We Need to Prioritize Digital Identities

post by cheeseglacier · 2025-04-01T21:14:11.651Z · LW · GW · 0 comments

Contents

No comments

LLMs are getting better and better at evading CAPTCHA’s and generating human-like text. If we can't verify human users, online discourse might soon be dominated by those who have the computing power to run the most bots. Our views and values are shaped by social consensus, so we need prevent bots from overwhelming online platforms

It baffles me that we don’t have an easy way to verify unique human users online yet. We're stuck in a situation where you can never fully trust a given poll/upvote count/comment section/recommendation feed since it might have been spammed by alt accounts, bots, or click-farms. The web of verifications we use online between google, twitter, and facebook accounts might not hold up to attacks from advanced LLM's that can create multiple profiles and cross-verify them.

The policy solution to bot manipulation is simple: develop a robust digital identity system that integrates with online platforms. It should take no more than two clicks to login to twitter/lesswrong/substack and verify that you're a real person from a real country. In the United States, proving your identity means uploading an ID or social security number to an untrusted site, which is cumbersome, insecure, and privacy invasive; most sites don't ask you to verify your identity except for where it's legally necessary like for banking. But with a private, simple system, I think a lot more sites would be able to verify human identities and prevent bot manipulation.

Verifying human users will also help us align AI models. Currently, LLM’s are pre-trained on an increasingly an AI generated internet, which might introduce idiosyncratic values and beliefs. If AI companies want to make sure their models are aligned with “human values,” they’re going to need a way to accurately measure them online.

Note: to allay the concerns of civil libertarians, the government can put in safeguards to prevent government overreach; the system could be implemented at the state or local level, be open source and auditable, have selective privacy capabilities, and could be implemented alongside civil protections on its usage by governments. To protect whistle-blowers and anonymity, the system can be built to verify only traits you want like your citizenship without giving your personal identity or name to the platform.

0 comments

Comments sorted by top scores.