Anthropic has started watermarking text produced by Claude, embedding an invisible signal that a detection tool can test for later. The company announced the change on 14 August 2026. It covers all future Claude models, and a rollout to models released before 2 August 2026 is underway.
For iGaming, where affiliate pages, game descriptions, bonus terms, localised copy and player messaging are increasingly written or drafted with large language models, the practical effect is that a large share of that output now carries a machine-readable marker of AI involvement.
How the watermark works
The watermark changes the source of randomness Claude uses when it picks the next word. Instead of drawing on ordinary random number generation, the model uses a cryptographic key combined with the preceding words to steer the choice between options that are equally valid. The wording still reads as normal text to a person, but the sequence can be checked against the key.
Anthropic says the watermark adds no latency, generates no extra tokens and does not increase cost. It also states that the change does not alter ownership of the output or legal responsibility for it.
Coverage extends beyond prose. Claude applies C2PA content credentials to metadata for supported file types, including .png, .jpg and .svg. Translations are watermarked. Code is watermarked only in a limited way: comments carry the signal, logic sections carry much less of it, because the range of valid word choices is narrower.
There is no opt-out. The watermark is applied globally at launch.
Why it has arrived now
Article 50 of the EU Artificial Intelligence Act (AI Act) applies from 2 August 2026. It requires providers of generative AI systems to mark outputs in a machine-readable format that is detectable as artificially generated or manipulated, using solutions that are, in the Commission’s wording, “effective, interoperable, robust” and in line with current technical standards.
The European Commission published the final Code of Practice on Transparency of AI-generated Content on 10 June 2026. By the end of July 2026, around 190 companies and organisations had signed it, Anthropic among them.
The Act places a second obligation on deployers, not just model providers. Deployers must disclose AI-generated text published to inform the public on matters of public interest, unless the content has been through human review and someone holds editorial responsibility for it. That carve-out matters for gambling media, affiliate news desks and operator blogs that publish regulatory and market commentary.
What it changes for operators and affiliates
The immediate change is verification. Anthropic says a watermark detection API is coming soon. Once it is live, a third party who can access a piece of text can test whether Claude was likely involved in producing it.
That shifts the position of anyone who commissions content. Operators buying SEO pages from affiliate networks, affiliates buying copy from freelancers and agencies, and B2B suppliers buying documentation or marketing collateral can all check what they paid for. Contracts that promise human-written content become testable rather than a matter of trust.
Regulators and platforms gain the same option. Gambling advertising is already under close supervision in several markets, from Australia’s agreed advertising reforms to the UK Gambling Commission’s enforcement work against unlicensed promotion, including the 252 illegal lotteries it removed from social media. A detector that flags AI involvement gives supervisors a new data point when they examine marketing material, complaint responses or affiliate claims.
Localisation, code and player-facing messaging
Translation is the exposure most operators will underestimate. iGaming localises heavily, often across a dozen or more languages, and machine translation with an LLM has become standard practice for terms pages, game guides and CRM messaging. Anthropic confirms that translations are watermarked, so localised copy carries the marker even when the source text was written by a person.
Player-facing messaging raises a separate question. Operators using Claude to draft responsible gambling nudges, complaint responses or support replies now send out text that can be identified as AI-assisted. The watermark carries no identifying information about the user, organisation or session, so it does not create an audit trail. It only shows that the model was probably involved.
On the development side, the limited watermarking of code means game logic and platform code are barely affected, while generated comments and documentation are. Studios and platform teams reviewing supplier deliverables should treat those as separately detectable.
What the watermark cannot do
Anthropic sets out the limits directly. A positive result confirms only that Claude was likely involved. It cannot separate full authorship from heavy editing of a human draft, and it cannot identify a specific user, organisation or chat session.
Detection strength rises with length. Short passages carry little signal, and factual writing where word choice is constrained, such as bonus terms, payout tables or licence conditions, carries a sparse watermark. Light editing tends to preserve it. A full rewrite removes it. Text where Claude only proofread a human draft is not reliably detectable at all.
The watermark also says nothing about content produced by other models. A negative result is not evidence that a page was written by a person, only that this particular system did not leave its mark. That distinction matters for anyone tempted to treat a detector output as proof in a supplier dispute.
Anthropic distinguishes its approach from third-party AI detection software such as Pangram, which infers AI involvement from linguistic patterns rather than reading a cryptographic signal.
What happens next
The detection API has no published launch date, and until it exists the watermark is a marker nobody outside Anthropic can read. The work for operators, affiliates and suppliers falls in that window: mapping where LLM output already sits in published content and player communications, checking whether Article 50 disclosure duties apply to any of it, and putting human review and named editorial responsibility on the material that needs the exemption.
Content supply contracts are the other item to look at before detection becomes routine. Any clause that guarantees original human writing is now a clause a client can test.
Source: Anthropic
