7 ms·

Understanding and managing the impact of machine learning models on the web

138 points by kaycebasques 3y ago

kaycebasques 3y ago
This would have been a better link: https://www.w3.org/reports/ai-web-impact/ https://www.w3.org/reports/ai-web-impact/
- dang 3y ago
  Ok, we've changed to that from https://github.com/w3c/ai-web-impact https://github.com/w3c/ai-web-impact above. Thanks!
MacsHeadroom 3y ago
> the copyright system creates a (relatively) shared understanding between creators and consumers that, by default, content cannot be redistributed, remixed, adapted or built upon without creators' consent. This shared understanding made it possible for a lot of content to be openly distributed on the Web. That is not remotely a shared understanding, is wrong, and has nothing to do with making it possible for a lot of content to be openly distributed on the web. Content is distributed quite widely without concern for copyright. > A number of AI systems combine (1) automated large-scale consumption of Web content, and (2) production at scale of content, in ways that do not recognize or otherwise compensate content it was trained from. > While some of these tensions are not new (as discussed below), systems based on Machine Learning are poised to upend the existing balance. Unless a new sustainable equilibrium is found, this exposes the Web to the following undesirable outcomes: > Significantly less open distributed content (which would likely have a disproportionate impact on the less wealthy part of the population) That's even more ridiculous. The wealthy stand the most to gain from restricting the flow of information to channels which collect rent on behalf of their capital. It's the "less wealthy" who routinely find ways to distribute content outside of rent-seeking channels. It's the "less wealthy" who benefit the most from the commoditization of creative content via generative algorithms. Quite frankly, I expected better from W3C.
- munificent 3y ago
  > > the copyright system creates a (relatively) shared understanding between creators and consumers that, by default, content cannot be redistributed, remixed, adapted or built upon without creators' consent. This shared understanding made it possible for a lot of content to be openly distributed on the Web. > That is not remotely a shared understanding, is wrong, and has nothing to do with making it possible for a lot of content to be openly distributed on the web. Content is distributed quite widely without concern for copyright. I'm not sure if the switch from active voice in the original quote to passive in yours was deliberate or not, but "understanding between creators and consumers" is very different from your "content is distributed". It is the case, yes, that people widely distribute content on the web with no regard for copyright law. But those people aren't generally creators of that content. The article is talking about the incentives that the web places on content creators. If the result of AIs harvesting every bit of content on the web and regurgitating it without sending consumers over to the creator's website, then creators will stop putting stuff online. People cloning and resharing content without regard to copyright has not so far seemed to have systemic negative effects on the web. Search engines seem to be pretty good at pointing users to upstream original sources of copyright content, so plagiarism is commong but apparently not common enough to cause context authors to stop putting it online. AI risks tipping that balance such that content creators really might stop posting stuff online. Why waste a meaningful chunk of your life creating a thing and putting it on the web if the only thing that will ever see it and know that it came from you is an AI slurping it up? > It's the "less wealthy" who routinely find ways to distribute content outside of rent-seeking channels. Again, I think you're presuming a world where content magically exists a priori and the network is simply a mechanism for deploying it. The article is about what happens when the system discourages people from making at all. Poor people can find ways to pirate just about every book on Earth... except for those books that never ended up getting written because the incentives placed on the author didn't work out.
  - visarga 3y ago
    I think you are not taking into consideration the new content being created by AI with human in the loop, such as in the chatGPT interface. With 100M users OpenAI might be generating on the order of 1 trillion tokens per month. These logs are mixed AI and human text, the human part containing tasks and feedback. By transforming these chat logs into training examples and fine-tuning the model there is a way to integrate LLM and human signals. Of course it is necessary to be mindful of copyright and PII during this process, not everything is generally useful to be included in the model. But having hundreds of millions of people inputting feedback into the model can scale even more than unassisted content publishing. Besides the AI users, we have the social networks. Billions of people comment on the news, if you scrape a comment thread such as this one you can readily generate an article from it, grounded in human feedback (sample based on this thread https://pastebin.com/R229b41s https://pastebin.com/R229b41s) Ultimately I don't believe publishers will stop creating content because there are many reasons to do it besides collecting advertising fees, but as a backup we have LLMs continuously learning by assisting their users or reshaping informal communications into well formed content. It's a two way street - both the LLM and human get help from each other. In turn, LLMs learn from other LLMs, usually Mistral and LLaMA would learn from GPT-4, this process works very very well. So any skills learned by a SOTA model with massive human chat logs would eventually be extracted for small models. Feedback percolates back to all AI agents. tl;dr Turning comment threads or human-AI chat logs into articles is a possible solution.
pmayrgundter 3y ago
I agree with the general idea of tagging content to help classify. I'd given this some thought via MIME and ended up with a kind of BioNFT.. so named bc it uses NFTs piecewise, but tracking the creation events and agents types (bio, ai, etc) as part of the content lifecycle https://twitter.com/PMayrgundter/status/1638016474483683328 https://twitter.com/PMayrgundter/status/1638016474483683328 Highlight.. What if: - devices sign source creations with a biosignature - editing tools sign input - media types include that, effectively saying: ai_edited(human_created(photo)) and do this under the experimental namespace in MIME: image/x.bio(pablo@example.com/photo123).html image/x.adobe.photoai(http://x.bio(pablo@example.com/photo123)).html
- Retr0id 3y ago
  What you're describing here is basically what C2PA is
  - xeyownt 3y ago
    This C2PA looks flawed from the start. It just helps making the untrustworthy looks more trustworthy.
  - zerojames 3y ago
    Reference, for context: https://c2pa.org/ https://c2pa.org/ And: the BBC just started using C2PA across some content. The BBC's R&D team talking about it: https://www.bbc.co.uk/rd/blog/2024-03-c2pa-verification-news-journalism-credentials https://www.bbc.co.uk/rd/blog/2024-03-c2pa-verification-news...
    - rentnorove 3y ago
      Related critique of the BBC's use of C2PA, and C2PA in general: https://www.hackerfactor.com/blog/index.php?/archives/1024-IEEE,-BBC,-and-C2PA.html https://www.hackerfactor.com/blog/index.php?/archives/1024-I...
      - dns_snek 3y ago
barfbagginus 3y ago
Prediction: soon we will want to label things written by actually useful AI, because it will be more helpful than things written by humans and parrot tier AI

138 points by kaycebasques 3y ago

kaycebasques 3y ago

This would have been a better link: https://www.w3.org/reports/ai-web-impact/ https://www.w3.org/reports/ai-web-impact/

dang 3y ago
Ok, we've changed to that from https://github.com/w3c/ai-web-impact https://github.com/w3c/ai-web-impact above. Thanks!

MacsHeadroom 3y ago

> the copyright system creates a (relatively) shared understanding between creators and consumers that, by default, content cannot be redistributed, remixed, adapted or built upon without creators' consent. This shared understanding made it possible for a lot of content to be openly distributed on the Web. That is not remotely a shared understanding, is wrong, and has nothing to do with making it possible for a lot of content to be openly distributed on the web. Content is distributed quite widely without concern for copyright. > A number of AI systems combine (1) automated large-scale consumption of Web content, and (2) production at scale of content, in ways that do not recognize or otherwise compensate content it was trained from. > While some of these tensions are not new (as discussed below), systems based on Machine Learning are poised to upend the existing balance. Unless a new sustainable equilibrium is found, this exposes the Web to the following undesirable outcomes: > Significantly less open distributed content (which would likely have a disproportionate impact on the less wealthy part of the population) That's even more ridiculous. The wealthy stand the most to gain from restricting the flow of information to channels which collect rent on behalf of their capital. It's the "less wealthy" who routinely find ways to distribute content outside of rent-seeking channels. It's the "less wealthy" who benefit the most from the commoditization of creative content via generative algorithms. Quite frankly, I expected better from W3C.

munificent 3y ago
> > the copyright system creates a (relatively) shared understanding between creators and consumers that, by default, content cannot be redistributed, remixed, adapted or built upon without creators' consent. This shared understanding made it possible for a lot of content to be openly distributed on the Web. > That is not remotely a shared understanding, is wrong, and has nothing to do with making it possible for a lot of content to be openly distributed on the web. Content is distributed quite widely without concern for copyright. I'm not sure if the switch from active voice in the original quote to passive in yours was deliberate or not, but "understanding between creators and consumers" is very different from your "content is distributed". It is the case, yes, that people widely distribute content on the web with no regard for copyright law. But those people aren't generally creators of that content. The article is talking about the incentives that the web places on content creators. If the result of AIs harvesting every bit of content on the web and regurgitating it without sending consumers over to the creator's website, then creators will stop putting stuff online. People cloning and resharing content without regard to copyright has not so far seemed to have systemic negative effects on the web. Search engines seem to be pretty good at pointing users to upstream original sources of copyright content, so plagiarism is commong but apparently not common enough to cause context authors to stop putting it online. AI risks tipping that balance such that content creators really might stop posting stuff online. Why waste a meaningful chunk of your life creating a thing and putting it on the web if the only thing that will ever see it and know that it came from you is an AI slurping it up? > It's the "less wealthy" who routinely find ways to distribute content outside of rent-seeking channels. Again, I think you're presuming a world where content magically exists a priori and the network is simply a mechanism for deploying it. The article is about what happens when the system discourages people from making at all. Poor people can find ways to pirate just about every book on Earth... except for those books that never ended up getting written because the incentives placed on the author didn't work out.
- visarga 3y ago
  I think you are not taking into consideration the new content being created by AI with human in the loop, such as in the chatGPT interface. With 100M users OpenAI might be generating on the order of 1 trillion tokens per month. These logs are mixed AI and human text, the human part containing tasks and feedback. By transforming these chat logs into training examples and fine-tuning the model there is a way to integrate LLM and human signals. Of course it is necessary to be mindful of copyright and PII during this process, not everything is generally useful to be included in the model. But having hundreds of millions of people inputting feedback into the model can scale even more than unassisted content publishing. Besides the AI users, we have the social networks. Billions of people comment on the news, if you scrape a comment thread such as this one you can readily generate an article from it, grounded in human feedback (sample based on this thread https://pastebin.com/R229b41s https://pastebin.com/R229b41s) Ultimately I don't believe publishers will stop creating content because there are many reasons to do it besides collecting advertising fees, but as a backup we have LLMs continuously learning by assisting their users or reshaping informal communications into well formed content. It's a two way street - both the LLM and human get help from each other. In turn, LLMs learn from other LLMs, usually Mistral and LLaMA would learn from GPT-4, this process works very very well. So any skills learned by a SOTA model with massive human chat logs would eventually be extracted for small models. Feedback percolates back to all AI agents. tl;dr Turning comment threads or human-AI chat logs into articles is a possible solution.

pmayrgundter 3y ago

I agree with the general idea of tagging content to help classify. I'd given this some thought via MIME and ended up with a kind of BioNFT.. so named bc it uses NFTs piecewise, but tracking the creation events and agents types (bio, ai, etc) as part of the content lifecycle https://twitter.com/PMayrgundter/status/1638016474483683328 https://twitter.com/PMayrgundter/status/1638016474483683328 Highlight.. What if: - devices sign source creations with a biosignature - editing tools sign input - media types include that, effectively saying: ai_edited(human_created(photo)) and do this under the experimental namespace in MIME: image/x.bio(pablo@example.com/photo123).html image/x.adobe.photoai(http://x.bio(pablo@example.com/photo123)).html

Retr0id 3y ago
What you're describing here is basically what C2PA is
- xeyownt 3y ago
  This C2PA looks flawed from the start. It just helps making the untrustworthy looks more trustworthy.
- zerojames 3y ago
  Reference, for context: https://c2pa.org/ https://c2pa.org/ And: the BBC just started using C2PA across some content. The BBC's R&D team talking about it: https://www.bbc.co.uk/rd/blog/2024-03-c2pa-verification-news-journalism-credentials https://www.bbc.co.uk/rd/blog/2024-03-c2pa-verification-news...
  - rentnorove 3y ago
    Related critique of the BBC's use of C2PA, and C2PA in general: https://www.hackerfactor.com/blog/index.php?/archives/1024-IEEE,-BBC,-and-C2PA.html https://www.hackerfactor.com/blog/index.php?/archives/1024-I...
    - dns_snek 3y ago

barfbagginus 3y ago

Prediction: soon we will want to label things written by actually useful AI, because it will be more helpful than things written by humans and parrot tier AI