Do creators agree with CC-BY being satisfied if their contribution is attributed at the model level and not at the output level?

[drawnto]

According to the current state of "Using CC-Licensed Works for AI Training" putting a link inside the dataset is sufficient for attribution.

All CC licenses require attribution to the creator(s) of the licensed material. For AI model training, attribution could be a simple link to the source of the dataset used to train the model. Where retrieval-augmented generation (RAG) or other methods are available, providing attribution to the CC-licensed work tied to the particular model output with a link to the source is ideal.

I find this highly surprising and will not use CC-BY for works i want attribution for if that is the recommended interpretation. I am against both biological and silicon chauvinism. I find this idea to lead to a double standard. The biological equivalent would be, if some flesh brained person publishes their browsing history somewhere on their website they are allowed to republish any CC-BY without link to the original or any attribution.

In my understanding if an AI is in any form creating derivatives or reproductions of a CC-BY licensed work it MUST reproduce the correct CC-BY attribution, just like it is with humans. If current AI model approaches can not guarantee that property mathematically they MUST NOT train on CC-BY information.

What do others think about this?

Are you an user of CC-BY (feel free to link some creations) and do you feel happy with a link to a dataset if the AI is trained on your CC-BY creations as attribution?

[greg-kennedy]

I have the same reservation. The dataset is not the work - it merely contains the works, which need to be individually attributed, else it is a license violation.

[wumpus]

When I was a kid, I went to the library and read thousands of books. When I later wrote my first research paper, I credited particular books for particular things I wanted to footnote, and did not give credit to the thousands of books of background reading. For example, I would mention the year that Poland was invaded at the start of World War II, without footnoting it. If I recalled something but thought it needed a footnote, I'd go find a book with that information to cite, even if it wasn't the same book I originally read.

In most LLMs, the pre-training data is "a bunch of Common Crawl", and if you want to footnote a particular thing, you can quiz the LLM for a concrete source.

Seems like CC-BY would work the same in both situations.

BTW the dataset world is working towards full lists of sources, via nesting Croissant metadata files.

[greg-kennedy]

I don't believe you can "quiz the LLM for a concrete source" actually. It is likely to misattribute or invent a source to please the user than actually preserve the lineage of sources through the transformation into model weights.

…

On Fri, Jun 27, 2025 at 3:21 PM Greg Lindahl ***@***.***> wrote: When I was a kid, I went to the library and read thousands of books. When I later wrote my first research paper, I credited particular books for particular things I wanted to footnote, and did not give credit to the thousands of books of background reading. For example, I would mention the year that Poland was invaded at the start of World War II, without footnoting it. If I recalled something but thought it needed a footnote, I'd go find a book with that information to cite, even if it wasn't the same book I originally read. In most LLMs, the pre-training data is "a bunch of Common Crawl", and if you want to footnote a particular thing, you can quiz the LLM for a concrete source. Seems like CC-BY would work the same in both situations. BTW the dataset world is working towards full lists of sources, via nesting Croissant metadata files. — Reply to this email directly, view it on GitHub <#41 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABFYTLXV4CYFG3OEX36G6NT3FWRT3AVCNFSM6AAAAACAGVOQBCVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTGNRQGEYTMMY> . You are receiving this because you commented.Message ID: ***@***.*** com>

[wumpus]

I do it frequently.

By examining the source, you can discover if it was a misattribution or invention.

That's the same way you should evaluate any claimed source, even those supplied by humans.

[greg-kennedy]

That is not how it works. You are anthropomorphizing the computer by applying human learning concepts to machine inference, which do not work in the same way. Model training destroys the relationship between source and claim. You can certainly prompt it afterwards for a source, and it will generate one, and then you can go find the claim within the source if it exists. But there is no connection between the source it gives you and the claim it made 1 sentence ago.

[greg-kennedy]

Anyway, you are a Common Crawl employee with a vested interest in mass scraping, so I won't be continuing this conversation with you any further.

[wumpus]

This is me as a high school student:

For example, I would mention the year that Poland was invaded at the start of World War II, without footnoting it. If I recalled something but thought it needed a footnote, I'd go find a book with that information to cite, even if it wasn't the same book I originally read.

[mirabilos]

We most certainly are not. See also: https://evolvis.org/~tg/cc.htm

[CC-signals]

We’re curious what people think! It is not generally possible to trace an output back to specific inputs in a foundation model’s training data; at the same time, attribution may be more possible for certain other uses of data (e.g. Retrieval Augmented Generation). Even in that instance, should such attribution be required? Can it be sufficient to simply provide attribution at the model level? Or should each output of the model also incorporate a form of attribution?

[drawnto]

It is not generally possible to trace an output back to specific inputs in a foundation model’s training data;

With current architectures, yes ot at least it is not done. This is a constraint which would be imposed on architectures for public benefit or to be able to incorporate CC-by training data without it being jucky by society. However it is often possible to have foundation model´s produce 1-to-1 memorized copies of copyrighted work, when it comes to code even code including the license headers.

Attribution at model level is close to useless. If models could consistently tell which source they are pulling from this would serve as advertising for those sources it pulls information from. If models could tell how close the produced content is to existing material that would allow people working with AI to actually appreciate their "model`s inspiration" rather then remaining ignorant towards it. I think this attribution would be net beneficial to society. That this isn't done is entirely explained by the motivation and incentives of companies training AI models.

Each output should approximate (as close as possible) the attributions it is pulled from. This could be done well for LLM in particular whose architectures includes essentially a differentiable key value look up. Essentially this essentially approximates a vector-database when the attribution information is available during training this could also be embedded there and one could sparsity the updates during training to limit mixing off attributions.

[stkdiretto]

I believe that the proposed justification for crediting a source when using RAG is lacking direction. I currently see two justifications of crediting a source in this case:

1. There is a moral right, based on the creative work of the author of the original source. This line of thinking is very closely tied to droit d'auteur.
2. A source is explicitly cited in order to make a conclusion reproducible and verifyable (if one bothers to actually check the original source)

Those two are very different animals. I see the former in stark contrast to your aim to not expand copyright into realms where it ought not belong. And the latter is not grounded on an individual's wishes to be cited, but on a principle of justified, verifiable knowledge, regardless of the mechanism and context.

Attribution on a foundational model level is an… interesting question which leads to the question I asked in #29 : How would you practically do that? Attach 3000 pages of credits? To whose benefits? If not: Why propose the Credit Signal anyway?

[mirabilos]

On Wed, 2 Jul 2025, Creative Commons wrote: It is not generally possible to trace an output back to specific inputs in a foundation model’s training data

It sufficiently is, people have proven that over and over again, and the “AI” operators merely put restrictions over the API to the model so people don’t exploit it as they get aware of the ways to do that, but raw access to the model generally means that, yes, you can sufficiently extract “training data”.

[cadadr]

I apply a licence with the intention that it's a document that describes the interaction between me and the person and/or institution that handles the licenced (a better term escapes me but..) product. An "AI" is a software package, an inanimate object lacking any personhood.

The reuse of the work happens during the training process, and is something that the creator of the model does. That's one place where attribution must happen. The other place is when such model is used and it regurgitates something for its user. Then in effect the said user is reusing the work, and they must correctly attribute the source. The LLM is but a means of retrieval, (unfaithful/malicious) paraphrasing, and synthesis. It cannot do attribution, as "reusing" or "creation" are not things it may do, unless we're to believe a text editor is reusing my past work to creatively produce a perfect facsimile of it at the moment of use.

Anthropomorphising software and computers will cause nothing but confusion and misunderstandings, at best.

Anyways, in that light, if the model output excludes attribution, then it's basically either causing its user to engage in plagiarism, if they pass the output as their creation, or else it's causing the creator/operator of the model and its user interface to engage in plagiarism, as those entities are passing the work off as theirs.

Attribution is not a formality, it's a means of verification, of tracing the sources of information, and of telling what's the original contribution of a work is. Plagiarism does not merely affect plagiarised authors/creators, but it is a threat to society as it allows for pseudo-experts to build an affect of authority where they merit none on the basis of their knowledge and skills.