Is It Legal to Remove an AI Watermark? What the Law Actually Says

A free tool for stripping the invisible watermark Anthropic added to Claude's text output appeared on GitHub within 24 hours of the feature going live, and crossed roughly 11,700 stars within six days. The obvious follow-up question — is any of this actually legal — turns out to have a more layered answer than either "obviously fine" or "obviously not."

Who the Law Actually Targets

The EU AI Act's Article 50 marking obligation is a provider-side duty — it requires companies like Anthropic, OpenAI, and Google to mark their AI output as machine-detectable, with fines up to €15 million or 3% of global turnover for failing to do so. End users interacting with or removing marks from AI systems don't face legal obligations under Article 50 itself; the law was written to regulate the companies building AI systems, not the people using their output.

What US Copyright Law Actually Prohibits

In the US, the relevant statute is DMCA §1202, which covers "copyright management information" — and the trigger for a violation isn't the act of removal itself, but doing so knowing it will "induce, enable, facilitate, or conceal" copyright infringement. Removing a watermark from content that's genuinely yours, with no intent to conceal infringement, doesn't fit that trigger — the statute was built around protecting copyright holders' attribution, not policing every edit made to a file someone has the right to modify.

The consistent thread across US and EU frameworks is that the law cares about what the removal is used to accomplish, not the mechanical act of stripping a mark. Cleaning a watermark off your own AI-generated draft sits in a different legal category than removing one to pass stolen work off as someone else's, or to dodge a disclosure requirement that specifically applies to that content.

Where It Actually Gets Risky

The genuine legal exposure shows up in two specific situations: removing copyright management information to conceal that content was copied from someone else, and removing an AI disclosure mark specifically to misrepresent AI-generated content as human-made in a context where disclosure is legally required — political advertising in some jurisdictions, certain journalism contexts, or academic submissions with their own separate honesty rules. In both cases, the underlying deception or infringement is what creates liability, not the technical act of stripping a mark by itself.

The Watermark Was Never a Complete Technical Lock

Independent researchers have pointed out that text watermarking, specifically, faces a structural weakness that image and video watermarking doesn't share as severely — running AI-generated text through a second AI model to paraphrase it tends to break the statistical pattern the original watermark relied on, since the words themselves change rather than surviving as pixels or audio samples would. This is a separate issue from the legal question: even where removal might be perfectly legal, it may also be trivially easy, which is part of why researchers have been skeptical of how much protection text watermarking actually provides.

Why the Legal Ambiguity Might Be the Point

Regulatory experts have described AI watermarking less as an unbreakable enforcement mechanism and more as a compliance paper trail — a way for both the AI Act and individual companies to demonstrate a good-faith transparency effort, shaping expectations and providing a compliance record rather than functioning as an unbeatable technical lock. That framing helps explain why the law focuses on providers marking output rather than criminalizing removal by end users: the goal was establishing a default of labeled content, not building a system assumed to be tamper-proof from day one.

What This Means Practically

For someone removing a watermark from their own AI-generated draft for personal or professional use, current US and EU law doesn't create a specific prohibition on that act alone — the legal risk attaches to what happens afterward, specifically deception or infringement, not the edit itself. That's a meaningfully different situation from a company failing to mark its own AI output in the first place, which is the actual violation Article 50's fine structure was built to address.

---

The law here draws a line most people don't expect: it's the companies building AI systems who face real regulatory exposure for not marking their output, while individuals stripping a mark from their own content generally aren't breaking a specific law by doing so — unless that removal is used to conceal infringement or to deceive someone in a context where disclosure is legally required. Understanding that distinction is more useful than either assuming removal is always fine or assuming it's automatically illegal.