Anthropic has implemented a watermarking system for its Claude chatbot that marks text generated or processed by the AI. The system works by making subtle changes in the AI's word choices across generated text — changes that are imperceptible to humans but detectable with the right tool[reference:71]. Anthropic says the watermark was implemented in response to the European Union's landmark AI Act, which requires AI companies to mark content generated or edited by their systems[reference:72].
The news kicked off a firestorm on social media, with users panicking that AI-assisted writing would now bear a permanent "scarlet letter"[reference:73]. But the real concern goes deeper: the watermark indicates Claude processed text, not that it was the original author[reference:74]. If Claude proofreads, translates, summarizes, or edits content, the watermark remains[reference:75].
The User Backlash
Dozens of users have posted on X that they canceled their Claude subscriptions, citing the watermark[reference:76]. Vladislav Rajtmajer, a freelance developer in the Czech Republic, told Business Insider he canceled his Claude Max subscription because he worried an AI label on code shipped to clients could raise questions about authorship or trigger contract penalties[reference:77]. Richard Echols, an AI consultant in Georgia, also canceled his Claude Max subscription, saying he uses Claude to create documentation for businesses and worried the mark could appear when he had written the underlying material himself and used the chatbot only for edits[reference:78]. "Even if you edit your own work that you wrote, they add the watermark anyway," Echols said[reference:79].
A Broader Trend
Anthropic isn't the first AI lab to use watermarks. Google uses its SynthID technology to watermark AI-generated content, and OpenAI uses SynthID for supported images and audio[reference:80]. Elon Musk's social media platform X also adds a "Made with AI" tag on content it determines to be AI-generated or manipulated[reference:81].
But Anthropic's implementation is different because the watermark applies to any text the AI processes, not just text it generates. That means even human-written content could carry the watermark if the user asked Claude to proofread or edit it.
The Other Side: A Benefit for AI Training
There may be an upside to watermarks. They could help AI companies stop training their models on AI-generated text — a growing problem as the internet becomes saturated with synthetic content. If watermarks are reliably detectable, companies could filter out AI-generated text from their training datasets.
What It Means for Users
For now, Anthropic users face a choice. They can accept the watermark and its implications for authorship and privacy. Or they can cancel their subscriptions and look for alternatives. The company has not indicated any plans to change its approach, and with the EU AI Act as a driver, watermarking may become standard across the industry.
The backlash, however, suggests that users are not willing to accept watermarks without a fight. And as more companies adopt similar measures, the debate over AI watermarking is only just beginning.