Anthropic Embeds Machine-Readable Watermarks in All Claude Outputs

From August 2, 2026, Anthropic began embedding machine-readable watermarks in all Claude-generated outputs — a move that aligns with EU AI Act Article 50(2)'s requirement that generative AI providers mark outputs in a detectable, machine-readable format. The watermarks apply to any Claude-powered tool, including the growing ecosystem of legal tech products built on Claude as a foundation model. For legal tech operators deploying Claude-based solutions, the watermark infrastructure solves one Article 50(2) compliance obligation — but also creates a disclosure audit trail that will surface in any regulatory inspection. Operators who have rebranded or customized Claude-based solutions should review whether they are now classified as providers rather than deployers under the EU AI Act's substantial-modification test, a reclassification that carries substantially higher compliance obligations.