Anthropic has published a technical explainer on how its upcoming text watermarking system will work, a feature designed to help determine the likelihood that Claude was involved in writing a given text. The move aligns with the EU AI Act, which requires AI providers serving the European market to mark AI-generated content as of August 2. Several other major AI developers have signed the same Code of Practice and will implement their own watermarking methods.
The watermarking technique, based on Google DeepMind's SynthID-Text approach, exploits the inherent randomness in how large language models choose words. When generating text, Claude selects each word from a list of plausible candidates. In low-stakes choices where multiple words are equally acceptable, the model uses a random number to decide. Watermarking replaces this random source with a key-based sequence, creating a subtle pattern detectable only with the key.
Anthropic emphasizes that the watermark is invisible to readers and has no practical impact on output quality, creativity, or readability. Internal testing showed no degradation, and Google DeepMind's SynthID-Text paper reported no statistically significant differences in user ratings between watermarked and unwatermarked models. The watermark adds no hidden characters, requires no extra tokens, and does not increase cost or latency.
The watermark does not carry identifying information about users or organizations and cannot be traced to a specific person or chat. It only indicates the likelihood that Claude generated the text, not whether it was human-written or produced by another AI. Detection is less reliable on short passages or highly factual content where word choices are constrained, such as code or proofreading, but improves with longer, more creative text.