It feels like we have a new privacy threat that’s emerged in the past few years, and this year especially. I kind of think of the privacy threats over the past few decades as happening in waves of:
- First we were concerned about governments spying on us. The way we fought back (and continue to fight back) was through encrypted and secure protocols.
- Then we were concerned about corporations (Big Tech) taking our data and selling it to advertisers to target us with ads, or otherwise manipulate us. This is still a hard battle being fought, but we’re fighting it mostly by avoiding Big Tech (“De-Googling”, switching from social media to communities, etc.).
- Now we’re in a new wave. Big Tech is now building massive GPTs (ChatGPT, Google Bard, etc.) and it’s all trained on our data. Our reddit posts and Stack Overflow posts and maybe even our Mastodon or Lemmy posts! Unlike with #2, avoiding Big Tech doesn’t help, since they can access our posts no matter where we post them.
So for that third one…what do we do? Anything that’s online is fair game to be used to train the new crop of GPTs. Is this a battle that you personally care a lot about, or are you okay with GPTs being trained on stuff you’ve provided? If you do care, do you think there’s any reasonable way we can fight back? Can we poison their training data somehow?
The biggest problem to me is what I just saw you post in another reply, that these models built upon our knowledge exist almost solely within proprietary ecosystems.
The Washington Post published a great piece which allows you to search which websites were included in the “C4” dataset published in 2019. I searched for my personal blog
jonaharagon.comand sure enough it was included, and the C4 dataset is practically minuscule compared to what is being compiled for larger models like ChatGPT. If my tiny website was included, Mastodon and Lemmy posts (which are actually very visible and SEO optimized tbh) are 100% being scraped as well, there’s no maybe about it.Thanks for linking to that, I hadn’t seen that article before. Interesting seeing it broken down like that and being able to search for a website to see if it was part of the training data