Claude's new Scarlet Letter watermark is invisible — for now

Anthropic to Add Machine-Readable Watermarks to Claude Outputs for EU AI Act Compliance

Anthropic has announced that it will apply machine-readable watermarks to content processed by its Claude models. The company stated in a support article that the measure is intended to meet requirements under the European Union’s AI Act. The regulation requires providers of AI systems to apply watermarks to AI-generated or AI-manipulated audio, image, text, and video outputs. The obligation applies to models released after August 2, with a compliance deadline of December 2026 for earlier models.

Anthropic said all new models released globally will include the watermarks from the time of launch. Text outputs will contain embedded marks that are not visible to users. Other generated files will include digitally signed provenance metadata where the format supports it. The company indicated that the watermarks will be applied to all processed content where technically feasible, including cases in which the model performs limited editing functions.

The EU AI Act exempts certain assistive uses, such as grammar correction, when the system does not substantially alter the user’s text or meaning. Because the watermark is applied at the model level, it cannot distinguish between full generation and minor edits such as the insertion of a comma. Anthropic has not yet released a detection tool, so the precise scope of watermark application remains to be verified through testing. The company stated that it intends to publish detection details to support the technical requirements of the EU regulation.

Anthropic noted that the watermarks will not function on some platforms or features that lack the necessary support. For non-text content, the company will use the C2PA metadata standard to record provenance information.

Technical characteristics and limitations of the watermark

The watermarking method described by Anthropic and required by the EU AI Act works by influencing token selection across a document in a statistical pattern. Detection requires specialized software that examines the full text. Anthropic stated that the marks are designed to remain after copying and pasting and to survive some forms of editing. The company also indicated that the marks may be removed if the text is processed by another system that does not preserve the statistical pattern.

For images and video, standard operations such as screenshotting, screen recording, or metadata stripping can remove the associated information. Once detection methods are made public, additional removal techniques can be developed. Anthropic described the watermark as providing a signal that content was processed by Claude, while noting that the signal is not conclusive. The mark may appear on material that was not generated by Claude, and the absence of a mark does not confirm that content was created without AI assistance.

Scope of application beyond explicit legal requirements

Anthropic acknowledged that its implementation may apply watermarks in situations the EU AI Act does not require. Users frequently employ Claude for proofreading, translation, summarization, or file conversion. In these cases, the output may receive a watermark even when the underlying text or data originated from a human source or another system. The company stated that the mark indicates only that content was processed by Claude.

Under Article 50(4) of the EU AI Act, providers of AI-generated text intended to inform the public on matters of public interest must disclose the use of AI unless a human editor has reviewed the material and is identified. In most other contexts, such as novels or marketing copy, no reader-facing label is required. The model-level watermark therefore operates independently of the disclosure rules that apply to publishers.

Detection timeline and outstanding questions

Ars Technica contacted Anthropic for information on the schedule for releasing detection specifications and any internal test results on false-positive or false-negative rates. The company was also asked how the watermarking system interacts with the editing exemptions in the AI Act. Anthropic did not provide a response by the time of publication.

Regulatory background and enforcement

The EU AI Act states that transparency measures are intended to support the integrity of the information ecosystem and to reduce risks associated with large-scale generation of synthetic content. The European Commission has indicated that individuals should be able to identify when they are interacting with AI systems or AI-generated material. The regulation places responsibility on AI providers to develop reliable, interoperable, and robust watermarking methods to the extent technically feasible.

Anthropic stated that it will continue to refine its watermarking and detection systems to align with these requirements. Noncompliance with the relevant provisions can result in administrative fines of up to 15 million euros or 3 percent of a company’s worldwide annual revenue, whichever is higher.

Ken Fisher, editor-in-chief of Ars Technica, contributed to this report.

Ashley Belanger is a senior policy reporter for Ars Technica who covers the social impacts of technology policy and emerging regulatory frameworks. She is based in Chicago and has worked as a journalist for two decades.