Claude Text Watermark Expands to Sonnet 5 and Opus 4.8

Most of the debate about AI-generated text has focused on what these systems can write. Less attention goes to a quieter question: whether anyone can later tell that a machine wrote it. Anthropic has now given its answer for a larger part of its Claude lineup.

From 30 September 2026, the company will apply its previously announced text watermark to three more models: Claude Fable 5, Claude Sonnet 5 and Claude Opus 4.8. Anthropic shared the news with customers by email. Its stated reason is compliance with the transparency requirements of the EU AI Act.

What the watermark actually is

The mark cannot be seen. Anthropic describes it as a statistical pattern built into the model's word choice. It is not a label or a line of metadata attached to the output. Instead, it is a signature spread across the text itself, invisible to a reader but detectable with the right tools.

According to Anthropic, the watermark does not affect quality, readability, latency or pricing. The company also says it carries no user data. The mark signals that a Claude model produced the text. It does not carry any information about who asked for it.

The newer models are already covered. Claude Fable 5.1, Claude Opus 5.5 and Claude Opus 5 carry the watermark today, so the September change brings three older models into line with the rest of the family.

Global by design, not by choice

The mark is embedded at the model level, and that has a practical consequence. The watermark is not limited to European users, even though European regulation is the reason for it. It applies worldwide, on every platform where these models run.

For customers, this works in two directions. They do not have to change settings or update integrations, because the watermark simply arrives with the model. They also cannot turn it off. According to Anthropic, the only ways around it are to use an older Claude model or move to another provider.

That leaves some businesses with a real decision. A team that relies on Claude for drafting, summarising or customer communication now produces text that can be traced back to an AI system, whether or not it wants that. For some, this will be a welcome sign of accountability. Others may see it as a constraint they never asked for.

Who gets to check

A watermark only matters if someone can detect it. Anthropic has opened a detection API for public authorities, media organisations and fact-checkers. These groups are the most likely to need a reliable way to judge whether a piece of text came from a Claude model.

This choice of users says something about the purpose. The tool is aimed at institutions that verify information, not at the general public. It fits the transparency logic of the EU AI Act, which asks that AI-generated content can be recognised as such.

The quality question

Anthropic's assurance that the watermark leaves output unchanged has not convinced everyone. Some doubt that the text is fully unaffected. The concern is easy to follow: if the pattern lives in word choice, the question is whether steering those choices can ever be entirely neutral.

A contested field

AI text detection is a disputed topic in any case. Part of the tension comes from an assumption that often travels with it: that text written with AI is text written without human intent or effort. Many people who use these tools would reject that idea. A report shaped over several rounds of prompts and edits reflects the judgement of the person who guided it, even if a model produced the sentences.

A watermark cannot capture that distinction. It can indicate that a model was involved. It cannot say how much thinking, revision or responsibility a human put into the final result.

What changes on 30 September

For most users, very little will look different. The responses will read the same, cost the same and arrive just as fast, if Anthropic's claims hold. What changes is what happens later. Text from these three models will carry a signal that authorities, newsrooms and fact-checkers can read.

The bigger question is not technical. As regulation pushes providers toward traceable output, users will have to decide how they feel about AI text that identifies itself, and whether being able to detect AI involvement also tells us who is responsible for the words.