Anthropic Introduces Invisible Watermarks for AI-Generated Text That Survive Edits

Anthropic, the developer of the AI chatbot Claude, is introducing invisible watermarks for text generated by artificial intelligence.

Anthropic announced on the 11th (local time) that content generated by its AI models now includes machine-readable transparency markers.

This measure complies with the European Union's AI Act, which took effect on the 2nd.

Consequently, watermarks have been applied to the outputs of Claude models released after the law's implementation, and the feature is being sequentially added to previously released models.

The company explained that adding the watermarks does not compromise the meaning, quality, or readability of the responses.

Anthropic noted that because these watermarks are integrated directly into the text, they remain intact even if the text is copied and pasted elsewhere or partially edited.

However, the watermarks may not be detected if the text is too short, or if users extensively rewrite or translate sentences generated by AI.

Anthropic also clarified that the presence of a watermark does not necessarily mean Claude is the original author of the text.

This is because watermarks can also be applied when users have Claude proofread, translate, or summarize text originally written elsewhere.

Anthropic added that it also attaches provenance metadata to outputs generated as images or files, alongside text.

The company stated that it will apply these transparency markers consistently worldwide where Claude is serviced, not just in the EU region.

However, Anthropic has not yet released a tool to detect these watermarks.

Previously, Google also developed and implemented a digital watermarking technology called SynthID for its AI outputs.
※ Please note: This article was translated by AI and may contain errors.