Anthropic Begins Watermarking Claude Text Outputs

Anthropic recently announced the quiet rollout of text watermarking for its Claude models to comply with the European Union Code of Practice. The mandate, signed by major Western AI laboratories, requires future models to implement these tracking features. While OpenAI intends to follow suit but is expected to miss the deadline, Google has already deployed similar watermarking technology across its models, including Gemini 3.7 Flash, since 2024. Google validated the approach through a massive test of 20 million users, which confirmed that the watermarking process caused no discernible difference in user feedback.

The underlying watermarking method, largely developed by Scott Aaronson and Hendrik Kirchner during Aaronson's tenure at OpenAI, leverages the non-deterministic nature of large language models. Instead of selecting tokens purely at random, the system uses a private source of pseudo-randomness tied to a secret cryptographic key. By analyzing the resulting text against this key, an application programming interface can determine if the content originated from the AI. Because the mathematical distribution of the token choices remains unchanged, the technique has a marginal cost of near zero and does not degrade the quality of the generated text.

For practitioners and developers, this development introduces a reliable way to detect AI-generated text without sacrificing performance. However, the rollout has sparked intense debate. Critics express concern over potential loss of anonymity, the precedent of regulatory overreach by Brussels, and the risk of false positives in collaborative workflows, such as when a human edits or translates text using Claude. Additionally, because the watermark is proportional to the number of original AI choices preserved, users can still strip the signature by heavily rewriting the text in their own words.