15

When the company announced the partnerships with Google and OpenAI they made a big deal out of their commitment to "socially responsible AI" (see this blog post for details). A big part of this seems to be the attribution requirements of the Creative Commons license.

Now, the OpenAI agreement is just days old and the Google agreement was a few months ago, so I don't necessarily expect some immediate impact here. But as far as I can tell Google Gemini isn't citing any sources yet.

I have a bit of trouble believing that SE has the power to get huge companies with many billions on the line to fundamentally change the way their hottest product right now would work. There's also some technical arguments here that LLMs fundamentally cannot provide attribution. I don't feel entirely competent to judge those arguments, and the way AI tools work might change in the future so while I wouldn't rule out that attribution is possible, I think it is at least very, very hard and probably impossible for at least some use cases.

So what actual, tangible effects will the "socially responsible AI" policy have on AI tools of the cooperation partners? What does SE here expect, and does this policy actually mean that ChatGPT and Google Gemini are expected to attribute content from SE sites in the reasonably near future?

2
  • 4
    I expect a lot of people want the answer to this question more than most other questions, but it depends how they implement it and what constraints are imposed on them by their contract-partners (and if they pay attention to the community's wishes). They've not told us about that yet.
    – W.O.
    Commented May 9 at 10:34
  • A related answer from a while back: meta.stackexchange.com/a/388569/273494 genAI is perfectly capable of providing plausible looking citations. That doesn't mean they are relevant. If you look at the citations in Jon's answer, they have multiple problems. The underlying problem is that we have no way of knowing if the citation/attribution is correct because we have no visibility into how it was determined to be relevant.
    – ColleenV
    Commented May 9 at 12:42

1 Answer 1

8

Gemini is, and has been, citing sources.

Here is an example of Gemini citing sources, specifically a Stack Overflow question:

Gemini Citing Sources

And here is an example of Google's AI Search / SGE citing sources, specifically a Software Engineering Stack Exchange question:

Google SGE Gemini Citing Sources

It's important to distinguish between "citation" and "attribution". In accordance with the CC license, attribution is required when resharing or adapting (remixing, transforming, etc.) the material.

Citations allow for ideas to be traced to their sources for the purpose of honesty and ethical behavior. Since I do not believe that Gemini's output can be considered resharing or adapting, citation is appropriate and Gemini appears to be creating an acceptable citation for the material. In a formal paper, someone may want to use a formal citation format, but that isn't necessarily a requirement to cite sources.

However, if we assume that attribution is required, we can ask if what is provided meets the requirements of the Creative Commons license. As much as I personally don't like it, I believe it does. According to the CC BY-SA 4.0 license, attribution may be provided in "any reasonable manner based on the medium, means, and context". It even gives the example that "it may be reasonable to satisfy the conditions by providing a URI or hyperlink to a resource that includes the required information". By giving a hyperlink to a Stack Exchange question that contains supporting information and the SE question page containing all of the elements of attribution, the CC requirements appear to be satisfied.

Currently, it doesn't look like OpenAI's ChatGPT, at least the freely available one, has similar functionality. However, if SE concurs that what Google is doing with Gemini and SGE meets the requirements of the CC license, I would expect that similar functionality implemented by OpenAI would also be seen as meeting the requirement.

There is also a question about attribution regarding training material. In the United States, there are ongoing court cases and pending legislation that may put requirements on identifying material used for training models and limiting what licenses for content can be used to train models. It may take many months, or even years, for these concerns to work their way through the courts and legislatures for final decisions.

I don't think that SE's actions will significantly impact the work of any partners. Court cases and legislation, both in the US and around the world, will have a greater impact. Unless SE starts to advocate for changes to global intellectual property laws and the CC licenses, it seems like the current status quo is acceptable to them.

32
  • 5
    unless the relevant parts have been cut-off from your picture, in my opinion that is NOT an example of "Gemini providing attribution". Gemini did two things: first, it used Google Search to provide a link to a only tangentially related SO question (notice that it clearly stated "Google search found", NOT "the following Gemini answer was based on"). Second, it used Gemini LLM to create a statistical word chain based on the question. And if you notice, the content it generated is not found in the SO question it linked as similar in the first place.
    – SPArcheon
    Commented May 9 at 11:49
  • 2
    @SPArcheon You are confusing "citation" and "attribution". Citation is an ethical concern to trace a potential source of an idea. Attribution is a legal concern regarding identifying content used, adapted, remixed, transformed, etc. The examples clearly show Gemini citing sources. I don't see any content sufficiently identical to warrant attribution, so citation is the proper thing to do here. However, even if the content was used in a manner that required attribution, what Google is doing does meet the requirements of the CC Attribution clauses. Commented May 9 at 11:58
  • 3
    that even if you want to differentiate between citation and attribution (I don't think it really matters on this topic, but you are correct in stating them being sightly different concepts) what Gemini did there by linking an unrelated question that doesn't anywhere express the concept found in the LLM generated answer is neither Citation nor Attribution. To be clear, Gemini answer does not look like it was based on the "cited" content.
    – SPArcheon
    Commented May 9 at 12:09
  • 1
    @SPArcheon You're referring to the technical correctness of the citation. I looked at both examples again, and I don't see issues. The Gemini example that cites the SO question does have an answer that claims that one return comes from structured programming. The SGE example points to a Software Engineering question that has several answers that refer to readability, understandability, and control flow. I don't agree that the generated content doesn't include concepts expressed in the linked sources. Commented May 9 at 12:29
  • 3
    That AI is making up a somewhat plausible citation is actually worse than none at all. I am highly skeptical that the search results it provided have anything to do with how the algorithm arrived at its answer. Some of them look like AI generated SEO content, but since you only provided a screenshot I can barely read, it's difficult to tell.
    – ColleenV
    Commented May 9 at 12:34
  • 2
    @SPArcheon I was always taught to cite what you read. If you read the Agile Manifesto, you cite the Agile Manifesto. If you read my answer on SE or PM that summarizes and interprets the Agile Manifesto and didn't read the original content, then you cite my answer. And if you read both, then you cite both. Of course, this gets fuzzy when you talk about a system that is capable of "reading" and "remembering" everything, what do you need to cite? I've read lots of things that I'm sure influenced my thinking, but I can't cite every single book or post I've ever read on a given topic. Commented May 9 at 12:43
  • 1
    I can search Google and give you "citations" that are relevant. That doesn't make them accurate and doesn't help find where the idea came from. It's just a chain of hearsay about where the idea came from that misleads people into believing they know a fact by the way it's presented. This is how we get crazy folk etymologies about the origin of idioms. It is the opposite of socially responsible.
    – ColleenV
    Commented May 9 at 12:47
  • 2
    You assume because they're relevant and confirm what you believe to be true about where that idea came from that they're correct. Why would you even need a citation in that case? Where did the person that wrote the SE post get the idea? These aren't citations. They're search results.
    – ColleenV
    Commented May 9 at 12:55
  • 1
    How do search results match that definition of citation? None of those things cited are the origin of the concept. The AI is not citing its sources. It's giving you search results. It's not even putting them in context. It's lying because it doesn't know how it arrived at the information it provided. It just knows the information was somewhere on the internet when it scraped information for its model.
    – ColleenV
    Commented May 9 at 13:00
  • 1
    How do you know it's not lying? You believe because an SE post was returned that it's properly attributed. What if it's not and that post had nothing to do with the information it generated? There's no way to prove that citation is accurate. That it's relevant is not proof. We don't know what data is in the model or how it's tagged.
    – ColleenV
    Commented May 9 at 13:06
  • 2
    My father wrote his PhD years ago. For every note he took from the literature he wrote up a card that told where it came from. That is a "citation". Taking some text and re-running your stochastic search on your corpus with a prompt to "produce something that looks like a citation" is not a citation. This answer is fundamentally wrong.
    – JonathanZ
    Commented May 9 at 14:28
  • 1
    Google's search results tags every result with the link it was found at. (Or at least it did for the first twenty years.) That again is a citation. You cannot compare rerunning your stochastic next-word selector to produce something that looks like a citation with an algorithm that actually tracks where it got something from. There is no rule that would prevent someone from creating a LLM that tracked this, but current ones most definitely do not, it's hard to see how to tweak them to do so, and we should not describe what they produce as "citations" until they do.
    – JonathanZ
    Commented May 9 at 14:42
  • 1
    @JonathanZ Except Google's SGE actually has functionality that tracks what content was used to produce the summary. Gemini and SGE are not just "stochastic next-word selectors". That is only a fraction of what they do. When you do a search, SGE receives the top search results and uses the LLM to summarize content and then structure it as a nice, human-readable summary. Again, you could argue about how the citations are presented, but you cannot say that what it doing is not a citation. Commented May 9 at 15:02
  • 1
    @JonathanZ Again, you are incorrect. Gemini links generated statements - often one or a few sentences - to specific sources that support the ideas summarized in the sentence(s). SGE does something similar. How it picks is irrelevant. Why did I read and cite Bob's book on the topic instead of reading and citing Alice's book on the same subject? It doesn't matter. And I'm not claiming the LLMs produce citations - they don't and can't without additional stuff around them. Gemini and SGE can and do produce valid citations for content because they have more stuff than just the underlying LLM. Commented May 9 at 15:43
  • 2
    The pushback on this is also interesting to me; I've been using the Search Experiments AI tool for a bit now, and I have yet to find an attributed response that doesn't reasonably summarize the cited source... personally, I just haven't seen any reason to doubt the "citations" (whether in the proper sense or not is besides the point) provided, because they've always matched up with the source when I click through. Now, sometimes that source is just bad or contains wrong info, but that's not an AI summary problem. It doesn't seem to be generating that data, it seems to be summarizing.
    – zcoop98
    Commented May 9 at 16:10

You must log in to answer this question.

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