5 ms·

It's going to be hard to remove every single "shorthand descriptions of well-known entities" or other prompts that can be used to generate copyrighted or tradem

It's going to be hard to remove every single "shorthand descriptions of well-known entities" or other prompts that can be used to generate copyrighted or trademarked content. Sure, if you're not deliberately trying to generate infringing content, you can probably remove or discard those results, the trouble is the people who will try to trick the AI to generate this content, blocking those people is going to be impossible, without excluding any copyrighted or trademarked training material. Another issue for generative AI is mentioned in the article: "Systems like DALL-E and ChatGPT are essentially black boxes." What happens when an AI is used to make decisions where the user/victim is entitled to know exactly why the AI did what it did? From a business and legal perspective I think the current AI solutions are dangerous and should be used very sparsely, exactly because even the creators can't point to the exact pieces of information that caused the AI to make the choices it did.

keiferski 3y ago
But does this actually matter if the people are only generating images for their own use? Does Photoshop prevent people from making drawings that look like Mickey Mouse? Of course not. I think it will be easy to prevent the obvious copyrighted stuff via the method I mentioned. People going around those restrictions are subject to the same rules as someone drawing the copyrighted image from scratch.
- mrweasel 3y ago
  > But does this actually matter if the people are only generating images for their own use? Arguably that might actually be a very small issue, but what happens when it happens on a larger scale? Disney might not care, they can easily fight you in court if your DALL-E generated comics looks like Mickey Mouse and Nintendo will make sure that your video game about a electrician named Marvin from the Bronx, but who looks a lot like Mario is never going to get featured on Steam. The issue is the smaller artists that might not have the resources to fight AI content in court. There's also the issue with using LLMs to "white-wash" articles and books. There will be people who will just run articles through ChatGPT and claim that it's AI generated content that was in no way stolen from The Barents Observer. The absolute massive volume, lose in revenue and cost to fight this in court could make running an investigating newspaper impossible and leave us without any actual reporting. Not thinking ahead and having a plan for copyrighted material was an oversight by the current AI companies, but they are arrogant and just assumed that it would be a detail and anyway "disruption" so screw it. There has been zero consideration to the fact that their product is useless without the previous work of millions of people. My concern is that AI takes over content generation to the point where we actually run out of human generated content to train future AIs on. We need to be incredibly careful about implementing AI and ensure that we do not pollute future training data, but people don't care, because they want profit now.
- loki-ai 3y ago
  They are selling these pictures for $20/mo, so it definitely matters. If people will hang them at their refrigerator has nothing with this case.
losvedir 3y ago
Ironically, I don't think it would be that hard with LLMs. I tried asking ChatGPT to describe what copyrighted characters each description is alluding to and it had no problem doing so.[0] [0] https://chat.openai.com/share/e8256470-8e45-4f36-9c84-026be1164ada https://chat.openai.com/share/e8256470-8e45-4f36-9c84-026be1...
numpad0 3y ago
I don't understand why some peoole thinks any infringing content can be singled out and removed. Aren't LLMs giant coefficient matrices like, a punched out croissant dough, made of all training data plyed over? How can you say you can remove one specific ply out of dough and declare that ever potential effect that the offending ply had created is now completely removed?
- keiferski 3y ago
  I’m no LLM expert but I think there is a distinction to be made between the LLM dataset and the output it gives to the user. What you’re suggesting is that it’s difficult to remove something from the dataset, which may be the case. But that doesn’t mean the user will necessarily be able to access it. My guess is that this is much easier to attack from the user end.
  - numpad0 3y ago
    Removing something from the dataset requires full training from scratch(~$100 million for base unaligned GPT-4). You can't like, edit the database file and keep the AI. the database file _is_ the AI.
- selimthegrim 3y ago
  I think given the use of ply for AlphaGo/chess engines this is a pretty cool metaphor.
- anonymousab 3y ago
  The "reasonable" singular removal is more about coming up with ways to block prompts that can produce infringing content, and having filters on the other end to catch infringements before they are published to the user. It's an endless whack-a-mole that never actually addresses the problem but might look good enough to the legal system or to keen supporters. Barring some major breakthrough, the actual answer is to train a new model without the infringing data. I think some of the people saying "remove it from your model" are aware of this and are simply being glib and needling; "you've created this infringement monstrosity, so surely you made sure to include a way to deal with this problem without throwing away all of your work, right?"
bbor 3y ago
What happens when an AI is used to make decisions where the user/victim is entitled to know exactly why the AI did what it did? From a business and legal perspective I think the current AI solutions are dangerous and should be used very sparsely, exactly because even the creators can't point to the exact pieces of information that caused the AI to make the choices it did. I totally agree that we need explainability (probably through symbolic systems, not fancier models), but I think you’re overestimating how much more satisfying explanations from more traditional AI are. “My rules told me to do X” is a bit more helpful for a troubleshooting engineer than “my training data trained me to do X”, but from a ‘business and legal perspective’ the difference is much less pronounced IMO. Both answers mean you did something wrong in creating the machine. The fault will always lie with the creator.

keiferski 3y ago

But does this actually matter if the people are only generating images for their own use? Does Photoshop prevent people from making drawings that look like Mickey Mouse? Of course not. I think it will be easy to prevent the obvious copyrighted stuff via the method I mentioned. People going around those restrictions are subject to the same rules as someone drawing the copyrighted image from scratch.

mrweasel 3y ago
> But does this actually matter if the people are only generating images for their own use? Arguably that might actually be a very small issue, but what happens when it happens on a larger scale? Disney might not care, they can easily fight you in court if your DALL-E generated comics looks like Mickey Mouse and Nintendo will make sure that your video game about a electrician named Marvin from the Bronx, but who looks a lot like Mario is never going to get featured on Steam. The issue is the smaller artists that might not have the resources to fight AI content in court. There's also the issue with using LLMs to "white-wash" articles and books. There will be people who will just run articles through ChatGPT and claim that it's AI generated content that was in no way stolen from The Barents Observer. The absolute massive volume, lose in revenue and cost to fight this in court could make running an investigating newspaper impossible and leave us without any actual reporting. Not thinking ahead and having a plan for copyrighted material was an oversight by the current AI companies, but they are arrogant and just assumed that it would be a detail and anyway "disruption" so screw it. There has been zero consideration to the fact that their product is useless without the previous work of millions of people. My concern is that AI takes over content generation to the point where we actually run out of human generated content to train future AIs on. We need to be incredibly careful about implementing AI and ensure that we do not pollute future training data, but people don't care, because they want profit now.
loki-ai 3y ago
They are selling these pictures for $20/mo, so it definitely matters. If people will hang them at their refrigerator has nothing with this case.

losvedir 3y ago

Ironically, I don't think it would be that hard with LLMs. I tried asking ChatGPT to describe what copyrighted characters each description is alluding to and it had no problem doing so.[0] [0] https://chat.openai.com/share/e8256470-8e45-4f36-9c84-026be1164ada https://chat.openai.com/share/e8256470-8e45-4f36-9c84-026be1...

numpad0 3y ago

I don't understand why some peoole thinks any infringing content can be singled out and removed. Aren't LLMs giant coefficient matrices like, a punched out croissant dough, made of all training data plyed over? How can you say you can remove one specific ply out of dough and declare that ever potential effect that the offending ply had created is now completely removed?

keiferski 3y ago
I’m no LLM expert but I think there is a distinction to be made between the LLM dataset and the output it gives to the user. What you’re suggesting is that it’s difficult to remove something from the dataset, which may be the case. But that doesn’t mean the user will necessarily be able to access it. My guess is that this is much easier to attack from the user end.
- numpad0 3y ago
  Removing something from the dataset requires full training from scratch(~$100 million for base unaligned GPT-4). You can't like, edit the database file and keep the AI. the database file _is_ the AI.
selimthegrim 3y ago
I think given the use of ply for AlphaGo/chess engines this is a pretty cool metaphor.
anonymousab 3y ago
The "reasonable" singular removal is more about coming up with ways to block prompts that can produce infringing content, and having filters on the other end to catch infringements before they are published to the user. It's an endless whack-a-mole that never actually addresses the problem but might look good enough to the legal system or to keen supporters. Barring some major breakthrough, the actual answer is to train a new model without the infringing data. I think some of the people saying "remove it from your model" are aware of this and are simply being glib and needling; "you've created this infringement monstrosity, so surely you made sure to include a way to deal with this problem without throwing away all of your work, right?"

bbor 3y ago

What happens when an AI is used to make decisions where the user/victim is entitled to know exactly why the AI did what it did? From a business and legal perspective I think the current AI solutions are dangerous and should be used very sparsely, exactly because even the creators can't point to the exact pieces of information that caused the AI to make the choices it did. I totally agree that we need explainability (probably through symbolic systems, not fancier models), but I think you’re overestimating how much more satisfying explanations from more traditional AI are. “My rules told me to do X” is a bit more helpful for a troubleshooting engineer than “my training data trained me to do X”, but from a ‘business and legal perspective’ the difference is much less pronounced IMO. Both answers mean you did something wrong in creating the machine. The fault will always lie with the creator.