Hidden Watermarks Will Track AI-Generated Text

To help you understand the trends surrounding business and technology and what we expect to happen in the future, our highly experienced Kiplinger Letter team will keep you abreast of the latest developments and forecasts. (Get a free issue of The Kiplinger Letter or subscribe.) You’ll get all the latest news first by subscribing, but we will publish many (but not all) of the forecasts a few days afterward online. Here’s the latest…

Wondering if a piece of writing was generated by artificial intelligence? Anthropic has a solution: Watermarks.

But the company’s recent update is not based on market demands or innovation. It stems from the European Union’s Artificial Intelligence Act. The strict rules prohibit all sorts of conduct and come with a slew of new requirements for AI deemed high-risk. Rules for “high-risk” AI went into effect this month.

Prohibitions on AI systems include “deploying subliminal, manipulative, or deceptive techniques to distort behavior and impair informed decision-making, causing significant harm.” AI systems also can’t “[infer] emotions in workplaces or educational institutions, except for medical or safety reasons.” And much more.

The EU AI Act’s most severe penalties are fines of up to 7% of global revenue for companies found in violation. It’s sure to make the EU a tougher place for American AI companies to do business. But the impact will be global, as seen by Anthropic rolling out watermarks in all countries, noting that the change is related to the law’s transparency requirements.

Anthropic has detailed how watermarks work, adopting a method developed and already used by Google. The process involves how the AI model chooses specific words and word fragments within the text. Anthropic has a key that involves two lists of words and the text generated must include enough of the words from one list to be statistically significant. The method has limitations, such as not working well on shorter passages and only revealing the “likelihood” of being written by AI.

“Because the watermark is part of the text, it will travel with the text when it’s copied and pasted elsewhere, and may persist through some editing,” according to Anthropic. “Watermarking will be applied at the model level, which means it will be present no matter which Claude product or surface the text comes from.”

If it’s working well, readers should not notice. “You won’t see it, and it doesn’t change the meaning, quality, or readability of Claude’s response,” according to the company. Theoretically, watermarks could help identify AI-generated text anywhere. But the move is likely to stir up concerns about Anthropic’s power over users’ text output and what it means for intellectual property.

Expect Anthropic’s adoption of watermarks to ignite more pushback from the Trump administration, with concerns about thwarting tech innovation and harming U.S. tech giants. President Trump said last month the administration will conduct a formal review to retaliate against the EU’s “discriminatory” digital practices, stemming from another EU digital law that has led to huge fines against U.S. tech giants. The growing backlash will also focus more attention on Google’s use of text watermarks for its AI model Gemini.

With the EU’s AI Act gradually coming into full force and American AI giants working to comply, the U.S. relationship with Europe is only set to get more tense.