Sunday 16 August 2026London --:--Frankfurt --:--Zurich --:--
Business

Anthropic to watermark 100 % of Claude text outputs globally

Anthropic will embed invisible, machine‑readable watermarks in every piece of text processed by Claude models worldwide, a move that goes beyond the EU AI Act’s labeling rules.

By
Anthropic to watermark 100 % of Claude text outputs globally

100 % of Claude text outputs will carry an invisible, machine‑readable watermark, Anthropic announced on Tuesday.

Why the rollout matters

The EU AI Act, which applies to AI systems released after 2 August 2026, obliges providers to label AI‑generated or manipulated text, audio, image and video. Anthropic says the watermark will satisfy that requirement and will be embedded in all new Claude models offered worldwide from day one.

Anthropic’s “nuke‑it‑from‑orbit” approach

“We are applying watermarks to all processed content where technically feasible, even in assist‑assistive editing scenarios that the law does not require,” the company wrote in a support article.

The company adds that the marks are invisible to users and can only be detected with specialised tools. They will not function on platforms that lack support for the provenance metadata.

How Anthropic’s policy differs from the law

EU AI Act requirement vs Anthropic’s implementation (source: Ars Technica, 13 Aug 2026)
Aspect EU AI Act Anthropic
Scope of labeling Only AI‑generated or manipulated text All processed text, including assist‑assistive edits
Geographic coverage Models released in the EU after 2 Aug 2026 All new Claude models worldwide
Implementation deadline Compliance by Dec 2026 (grace period) Immediate on launch of each new model

Impact for users and regulators

For European businesses that integrate Claude, the watermark will appear in every piece of text the service touches, regardless of whether the AI merely refines existing content. Regulators will be able to detect the mark with dedicated tools, but the signal is not conclusive proof of AI involvement. Platforms that cannot support the metadata will deliver unmarked outputs.

Anthropic’s pre‑emptive stance may set a de‑facto standard for AI provenance, prompting other providers to adopt similar blanket watermarking ahead of the EU deadline.

What’s next

Anthropic plans to roll the feature out as new Claude models are released. The company has not disclosed a timeline for retrofitting existing models, and it remains to be seen whether European watchdogs will require broader adoption across the sector.

More from Business