Anthropic adds machine-detectable invisible watermark to Claude text
On Aug 25 Anthropic added machine-detectable invisible watermark to Claude text outputs to combat AI content proliferation; researchers question whether it can be bypassed.
On August 25, reports emerged that Anthropic has added machine-detectable invisible watermarks to Claude text outputs, aimed at reducing AI-generated spam and abuse.
The watermark is embedded in text in a way that does not affect user reading experience, and external detection tools can identify Claude-generated content via specific algorithms. Anthropic has not disclosed the specific watermark algorithm but confirms its detection accuracy is significantly higher than existing solutions.
Academic researchers question whether the watermark can be rewritten or bypassed, and argue it cannot single-handedly solve the proliferation of AI-generated content. OpenAI and Google have previously added similar watermarks to their model outputs, but effectiveness drops significantly after rewriting or translation.