Claude watermarks all processed text to comply with EU AI Act

Anthropic has announced plans to apply invisible, machine-readable watermarks to all content processed by Claude, including text that Claude has only edited rather than generated from scratch. The company made this disclosure in a support article explaining its response to the European Union's AI Act, which mandates that all AI system providers watermark AI-generated or manipulated audio, image, text, and video outputs. The law applies to any AI model released after August 2 and grants providers until December 2026 to update previously released models; Anthropic confirmed that all new models offered globally will carry these marks "from day one." Text outputs will include embedded watermarks invisible to users, while generated files in other formats will include digitally signed provenance metadata where the technology supports it. Anthropic is taking what the article describes as a "nuke it from orbit" approach by watermarking all processed content where supported, even though the EU AI Act does not require watermarks for cases where an AI system performs "an assistive function for standard editing" (such as grammar correction) or does not "substantially alter" the user's text or its meaning. The company explained that users commonly employ Claude for proofreading, translation, summarization, and file conversion, and all such outputs "can carry a Claude mark even if the underlying ideas, text, or data originated from another source." Technically, watermarks work by biasing the model's word choices in a pattern distributed across an entire document, detectable only in aggregate by the correct tool. However, this approach carries significant limitations. The watermark is trivially easy to defeat: text can be copied and pasted into another system, which may destroy the watermark if that system edits it, and metadata can be removed via standard editing tools. Critically, the watermark is not particularly informative: Anthropic acknowledged that a detected mark "provides a signal that content was processed by Claude, but is not fully conclusive" and that "the content may have been processed by Claude," while also noting that "the lack of a detected mark doesn't mean the content wasn't AI-generated or processed." Anthropic said it plans to eventually share details on how to detect watermarks in order to provide the technical support the EU law requires. The company also stated that watermarks won't work on "some platforms or features" that don't support them. When Ars Technica asked Anthropic about a timeline for detection tool release, potential false negatives or positives, or how its watermarks might conflict with editing exemptions in the AI Act, Anthropic provided a statement that did not directly address these questions. The broader EU regulatory context adds another layer of complexity: while the AI Act requires watermarking at the model level, Section 50(4) addresses how publishers must explicitly label content for public disclosure. Under this regime, wholly AI-generated text—such as an AI-written novel or marketing copy—does not require a label in most cases; labeling is mandatory only if the text is meant to "inform the public on matters of public interest" and has not been editorially reviewed by a known, accountable human. This creates a paradox: a model-level watermark on all content contrasts with a publisher-level rule that does not require labeling for wholly generated text if an editor has reviewed it. Violations of the EU AI Act carry steep penalties, including fines up to 15 million euros or 3 percent of a company's worldwide annual revenue. Anthropic stated that "other labs are taking similar steps" and that the company plans to continue working on watermarks and detection methods that meet the EU's demands.