Watermarks made by popular from Claude’s announcement are causing a stir within the AI industry
getty
A free tool that strips AI watermarks from Claude’s text appeared on GitHub within 24 hours of Anthropic’s watermarking feature going live. Within days, several more showed up, targeting everything from invisible Unicode markers to the signed metadata Anthropic attaches to generated files. If AI watermarks can be defeated that easily, are they built to last, or just for show?
That’s the question worth asking as Anthropic, Google, Microsoft, Meta and OpenAI all move to mark their AI output, and it has a more useful answer than “yes” or “no.”
Why AI Watermarks Are Suddenly Everywhere
Article 50 of the European Union’s AI Act became applicable on August 2, 2026, requiring companies that build generative AI systems to mark their output in a machine-readable, detectable format. Non-compliance carries fines of up to €15 million or 3% of a company’s global annual turnover, whichever is higher. About 190 organizations, including Anthropic, Google, Meta, Microsoft and OpenAI, signed the EU’s voluntary Code of Practice on Transparency of AI-Generated Content ahead of that deadline, a framework that gives signatories a presumption of compliance.
Nine days after the deadline passed, Anthropic announced Claude would embed AI watermarks directly into its text output, an invisible statistical pattern, plus signed C2PA metadata in generated files, rolled out globally rather than restricted to EU users. Google has watermarked AI-generated images since 2023 and has since extended that to text, audio and video. OpenAI reportedly built similar text-watermarking capability years ago and chose not to deploy it, reportedly over concerns about false positives and giving competitors a way to fingerprint ChatGPT usage patterns.
Why AI Watermarks Break So Easily
The technical problem facing every one of these companies is asymmetric. An AI watermark has to survive nearly anything a user might do to their text or file. An attacker only needs one method that works.
- Statistical marks degrade under editing. Anthropic’s own documentation acknowledges that proofreading, translation, heavy paraphrasing or short outputs can all cause its text watermark to go undetected.
- Removal tools move fast. The GitHub tool that appeared within a day of Anthropic’s announcement gained roughly 72 stars a day by running Claude’s output through one or two rewrite passes, enough to disrupt about 70% of the token sequences the watermark depends on.
- A detected watermark isn’t proof of authorship. Anthropic itself notes a match only shows Claude “may have processed” the content, not that Claude wrote it, an ambiguity platforms and employers are unlikely to account for when treating an AI watermark check as a verdict.
The Backlash Is Already Costing Anthropic Subscribers
The fragility of AI watermarks hasn’t stopped them from having a real effect on user behavior. Business Insider reported that dozens of users were canceling Claude subscriptions after the watermark rollout, and named several who followed through. An AI consultant told the outlet he canceled because the watermark can appear even on text he wrote and lightly edited himself, not just text Claude generated from scratch. A software engineer said the watermark confirmed a decision to leave he’d already made over other service concerns. A digital agency founder cited a broader worry: building a workflow around a vendor’s tools means that vendor can change the terms later, unilaterally.
Anthropic told Business Insider it hasn’t seen a measurable increase in cancellations tied to the announcement, and the company reports roughly 300,000 business customers as of last year, a scale at which a few dozen public complaints on X barely register.
Regardless of volume, the backlash is a useful data point regardless of its size. It shows that even an AI watermark easy enough to defeat with a free GitHub tool still changes how some users perceive the product they’re paying for. The mark doesn’t have to be technically robust to affect trust. It just has to be visible enough, or rumored enough, for people to feel like something changed.
Are AI Watermarks Here To Stay?
As a technical guarantee, AI watermarks are close to theater. They don’t hold up to a determined adversary, and Anthropic doesn’t claim they do. But as regulatory infrastructure it’s likely that they’re here to stay, at least in the near term.
Their job at this stage is not to be unbreakable, it is to be the default. The long-term goal is a paper trail for the cases that actually get scrutinized: a dispute, an audit, a lawsuit, where nobody thought to strip the mark first.
Regulators built the underlying framework around that assumption. The EU’s Code of Practice calls for a layered approach combining metadata, statistical marking and detection tools, precisely because no single layer was expected to hold on its own.
That’s also the pattern digital rights management and other “easily circumvented” provenance systems have followed for two decades: broken within days of release, and still standard practice years later, because the institutional weight sits behind the mark rather than the mark’s technical resilience.
What Are The Impacts Of AI Watermarks?
For businesses producing or relying on AI-assisted content, three effects follow directly from that reality:
- A new default, and a new tell. AI-generated content now arrives marked unless someone deliberately strips it, which means the absence of an AI watermark starts to look more deliberate over time, the same way scrubbed metadata on a photo reads as more suspicious than a photo that simply never had any.
- Trust erodes faster than the technology does. The Anthropic cancellations show that user perception moves ahead of technical reality. A watermark doesn’t need to work perfectly to cost a vendor goodwill, and it doesn’t need to be broken for a customer to feel surveilled by it.
AI watermarks aren’t just for show, but they’re not the finished answer either. They were built to be a default, not a lock, and the businesses that treat them as anything more are the ones that will get caught out first.
Leave a Reply