Anthropic has announced it will begin embedding invisible watermarks into text generated by Claude, its AI chatbot, as part of preparations to comply with transparency requirements under European Union artificial intelligence legislation. Unlike watermarks visible on images, readers will see no marking on the text itself.
The company will deploy SynthID, a technology developed by Google DeepMind. During text generation, the system subtly influences the AI model's word choices: when multiple suitable options exist for continuing a sentence, it causes Claude to select words according to a specific statistical pattern.
For instance, if the system must choose between "gray weather" and "overcast," it will consistently favor one option over the other. To readers, the result should appear entirely natural, but detection tools familiar with the pattern can analyze the word sequence and determine with high probability whether the text was produced by the model.
Anthropic emphasizes that the watermarking should not compromise the quality of responses or alter their meaning. Longer texts are easier to identify, while shorter passages yield lower certainty levels. Watermarking in code is also expected to be weaker, since the model has less freedom in word selection given the syntactic requirements of valid programming.
The company acknowledges that the watermark can be compromised through substantial rewriting of the text. Minor edits or replacing a few words may not suffice, but extensive rewriting could remove the pattern. In cases where Claude is used only for proofreading or light corrections of human-written text, the likelihood of the text being flagged as AI-generated is expected to be lower.
Anthropic notes that this move is not unique to the company, and that other AI firms are expected to adopt similar solutions amid growing demands to enable identification of artificially generated content.
