The AI industry has spent three years fretting about machines passing as people. The fix now arriving inverts the problem: marks on everything the machines make, whether the person using them likes it or not.
Anthropic said this week that Claude will apply invisible, machine-readable watermarks to the text and images it generates — and that the marking will extend back across older models, not just the newest. The stated reason is compliance with the European Union's AI transparency rules. There is no opt-out. Other AI companies are moving the same way, though it is Anthropic, by Forbes' account, collecting most of the public pushback for saying so plainly.
For the reader, the arrival of watermarked AI text is one of those changes that sounds abstract and is not. A very large number of people now draft things with Claude and its rivals — emails, reports, covering letters, essays, condolence notes — and quietly pass the results off as unassisted. Watermarking means that text may be identifiable as machine-made by anyone holding the right detection tool. Transparency for the reader arrives, inescapably, as exposure for the writer. Which of those you find reassuring depends on which of them you are on any given day.
The questions that matter next are ones the announcement, at least in what has been reported so far, does not answer. How durable is the mark — does it survive a paragraph's editing, or a paraphrase, or a pass through another tool? Who gets the detectors: employers, universities, everyone, or a regulator's chosen few? And what of false positives, the honest writer whose own prose trips the alarm? Watermarking schemes are only as trustworthy as their error rates, and error rates are precisely what launch announcements omit.
A note of history is useful here. Detection of AI text has been promised before, by tools that proved confidently wrong in both directions, and the arms race between marking and scrubbing is as old as watermarking itself. Axios calls this a new front in AI detection, which is right — a front, not a victory.
Our advice, in the meantime, is unglamorous. Assume anything you generate with a major AI model may eventually be attributable to that model, and let that assumption inform what you are willing to sign your own name to. That was always the honest policy. It is now becoming the enforced one.
