There was a point when we argued that artificial intelligence could take a writer's job. The basic justification was that if it could produce passable copy faster and cheaper, then there would be less reason to pay someone to do it.

Then it gradually became a thing where a writer who knew how to use AI effectively became a threat. Now, even that advantage is becoming harder to keep invisible because Claude is adding invisible watermark to AI-generated text.

Content now carries a machine-readable signal identifying that help was involved. Here's how it affects you.

Signed with love, Claude

Anthropic recently updated its content-marking system in response to its commitments under the European Union AI Act's transparency rules.

As AI-generated content becomes commonplace, greater transparency and signals about where content comes from can give people useful context about the information they consume. To support transparency and follow our legal obligations, Anthropic is working to include machine-readable marks in content that Claude generates.

Anthropic

Starting with Claude models launched on or after August 2, 2026, the company says generated content will carry machine-readable provenance marks from the moment the model produces it.

The new policy applies worldwide, across Claude's apps, API, Claude Code, and Claude Cowork. It also applies to supported Claude models accessed through Amazon Web Services (AWS), Google Cloud, and Microsoft Foundry.

You won't see the watermark while reading the response, nor does it affect the text's meaning or quality. The signal is merely embedded in the text and can travel when you copy and paste the response somewhere else. It may even survive some editing.

Generated or processed files like PNG, JPG, or SVG are not exempt. Claude will attach signed Coalition for Content Provenance and Authenticity (C2PA) metadata. Content Credentials Verify, from the Content Authenticity Initiative (CAI), can already inspect C2PA credentials attached to such files. Adobe Content Credentials is another compatible tool.

However, detection mechanisms for these marks are in the works. That said, they're likely to only indicate that Claude may have processed the content. They don't prove the tool wrote the original material.

You could write something yourself and ask Claude to proofread or translate it, for example, and the resulting content could still carry the mark.

It doesn't sound good for workers whose employers use AI detectors, which already have a questionable track record. They produce false positives and can't reliably prove AI involvement either way. But Anthropic itself revealed that there are ways to make the watermark undetectable.

You can remove the watermarks (sort of)

According to Anthropic, your Claude-generated piece of content might escape its new detectable mark for several reasons. You may use a model it released before introducing the system.

Heavy editing, paraphrasing, translation, or mixing Claude's output with other writing can also alter the text enough that the watermark is undetectable. If you type very short passages, the AI may conclude that they contain too little text for the detection system to identify any signal.

Because provenance information is stored in metadata, converting or resaving a file, taking a screenshot, or manually stripping the metadata can remove the information.

Irene is an experienced writer with five years' experience in the Tech, E-commerce, and Food niches. She currently writes for Talk Android and Android Police, where she geeks out about the latest innovative practices, solutions, and gadgets. When not writing, Irene embraces her inner couch potato, finding joy in gaming, singing, music, and of course, eating.

Previous Post