Anthropic, along with several other major AI providers and around 190 total signatories, signed the EU Code of Practice on Transparency of AI-Generated Content this year. The practical result is that Claude’s outputs now carry a watermark, a marking method required under the EU AI Act to identify content as AI-generated.
What the watermark actually is, and isn’t
It is worth being precise here, because this detail gets misreported often: the watermarking applies to the content Claude produces, not to individual users. There is nothing in the watermark or its underlying key that would let anyone recover information about the user, their organization, or their conversations with Claude. This is a content-labeling requirement, not a user-tracking mechanism.
Why this matters for your content workflow even outside the EU
Regulatory requirements built for one jurisdiction have a habit of becoming the practical global standard, because it is simpler for a platform to apply one policy everywhere than to maintain separate output pipelines by region. If your marketing team uses Claude or other signatory models for content drafting, assume watermarking behavior may apply to your output regardless of where your business operates, and plan accordingly.
How watermarking actually works, at a high level
Without getting deep into the cryptography, the general approach behind most AI content watermarking is to subtly bias the statistical patterns in generated text — things like word choice or token sequencing at a level invisible to a human reader — in a way that’s detectable by a tool with the right key, without changing how the content reads or performs. It’s not a visible tag, a metadata field you could strip out, or anything that changes your workflow in the moment of generating content. That’s exactly why it doesn’t require you to change your prompting or editing process to comply — the labeling happens at the model level, automatically, regardless of what you do downstream. Where it does matter is if you’re editing AI-drafted content heavily before publishing, which is standard practice and should stay standard practice: heavy human editing can reduce or eliminate the statistical signal the watermark relies on, which is a separate, useful fact if you’re ever trying to reason about whether a given piece would even register as AI-assisted under this kind of detection.
The disclosure question this actually forces
The more practically important question this raises isn’t technical, it’s a decision your brand hasn’t necessarily made explicitly yet: what’s your policy on disclosing AI involvement in published content, and is it written down anywhere a new team member could find it. Most marketing teams have an implicit, ad hoc answer — “we don’t really talk about it either way” — rather than an actual policy. That’s a reasonable place to have started, but as regulatory and platform-level disclosure requirements keep expanding (see the AI marketing backlash data on how audiences already react to visible AI use), an implicit non-policy stops being a safe default and starts being a gap someone eventually has to explain under pressure. Writing the actual policy down — what gets disclosed, where, and in what language — takes an afternoon and removes the scramble later.
What to actually check
Understand how watermarking interacts with your publishing platform. Some content platforms and social networks have their own AI-disclosure requirements layered on top of what the model provider does. Know both layers, not just one.
Revisit your own AI-use disclosure policy. If your brand publishes AI-assisted content without disclosure and a regulatory or platform requirement changes that assumption, you want to have already decided your position rather than reacting under pressure. Given the AI marketing backlash data circulating this year, proactive disclosure is often a trust-building move rather than a liability, not something to avoid.
Watch for this pattern to spread. The EU AI Act is one of the more aggressive regulatory frameworks globally, and other jurisdictions tend to reference it when drafting their own AI rules. Expect content transparency and disclosure requirements to keep expanding rather than staying isolated to one region.
The direction here is not ambiguous: AI-generated content transparency is moving from a voluntary best practice toward a regulatory baseline. Marketing teams that build disclosure into their process now, rather than treating it as a future compliance project, will have an easier time adapting as more jurisdictions catch up to the EU’s approach.