4 ms·

Looks to me like these "watermarks" are embedded in monetary numbers, acronyms etc. Maybe to stop them breaking into different lines?

by mhatma 1y ago

niel 1y ago
This is the most likely explanation. I mean, sure, these characters could be used to help estimate the likelihood text was generated (because human writers might be less likely to add proper non-breaking spaces), but I doubt these are watermarks.
rep_lodsb 1y ago
Yeah, it's probably not an intentional watermark, just something the model has been trained to do. Maybe some professionally written news articles already use them for the same purpose? Still hope HN adds a filter to block any comment with those characters in it :)
- johnisgood 1y ago
  It is very easy to filter those out from the output of GPT, though, using basic UNIX utilities. In fact, many methods don't survive reformatting or copy-pasting, not requiring filtering at all. It is a very basic watermark technique (text steganography) if it indeed is supposed to be one. A more advanced one would be a linguistic (grammar-based) one, but I am not going to give any more ideas. :D
  - rep_lodsb 1y ago
    It's easy to remove those characters, but that still requires being aware of them, and an intent to deceive. So many people just copy LLM output here because they (wrongly) believe it adds something of value to a discussion.
    - johnisgood 1y ago
      I do not think either that pasting output of LLM typically adds anything to the conversation. It might, usually it does not.
crooked-v 1y ago
U+202F in the screenshot, between "FY" and "2024", is the "Narrow No-Break Space ". Similarly, U+A0 is the "No-Break Space" (aka  ). It's not watermarks, it's just scraped typography.
helsinkiandrew 1y ago
The examples they give all look like valid uses of different Non-breaking spaces, with width hints for their use/location, this might be a little overzealous if written by a human but perhaps not for a machine. https://en.wikipedia.org/wiki/Non-breaking_space https://en.wikipedia.org/wiki/Non-breaking_space Different display apps may "display them identically" but others and typesetters/printing apps might not.

Looks to me like these "watermarks" are embedded in monetary numbers, acronyms etc. Maybe to stop them breaking into different lines?

by mhatma 1y ago

Looks to me like these "watermarks" are embedded in monetary numbers, acronyms etc. Maybe to stop them breaking into different lines?

niel 1y ago

This is the most likely explanation. I mean, sure, these characters could be used to help estimate the likelihood text was generated (because human writers might be less likely to add proper non-breaking spaces), but I doubt these are watermarks.

rep_lodsb 1y ago

Yeah, it's probably not an intentional watermark, just something the model has been trained to do. Maybe some professionally written news articles already use them for the same purpose? Still hope HN adds a filter to block any comment with those characters in it :)

johnisgood 1y ago
It is very easy to filter those out from the output of GPT, though, using basic UNIX utilities. In fact, many methods don't survive reformatting or copy-pasting, not requiring filtering at all. It is a very basic watermark technique (text steganography) if it indeed is supposed to be one. A more advanced one would be a linguistic (grammar-based) one, but I am not going to give any more ideas. :D
- rep_lodsb 1y ago
  It's easy to remove those characters, but that still requires being aware of them, and an intent to deceive. So many people just copy LLM output here because they (wrongly) believe it adds something of value to a discussion.
  - johnisgood 1y ago
    I do not think either that pasting output of LLM typically adds anything to the conversation. It might, usually it does not.

crooked-v 1y ago

U+202F in the screenshot, between "FY" and "2024", is the "Narrow No-Break Space ". Similarly, U+A0 is the "No-Break Space" (aka  ). It's not watermarks, it's just scraped typography.

helsinkiandrew 1y ago

The examples they give all look like valid uses of different Non-breaking spaces, with width hints for their use/location, this might be a little overzealous if written by a human but perhaps not for a machine. https://en.wikipedia.org/wiki/Non-breaking_space https://en.wikipedia.org/wiki/Non-breaking_space Different display apps may "display them identically" but others and typesetters/printing apps might not.