Is there any benefit to the author of letting a for profit LLM like OpenAI spider and consume their writing? I can’t think of any.
If that’s true, then should we be rethinking Creative Commons licenses? I’m wondering if the current one we use at Medium is obsolete.
At least with something like Google there was an exchange of value: contribute to their search results and get traffic back.
@coachtony not that I’d do it, but there is one thing to consider. Much of what was fed into the AI was likely made by white, cisgender, heterosexual sexual, non-poor men which means that’s the lens all of the resulting output will be framed through.
For those of us who don’t fit that category we can get screwed with our pants on or be made more invisible.
So, a typical choice.
@hramrach @soaproot @coachtony The problem is, if ML training is fair use, it is legally allowed to "break" all licensing terms. The license may say "you can't use this content at all for anything, ever, nonono, go away!" and it will be perfectly legal to use that content within limits of fair use.
https://felixreda.eu/2021/07/github-copilot-is-not-infringing-your-copyright/
@anildash @coachtony does there need to be direct financial benefit here? Not every use of creative material needs to be a transaction. The theory of Creative Commons was to literally make a commons of creative works. Why is it a problem if software is also benefiting?
(I'm motivated by maximizing the current boom in AI research and development. I do worry that these systems benefit large commercial players right now but am hopeful that will change.)
@coachtony you may appreciate my thread from yesterday on this topic. i focused on source code + how the license under which it's released impacts this:
https://elk.zone/toot.bldrweb.org/@jbminn/110084548838376306
1/ a thread on AI code generators the longer i'm around, the less inclined i am to profess expertise on anything. but i do have lots of experience w/ source code hosting, licenses that apply to it + associated issues. so i'm expert-y i guess. scraping code from a public repo + using it to generate derivative works may be permissible if the original work was licensed to explicitly allow such use. the orig author(s) may be unhappy, but that's how licenses work (generally).
@coachtony Well, AI can't create copyrightable works and learning their database is not a copyrightable or licenseable task, since otherwise school textbook publishers would be entitled to compensation (i.e. their income in lieu of a proper license) from every wageworker who learned from their works.
Because that's the de-jure equivalent w/ pirated Software...
@coachtony @kkarhan A court ruled that they don't, and until it's repealed that's how it is.
Nonetheless, it ruled that a comics made of AI generated art is not subject to copyright which is a collection of works which is in turn subject to copyright which is a contradiction 🤷
However, the material that went into the LLM is for the most part copyrighted, and if the AI work is not then the result is pretty much solely the aggregation of these works that went in. Want to sue Microsoft?
@coachtony My open source code is licensed in such a way that no AI can be permitted to casually ingest and regurgitate.
Attribution is mandatory.
The whole "Wild West" of open ingest can not last. These can and must not be opt-out scraping processes.
As it stands, I literally can not touch these tools in relation to anything at work. I'm not even allowed to copy anything out of Stack Overflow, by Romulus.
@coachtony Heh, amusingly, it's still a strong concern for the B2B multinational I work for. When my prior company (+ my project + me) were being acquired, we sent the codebase off to a pair of third-party auditors, one for pen-testing the other for license auditing.
They found a line of Stack Overflow code. I found back where I got it, searched down the answerer's GitHub, and was happy to discover he had packaged up the solution with a proper license.
Panic avoided. 😜
@coachtony Wasn’t my panic; I mean, in that instance I had just been lazy. It was a line or two of Python. Execs, though, saw a blocker to that acquisition.
Having been lead developer *and* product owner for 8 years gave me some interesting perspective across the realms of engineering and management. 😝
Loved that entire team. 🥰
@coachtony
I feel this line of thinking is much like open source developers who change their license away from an open one when Amazon or Google start offering it as a service.
Does it feel fair? No! Should you change your license? I don't know maybe.
But, you choose to release it under an open license, now there is a business model that uses that data, there isn't really putting that genie back in the bottle.
Now that it's here you do need to do the math if you want to keep participating.
@coachtony Would you prefer if Disney trained an AI on only images and text that Disney owns, without paying anything to the authors or getting their permission?
Like, I don't get why a world where Disney has an AI generator and nobody else is allowed to make one is somehow better than a world where everybody can make an AI generator.
@183231bcb Are these the only two options? I don't think option two even accurately reflects the status quo. Other options:
* No LLMs at all. This is the Butlerian Jihad of Dune.
* LLMs pay and attribute their source material.
* LLMs have more limited opt-in coverage.
Etc.
@coachtony As long as corporations can own copyright, LLMs "paying and attributing source material" just means Disney training an LLM on media they own and don't have to pay royalties on. Disney can "opt in" their own media as training data for their own LLM.
Maybe you could try convincing legislators to abolish corporate-owned copyright, but that seems very unlikely to succeed.
@mcc I wonder if any other content companies, i.e. WordPress, Substack, StackOverflow, care?
In matters of law there are almost always two alternative paths: relationship or power.
I don't think the platforms are powerless here.
@BartV @coachtony It seems to me that if a court held a derivative work of your scraped data was unbound by license requirements like attribution unless you add a magic additional "noAI" tag (which there probably wasn't even a standard for at the time the data was uploaded), this would be utterly absurd.
I am also not aware of any lawsuits proving any of these scraped large models are *legal*.
Tony Stubblebine
Mike Lapidakis
Kevin Karhan 
Kommadieb 🌆
Alex Zani
Kelli Murbach
Werner Mesch
soaproot
hramrach
Kornel
Jason Weiland
Anil Dash
Nelson Minar
skry
Erik Haugen
Ethan White
paulmwatson
John Minnihan
Jeff Watkins
Roy Wiggins
Rev. GothAlice
MadVikingGod 
varx/social
Francesco Orsenigo (he/him)
bettybarcode (she/her) 🚲
183231bcb
mcc
Bart Veldhuizen 🚀
Matunos
Andy Baio
Michal Migurski 🍉