Leadership3 distinct publishers3 min readPublished Updated
Anthropic says the mark exists to satisfy the EU AI Act and that it has no durable way to scope it by region, so every customer gets it. OpenAI's support page says text watermarks are coming too.
The Board Room · Leadership desk
Compiled by The Board RoomSomething wrong?How this is made
Anthropic has been embedding an imperceptible watermark in Claude-generated text on supported models globally since August 2, and it told Business Insider the reason is compliance with the European Union's AI Act [1][2]. In a blog post on Friday defending the feature, the company said it is "applying watermarking globally at launch because we don't yet have a durable way to scope it by region" [3][4].
That is the clearest case yet of an EU obligation shipping as a worldwide product default. The trigger is not where the customer sits or which law binds them; it is which model handled the text [1].
Anthropic's post also said its provider peers will have to do the same, which Business Insider read as a likely nod toward OpenAI [5]. OpenAI did not immediately respond to Business Insider's request for comment, though a support page updated two weeks earlier says it also plans to add watermarks to text [6][7]. Google already watermarks AI-generated content using SynthID, and OpenAI uses SynthID for supported images and audio [8][9].
The technique, drawn from a 2024 Google DeepMind paper, works on word choice [10]. Where a model could pick "joyful" or "cheerful" at random, the watermark biases that selection so that holders of a key can detect the pattern [11]. There are no hidden characters, unusual fonts, or secret text, and Anthropic says readers will not be able to tell the difference [12]. The key has not been released; Anthropic says it will offer a free API so users and third parties can check text themselves [13][14].
The limits matter more than the mechanism for anyone relying on this. Anthropic's post says that with its key, "one can only answer the question 'What is the likelihood this was partly written by Claude?'" [15]. Longer passages are easier to detect because more word choices were made, factual passages get fewer marks to protect accuracy, and code gets fewer still because a wrong term can stop software running [16][17]. Light editing probably will not remove the mark; a complete rewrite will, at which point Anthropic notes it is arguable whether the text is AI-generated at all [18].
The objections are governance objections, not aesthetic ones. Investor Bill Gurley said that if only Anthropic can identify the mark, the company becomes "judge, jury, and prosecutor" [19]. Former Microsoft executive Steven Sinofsky raised data retention and "your right to private thoughts free of a digital trail" [20]. Trainer John Crickett questioned whether watermarked code complicates copyright claims where authors cannot show sufficient human input [21]. Klick's Simon Smith warned that lightly assisted work, such as a grammar check, could be read as AI-authored [22]. Anthropic's answer is that a watermark shows Claude processed the content, can survive proofreading, translation or summarising, and does not change a user's rights or ownership under its terms [23][24].
Commercially, some customers told Business Insider they cancelled Claude subscriptions over the watermark, and Anthropic told the publication it has not seen an uptick trend in cancellations [25][26].
Watch three things. Whether "durable" regional scoping arrives, because that determines if global marking was a transitional constraint or the new baseline. Whether the free detection API and key actually ship, since the credibility of the whole scheme rests on third parties being able to check [13]. And whether provenance labels hold up operationally: X's "Made with AI" tag appeared on outgoing White House Press Secretary Karoline Leavitt's resignation post on Wednesday, then disappeared [27]. Anthropic says older models get the watermark in the coming months, so procurement teams have a short window to fix ownership and disclosure language in contracts before the mark is everywhere [28].
Ranked by verification strength, evidence, and original report placement.
Anthropic told Business Insider that the watermark has applied to Claude-generated text globally on supported models since August 2.
Anthropic told Business Insider it is adding the watermark in order to comply with the European Union's AI Act.
Anthropic published a blog post on Friday that defends and explains the watermarking technology and outlines its limitations, after concerns from Claude users.
Anthropic cites a 2024 Google DeepMind paper that established the watermarking technique it will apply to Claude outputs.
The watermark works through word choice: where a model could randomly select between synonyms such as "joyful" or "cheerful", the watermark acts as a different type of random number generator, and a holder of a key can detect whether text follows Claude's subtle patterns of choice.
Anthropic has not released the key, and says it will offer a free API allowing users and third parties to check text for Claude's watermark themselves.
Follow any of these and your For You feed starts watching them — no settings page required.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
On-record but first-party and unverifiable
The load-bearing facts — global application on supported models since August 2, the EU AI Act rationale, the lack of regional scoping, the mechanism, and the stated limits — are all attributable and quoted, either to Anthropic's blog post or to a spokesperson. What is missing is anything outside Anthropic: the detection key is unreleased, full technical details were not published, and neither article reports independent testing of imperceptibility, output-quality neutrality, or resistance to editing. All of it also arrives through one publisher.
Default-on in production, detection side still absent
Adoption of the marking itself is high and involuntary: it has been live globally on supported models since August 2 and is not something customers opt into, with older models slated to follow. But the ecosystem half of the deployment does not yet exist — no key, no detection API — so no measurable checking behaviour is reported, and the only demand-side signals are unquantified cancellation anecdotes contradicted by the company.
Modestly overstated capability, honestly stated limits
Anthropic's own framing is unusually hedged — it discloses that detection identifies Claude rather than AI generally, that factual text and code receive fewer marks, and that a full rewrite defeats the mark — and the coverage carries those limits forward. The residual gap comes from capability asserted ahead of anything checkable: an 'imperceptible' mark that changes nothing about quality, and a free third-party detection API, are both promised while the key stays private, so the practical usefulness of the mark to anyone outside Anthropic is currently zero.
Compliance-driven, reputation-managing, single-outlet
Incentives are visible on every side. Anthropic's stated motive is regulatory — EU AI Act compliance — and it published the explainer only after user backlash, while simultaneously asserting that peers will have to follow, which frames a cost it is absorbing as an industry-wide obligation. It also retains sole possession of the detection key, giving it an interest in how detection is characterised. Critics carry their own positions: an investor, a former platform executive, and vendors working in AI-adjacent businesses. Both articles come from the same outlet, whose second piece is largely a channel for the company's own post.
Solid on what was said, thin on what is true
Confidence is high that Anthropic said and shipped what is described: dates, quotes, mechanism, rollout order and limits are consistent across two same-publisher items and largely traceable to a published company post. Confidence is materially lower on effects — quality neutrality, removal resistance, detection accuracy and churn — because there is one publisher, no independent verification, no released key, and disputed, unquantified customer-impact evidence.
product
A five-hour script beats Claude's watermark, so stop treating it as provenance3 distinct publishers
leadership
Anthropic's invisible watermark lands hardest on the customers paying $100 a month1 distinct publisher
build
A 14,000-star watermark remover, and no detector to test it against1 distinct publisher
leadership
Anthropic watermarks all Claude text, and the disclosure problem lands on employers4 distinct publishers
Distinct publishers with included, body-backed reporting in this cluster.
businessinsider.com
5 articles · August 18, 2026
forbes.com
5 articles · August 20, 2026
theguardian.com
1 article · August 17, 2026