I have been thinking a lot about how this recent LLM-explosion affects the way the internet works, and one worrying aspect is a loss of trust that the content was written by a fellow human being.I believe that content produced by bots should be clearly labeled/marked as such. Unfortunately given the incentives at play this is not a feasible solution, so I would choose the next best version, verified humans’ content should be labeled (similarly to Twitter’s old blue checks), all across the internet.I don’t have all the technical details worked out, but let me give a summary of one potential model.It would be basically a simple API that would accept a website-user pair, and return whether it is a verified human being. For the verification algorithm I’m thinking of a chain of trust, similar to Lobste.rs’s invitation feature. That would also help scale out the “real-world verification” process, I can easily vouch for a couple of people, but not for hundreds/thousands of users. Also, could aid in banning a network of malicious users (e.g. someone inviting a lot of accounts that turned out to be bots, more on this later). Maybe a public key could also be registered for a database entry, so that platformless content could also be validated by signing.In the early, growing phase, it could be as simple as a browser extension that given a website it supports would simply insert a checkmark next to the usernames of verified users.Who can be considered verified is not a trivial question either. It definitely has to be private, and we wouldn’t want to link together a single person’s multiple accounts (neither within a single website, or multiple ones) as that might even end up exposing someone’s identity. Maybe the simple way of semi-anonymous users asking for an invitation from already verified users, who briefly review their comment history would be a sufficient gate for now. Perhaps higher certainty levels could be differentiated later, e.g. a public figure on a platform.Moderation is a necessary feature, perhaps something as simple as being able to flag users with questionable content (guidelines have to be written, but I believe that comments (or parts of them) that are written by AI but marked as such, as well as certain non-marked uses are more than fine) and after multiple issues their validation might be revoked (either just removed from the database, or adding another state to the above boolean function that would mark such “users” on websites as bots). Other metrics could also be used to make a decision, such as post-frequency.All in all, user verification can arguably accept some false positives/negatives and the database doesn’t have to be completely in sync, so it is highly cache-able and can be served from many different servers, so scaling should not be an issue.I would have liked to post a working POC here, but I am interested in its reception, and I’m certain that there are many aspects I haven’t thought of.
via by gf0