A European regulation just changed what your AI tools leave behind. To comply with the EU AI Act's Transparency Code, Anthropic has begun watermarking the text its Claude models generate — an invisible signal, woven into the words themselves, that marks the content as AI-produced. Every major model developer that signed the same Code of Practice is now on the hook to do it too.
The reaction was not quiet. Within days, users were calling it surveillance, and some announced on X and Reddit that they were canceling their subscriptions. Much of the pushback came from users who didn't want their AI use marked.
For marketing departments, the issue is knowing where AI already participated in the work your team produces and whether you could account for that involvement.
Claude models launched on or after August 2, 2026 carry the watermark from release. It applies worldwide — across the Claude app, API, Claude Code, Cowork, and Tag, and through cloud providers like AWS, Google Cloud, and Microsoft Foundry. Because it's applied at the model level, there's no documented opt-out, and Anthropic says it's working to add the mark to older models over the coming months.
The immediate change affects Claude. But beyond Claude, AI's role in the things companies produce is becoming more visible. As AI becomes embedded in everyday workflows, the finished product doesn't necessarily reveal how much of it AI produced. Watermarking begins to make some of that involvement detectable.
How Claude's AI Text Watermark Works
Claude's text watermark isn't a hidden character or piece of metadata that disappears when someone copies the response into another document.
Anthropic is using a version of Google DeepMind's SynthID-Text method. When Claude has several equally reasonable words to choose from, the system uses a private key to influence the selection.
A sentence might work equally well with "overcast" or "grey." The watermark affects how those low-stakes choices are made. Repeated across a longer passage, those choices create a statistical pattern that Anthropic can detect.
The reader can't see the pattern. Anthropic says it doesn't change the meaning, quality, or readability of Claude's response.
Supported PNG, JPG, and SVG files are handled differently. Those files receive signed provenance metadata using the C2PA industry standard, indicating that Claude created or processed the file.
Claude is also unlikely to be alone. Anthropic says other major AI providers have committed to marking AI-generated content as part of new transparency requirements in the European Union. AI provenance is becoming an industry-wide concern, and watermarking is one way providers are responding.
Can Claude's AI Watermark Be Removed?
The text watermark survives copying, pasting, and reformatting. Anthropic says it may also remain after light editing.
A complete rewrite can remove it. Developers have also claimed they can erase the watermark by using another model to paraphrase Claude's output. Those claims are hard to evaluate from the outside. Anthropic has opened watermark detection only to a narrow set of eligible organizations — regulators, media, fact-checkers, researchers, and enterprises with their own compliance obligations — and hasn't released it to the general public.
The watermark also has limits. Detection works better on longer passages. Factual text, code, and light proofreading give the system fewer word choices, which can make the watermark weaker.
Most importantly, detecting a watermark doesn't prove that Claude wrote an entire article. It only indicates that Claude was likely involved at some point. It cannot identify the user, company, or conversation behind the text.
AI involvement exists on a spectrum. A marketer might use Claude to fix grammar in a draft, reorganize an argument, generate campaign variations, translate finished copy, or produce a complete first draft. The finished content may look equally polished in every case, but the role AI played is very different.
What the watermark leaves unanswered is how AI was actually used.
Marketing Teams Need Visibility Into the Process
Marketing teams rarely keep a record of how much AI shaped a piece of work in the first place.
This becomes especially important when the content depends on a person's identity or expertise. A CEO's commentary is the clearest case. AI can transcribe the conversation, organize the CEO's ideas, research support, or tighten a draft. But there's a meaningful difference between AI helping communicate someone's perspective and AI generating that perspective on their behalf. The same line runs through crisis communications, sensitive customer messages, original reporting, and anything built on personal experience: the audience is trusting the judgment behind the words, not just the words.
Other work sits at the opposite end. Content variations, internal summaries, routine updates, early drafts — all fair game for heavier automation. The challenge is knowing the difference.
AI Provenance Is Becoming Part of Marketing Infrastructure
For years, companies could adopt AI primarily by deciding which tools employees were allowed to use. That becomes harder as AI moves inside the software teams already use and participates in more stages of everyday work. Approving a list of tools is no longer enough; companies also need visibility into how AI moves through them.
Watermarks won't solve that problem on their own, but they point toward a future in which AI involvement becomes more traceable.
Traceable isn't the same as trustworthy, and that's the gap RAD cares about. A watermark can confirm AI was involved. It can't tell you whether the content is authentic, whether it will land with the people you're trying to reach, or whether it's defensible when someone asks you to stand behind it. Provenance is the easy half of the problem — the industry is only now treating it as urgent. The harder half, the one that actually decides whether the work is any good, is knowing what's real before it goes out. It's the question RAD has been sitting with the whole time.
For marketing departments, that makes provenance part of the infrastructure around AI adoption. Companies may need clearer internal definitions of AI-assisted and AI-generated work. They may need records of which systems contributed to sensitive content and expectations for when human review is required. Most importantly, they need people who remain accountable for the work regardless of how much technology helped produce it.
The goal isn't to prove that every sentence was written by a person. AI is becoming too useful and too integrated into modern work for that to be a practical standard.
The more useful standard is knowing how the work came together. As human + AI becomes an ordinary way of working, companies will increasingly need to understand what happened on both sides of the plus sign.
