Claude's text watermark, explained

Short answer: text from newer Claude models now carries an invisible watermark, a pattern woven into the words that does not change what the response says. For most Australian businesses nothing changes and there is nothing to do. If you rely on the watermark, for example because some of your AI-generated content reaches the EU, check which models your builds call and leave the sampling settings at their defaults on the models Anthropic flags.

We read Anthropic’s help page on how Claude marks AI-generated content and the Claude Platform release notes on 5 November 2026. Here is what we found, and what to check.

What is Claude’s text watermark?

It is a mark built into the text Claude writes. Anthropic says Claude uses two techniques side by side: an invisible watermark in generated text, and signed C2PA Content Credentials on certain files Claude creates, such as images. In Anthropic’s words, the text watermark is woven into the text itself:

You won’t see it, and it doesn’t change the meaning, quality, or readability of Claude’s response.

It is added by the model, so it travels with the text wherever the model is used. The help page lists the Claude Platform (API), Claude, Claude Code, Claude Cowork and Claude Tag, and says supported models carry the watermark when you reach them through AWS, Google Cloud or Microsoft Foundry too. The 1 September release note adds that “marking requires no changes to your requests or response handling.”

Which Claude models are watermarked?

Anthropic says Claude models launched on or after 2 August 2026 support marking at launch, and that it is working to add marking to models released before that date. The release notes give the dates so far.

The help page has a table of every model and platform. It changes as Anthropic adds models, so treat it as the source of truth and check it before you rely on a particular model.

Can a setting stop the watermark applying?

Yes, on some models. The help page says content may not carry a detectable mark if it “was generated by Claude Opus 4.6 or Claude Sonnet 4.6 with a non-default temperature, top_p, or top_k.” The 3 November release note says the same for Sonnet 4.6.

Those three settings control how much variety the model puts into its wording. Builders often turn temperature down so an agent gives steadier, more predictable answers, for example when it fills in CRM fields. That is a reasonable choice, but on the models above it can mean the output is not marked.

So if you need the mark, leave those settings at their defaults on those models. If you do not need it, nothing about your build has to change.

Newer tooling is moving away from these settings anyway. Anthropic’s Python SDK v1.0, released on 20 August 2026, removed temperature, top_p and top_k from its Messages methods, and on Haiku 5.5 each of them can return an error. The builds most likely to still set them are older ones.

Why is Anthropic watermarking Claude?

The EU AI Act. Anthropic says it has signed the EU AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content, and that the law includes a transition period for AI systems launched before 2 August 2026. That is why new models are marked at launch and older ones are being added.

The Act is European law, so it does not apply to an Australian business just because it uses Claude. If your business sells into the EU or publishes AI-generated content there, ask your lawyer whether you have obligations of your own. For the Australian side, see our guide to AI governance in Australia.

Who can detect the watermark?

Not everyone yet. Anthropic says watermark detection is in private preview, available to organisations eligible under EU law, such as regulators, law enforcement, media, fact-checkers, independent researchers, educational organisations and EU civil society groups. It is also available to businesses that must verify watermarking for their own compliance with the Act. Anthropic says it plans to widen access over time, and the help page links to a request form.

A detected mark is evidence, not proof. Anthropic says it indicates the content “may have been generated or processed by Claude.” The help page also lists cases where Claude-written text may not carry a detectable mark, so a missing mark does not prove a person wrote something either.

What should your business do this week?

• List the models each build calls. Your agents, workflows and any custom code that sends requests to Claude. Then check each one against the help page table.
• Search your code for the three settings. Look for temperature, top_p and top_k, especially on Opus 4.6 and Sonnet 4.6.
• Decide which outputs need the mark. Usually that is content going to customers or the public in the EU. Internal notes rarely need it.
• Plan any model moves. Claude Sonnet 4.5 retires on 30 November 2026, so anything still on it needs a new model this month whatever you decide about the watermark.
• Update your AI use policy. Say when your team should disclose AI-assisted writing. The watermark does not replace a person checking the work.

Our take

We think the watermark is a sensible default. It changes nothing a reader can see, and it makes the honest answer to “did AI help write this?” easier to stand behind. The risk for a business was never that AI helped. It is AI writing that nobody checked before it went out. That is a process problem, and a watermark does not fix it either way.

Where we fit

We build custom AI on Claude for Australian companies: assistants that answer from your own records, pipelines that read the documents your team still opens by hand, and agents that act inside your security rules. A watermark is one more setting in a build like that, and every setting should have a reason someone can point to.

So before anything goes live, we test it on your own cases and write down what it is allowed to do. Routine steps run and are logged, sensitive ones notify a person, and anything irreversible waits for a human to say go. When someone asks what the system did, you read the log. Your PM is your builder, and we stay on the hook once it is in production.

Neighbourhood is Certified in the Claude Partner Network Services Track. If you are still working out which jobs AI should take on, start with AI consulting. If you know the job and want it built, see AI engineering. For costs, our Claude pricing guide has every plan in Australian dollars.

Frequently asked questions

Can you see Claude’s watermark? No. Anthropic says the watermark is woven into the text and does not change the meaning, quality or readability of the response.

Which Claude models are watermarked? Models launched on or after 2 August 2026 support it at launch. Claude Mythos 5, Opus 4.7 and Sonnet 4.6 were added on 3 November 2026, and Anthropic is adding other earlier models. The table on Anthropic’s help page is the current list.

Does changing the temperature remove the watermark? On Claude Opus 4.6 and Sonnet 4.6, Anthropic says text generated with a non-default temperature, top_p or top_k may not carry a detectable mark. If you rely on the mark, leave those settings at their defaults.

Who can check whether text has Claude’s watermark? Detection is in private preview for organisations eligible under EU law, such as regulators, media, researchers and educational organisations, and for businesses that must verify watermarking for their own EU AI Act compliance.

Do Australian businesses need to do anything? Most do not. Marking needs no changes to your requests. If some of your AI-generated content reaches the EU, check your models and sampling settings, and ask your lawyer about your own obligations.

Sources

• How Claude marks AI-generated content, Claude Help Center, checked 5 November 2026, for what the watermark is, supported models and platforms, the sampling settings limit, the EU AI Act and detection access.
• Claude Platform release notes, Anthropic, checked 5 November 2026, for the 1 September and 3 November watermark entries, the Python SDK v1.0 change on 20 August, the Haiku 5.5 sampling settings note and the Sonnet 4.5 retirement date.

---