Anthropic will watermark text and files processed by Claude

In short: Anthropic says it will start adding invisible watermarks to content that Claude generates and also content that Claude edits.

What happened

Anthropic published a support article explaining that it is rolling out “machine readable” watermarks for Claude outputs. A machine readable watermark is a hidden signal that software can check, like an invisible stamp on a document that a special scanner can spot.

Anthropic says the change is meant to help it comply with the European Union’s AI Act. This law requires AI providers to watermark or label AI generated or AI manipulated audio, images, text, and video. The law applies to new AI models released after August 2, and it gives companies until December 2026 to update older models.

Anthropic says all new Claude models will include these marks globally, not only in the EU. For text, Claude will embed an invisible watermark. For other file types, Claude will add “provenance metadata,” which is extra information attached to a file that can describe where it came from, like a luggage tag on a suitcase.

A key detail is that Anthropic plans to mark content that Claude “processed,” not just content it fully wrote. That could include proofreading, translation, summarizing, and formatting. Anthropic also says the mark is not fully conclusive, it only suggests the content may have been processed by Claude, and the lack of a mark does not prove Claude was not involved.

Why it matters

These watermarks could affect everyday situations where people use Claude for light editing. A teacher, employer, or platform might see a detection result and assume something was fully AI written, even if a person wrote it and only used Claude for grammar fixes. Anthropic has not yet released a public detection tool, and it also notes the marks may not work on every platform or feature.

Source: Arstechnica