Anthropic has started giving outside groups a way to check whether a piece of text came from its Claude AI. The access goes to regulators, law enforcement, journalists, fact-checkers, researchers, teachers, and civil society groups in Europe, plus any business that needs to check its own compliance. For now it is a private preview, so you cannot simply paste text into a public website; approved organizations have to request access directly.
The reason this exists is a European rule. Article 50 of the EU AI Act took effect on August 2, 2026, and it requires companies to mark AI generated text, images, and audio so people can tell the content was not made by a person. Miss this and a company can be fined up to 15 million euros, or 3 percent of its entire worldwide sales for the year, whichever number is bigger. That is a real number a compliance officer has to plan around, not a symbolic penalty.
The method behind the watermark was built by Google. It is called SynthID, and it works by slightly adjusting how the AI picks between similar words while it writes. A reader sees nothing odd, but a computer checking the text later can spot a statistical pattern that would almost never happen by chance. Anthropic says this pattern can survive some light editing, which would make it sturdier than older AI detection tools that guess based on writing style and get fooled easily by a quick rewrite.
Not everyone buys the claim that this comes free. If the AI is choosing words partly based on a hidden code rather than picking the word that best fits the sentence, some researchers argue the writing quality has to suffer somewhere, even if it is hard to notice. Anthropic disputes this and says quality and content are unaffected. There is no independent way for an outsider to settle that argument yet, since the detection tool itself is still restricted to a small group of approved organizations.
There is also a business wrinkle worth knowing about. Trade publication Artificial Lawyer pointed out that a permanent, hidden AI signature could create awkward moments in industries like law, consulting, or any field where contracts explicitly forbid AI generated work, or where clients negotiate fees partly based on how much human effort went into a document. A watermark that survives editing means a client, a court, or a counterparty could later confirm AI was involved in a report or contract clause, even if the company using Claude never mentioned it. Nothing about this is illegal, but it changes the calculation for anyone who prefers to keep their AI use quiet.
The bigger pattern here is that watermarking is becoming a compliance requirement, not a nice-to-have feature. Any business operating in Europe that uses AI to write customer communications, reports, or marketing copy should assume that text produced by major AI tools may carry a hidden, checkable tag going forward. The practical move is to treat AI generated drafts the way you would treat any output you plan to publish under your own name: read it, edit it meaningfully, and decide upfront whether disclosing AI involvement is a problem for that specific contract or client relationship, rather than finding out after the fact.