5 ms·

> Yet I believe OpenAI isn't using data from Dropbox to train their models without users' consent. why? they trained on my code without my consent, why is user

> Yet I believe OpenAI isn't using data from Dropbox to train their models without users' consent. why? they trained on my code without my consent, why is user data any different? training is either fair use or it isn't and high growth Silicon Valley companies aren't known for the adherence to the spirit of the law

vorticalbox 3y ago
Just out of interest I have a few questions. 1. How do you know for sure it was trained on your code? 1.1 if you saw it reproduce you code verbatim how do you know it didn't just hallucinate it ? 2. Was you code publicly available? 3. What licence was it under?
- spatley 3y ago
  Well Github says it out loud: “ GitHub Copilot is trained on all languages that appear in public repositories.” Note it says “public” and not “open source” https://docs.github.com/en/copilot/overview-of-github-copilot/about-github-copilot-individual# https://docs.github.com/en/copilot/overview-of-github-copilo...
d4mi3n 3y ago
Training against your code shouldn't be kosher if your code isn't open, you retain a copyright, and you didn't consent. Even if training on copyrighted, private code _was_ fair use, training on PII without consent is _still_ problematic. It's a huge violation of privacy and adds all kinds of legal/regulatory/ethical risks that potential consumers of said models won't be on board with. The language and spirit of the law certainly hasn't caught up, but I feel like in some cases (unconsenting use of PII, medical information, etc), the laws and ethics of how such data is used is pretty well established.
- lacker 3y ago
  AI systems are trained against private, PII, copyrighted data all the time without explicit consent. For example, consider the spellcheck in Google search. Every query you make will go into the training for that system, along with your preferred language and country of origin. You can certainly argue that generative AI systems are different than previous AI systems and should be treated differently. But the current situation is basically that you are allowed to train an AI on any data you have, regardless of copyright or consent. I wouldn't be surprised if that ends up being considered legally and ethically okay, because it's the status quo, and because it's hard to define "what counts as AI".
  - d4mi3n 3y ago
    The point about spell check is a very good one. I think one big differentiator is that the attack/risk surface for something as complex as a LLM is much higher and that much, much more information is encoded in an LLM than a spell checking dictionary. For example, it's possible to extract training data from an LLM—which could include PII/medical data/etc. Those risks don't exist with spellcheck as far as I'm aware. To your point about what is "AI", I'd state that AI is a misnomer. What we're really talking about are generative large language models (LLMs). What an _can_ be considered an LLM is definitely up for debate, but if you were to describe one in general terms I think we could reasonably say that most (or all) things we consider LLMs are: 1. A probabilistic model of a natural language 2. Have the ability to interpret and generate natural language text 3. Typically encode a large volume of training data I'd love to hear other thoughts on how one would define an LLM in practical, simple language. I imagine doing so would be a pre-requisite to any effective legislation.
bambax 3y ago
I don't think they do, but I don't know for sure. But my point is, this (training) isn't the main problem.
simonw 3y ago
> why? they trained on my code without my consent, why is user data any different? It's different. Code and other content that you shared online - no matter what license you shared it under - is still fundamentally different from user data that you never shared anywhere at all. That difference is really important, to me at least. I would be absolutely furious if I found that someone had trained an AI model on my private data in Dropbox. I personally have no problem at all with someone training an AI model on content I have posted to my blog.
- __loam 3y ago
  A huge problem with the art community right now is if you're a professional artist, you basically need to maintain a public portfolio and social media presence to find work. That ecosystem was developed and existed before image generation AI was a commercial thing, yet artists are finding out retroactively that their data got pulled into these training sets without their knowledge or consent. Even if it's legal, it's still pretty gross imo.
- JohnFen 3y ago
  The two things aren't the same, and one is more egregious than the other. But I do think that hoovering up all that web data to train models was way over the line. Worse, there's nothing I can do to prevent it. That's why I felt forced to remove my websites from the public web. I can't think of any other way to defend myself from these entities.

vorticalbox 3y ago

Just out of interest I have a few questions. 1. How do you know for sure it was trained on your code? 1.1 if you saw it reproduce you code verbatim how do you know it didn't just hallucinate it ? 2. Was you code publicly available? 3. What licence was it under?

spatley 3y ago
Well Github says it out loud: “ GitHub Copilot is trained on all languages that appear in public repositories.” Note it says “public” and not “open source” https://docs.github.com/en/copilot/overview-of-github-copilot/about-github-copilot-individual# https://docs.github.com/en/copilot/overview-of-github-copilo...

d4mi3n 3y ago

Training against your code shouldn't be kosher if your code isn't open, you retain a copyright, and you didn't consent. Even if training on copyrighted, private code _was_ fair use, training on PII without consent is _still_ problematic. It's a huge violation of privacy and adds all kinds of legal/regulatory/ethical risks that potential consumers of said models won't be on board with. The language and spirit of the law certainly hasn't caught up, but I feel like in some cases (unconsenting use of PII, medical information, etc), the laws and ethics of how such data is used is pretty well established.

lacker 3y ago
AI systems are trained against private, PII, copyrighted data all the time without explicit consent. For example, consider the spellcheck in Google search. Every query you make will go into the training for that system, along with your preferred language and country of origin. You can certainly argue that generative AI systems are different than previous AI systems and should be treated differently. But the current situation is basically that you are allowed to train an AI on any data you have, regardless of copyright or consent. I wouldn't be surprised if that ends up being considered legally and ethically okay, because it's the status quo, and because it's hard to define "what counts as AI".
- d4mi3n 3y ago
  The point about spell check is a very good one. I think one big differentiator is that the attack/risk surface for something as complex as a LLM is much higher and that much, much more information is encoded in an LLM than a spell checking dictionary. For example, it's possible to extract training data from an LLM—which could include PII/medical data/etc. Those risks don't exist with spellcheck as far as I'm aware. To your point about what is "AI", I'd state that AI is a misnomer. What we're really talking about are generative large language models (LLMs). What an _can_ be considered an LLM is definitely up for debate, but if you were to describe one in general terms I think we could reasonably say that most (or all) things we consider LLMs are: 1. A probabilistic model of a natural language 2. Have the ability to interpret and generate natural language text 3. Typically encode a large volume of training data I'd love to hear other thoughts on how one would define an LLM in practical, simple language. I imagine doing so would be a pre-requisite to any effective legislation.

bambax 3y ago

I don't think they do, but I don't know for sure. But my point is, this (training) isn't the main problem.

simonw 3y ago

> why? they trained on my code without my consent, why is user data any different? It's different. Code and other content that you shared online - no matter what license you shared it under - is still fundamentally different from user data that you never shared anywhere at all. That difference is really important, to me at least. I would be absolutely furious if I found that someone had trained an AI model on my private data in Dropbox. I personally have no problem at all with someone training an AI model on content I have posted to my blog.

__loam 3y ago
A huge problem with the art community right now is if you're a professional artist, you basically need to maintain a public portfolio and social media presence to find work. That ecosystem was developed and existed before image generation AI was a commercial thing, yet artists are finding out retroactively that their data got pulled into these training sets without their knowledge or consent. Even if it's legal, it's still pretty gross imo.
JohnFen 3y ago
The two things aren't the same, and one is more egregious than the other. But I do think that hoovering up all that web data to train models was way over the line. Worse, there's nothing I can do to prevent it. That's why I felt forced to remove my websites from the public web. I can't think of any other way to defend myself from these entities.