20

I have read the blog post about Community is the future of AI and I could not find any serious answer to this question.

Here is the way I see it: The knowledge in Stack Exchange (and Stack Overflow in particular) is built based on the assumption that, as a contributor, you spend your time and precious knowledge, mostly for the love of helping others, but also expecting at least to get the proper attribution and credit for your efforts, regardless of whether this has any materiel benefit for them or not.

Once this knowledge is used for AI training that will provide answers based on it, the link between the contribution and the credit for it is somewhat broken, which raises two main problems:

  1. While this form of using the contribution is arguably legal, I doubt whether it is moral to use knowledge that was meant to be used in a different way.
  2. In the lack of any attribution or credit, I suspect many people will be less inclined to spend the time and knowledge to make further contributions the way they do now

I'd be happy to hear from both users and SE personnel if they share these concerns and how they see it addressed.

Improve this question
7
  • 3
    not that necessarily answers your question, but have you seen Is SE [going to be] selling our content for AI model training? And what exactly does "reinvest back into our communities" mean? (and Philippe's answer post there?)
    – starball
    May 16 at 0:39
  • 1
    Re "mostly for the love of helping others": Yes, all future readers (potentially in the millions).
    – This_is_NOT_a_forum
    May 16 at 0:57
  • When you publish something publicly on the Internet, why do you expect to have a say in who reads it or what they do with the information? If I learn something from an answer on SE, I am not obligated to attribute that knowledge to the author every time I use it. An AI model is similar in retaining the "knowledge" without retaining the actual training text or image. I sympathize with people who are upset their work is training AI, but your content has been scraped and used long before ChatGPT got popular.
    – ColleenV
    Jun 15 at 13:03
  • 1
    @ColleenV, are we in agreement that when you publish a song, when you publish code or when you upload a clip to youtube. The fact that anyone can access it doesn't mean anyone can do whatever they want with it.
    – Gal Weiss
    Jun 16 at 14:21
  • @GalWeiss But I can listen to that song, or look at that code and incorporate the ideas into my own stuff. When you share your knowledge publicly, you have no control over who learns from it or how they use that knowledge. You can't copyright knowledge. There's a difference as well between derivative and transformative use. I can't remember where I learned how to handle exceptions in Python, but I use it all over my code. Do I really need to credit some source for it? I'm sure somewhere on SO there's a post with except Exception as e... Did I plagiarize it?
    – ColleenV
    Jun 16 at 14:54
  • You think you feel bad about AI training on your posts? What about all the creators that submitted stuff to Adobe's stock photo service, who not only made high quality artistic images available to Adobe under a global irrevocable license, but tagged the content so it was a good quality data set. Now Adobe can generate images that compete with the creators whose content they trained their AI on. It is absolutely unfair and awful, but if you don't want people extracting value from your work w/o compensating you, don't give it away for free.
    – ColleenV
    Jun 16 at 15:00
  • @ColleenV, I don't feel bad about my stuff, there's not enough of it to be sorry about. my point is that no one that published his work (free or paid, it doesn't matter) didn't expect it to be used in this manner. And that is (a) maybe illegal and (b) definitely unfair.
    – Gal Weiss
    Jun 16 at 20:59

3 Answers 3

Reset to default
12

From a legal perspective in the US, the lawsuit against Microsoft involving Copilot is what to watch.

I'm not sure what content creators or platforms that host content can do. Much of the obligations - both legal with respect to licenses as well as ethical with respect to crediting the authors of work - is on the creators of the models and the tools that use those models.

Historically, the company has directed contributors to information or to seek out their own legal advice when it comes to protecting their contributions. Unless there is an announcement by the company, I would expect that this holds true for companies building models based on content on Stack Exchange made available under the CC BY-SA license.

Improve this answer
1
  • 1
    Here is an update on the lawsuit concerning possible dismissal: "Taking the facts of the complaint as true and construing all inferences in the Plaintiffs’ favor, the Court can reasonably infer that, should Plaintiffs’ code be reproduced as output, it will be reproduced in a manner that violates the open-source licenses under which Plaintiffs published their code."
    – Karsten
    Jun 15 at 20:09
6

While this form of using the contribution is arguably legal, I doubt whether it is moral to use knowledge that was meant to be used in a different way.

I'm not sure whether the legality has been tested. We contribute with a CC BY-SA 4.0 license. So contributions have to be attributed when they are used in derivative work, and that derivative work has be be made available under the same licensing terms.

Does reading the entire internet to train a large language model mean that you have to attribute the entire internet (with a list of specific sources used)? I don't think there is an answer for this yet. When you ask ChatGPT about the sources it has been trained with, it claims it has no self-awareness of it, and that the exact sources used are a trade secret. I see a conflict with CC BY-SA 4.0 there, but I'm no intellectual property lawyer.

If someone starts a class-action lawsuit on this topic, I would consider being part of it.

Improve this answer
1
  • 6
    If it's technically not possible to add attribution to every text output, they still need to disclose sources they use for training somehow (at least in aggregation, i.e. how many pages they used from what domain). Trade secret is bullshit, if they do derivative works, they need to follow license.
    – MSDN.WhiteKnight
    Jun 15 at 14:07
-5

While this form of using the contribution is arguably legal, I doubt whether it is moral to use knowledge that was meant to be used in a different way.

One can argue that providing some AI systems help societies develop faster, hence it's moral to use SE as training data. E.g., if an AI system, such as a code generation system or a question-answering system, helped scientists develop cancer detection programs and or some medical treatment, that doesn't strike me as being overly immoral even if attribution is lacking due to the AI not being capable to provide attribution.

FYI: Is it illegal for a firm to train an AI model on a CC BY-SA 4.0 corpus and make a commercial use of it without distributing the model under CC BY-SA?

In the lack of any attribution or credit, I suspect many people will be less inclined to spend the time and knowledge to make further contributions the way they do now

AI models have been trained on web data without attribution for decades. I am not aware of any report indicating this caused a decrease in contributions.

How do we make sure that contributors get the credit and attribution they deserve when their work is used for training AI models?

Hopefully researchers can find out ways to develop AI that can better provide attribution.

Improve this answer
4
  • 1
    AI models have been trained on wikipedia and other sources for decades, but just a year ago, there was nothing even remotely close to chat GPT publicly available. So I do believe this is a new problem we are facing
    – Gal Weiss
    May 16 at 15:33
  • @GalWeiss"there was nothing even remotely close to chat GPT publicly available" true. People will have to get used to it until better attribution models show up.
    – Franck Dernoncourt
    May 16 at 16:14
  • @GalWeiss there was nothing even remotely close to chat GPT publicly available If "AI" didn't exist there was already real (human) intelligence. I guess the analogy with wikipedia was humans learning something somewhere, then writing it on wikipedia without attributing their sources!
    – ChrisW
    May 17 at 3:18
  • 4
    @ChrisW quite the opposite, any trustworthy value in wikipedia should contain references to relevant sources.
    – Gal Weiss
    May 17 at 7:46

You must log in to answer this question.

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