Anthropic has announced it will begin embedding invisible watermarks into text produced by its Claude AI models, in response to new transparency requirements under the EU AI Act, the world's first comprehensive AI regulation. The watermarks are designed to make AI-generated content easier to identify, but the system will also flag text originally written by humans that has subsequently been edited or processed by Claude — a scope that could complicate efforts to determine whether material was genuinely created by AI.

The San Francisco-based AI company stated on Tuesday that Claude models launched in the EU from 2 August onward will include machine-readable watermarks from release. Anthropic is also working to extend the technology to older models.

Although the watermarks were introduced to comply with the EU AI Act, Anthropic confirmed that the markings will apply globally wherever supported Claude models are available, covering Claude, Claude Code, and its developer platform.

Text generated by the models will carry an embedded watermark that Anthropic said should persist through copy-and-paste operations and may survive at least some degree of editing. Images and other supported file types will carry signed metadata indicating they have been processed by Claude.

Anthropic plans to release tools enabling users and third parties to detect its watermarks, though technical specifications have not yet been disclosed.

The announcement comes amid mounting pressure from regulators and publishers for AI companies to make synthetic content more identifiable, following a surge in machine-generated material across the internet. Under the EU AI Act's transparency provisions, all major AI providers operating in the bloc — including OpenAI, Google, and Meta — face comparable obligations to mark AI-generated output, and several jurisdictions beyond the EU, including the UK and individual US states, are advancing their own AI labelling requirements. City AM has previously reported on the spread of AI-generated "slop" in newsrooms, including opinion pieces and pitches deceptively presented as the work of executives and other contributors.

Human-Written Text May Also Be Flagged

Anthropic acknowledged that detecting one of its watermarks would not necessarily mean Claude authored the original content. A user could write an article independently and then ask Claude to proofread, translate, or edit it — and the resulting text could still carry Claude's watermark, even though the underlying content and ideas originated with the human author.

Article 50 of the EU AI Act stipulates that marking requirements do not apply when an AI system performs "an assistive function for standard editing" or does not substantially alter the material or its meaning. Anthropic has signed the EU's voluntary Code of Practice outlining how companies can comply with transparency obligations, but its decision to apply watermarks across all supported Claude output means the system could capture material beyond what the legislation strictly mandates.

The company also cautioned against treating the watermarks as definitive proof of AI authorship. Heavy editing or paraphrasing can render a watermark undetectable, and short passages may contain insufficient text to produce a reliable signal. Metadata attached to images and other files can also be stripped when files are converted or screenshotted.

Detection Technology Still Imperfect

Alex Cui, chief technology officer at AI detection firm GPTZero, explained in a post on X that text watermarking typically works by subtly altering the probability of word selections within a model, creating a statistical pattern that can later be detected.

However, Cui noted that the technology remains imperfect, observing that watermarks may struggle to withstand "intense paraphrasing." He also warned that publicly releasing detection tools could make it easier for users to figure out how to circumvent them.

"The free paraphrasers I've tried have quickly bypassed Google DeepMind's SynthId for what it's worth," he wrote.

A recent study involving 1,682 adults, published in the journal Judgment and Decision Making, found that participants rated stories generated by ChatGPT as more engaging and of higher quality than human-written stories. However, respondents gave higher ratings to stories they believed had been written by humans. In two additional experiments, researchers found no general evidence that participants could reliably distinguish between human-authored and AI-written stories.

Anthropic stated that its watermark will not visibly alter Claude's output or affect its meaning or readability. The company conceded that the system provides a signal rather than conclusive evidence of a piece of content's origin, leaving businesses and publishers to determine how much weight to assign the new marks.