Anthropic has announced it will apply machine-readable watermarks to content processed by any of its Claude models — including text that was only edited, not generated — in a rollout that goes further than the European Union's AI Act requires. The company confirmed that all new models offered globally will mark AI-generated content from the time of release, with previously released models given until December 2026 to comply under the law's grace period.
Anthropic to watermark all Claude-processed content globally, going beyond EU AI Act requirements
Anthropic has announced it will apply machine-readable watermarks to content processed by any of its Claude models — including text that was only edited, not generated — in a rollout that goes further than the European Union's AI Act requires. The company confirmed that all new models offered globally will mark AI-generated content from the time of release, with previously released models given until December 2026 to comply under the law's grace period.
The EU AI Act, which applies to AI models released after August 2, requires providers to watermark AI-generated or manipulated audio, image, text, and video outputs. The law explicitly exempts cases where an AI system performs 'an assistive function for standard editing' — such as grammar correction — or where it does not 'substantially alter' the user's text or its meaning.
Anthropic stated in a support article that text outputs will carry embedded watermarks invisible to the user, and that other generated files will include digitally signed provenance metadata where supported. For non-text content, the company said it will use the C2PA metadata standard to record provenance.
Because the watermarks are applied at the model level, the system cannot distinguish between wholesale text generation and minor edits such as a comma correction. Anthropic acknowledged that 'people often use Claude to proofread, translate, summarize, or convert files,' and that 'the output can carry a Claude mark even if the underlying ideas, text, or data originated from another source.'
Anthropic noted that the watermarks will not function on 'some platforms or features' that do not support them. The company said it plans to eventually share details on how to detect the marks, as required by EU law, but had not released a detection tool as of publication. Ars Technica said it asked Anthropic for a timeline on detection details and testing results, and that Anthropic did not immediately respond.
Text watermarks work by biasing a model's word choices across a document in a pattern detectable only in aggregate by a purpose-built tool. Anthropic noted that the marks 'will travel with the text when it's copied and pasted elsewhere, and may persist through some editing,' but that pasting watermarked text into another system that edits it could destroy the mark. For image and video content, screenshotting, screen recording, or metadata editing tools can remove the information.
A separate provision of the AI Act, Article 50(4), addresses public disclosure obligations differently: wholly AI-generated text intended to inform the public on matters of public interest must carry a reader-facing label unless a known, accountable human editor reviews it, while AI-generated novels or marketing copy carry no such requirement under that section.
- The source does not state a specific date on which Anthropic's watermarking rollout will begin.
- The source does not state whether Anthropic has conducted any testing to assess the rate of false positives or false negatives in its watermark detection.
- The source does not establish what detection tool, if any, Anthropic plans to release, or when details on detection methods will be published.
- The source does not state Anthropic's annual worldwide revenue, which would determine the maximum fine under the AI Act's 3 percent penalty provision.
- The source does not explain which specific platforms or features will not support the watermarks.
- The source does not state whether any other major AI providers have announced comparable watermarking approaches under the same EU AI Act requirements.
Read the original at Ars Technica