@Li @cwebber That's what I thought rightaway after I came up with this idea with overattribution. Obvious for me, and I wanted to see if it is the same for others. (It is not).

For the attribution context, one can always invent new word for it (like "statistically inspired" of whatever).

The bigger issue is for authors who may not even care about open source, but just want attribution to see usefulness of their work. Over-attribution is an insult for them.

@Li @cwebber Let's me give you exaggregated example - what if every slop code starts to attribute *every possible website on the internet*. Or even non-slop code using a code licensed with attribution does not attribute just to that code, but also to 1000 other codes (that have very little to do with this non-slop code). That feels not exactly right, although might technically fulfill the license. But maybe that's just me.

@Li @cwebber I conclude from your answer that referring to all sources, even not really used in the specific output is practically solving attribution (at least, from your perspective). I'd like to point out that may or may not be enough for other people who are looking for their work to be attributed.

So, no, I did not say "consider one issue, then solve that issue". What I proposed may or may not be a solution.

@Li @cwebber I did not imply that solving attribution automatically resolves every other issue. I am asking about specific thing - AI not propertly attributing its sources used for the output. That's a concrete problem. I asked if attributing to *all* sources is good.engough to solve this particular problem.

@Li @cwebber Yet, the question about attribution can be addressed independently. I am not sure what is dishonest about this.

@Li @cwebber I am not interested to discuss other problems with LLMs here. I am only looking to see about attribution specifically.

@Li @cwebber Sure. But I am asking about very specific issue - attribution.

@Li @cwebber Not all the training data is putting strain. There are training datasets that were crawled respecting robots.txt.

I am most interested in the scenario when there is a training dataset (crawled once without hitting servers repeatedly), and used for open weight model.

Open data+weight still does not address attribution. That's why my question.

@cwebber What if the attribution is then done to every piece of training data? That's not impossible if the training data is known, but will it be acceptable?