Hacker News Reader

The question is whether LLM providers are going to do this. Anthropic currently says they don’t, but it’s impossible to verify:

> Watermarking carries no identifying information and can’t be traced to a specific person, organization, or chat;

https://www.anthropic.com/news/claude-text-watermark

View on HN

Probably just means they haven't (yet) figured out how to do the watermarking carrying enough bits of data to encode a GUID.

It could also mean they are lying. It could be they are being forced to downplay how good the watermarking is because gov agencies want that tool available to them without people knowing about it. We'll have to wait for the next Snowden to find out.

I still don't see how this is supposed to work for text content that I copy and paste somewhere else. And with models getting better and better to "human language", wouldn't the false positives also rise as "real text" is flagged to be AI generated?

I don't know either, and maybe it doesn't or can't work. But also maybe it does and can. I do not have any where near the expertise to know which is the case.

In other words, plausible deniability. “We aren’t watermarking your text.” (Today, and we’ll start tomorrow.)

At least, according to the very trustworthy Claude.

Reply on news.ycombinator.com