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?
I’m considering using Power Delete Suite to delete my account, overwrite my previous comments, and maybe leaving a couple of my top comments up regarding tech support so people can still find information on troubleshooting
The issue is that most of the content posted is archived fairly quickly. Deleting/rewriting only hurts the humans that might have gone looking for it. The way I look at it is, if the data is searchable/indexable by search engines (as a proxy for all other tools) at any point of its life cycle then it’s essentially permanent.
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.
and maybe even our Mastodon or Lemmy posts!
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.com
and 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.Removed by mod
Exactly. If you’re posting something on the internet for the world to see, you can’t get upset when people or in this case AI read it.
It depends on if the data is suitably anonymized or not. If my data isn’t able to be reconstructed word for word in a way to directly links back to me? I don’t know if I mind that anymore then I’d mind someone reading content I wrote and taking inspiration from that.
On the topic of privacy - how do people feel Lemmy compares to Reddit for privacy? I don’t really like the way Lemmy handles deleted content for example.
Regardless of how anyone feels about their writing being used for model training, there’s definitely nothing anyone can do to prevent it other than just not writing anything visitble to the public.
Not yet, I think. If AI as regulated more strictly, users might get the chance of putting permission on their data. However that well look like. I hope it’s better than the cookie opt-out or do-not-track setting in your browser though.
I’ve been posting publically for years. I expect when I do, it was viewed and used by anyone any time for anything. AI hasn’t changed that.
Do I care? Sure, a little, someone is going to get paid and it’s not going to be me. There’s nothing I can do about it and my boss gets paid for my work too.
I’m okay with it as long as I’m aware of it. If the platforms are up front about it, then users can choose for themselves whether they want to potentially contribute to training data. It will be interesting to watch the next few years.
deleted by creator
I agree. I’ve always thought that AI would be our successor species, humanity’s child. I like that I might be some small part of its heritage.
I have a similar kind of idea. I think if it had been a free/open source/community project that made the headlines I would have been all like “this is so awesome”.
I guess what I don’t like is the economic system that makes that impractical. In order to build one of those giant GPTs, you need tonnes of hardware (capital), so the community projects are always going to be playing catchup, and I think quite serious catchup in this arena. So the economic system requires that instead of our posts going to a “collective hive mind” that aid human knowledge, they go to some walled garden owned by OpenAI, which filters and controls it for us, and gives us little bits of access to our own data, as long as we used it only in approved ways (i.e., ways that benefit them).
deleted by creator
Most of the data used in training GPT4 has been gathered through open initiatives like Wikipedia and CommonCrawl. Both are freely accessible by anyone. As for building datasets and models, there are many non-profits like LAION and EleutherAI involved that release their models for free for others to iterate on.
While actually running the larger models at a reasonable scale will always require expensive computational resources, you really only need to do the expensive base model training once. So the cost is not nearly as expensive as one might first think.
Any headstart OpenAI may have gotten is quickly diminishing, and it’s not like they actually have any super secret sauce behind the scenes. The situation is nowhere as bleak as you make it sound.
Fighting against the use of publicly accessible data is ultimately as self-sabotaging ludditism as fighting against encryption.
GIGO - Garbage In, Garbage Out. I asked ChatGPT to write a short essay and include a bibliography with URL 's. Every URL was a 404, and when looking up the bibliographic entries, they were nonexistent as well.
That’s because you don’t understand the tool you are using and use tech-sounding language in the wrong context to look like you do.
GPT models generate text based on the patterns of the tokens it learned during training. The URL it gives you doesn’t work because they have to only look legit. It’s all statistical patterns.
It’s not because they fed it garbage during the semi-supervised training, it’s because that literally is what the tool is meant for. Use the right tool like google scholar if what you need are sources.