0

I read Is SE [going to be] selling our content for AI model training? And what exactly does "reinvest back into our communities" mean?. SE Inc. indicates that the will charge some firms if the firms train an AI model on the SE data dump, because:

The community is - you are - being denied your rightful attribution as it stands right now. Prashanth is saying that you should at least get to benefit from the financial impact. This is about protecting your interest in the content that you have created.

I think a company can very well attribute all authors (just point to the SE user list). Therefore, I wonder: if a company makes a commercial use of an AI model trained on SE data and attribute all authors, will SE Inc. not charge them?


To attribute all authors, 2 options:

  1. Link to a webpage listing the usernames of all Stack Exchange users who made at least one edit to Stack Exchange (and possibly adding hyperlinks to the corresponding Stack Exchange profile pages). Since typically all training samples contributed to the model weights, therefore to the LLM output, that would satisfy the BY terms of the CC-BY-SA license that SE uses.

  2. Some LLMs try to track which training samples contributed to the output: How can a language model keep track of the provenance of the main knowledge/sources used to generate a given output. Therefore, an LLM output can be given along with the list of Stack Exchange users who were the most influential (or the only influential) for the LLM output.

6
  • 4
    I'm not aware of an AI being able to point to specific pieces of training data in its answers. Commented Apr 22, 2023 at 2:42
  • 3
    @RobertLongson Just cite all authors and that's still in the CC-BY spirit, since typically all training samples contributed to the model weights, therefore to the LLM output. Note that some LLMs can track provenance to some extent: How can a language model keep track of the provenance of the main knowledge/sources used to generate a given output?. Commented Apr 22, 2023 at 2:44
  • Can you define "all authors"? Every user on Stack Overflow that has contributed a single byte to any post that is used by LLM? That would be, well, an impossibly large list. Would that go in a footnote? Commented Apr 22, 2023 at 13:13
  • 1
    And how would they keep it up to date? If it's just a literally generic "all the folks contributing to Stack Overflow contributed to this" then that's really no different than "everyone in the world contributed to this" - attribution isn't really attribution if it's just aimed generally and anonymously at a large population. Commented Apr 22, 2023 at 13:20
  • @Stuckat1337 "define all authors" -> added in the answer. "they keep it up to date" -> update the list when LLM is retrained on new SE posts. " - attribution isn't really attribution if it's just aimed generally and anonymously at a large population." -> it's not anonymous. Where did you read the list of contributors can't be large? Commented Apr 22, 2023 at 21:56
  • 2

1 Answer 1

5

Well - they'd be respecting the 'primary' licence that our contributions are under.

Assuming that they attribute the correct author for the referenced source - and I'd presume the CC licence the specific post is under, as per the licencing page.

Stack Exchange cannot 'deny' a specific use case use of the licence compliant to the CC-By-SA version the post/post revision is under.

There's various ways SE could charge a organisation building an AI model - either for access to the API for 'fresh' data or even more optimised data dumps, but nothing's stopping someone from grabbing a data dump and using it free in a licence compliant manner

4
  • Thanks, agreed. I want SE to state they agree to let company make a commercial use if they attribute all authors, since SE claims that the only issue they have is lack of attribution. But whatever SE says, the CC BY-SA license will anyway let companies train and commercially use models without getting robbed by SE IMHO. Commented Apr 22, 2023 at 22:19
  • 1
    I wouldn't say 'robbed' - I feel its more about SE' s obligations to its users. Companies can afford to pay :) Commented Apr 22, 2023 at 22:48
  • Company expenses will be paid by company users aka us. Commented Apr 22, 2023 at 22:49
  • Other companies using the data that is. SE isn't robbing them - merely getting compensated Commented Apr 22, 2023 at 22:50

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .