Anthropic has laid out how it plans to watermark text produced by future versions of its Claude AI assistant — and some paying customers are already heading for the exit.
According to a report carried by MSN, Anthropic has detailed a text-watermarking system planned for future Claude models. The system is designed to offer a way to estimate whether the AI assistant was involved in writing or editing a given passage. In other words, text that passes through Claude would carry a signal that can later be checked, rather than being indistinguishable from anything a person typed themselves.
The reaction from users has not been warm. BetaNews reports that Claude's watermark is leading paying users to cancel their subscriptions.
That tension is the whole story in miniature. Watermarking is one of the main technical answers the AI industry has offered to a growing problem: nobody can reliably tell what a machine wrote. Schools, publishers, courts, and employers have all been asking for a way to check. A watermark embedded at the point of generation is a far more dependable signal than the guesswork-based "AI detectors" that have circulated so far.
But the same feature reads very differently from the customer's side of the transaction. People paying for a writing assistant may see an invisible marker attached to their output as something closer to a tracking device on their own work — a label they did not ask for, applied to documents they consider theirs.
Anthropic has not, in these reports, said the feature has shipped; the watermarking is described as planned for future models.
This matters because it is an early test of whether AI companies can deliver the transparency that institutions are demanding without alienating the paying customers who fund the products.