4 ms·

> they weren't given license/permission to use Do we know what data was used? And what the constraints were around it? Do we know it was used without permissi

> they weren't given license/permission to use Do we know what data was used? And what the constraints were around it? Do we know it was used without permission or are we just jumping in the “ai bad” bandwagon?

ChatGTP 3y ago
No, but their products regurgitate copyrighted material so I guess they either come clean about the data they used or we have to assume they stole it.
- jameshart 3y ago
  I can quote the entire script of Monty Python and the Holy Grail - would you assume I stole it?
  - notahacker 3y ago
    No, but then as you're (presumably) a person rather than an information retrieval system you would be legally responsible for ensuring you had performance rights and paid royalties if you were quoting that script as part of a commercial service. That responsibility rests with you, not whoever gave you access to the film Conversely, photocopiers and text-to-speech engines and LLMs don't exercise choice over whether they reproduce copyrighted material and so can't be held responsible, so responsibility for clearing rights to redistribute/transform in that format clearly lies with the people inputting the copyrighted material. Obviously, OpenAI has tended to avoid making any attempts to secure those rights whatsoever
    - jameshart 3y ago
      Most libraries have photocopiers for their patrons to use. It’s their patrons’ responsibility to determine if any copying they do is permissible under fair or personal use rules. The library doesn’t know what you’re planning on doing with the information they shared with you.
      - notahacker 3y ago
        Few libraries use the scanner themselves to make a digital copy of every single book to import into their proprietary information retrieval service. The ones that do secure permission where the works involved are subject to copyright.
  - dspillett 3y ago
    If you presenting it as anything like original work, or have it to someone else without appropriate credit/attribution so that they do, yes.
dspillett 3y ago
> Do we know what data was used? And what the constraints were around it? The fact that the question/accusation has been raised a great many times and they have not stated "we know we haven't used information without licence because we had procedures to check licensing for all the data used to train our models", would certainly imply that they scraped data for training without reference to licensing, which makes it very likely that the models are based significantly on copyrighted and copyleft covered information. > Do we know it was used without permission No, we don't know for sure. But the balance of probabilities is massively skewed in that direction. There are enough examples of image producing AIs regurgitating obvious parts of unlicensed inputs, as an indication of the common practise of just scraping everything without a care for permission. So asking for those with other models to state how they checked for permission for the input data is reasonable.
dns_snek 3y ago
Yes, we've known for a long time that they don't shy away from taking any old code on GitHub and regurgitating it without explicit permission. They don't have benefit of the doubt anymore.
- jameshart 3y ago
  It’s not as cut and dried as you’re making it out to be. For years we’ve accepted that search engines, for example, can grab all the code on GitHub and use it to build a search index. Google image search, in particular, ‘regurgitates’ all the images it has indexed when it thinks they match a search term. It has a little disclaimer it shows next to the results saying that ‘images may be copyrighted’ - figuring out if they are copyrighted and if so by whom is left as an exercise for you the searcher. Depending on what you are using the image search for, the copyright of the images may, after all, not be relevant. Like, if you’re using a Google image search to get inspiration for home decor designs, do you care who owns the copyright of each image? Should Google? GPT poses similar risks to that. It has the explicit disclaimer that things it produces might be subject to copyright. Depending on what you’re using the output for, the copyright may or may not be relevant.
  - notahacker 3y ago
    There seems to be a fairly clear distinction between importing copyrighted material to make an index for the narrow purpose of directing people towards the copyrighted material at its original location in its original form, and importing copyrighted material to make an index which improves their own service's ability to generate unattributed derivative work. It's a bit muddier when it comes to things like image searches, but they're not exactly difficult to opt out of. Google actually pays license fees to News Corp to excerpt their content in Google News following a legal challenge so it's not exactly conclusively established that search engines have global rights to do what they do anyway. But search engines are mostly beneficial to contact creators rather than mostly competitive with them.
    - jameshart 3y ago
      Google does its best, but there are limits to “directing people towards the copyrighted material at its original location in its original form” - not everything in the intellectual property world is ‘internet native’. The original form of a song lyric or a movie screenplay doesn’t have an ‘original location’ you can be directed to. You can be directed to various online sources that may or may not accurately reproduce the original, and may or may not be scrupulous about attribution, and may or may not have legitimate copyright license to distribute it in the first place.

ChatGTP 3y ago

No, but their products regurgitate copyrighted material so I guess they either come clean about the data they used or we have to assume they stole it.

jameshart 3y ago
I can quote the entire script of Monty Python and the Holy Grail - would you assume I stole it?
- notahacker 3y ago
  No, but then as you're (presumably) a person rather than an information retrieval system you would be legally responsible for ensuring you had performance rights and paid royalties if you were quoting that script as part of a commercial service. That responsibility rests with you, not whoever gave you access to the film Conversely, photocopiers and text-to-speech engines and LLMs don't exercise choice over whether they reproduce copyrighted material and so can't be held responsible, so responsibility for clearing rights to redistribute/transform in that format clearly lies with the people inputting the copyrighted material. Obviously, OpenAI has tended to avoid making any attempts to secure those rights whatsoever
  - jameshart 3y ago
    Most libraries have photocopiers for their patrons to use. It’s their patrons’ responsibility to determine if any copying they do is permissible under fair or personal use rules. The library doesn’t know what you’re planning on doing with the information they shared with you.
    - notahacker 3y ago
      Few libraries use the scanner themselves to make a digital copy of every single book to import into their proprietary information retrieval service. The ones that do secure permission where the works involved are subject to copyright.
- dspillett 3y ago
  If you presenting it as anything like original work, or have it to someone else without appropriate credit/attribution so that they do, yes.

dspillett 3y ago

> Do we know what data was used? And what the constraints were around it? The fact that the question/accusation has been raised a great many times and they have not stated "we know we haven't used information without licence because we had procedures to check licensing for all the data used to train our models", would certainly imply that they scraped data for training without reference to licensing, which makes it very likely that the models are based significantly on copyrighted and copyleft covered information. > Do we know it was used without permission No, we don't know for sure. But the balance of probabilities is massively skewed in that direction. There are enough examples of image producing AIs regurgitating obvious parts of unlicensed inputs, as an indication of the common practise of just scraping everything without a care for permission. So asking for those with other models to state how they checked for permission for the input data is reasonable.

dns_snek 3y ago

Yes, we've known for a long time that they don't shy away from taking any old code on GitHub and regurgitating it without explicit permission. They don't have benefit of the doubt anymore.

jameshart 3y ago
It’s not as cut and dried as you’re making it out to be. For years we’ve accepted that search engines, for example, can grab all the code on GitHub and use it to build a search index. Google image search, in particular, ‘regurgitates’ all the images it has indexed when it thinks they match a search term. It has a little disclaimer it shows next to the results saying that ‘images may be copyrighted’ - figuring out if they are copyrighted and if so by whom is left as an exercise for you the searcher. Depending on what you are using the image search for, the copyright of the images may, after all, not be relevant. Like, if you’re using a Google image search to get inspiration for home decor designs, do you care who owns the copyright of each image? Should Google? GPT poses similar risks to that. It has the explicit disclaimer that things it produces might be subject to copyright. Depending on what you’re using the output for, the copyright may or may not be relevant.
- notahacker 3y ago
  There seems to be a fairly clear distinction between importing copyrighted material to make an index for the narrow purpose of directing people towards the copyrighted material at its original location in its original form, and importing copyrighted material to make an index which improves their own service's ability to generate unattributed derivative work. It's a bit muddier when it comes to things like image searches, but they're not exactly difficult to opt out of. Google actually pays license fees to News Corp to excerpt their content in Google News following a legal challenge so it's not exactly conclusively established that search engines have global rights to do what they do anyway. But search engines are mostly beneficial to contact creators rather than mostly competitive with them.
  - jameshart 3y ago
    Google does its best, but there are limits to “directing people towards the copyrighted material at its original location in its original form” - not everything in the intellectual property world is ‘internet native’. The original form of a song lyric or a movie screenplay doesn’t have an ‘original location’ you can be directed to. You can be directed to various online sources that may or may not accurately reproduce the original, and may or may not be scrupulous about attribution, and may or may not have legitimate copyright license to distribute it in the first place.