Can AI-generated text be identified even when there’s no visible label?
Anthropic says its Claude AI watermarking technology is designed to do exactly that.
The system creates a hidden statistical pattern through some of the word choices Claude makes while generating text. Readers shouldn't notice anything different, but compatible detection technology can analyze the content for the watermark.
Anthropic is using SynthID-Text, a watermarking technology developed by Google DeepMind, and also plans to introduce a detection API.
But there are important limitations.
Minor edits may leave some of the watermark signal intact, while substantially rewriting the content can weaken or remove it.
Coding presents another challenge because programming languages offer fewer flexible word choices than normal writing.
So, could invisible watermarking become a standard way to improve transparency around AI-generated content?
Read the full story on WikiGlitz
Top comments (0)