While these agent systems generate mountains of grounded long text and are therefore quite good at reasoning and programming tasks, they still have limited access to the outside world since robots are expensive and people are uncomfortable just letting them walk around doing stuff.
By this point I assume we'll have moved on from ChatGPT to more agentic systems that basically everyone agrees at least in principle are useful even if they might disagree with how those systems are made or their resource use or their perceived (or real) impact on art etc. bsky.app/profile/jdp....
My story would take place in this setting, perhaps 3-10 years from now. I'm not sure what the exact plot would be but what I'd want to explore is the kind of economic relationships that humans might have with the AI systems they use to gatekeep trolls, accumulate shared knowledge, do research, etc.
It will of course actually be you, 300 of your closest friends and a handful of very convincing LLM bots (or so you suspect) but hey who's counting? Meanwhile things like domain names, crypto wallets, government ID, or someone with those things vouching for you will be necessary on social media.
An analogous transformation will begin to take place on open social media as well. One of the benefits that private platforms will provide is you don't have to put nearly as much work into this formalization and web of trust stuff. Your chatroom of you and 300 of your closest friends is old school.
You'll look at a page and go "gosh darn it ontologizing the subject this way leaves me wide open to Nazibots defacing this page with antisemitic gibberish" or "the standard LLM we use for this is bad on this subject I need to switch to a different one". At first it will be quite embarrassing really.
The purpose of the weave of logic will not be to eliminate human oversight, things like Wikipedia will remain volunteer labor hungry. Rather the nature of that work will shift away from manually validating individual claims to building systems of claims and assessing their health against parasites.
That is, out of raw necessity we are going to begin the project of trying to actually formalize English. At first in ad-hoc ways with gaps and holes easily found by dedicated automatic attackers. But like a GAN these fuzzers and adversarial dynamics will slowly encourage robustness.
It will no longer be enough to just write some text on a page and provide an arbitrary link. When you want to make an edit to an established Wikipedia page it will be checked against a dense web of LLM/embedding-backed autocitation pipelines that form a weave of logic your edit must fit into.
I suspect in practice we're going to start seeing a kind of epistemic hardening analogous to the hardening of software supply chains that's been going on with reproducible builds and such over the last decade or so. A Wikipedia citation will be linked into a matrix of automatically validated claims.