Anthropic explains exactly how it will identify AI-generated text
Anthropic has detailed its implementation of watermarks for Claude-generated texts. Using Google's SynthID technology, the system embeds statistical patterns into token sequences to identify AI content while maintaining text quality.
Over the weekend, Anthropic published a blog post detailing its decision to integrate "watermarks" into texts generated by the Claude bot. While watermarks have previously been used primarily in image generation models, the company is now applying them to text to comply with the European AI Act's transparency requirements.
Google Technology
Anthropic explains that the model does not replace words randomly; instead, it intervenes during the selection of the next token based on probabilities. Using a "pseudo-random key," the system creates a subtle pattern in its responses. This pattern is imperceptible to human readers but detectable to those holding the decoding key.
In "low-risk" scenarios—where word choices are predictable, such as "The weather in London is..."—the entropy is low, allowing the model to embed a marker. In high-entropy situations, where many completion options exist, the system can hide the watermark without affecting syntactic logic.
The company utilizes the SynthID technology developed by Google's DeepMind team in 2024. The system performs a mathematical calculation on the sequence of tokens. If the cumulative score crosses a predefined statistical threshold, the system identifies the text as AI-generated. Accuracy increases with the length of the text.
Can you remove the mark?
Light editing will likely not remove the watermark. A full rewrite, replacing nearly every word, might succeed, but Anthropic questions whether such a text would still be considered an AI product.
If Claude is used only for proofreading, the outcome depends on the extent of the editing. Minor changes, such as adding punctuation, leave little room for the system to embed the mark, meaning the text will not be identified as AI-generated. However, heavy editing will trigger an "AI-written" tag.
Regarding code, Anthropic emphasizes that it will contain fewer watermarks compared to regular text. The model is constrained by fixed syntax, limiting its freedom to choose between alternatives. Markers may appear in documentation, comments, or variable names. Finally, Anthropic notes that other AI labs are expected to implement similar watermarking standards soon.
