From 2nd August 2026, every large language model provider selling into the European Union must embed a hidden, machine-readable mark in the text, audio, images, and video their systems generate. The obligation stems from Article 50(2) of the EU AI Act. Existing LLMs that were on the market prior to 2nd August 2026 have until 2nd December 2026 to comply.

Last reviewed – 18th August 2026

The mark is not a watermark in the traditional sense, i.e. a stamp across a page. The process that creates it is mathematical, and only a machine (with the right key) can detect it (if you want a full explanation of how Anthropic creates a mark, you can read its blog post on the subject). A human being reading your content cannot see or detect the mark and neither can publicly available AI detection tools.

Anthropic (and presumably Open AI, Google, Perplexity etc) is currently developing a watermark detection API. This will mean the general public and institutions will be able to check if an LLM was involved in creating text, audio, images, or video.

The marking must be robust and reliable, meaning it should ideally survive light editing, copying, or formatting changes.

At present, due to technical constraints that prevent the mark from applying only to LLM customers in the EU, Anthropic has applied its watermark worldwide. Other LLM providers such as ChatGPT and Gemini likely face the same constraints and, like Anthropic, will keep the situation under review.

Article 50 sets a binding compliance obligation, carrying fines of up to fifteen million euros or 3% of global annual turnover for non-compliant providers.

How do I remove the watermark?

The mark’s persistence is directly tied to how much of the AI-generated wording survives into the final document. Anthropic states that light editing probably will not remove the watermark completely, while a complete rewrite in which every word is replaced will.

Longer passages carry a stronger signal for including a mark because the model has made more word-choice decisions, while tightly constrained factual or code content receives fewer marks, since altering exact wording risks introducing errors.

Anthropic states that its watermark only confirms Claude’s involvement in creating a document’s text, not how much of the text was created using Claude. This means that if you use Claude to draft a client’s contract and do not heavily edit it, the watermark will be present. However, if you draft the contract yourself and feed it into Claude to tighten and proofread the document, the watermark may also appear, depending on how much of the visible text Claude actually generated via the editing process versus the amount of visible text created by you. This basic principle is likely to apply regardless of which LLM model you use, but the amount of editing needed to weaken the watermark will depend on the provider. As at the time of writing, only Anthropic has disclosed this information.

Implications for Barristers and Solicitors

Most commentary on the AI Act watermark has focused on confidentiality and legal privilege risks from feeding client material into public models. To ensure SRA compliance and client confidentiality, no client information should ever be given to a public LLM. The SRA’s Warning Notice, issued yesterday (17 August 2016) provides:

“Client information should only be entered into AI systems where appropriate contractual, technical and organisational safeguards are in place to protect confidentiality. Firms and solicitors should satisfy themselves that client data remains within a secure environment, is not accessed by unauthorised third parties, is not used to train AI models except where explicitly authorised and appropriate to do so, and is not retained longer than necessary.”

We would strongly argue that the above could never be guaranteed when using a public LLM such as ChatGPT so client information, even if redacted in the document, should not be fed into these systems. Your firm’s/chamber’s AI policy should reflect this, especially regarding LLM use by junior and trainee solicitors and clerks.

The most immediate consideration for legal professionals is in drafting court briefs and client documents. As LLM providers begin developing watermark detection tools, it may in future become possible for clients, court officials, and regulators to check for signs of LLM involvement in drafting with more statistical rigour than current AI-detection tools offer. Existing tools rely on stylistic guesswork and are widely criticised as unreliable, regardless of their accuracy claims. Cryptographic watermarking is a more robust signal in principle, but detection remains probabilistic rather than conclusive: a positive result indicates an LLM was likely involved in producing some of the text, not how much of it was AI-generated, and a negative result does not prove the opposite. The signal can also weaken or disappear entirely after substantive editing, paraphrasing, or rewording, so a heavily reviewed and revised brief or contract may show little or no detectable trace even if an LLM contributed to an early draft.

If a brief contains a hallucinated case or other fabricated content, courts have already shown they will scrutinise and sanction this through existing professional conduct channels, such as asking the lawyer directly and reviewing drafting records. This is likely to remain the primary checking mechanism even as watermarking tools become available. Separately, a client might one day point to a watermark as evidence that AI assistance was used on a matter, but this would be weak standalone evidence for disputing time billed or invoice amounts, since a watermark shows involvement, not the extent of the barrister or solicitor’s research, verification, and/or editing.

However, regardless of the above, the development of watermark detectors could lead to uncomfortable conversations that risk wasting considerable resources in defending your position. Therefore, it is important to update your AI usage policy document accordingly. It is also crucial to ensure all documents, whether they be briefs, contracts, or research, are created in a way that demonstrates and proves the extent of human input involved.

We will provide updates on this issue as more information becomes available.

Warm wishes,

Paul and Corinne McKenna

www.lawtelligence.co.uk

Author Corinne McKenna is the co-founder and director of Lawtelligence, a specialist legal marketing agency serving UK solicitors and barristers. With an LLB degree from the University of Canterbury and over 27 years’ experience in legal services sales and marketing, Corinne brings substantive legal knowledge to marketing strategy and brand development. Her background includes roles at LexisNexis in the UK and New Zealand, where she managed key legal accounts and delivered training to law firms. Corinne has authored widely on legal marketing topics for publications including Today’s Conveyancer and Solicitors Journal, with particular expertise in E-E-A-T principles, AI-optimised content, and SEO strategy for legal services.