Anthropic watermarks all Claude text, and the disclosure problem lands on employers

The mechanism decides how much of this a company has to care about. Anthropic's watermark steers Claude's word choices into a detectable pattern, and it is built to stay detectable through copying and pasting, appearing in translations, summaries and more extensively edited text ^[6][12]. The signal therefore tracks how many of the finished words the model picked. An Anthropic spokesperson told Business Insider that lighter editing is less likely to trigger it because "there's very little (if anything) for the watermark to attach to" ^[15]. That is a description, not a threshold, and it cannot be written into a staff handbook. Caroline De Cock of the Lisbon Council put the gap more bluntly to the same publication: Article 50 "called for a scalpel to catch deepfakes and mass synthetic generation, but Anthropic is using a sledgehammer that hits simple proofreading and translation" ^[13].

De Cock's second observation is the one for the operating committee. The approach can reduce Anthropic's own regulatory risk by shifting the burden of filtering and interpreting those labels onto downstream users and external platforms ^[17]. Downstream means the firm that sends the deliverable.

The vendor estate makes that harder before it gets easier. Of the four rivals Business Insider approached, one, Google, says it already watermarks text worldwide using SynthID ^[19][24]. OpenAI pointed to an FAQ saying it plans to extend labels to text but named no date ^[20]. Meta pointed to a July blog post and has so far worked on images, with no text plan disclosed ^[21]. xAI did not respond, having recently tagged a White House press secretary's departure message on X as AI-generated, after which the tag disappeared ^[22]. The EU rules apply to new models launched in Europe from 2 August, with older models given until December ^[9]: about four months ^[23] in which one client memo assembled from three tools can carry three provenance regimes and a single author line.

None of this depends on detection working well today. Governments and other qualified parties may already hold a detector, and how well such tools perform is untested ^[26]. The exposure is a document drafted this quarter and questioned in two years, against a policy the firm never wrote, on the ownership question the Forbes column raises without answering: an employee writes the memo, AI edits it, whose work is it ^[2].

Deepset chief executive Milos Rusic told Business Insider that these systems "can be useful evidence" but that he would be cautious about turning them into "an automated judgment about authorship or misconduct" ^[8]. Someone will make that judgment anyway, in procurement review or an internal inquiry, and it was employers who encouraged staff to use the tools in the first place ^[7]. Since providers can change watermarking methods at will ^[10], a policy pinned to what a detector can currently find has a short life. A disclosure line in the engagement letter does not.