Stealth Tracking: How Anthropic Secretly Stamps Claude Text Outputs

Anthropic published a new updates page on Friday to address growing questions about how its Claude assistant marks generated text. The company broke down how these invisible markers work, whether light editing wipes them out, and how the changes impact generated code.

Claude users started debating the feature after Anthropic revealed earlier in the week that it would add watermarks to satisfy European Union rules. The EU Transparency Code requires technology companies to build detection systems that identify machine-generated text.

The announcement triggered immediate pushback across online communities. Reddit users labeled the tracking system a blow against user privacy, while dozens of subscribers on X reported canceling their paid accounts.

Anthropic explained that the watermarking system works by tweaking low-stakes choices during text generation. When choosing between similar words, like picking overcast instead of cloudy, Claude creates subtle statistical patterns across its output. Anyone holding the corresponding decryption key can spot these patterns, while standard human readers see normal phrasing.

Anthropic emphasized that watermarking does not lower response quality. To an everyday reader, a watermarked passage looks identical to unwatermarked writing.

The technical setup relies on the SynthID-Text framework, an approach originally detailed by Google DeepMind teams in 2024. Anthropic also plans to release an official detection application programming interface later this year. The company noted that statistical watermarking works differently than standard AI detection tools like Pangram, which look for common writing habits or repetitive phrasing.

Many creators wondered if light editing removes the embedded signature. Anthropic confirmed that minor line edits leave the hidden patterns intact. However, a complete rewrite that replaces every word wipes out the signature entirely, making the resulting text functionally human-written.

For documents where Claude simply proofreads or makes minor grammar edits, detection depends on how much text changed. If a human writes most of the draft and Claude only fixes minor typos, very few mathematical markers attach to the final passage.

Computer code carries much lighter watermarking than standard prose. Because programming languages require precise syntax, the model lacks the freedom to swap out words without breaking execution. Watermarks only appear where arbitrary choices exist, such as variable names or comment lines, leaving actual code execution untouched.

Anthropic noted that Claude is not the only assistant adopting these tracking methods. Other major AI labs signed the same European code of practice and will roll out similar hidden signatures across their platforms soon.