Anthropic embeds invisible watermarks in Claude text to flag AI-generated content

Anthropic announced it will introduce subtle, invisible watermarks into text generated by Claude models released on or after August 2. The watermark is machine-readable but imperceptible to human readers and does not affect text quality or readability. The company says the watermark is designed to travel with text when copied and pasted and may survive some editing, though heavy rewrites or translations could remove it. Images generated or processed by Claude will also receive watermarks indicating that Claude handled the file and flagging any tampering since processing.

Because the watermark sits at the model level, it follows Claude's output everywhere users engage with the model—through the chatbot, the API, and third-party tools like Claude Code. The move is partly motivated by the EU AI Act's transparency rules, which took effect August 2 and require generative AI providers to make synthetic output machine-readable and detectable. While other AI companies have attempted watermarking, most have focused on images. Text watermarking is traditionally harder because text gets copied, paraphrased, translated, and incorporated into other writing constantly. Anthropic acknowledges that the watermark only shows Claude had a hand in something, not that a Claude model generated the entire output—even asking the model to proofread or translate a paragraph could leave a trace. Older Claude models will not yet carry watermarks, but Anthropic says it is still working on extending the feature.

The watermarking initiative occurs amid a wider industry push to combat "AI slop"—low-quality, often mass-produced AI-generated content clogging social feeds. YouTube clarified its "inauthentic content" policy last month to crack down on channels leaning on generic, templated AI output, including AI personas dispensing health, legal, financial, or political advice; those channels can lose monetization, though AI-assisted work like scripts or editing remains permitted. Substack has rolled out a reader-triggered AI scanner allowing readers to estimate how much of a post or comment was human versus AI-written, and writers can add "How I make this" disclosures explaining their process. However, watermarking alone will not solve the problem: anyone determined to disguise AI output has many ways to degrade or erase a statistical text watermark, and previous watermarking efforts have typically been easy to remove. Nevertheless, Anthropic's move reflects growing pressure from consumers and publishers for clearer ways to distinguish AI content from human content, with users increasingly frustrated that they bear the burden of detection when AI companies play a vital role in producing low-quality work at scale.