Is there presently AI watermark built into ChatGPT that we will know about prior 2028?
Basic
6
Ṁ187
2028
13%
chance

By AI watermark, I mean intentional mechanism that allows to distinguish the ChatGPT texts from other human texts. (E.g. preferring atypical words of some sort). If someone find a statistical method to tell the difference, it would not count, the mechanism needs to be intentional.

Also, if it is added later, it would not count, it needs to be present in ChatGPT around 1/1/2024 already. However, if there is public information that the watermark is there and it remains unknown since when, it will be assumed it was there from the beginning. (I will wait cca 1 month since the news for the information about when it was implemented. If there is no such information, the question resolves YES.)

Get
Ṁ1,000
and
S3.00
Sort by:

Does a database of text hashes count as a watermark? As in, openAI has a secret dataset of "hashes* of generated text that allow them to give a probability score that a sample of text was produced by one of their models, without actually storing the original generated text?

@RemNi I would count is as a watermark if the intention is to use it for text identification, or if it is used for that purpose. (However, it would not work if anyone makes even a small edit of the text, unless they also hash sub-strings). If they have it for other technical reasons and they do not use it for text identification, I would not count it.

@Irigi I think there are many ways of making this sort of thing robust to edits

© Manifold Markets, Inc.Terms + Mana-only TermsPrivacyRules