Or, maybe just require an open disclosure of where the bot pulled from and how much, instead of having it be a black box? "Text in this response derived from: 17% Wikipedia article 'Example', 12% Wikipedia article 'SomeOtherThing', 10%...".Current (ie. ChatGPT) systems doesn't work that way, as the source of information is lost in the process when the information is encoded into the model. The model is just a network of probabilities, and it is highly compressed compared to the original data. We are missing the point if we believe it is a copy of source data and not a tool to interact with information using natural languages.
Soon, tools can retrieve data from external sources and write answers based on them[1]. For example, in the Wikipedia context, this would be to use a search engine to find information automatically, summarize findings, and generate references for the results. Or vice versa, retrieve information from Wikipedia or Wikidata. Then we will get source data, too, but the LLM model's internal reasoning will still be fuzzy.Br,-- Kimmo Virtanen_______________________________________________On Sun, Mar 19, 2023 at 8:24 AM Todd Allen <toddmallen@gmail.com> wrote:Or, maybe just require an open disclosure of where the bot pulled from and how much, instead of having it be a black box? "Text in this response derived from: 17% Wikipedia article 'Example', 12% Wikipedia article 'SomeOtherThing', 10%..."._______________________________________________On Sat, Mar 18, 2023 at 10:17 PM Steven Walling <steven.walling@gmail.com> wrote:_______________________________________________On Sat, Mar 18, 2023 at 3:49 PM Erik Moeller <eloquence@gmail.com> wrote:On Fri, Mar 17, 2023 at 7:05 PM Steven Walling <steven.walling@gmail.com> wrote:
> IANAL of course, but to me this implies that responsibility for the *egregious* lack
> of attribution in models that rely substantially on Wikipedia is violating the Attribution
> requirements of CC licenses.
Morally, I agree that companies like OpenAI would do well to recognize
and nurture the sources they rely upon in training their models.
Especially as the web becomes polluted with low quality AI-generated
content, it would seem in everybody's best interest to sustain the
communities and services that make and keep high quality information
available. Not just Wikimedia, but also the Internet Archive, open
access journals and preprint servers, etc.
Legally, it seems a lot murkier. OpenAI in particular does not
distribute any of its GPT models. You can feed them prompts by various
means, and get responses back. Do those responses plagiarize
Wikipedia?
With image-generating models like Stable Diffusion, it's been found
that the models sometimes generate output nearly indistinguishable
from source material [1]. I don't know if similar studies have been
undertaken for text-generating models yet. You can certainly ask GPT-4
to generate something that looks like a Wikipedia article -- here are
example results for generating a random Wikipedia article:
Article: https://en.wikipedia.org/wiki/The_Talented_Mr._Ripley_(film)
GPT-4 run 1: https://en.wikipedia.org/wiki/User:Eloquence/GPT4_Example/1
(cut off at the ChatGPT generation limit)
GPT-4 run 2: https://en.wikipedia.org/wiki/User:Eloquence/GPT4_Example/2
GPT-4 run 3: https://en.wikipedia.org/wiki/User:Eloquence/GPT4_Example/3
It imitates the form of a Wikipedia article & mixes up / makes up
assertions, but I don't know that any of its generations would meet
the standard of infringing on the Wikipedia article's copyright. IANAL
either, and as you say, the legal landscape is evolving rapidly.
Warmly,
ErikThe whole thing is definitely a hot mess. If the remixing/transformation by the model is a derivative work, it means OpenAI is potentially violating the ShareAlike requirement by not distributing the text output as CC. But on other hand the nature of the model means they’re combining CC and non free works freely / at random, unless a court would interpret whatever % of training data comes from us as the direct degree to which the model output is derived from Wikipedia. Either way it’s going to be up to some legal representation of copyright holders to test the boundaries here.
[1] https://arstechnica.com/information-technology/2023/02/researchers-extract-training-images-from-stable-diffusion-but-its-difficult/
_______________________________________________
Wikimedia-l mailing list -- wikimedia-l@lists.wikimedia.org, guidelines at: https://meta.wikimedia.org/wiki/Mailing_lists/Guidelines and https://meta.wikimedia.org/wiki/Wikimedia-l
Public archives at https://lists.wikimedia.org/hyperkitty/list/wikimedia-l@lists.wikimedia.org/message/CO3IJWXGHTBP3YE7AKUHHKPAL5HA56IC/
To unsubscribe send an email to wikimedia-l-leave@lists.wikimedia.org
Wikimedia-l mailing list -- wikimedia-l@lists.wikimedia.org, guidelines at: https://meta.wikimedia.org/wiki/Mailing_lists/Guidelines and https://meta.wikimedia.org/wiki/Wikimedia-l
Public archives at https://lists.wikimedia.org/hyperkitty/list/wikimedia-l@lists.wikimedia.org/message/4BZ5B4DFK3HTWM6CHPZ4Q4RDZIGIN26V/
To unsubscribe send an email to wikimedia-l-leave@lists.wikimedia.org
Wikimedia-l mailing list -- wikimedia-l@lists.wikimedia.org, guidelines at: https://meta.wikimedia.org/wiki/Mailing_lists/Guidelines and https://meta.wikimedia.org/wiki/Wikimedia-l
Public archives at https://lists.wikimedia.org/hyperkitty/list/wikimedia-l@lists.wikimedia.org/message/4YHAFKDLAPFCNRQGAY77KWRIOIBRWVUH/
To unsubscribe send an email to wikimedia-l-leave@lists.wikimedia.org
Wikimedia-l mailing list -- wikimedia-l@lists.wikimedia.org, guidelines at: https://meta.wikimedia.org/wiki/Mailing_lists/Guidelines and https://meta.wikimedia.org/wiki/Wikimedia-l
Public archives at https://lists.wikimedia.org/hyperkitty/list/wikimedia-l@lists.wikimedia.org/message/MTEQEC2VX77LNI6D2XXIPRLAVXY37DXL/
To unsubscribe send an email to wikimedia-l-leave@lists.wikimedia.org