Claude Will Watermark Its Output. Your Judgment Won't Carry a Label.
Anthropic announced this week that content produced by Claude will start carrying machine-readable marks. Generated text will contain an invisible watermark, and supported files like PNG, JPG, and SVG will carry digitally signed provenance metadata based on the C2PA standard. Models launched from August 2, 2026 support marking at launch, and Anthropic says it is working to extend this to earlier models. The rollout is global, across the Claude app, the API, Claude Code, and the major cloud platforms.
The trigger is regulatory. Anthropic signed the EU AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content. But rather than build an EU-only feature, they are shipping it everywhere. Expect other model providers to follow the same path.
What the marks actually tell you
Less than the headlines suggest. Anthropic itself is careful here: a Claude mark means the content may have been generated or processed by Claude. That is all.
Processed is doing a lot of work in that sentence. If you write a paragraph yourself and ask Claude to fix the grammar, the output can carry a mark. If you draft a proposal from your own twenty years of experience and use AI to expand it into clean prose, marked. Translate your own contract summary, marked. The ideas, the data, and the judgment can be entirely yours, and the watermark will not distinguish that from a fully machine-written piece.
The reverse is also true. Heavily edited or paraphrased text may lose the signal. Metadata gets stripped by screenshots, format conversions, and re-saves. So the absence of a mark proves nothing either.
What SMEs should do about it
First, do not panic about your marketing content being “outed” as AI-assisted. Within a year or two, a large share of professional content will carry some provenance signal. The stigma fades when everyone’s output is marked. What will not fade is accountability for what the content says.
Second, treat this as a free line item for your AI policy. If you have a one-page AI usage policy (and you should), add a sentence: we assume AI-assisted output may be watermarked, so we never publish anything we would be uncomfortable owning. That single assumption enforces better behavior than most review gates.
Third, flip it around: provenance is a tool for you, not just about you. When a supplier sends you a compliance certificate, a CV, or a signed-looking PDF, C2PA-style metadata gives you one more integrity check. Small firms are frequent targets of doctored documents. This helps.
The voice of reason still matters
Here is the part that gets lost in the watermarking debate. These marks answer the question “how was this made?” They say nothing about “is this true?” or “is this worth your time?”
A well-reasoned argument does not become weaker because AI helped phrase it, and a bad take does not become credible because a human typed every word. If AI expanded your reasoning, the reasoning is still yours, and so is the responsibility for it. The watermark labels the instrument, not the musician.
For SMEs, that is the practical takeaway. Provenance marking is infrastructure, like HTTPS was. Adopt it, disclose sensibly, keep a human accountable for every claim you publish, and compete on the one thing that will never carry a machine-readable label: judgment.
If you are working out what AI transparency rules mean for your business, start with an assessment. For how we think about AI oversight in operational systems, see governance-first agentic ERP and when not to give an agent write access.