Anthropic said it will change how its Claude AI model generates text to comply with an EU regulation requiring all AI-generated content to be watermarked starting in December. The company said the changes would alter the granular, random choices at which its model frames sentences and would be undetectable to most readers.
Anthropic's watermark will create a pattern detectable to the company itself and to those with a decryption key. Tech blogger John Gruber argued the watermark would constrain Claude by forcing less precise word choices, though the changes may not affect accuracy. Steven Murdoch, a computer science professor at University College London, said the change would "probably wouldn't have any noticeable impact."
Large language models rely on randomness to avoid repetition and getting stuck in loops, Murdoch said. "There's already randomness involved in any of these large language models," he said. "It's pretty essential to how they work." The watermark would make random choices statistically predictable rather than completely random, but users would see no practical difference, he said.
The regulation applies to all AI companies operating in the EU. Watermarking could make it harder for students, lawyers and university professors to pass off chatbot-written content as their own. Murdoch also cited another reason for watermarking: AI-generated content already circulating in large quantities could damage AI models themselves through "model collapse," in which models trained on AI-written content begin to confuse concepts.
