So, it seems the watermark is less what it sounds like (a stamp) and more an "imperceptible statistical pattern woven into the choice of words and sentence structures." Does that mean AI responses will sound even more "AI?" Like, will it become even easier to detect on a read-through because of the word choices and patterning? I see the word "imperceptible" there, but what does this mean in this context? My non-tech brain is kinda breaking here.

▲

xtiansimon a few seconds ago | parent | next [-]

> "imperceptible statistical pattern”

This is how I’m feeling right now.

A “watermark” is as an author’s mark. I struggle to understand how different texts will produce the same output. How big does a text have to be to generate this sign? I find ChatGPT to be overly loquacious, and my preference for Claude is the brevity of output. Does this mean I will now have to suffer Clause gibbering, too?

▲

josephg 24 minutes ago | parent | prev | next [-]

> Like, will it become even easier to detect on a read-through because of the word choices and patterning?

It's turned on right now. Can you tell a difference? I can't.

How many ways could I write this paragraph and still convey the same idea? Way more than we're aware of. Hundreds? Thousands? Maybe a lot more? The number of semantically similar variants increases exponentially with each word.

I suspect anthropic could turn their fingerprinting up or down if they want. If it were turned way up, claude would use weird phrasing but it would take very little text to tell if something were AI generated. If they turned it down, it would seem imperceptible to humans, but you would need a large sample to determine (with high accuracy) that a passage was AI generated. There's probably a very large middle ground where humans can't tell, and where it doesn't take a large text sample to know (with high probability) that some text was AI generated.

▲

floppydive 5 minutes ago | parent | prev | next [-]

I think its somewhat the opposite of easy to detect patterns in the watermarked text. Regular authors can be statistically fingerprinted and I think we do a rough version of this ourselves. An LLM with dense watermarking may sound less like one author we have a low opinion of and more like an encyclopedia set made by a mix of authors we have low opinions of.

▲

317070 32 minutes ago | parent | prev | next [-]

https://arxiv.org/html/2510.20075v6

It is quite counterintuitive, but you can hide texts the same size as the original text in imperceptible statistics of a text.

Compared to that feat, hiding a watermark is very easy.

▲

joenot443 10 minutes ago | parent [-]

The information being encoded (the watermark) is the _relative ranking of each token compared to other possibilities_. If our prompt was "Write a positive review for a restaurant" and the response began:

"The restaurant "

Our next set of predictions might be:

[was, had, offers]

So we append the rank/index of the next token (0, 1, or 2) onto the secret. Given a long enough response, that secret becomes unique enough to use as a watermark. This obviously relies on having full deterministic access to the LLM itself, i.e. I don't believe it will be possible for users to derive the fingerprint from text that they've generated, only Anthropic will be able to.

The immediate objection is that this runs the risk of degrading the quality of the response. I think that's totally valid and I'll be curious how Anthropic handles it.

That's my very rough understanding! If someone with more knowledge wants to expand, feel free.

	▲	ozgung 5 minutes ago \| parent [-]
		Do we need the original prompt to recover the watermark?

▲

iamacyborg 33 minutes ago | parent | prev | next [-]

Is it possible for them to sound even more AI?

▲

dgellow 30 minutes ago | parent [-]

Don’t forget the survivor bias is at play, you don’t see the ones that are good at passing for human written

	▲	iamacyborg 18 minutes ago \| parent [-]
		Sure, but if my experience (I know, I know) is anything to go by, it's increasingly difficult to get Claude to write in anything other than it's own house style.

▲

dgellow 32 minutes ago | parent | prev | next [-]

Unlikely, but I really hope so!

▲

hemkeshr 17 minutes ago | parent | prev [-]

[dead]