Most businesses using generative AI have been working on an unspoken assumption: once text leaves the chat window, nobody can reliably tell whether a person or a model wrote it. In August 2026, Anthropic announced that text generated by Claude now carries an invisible watermark, and that images and files it produces will carry signed provenance metadata. For any organisation that uses Claude to draft client communications, marketing copy, reports, contracts or product outputs, the practical question is no longer whether AI involvement can be detected, but who will be able to detect it, in what circumstances, and what that means for your contracts, your customers and your regulatory position in Europe.
This article explains what the watermark actually is, why it has arrived now, the range of ways it could affect commercial use of AI, and where regulation of AI-generated content is likely to go next.
What Exactly Is Claude’s Watermark?
A large language model writes one word at a time, and at many points in a sentence it has several options that would be equally good. Ordinarily, the choice between those options is settled by a random number generator. Anthropic's watermark replaces that randomness with a pattern derived from a secret cryptographic key and the few words immediately preceding each choice. To a reader, the output looks and reads exactly as it would otherwise. To someone holding the key, the pattern of choices across a passage is statistically distinctive enough to indicate that Claude was probably involved in producing it.
Anthropic explains that the technique builds on SynthID-Text, a method Google DeepMind published in Nature in 2024, and says its own testing found no measurable effect on quality, speed or cost. The watermark adds no hidden characters and no extra text, so there is nothing visible to strip out. For images and supported file types such as PNG, JPEG and SVG, Claude attaches Content Credentials based on the C2PA open standard, which is signed metadata recording where the file came from.
Several features of the system matter for businesses. According to Anthropic's own materials, the watermark contains no information identifying the user, organisation or conversation, so it signals that Claude was involved rather than who used it. It applies across Anthropic's own apps, the API and Claude as offered through major cloud providers, and it is being applied worldwide because Anthropic says it cannot yet reliably limit it to particular regions. New models launched from 2nd August 2026 carry both text watermarks and file credentials, while older models are being brought into scope by 2nd December 2026.
How Reliable Is Detection?
Anthropic is candid about the limits. Detection is probabilistic, and a positive result indicates that text may have been produced or processed by Claude rather than proving that it wrote it. Light editing will not remove the signal, but a thorough rewrite in which every word changes will. Short passages contain too few word choices to carry a strong signal, factual or highly constrained passages leave less room for watermarking, code is barely affected, and the system cannot tell whether text came from another AI provider. Equally, the absence of a watermark does not show that a person wrote something.
Access to detection is also restricted for now. A detection API is in private preview for organisations such as regulators, law enforcement, media organisations, fact-checkers, researchers, educational institutions, civil society groups and enterprises with compliance obligations, with broader access planned. A free Content Checker is available for verifying file credentials, and any C2PA-compatible tool can read them.
Why Is This Happening Now With Watermarking?
Article 50(2) of the EU AI Act requires providers of AI systems that generate synthetic audio, images, video or text to ensure outputs are marked in a machine-readable format and detectable as artificially generated. That obligation has applied since 2nd August 2026, and the Digital Omnibus on AI, which entered into force on 27th July 2026, gave systems already on the market a grace period until 2nd December 2026. That date matches Anthropic's own deadline for bringing older models into scope.
In July 2026 the Commission also backed a Code of Practice on transparency of AI-generated content, which sets out how providers can meet the marking obligation in practice, including by combining more than one technique such as signed metadata and imperceptible watermarking. Anthropic is one of the signatories. Breaches of the Article 50 transparency obligations can attract fines of up to €15 million or 3% of worldwide annual turnover, whichever is higher.
It is worth mentioning that the marking obligation falls on the provider of the generative system, which in this case is Anthropic, not on every business that uses Claude. Second, a separate obligation in Article 50(4) applies to deployers who publish AI-generated text to inform the public on matters of public interest, requiring them to disclose that it is AI-generated unless it has undergone genuine human review or editorial control by someone who takes responsibility for it. Watermarking and disclosure are therefore related but distinct duties, and a business using Claude may have obligations of its own even though the watermark is applied automatically.
What Could This Mean for Your Business in Practice?
The most immediate effect is on contractual commitments about AI use. A growing number of enterprise customers, particularly in financial services, the public sector and legal services, include clauses in their supplier terms restricting or prohibiting the use of generative AI, or requiring disclosure of it. Until now, compliance with those clauses has largely rested on trust. If your team uses Claude to prepare deliverables for a customer who has prohibited AI use, the customer or its auditors may in future be able to test the documents they receive. The time to reconcile your internal AI practices with your contractual promises is before that test, not after.
Professional services firms face a related question on pricing. Where AI shortens the time spent on a piece of work, clients who can see that AI was involved are likely to ask whether the fee reflects it. Firms that already explain how they use AI and how that is reflected in their pricing will find the watermark far less awkward than firms that have said nothing.
For SaaS providers building on the Claude API, the watermark travels with the output of your product. If your platform generates customer emails, reports, summaries or content for your users, that text may carry Claude's signal when it reaches their recipients. The watermark does not identify you or your customer, but it may reveal that your product relies on a particular model, which some vendors treat as commercially sensitive. It also means your customers could face questions from their own clients about AI use that originated in your product, so your documentation and customer terms should be clear about which features use generative AI and what that means for outputs.
Marketing and content teams should consider how watermarked copy interacts with their publication practices. Most marketing material will not fall within Article 50(4), which is aimed at text informing the public on matters of public interest, but thought leadership, news-style content and public statements can come closer to that line. If you intend to rely on the human review exception rather than labelling content, you will want a genuine editorial process, a named person who takes responsibility, and a record that the review happened.
Knowledge management is a less obvious area of exposure. Watermarked text copied into templates, precedent banks, policy documents and standard clauses will carry that signal forward into every document built from them. Over time, a firm's standard contract terms or client guidance may contain passages that test positive for AI involvement long after anyone remembers where they came from. That is not unlawful, but it can create confusion in disputes, audits or customer reviews unless you understand the provenance of your core materials.
Employment and recruitment raise a fairness issue. If an employer or university gains access to detection tools and treats a positive result as proof that a candidate or student did not write their submission, it risks drawing conclusions the technology cannot support. Anthropic itself states that detection cannot distinguish between text Claude wrote and text Claude heavily edited, and that it works poorly on short passages. Any organisation using detection to make decisions about people should treat the result as one piece of evidence, give the individual a chance to respond, and consider its obligations under data protection law, including the rules on automated decision-making.
Finally, there is the question of evidence in disputes. Detection results may start to appear in litigation, regulatory investigations and contractual claims, for example where one party alleges that a report was produced by AI contrary to an agreement or that a document was represented as expert human work when it was not. Due to the signal being probabilistic, being removed by heavy rewriting, and not being readable by the public at present, it is unlikely to be decisive on its own. Businesses should nonetheless expect it to be raised, and should keep their own records of how AI was used on significant pieces of work.
Does the Watermark Change Who Owns or Is Responsible for the Content?
No, the watermark indicates how content was produced. It does not transfer ownership, affect whether copyright subsists in the output, or shift liability for errors. If your business publishes or relies on AI-generated content, you remain responsible for its accuracy, its compliance with advertising and consumer law, and any rights it infringes, exactly as you were before. The watermark simply makes the production method more visible.
Where Is Regulation of Ai-Generated Content Heading?
Within the EU, the direction is towards consistent, machine-readable marking across all major providers rather than a patchwork of voluntary measures. Article 50 applies to every provider of generative AI placing systems on the EU market, so Claude's watermark should be seen as an early example of something that will become standard across the industry. The Code of Practice also contemplates work on interoperable detection, and as that develops, the practical ability of regulators, platforms and businesses to check content across several providers is likely to grow. Businesses should plan on the basis that AI involvement in published content will become progressively easier to identify.
The UK has no equivalent statutory marking obligation at present. However, UK businesses selling into the EU, and UK content that reaches EU audiences through global products, will be affected by the EU rules regardless. In the United States, California's AI Transparency Act (SB 942) takes a comparable approach by requiring large generative AI providers to embed latent disclosures and offer detection tools, which suggests some convergence between the two largest regulatory systems on the principle, if not the detail.
Over the longer term, we expect the regulatory focus to shift from whether content is marked to how businesses behave once it can be detected. Consumer protection regulators may look more closely at content presented as human or expert-written when it was not, sector regulators may set expectations about disclosure in regulated advice, and procurement rules may begin to require suppliers to state their AI use in a way that can be verified. Watermarking is the infrastructure that makes all of those developments enforceable.
What Should You Do Now?
The sensible starting point is to establish where Claude and other generative tools are being used across your organisation, including in customer-facing deliverables and in any product features built on third-party models. With that picture in place, review your customer contracts and supplier terms for clauses that restrict or require disclosure of AI use, and check that your practices match what you have promised. If you publish content that could fall within Article 50(4), decide whether you will label it or rely on the human review exception, and document the process you choose.
For product teams, it is worth confirming with your AI providers what marking they apply, how it will change as older models come into scope before 2nd December 2026, and whether anything in your product strips or alters metadata that customers may expect to remain intact. The position will vary depending on what your business produces, who it sells to and where, and where the answer is unclear, tailored advice on your contracts and product disclosures will help you get ahead of questions that customers and regulators are likely to start asking.
How Can Gerrish Legal Help?
Gerrish Legal is a dynamic digital law firm. We pride ourselves on giving high-quality and expert legal advice to our valued clients. We specialise in many aspects of digital law such as GDPR, data privacy, digital and technology law, commercial law, and intellectual property.
We give companies the support they need to successfully and confidently run their businesses whilst complying with legal regulations without the burdens of keeping up with ever-changing digital requirements.
We are here to help you, get in contact with us today for more information.