 splits obligations by role:

• Providers of AI systems that generate synthetic audio, image, video, or text — the labs building the models — "shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated." That's a duty on Anthropic, OpenAI, Google, and similar providers, not on you.
• Deployers of a system that generates deepfakes or manipulates image/audio/video content must disclose that the content is artificially generated or manipulated when they put it in front of people. If you're building a product on top of a model and shipping synthetic content to end users, this duty can land on you — but it's a disclosure duty on distribution, not a rule against ever touching a watermark.

Article 50 carries real penalties for non-compliance — up to €15 million or 3% of global annual turnover, whichever is higher — but those penalties attach to a provider or deployer failing to mark or disclose, not to a private individual who later strips a mark from content they already have. Our complete EU AI Act guide and what changes after enforcement both cover the fuller compliance picture if that's the duty you're actually asking about.

The distinction that matters: "a company must legally watermark its own AI output" and "an individual user removing that watermark afterward is itself illegal" are two separate legal questions, governed by different provisions and different actors. The Act answers the first one directly. It does not answer the second one at all — Article 50 has no clause criminalizing watermark removal by a downstream user.

C2PA's own spec treats stripping as routine, not sinister

It's worth grounding this in what the standard itself says, not just what statutes say about it. The Coalition for Content Provenance and Authenticity (C2PA) — the group behind the Content Credentials format Anthropic, LinkedIn, and others use for file metadata — designed the spec knowing manifests get stripped constantly, mostly by accident.

Re-saving an image through almost any editing tool, screenshotting it, converting the format, or uploading to a platform that recompresses images on ingest all silently destroy a C2PA manifest — a point we covered when Anthropic shipped its own C2PA metadata and again in how the removal tool handles it. The spec's own guidance treats an absent manifest as inconclusive, not as evidence of tampering — because in the overwhelming majority of real-world cases, the metadata is gone because of an ordinary re-save, not a deliberate strip.

That's a meaningful data point for the legal question, even though it isn't itself a legal ruling: a standard that expects and tolerates routine, accidental loss of its own credentials is a weak foundation for treating deliberate removal as inherently wrongful, absent some other bad act layered on top (concealing infringement, fraud, and so on). The C2PA project's own "Durable Content Credentials" work — pairing manifests with watermarks and fingerprints specifically because manifests alone don't survive the real world — is an admission that stripping is the expected failure mode, not an edge case.

Terms of service is not law — keep these separate

A huge share of the online "is this legal" debate is actually a debate about platform rules, not statutes, and collapsing the two leads to bad conclusions in both directions.

If you strip an AI watermark and re-upload the content to a platform in a way that violates that platform's content policy — claiming AI art as hand-drawn in a community that bans that, for instance — you've broken a contract (the platform's terms of service), not necessarily a law. The consequence is account suspension, content removal, or a ban, enforced by the platform, not a criminal or civil law claim brought by the state or a court.

This distinction is the single most common source of misleading claims about this topic online — a post breaking a platform's rules gets described as "illegal" when the actual exposure is a suspended account, and a post that's genuinely a legal problem (concealing infringement, fraud) gets waved off as "just a ToS thing." Get the category right before reasoning about consequences.

Where this actually turns legally risky

Pulling the threads above together, the honest, non-hedged version of the answer is: removing an AI watermark, by itself, is unlikely to be illegal in most jurisdictions most of the time — but pairing that removal with a separate bad act is where real legal exposure shows up. Concretely:

• Concealing copyright infringement. If you strip CMI specifically to hide that a work infringes someone else's copyright, that's the exact fact pattern § 1202(b) targets, and here the intent element is satisfied.
• Fraud. Passing off AI-generated content as human-made to induce someone to pay, invest, or rely on it — a fake "authentic" product photo, a fabricated "real" testimonial — can implicate ordinary fraud law, which has nothing to do with watermarks specifically and everything to do with the deception.
• Academic dishonesty. Stripping a mark to submit AI-written work as your own is almost always a violation of an institution's academic-integrity policy — enforced by the school, similar in category to a ToS violation, not a criminal statute (absent some other law, like the growing set of state AI-disclosure rules covered in our EU AI Act and US policy guide).
• Impersonation or defamation. Using unmarked synthetic content to impersonate a real person or spread false claims about them can trigger separate torts and, in some jurisdictions, criminal impersonation statutes — again, the watermark removal isn't the violation; the impersonation is.
• Sector-specific disclosure laws. Some jurisdictions now mandate AI disclosure in specific contexts regardless of watermarks — for example, New York's AI video disclosure law for synthetic performers. Removing a watermark doesn't remove your independent obligation to disclose under a law like that.

None of these turn on the watermark-removal act in isolation. They turn on what the removal was for.

Practical takeaways

1. Don't treat "I can strip it" as "I'm allowed to strip it," and don't treat "I stripped it" as automatically criminal either. Both extremes overstate a genuinely unsettled area of law.
2. The intent behind removal is what statutes like § 1202 actually look at, not the mechanical act of removal. Personal, privacy, or research use of your own content sits in a very different risk category than removal paired with concealing infringement or committing fraud.
3. Know which duty you're actually asking about. "Must a company mark its AI output" (EU AI Act, mostly a provider/deployer question) and "can I remove that mark afterward" (mostly untested, fact-dependent) are different questions with different answers.
4. A platform ban is not proof you broke a law, and a clean legal read is not proof a platform won't ban you. Check both separately.
5. If money, litigation, or regulatory exposure is actually on the line, this article is not enough. Get a lawyer licensed in your jurisdiction — see the FAQ disclaimer above, which is not boilerplate.

Related on explainx.ai

• A watermark removal tool just added OpenAI and Gemini support — the technical companion to this post, covering what watermarks-remover actually strips and what survives
• Anthropic is watermarking Claude text — the C2PA and text-watermark mechanisms this legal analysis applies to
• How AI text watermarking actually works — the statistical mechanism behind the marks discussed above
• EU AI Act and US policy: complete regulation guide — the fuller compliance picture behind Article 50
• What actually changes after EU AI Act enforcement — provider vs. deployer duties, in depth
• Are AI watermarks monetisable? — why no public detector exists to even test a removal against
• The case FOR AI watermarks — the provenance argument this legal analysis assumes as background
• New York's AI video disclosure law for synthetic performers — an example of a disclosure duty independent of any watermark

Primary sources: 17 U.S.C. § 1202 (Cornell LII) · 17 U.S.C. § 1203 (civil remedies) · 17 U.S.C. § 1204 (criminal offenses) · EU AI Act, Article 50 · C2PA specification and FAQ · Content Authenticity Initiative, "Durable Content Credentials"

This article is general information for a technical and educational audience, current as of August 18, 2026. It is not legal advice, does not create an attorney-client relationship, and should not be relied on as a prediction of the outcome in any specific case. Copyright and AI-transparency law is unsettled in several of the areas discussed, varies by jurisdiction, and continues to evolve — consult a qualified attorney licensed in your jurisdiction for guidance on your specific situation. Follow @explainx_ai for updates.