It's getting harder to know what's real on the internet. It's getting harder to know who's real on the internet. Large language (AI) models have enabled bad actors to create imposter accounts at an unprecedented scale, consuming our attention as they try to choke our discourse. I fear it's only just begun.
How, then, can we combat this? AI text detection tools are an appealing idea, yet I can't see them as a solution. Sussing out those patterns seems like an endless game of cat and mouse, likely to sweep up real people as false positives besides. The truth of whether an account is a bot isn't something we can expect technology to reveal for us. This problem is social. It's a question of our beliefs and relationships, our trust in each other.
Once, a blue check mark on Twitter indicated that a user had been verified. Seeing that mark meant Twitter was vouching for them. In this, Twitter was trustworthy. Verification wasn't for sale; it represented Twitter transmitting the trust people had in it to the people it verified. Despite occasional mistakes, Twitter broadly succeeded in creating a useful signal that helped users know who was real. This is a hint: a trusted organization was able to provide useful information to many, if only about a few notable users.
Yet Twitter is gone; X marks the spot where it's buried. One man gave the order to plunder that trust. X's reworked "verification" system cashed in on people's perceptions of that mark by selling it. That trust was betrayed, creating comical situations like when a "verified" Eli Lilly posted that insulin would now be free.
After seeing cases like this, people rightfully reduced their trust in X and its claims. That once-valuable signal is gone, replaced by a pale imitation.
Nonetheless, X's check mark still does correlate with whether an account is human. It's not a terribly strong signal, mind you, but $89/yr per account is a real cost for someone trying to make a massive AI army. While only a fool sees that mark and concludes the holder is human, being a little more inclined suspect so is reasonable. This too is a hint: these signals aren't all or nothing. People might reasonably differ on just how strong such an indicator is.
Here's a key insight: while the truth may be a binary, our estimation of it isn't. These systems are usually displayed as a badge that either exists or doesn't, poorly modeling the depth of this information space. Similarly, one-size-fits-all solutions are similarly poor models. Trust and belief are inherently personal. A single organization providing an authoritative measure can hardly hope to do much better than Twitter did. Such a system can only be applied to a small minority before encountering the costs and risks true scale brings.
That reflection of trust from a single institution onto a few public figures won't solve our problem, but we can learn from it. We already participate in a rich network of relationships and trust; it already informs who we think is real. Twitter's verification was a tiny corner of that network, uniquely useful due to how it was displayed. The rest of that information is mostly inaccessible. It travels slowly, if at all. It's almost certain that a short chain of connections link you and anyone else0, but finding that path isn't practical. We need something as conveniently informative as Twitter's blue check based on our relationships instead of just a single company.
We'll begin by publishing those connections on an open protocol, preventing a single company from gatekeeping access. We'll testify whether others are bots or humans then weigh that info according to how much we trust its source. If my trusted friend claims they know someone is human, that's meaningful. I have some level of trust in everyone that friend trusts, so their thoughts are valuable too. I might trust an institution that verifies a few notable identities. By analyzing this network, we can compute useful, personalized signals to help determine who's real.
I'd love to hear your thoughts. Drop a comment below or subscribe to get updates on the project. Collaborators are always welcome. I'll add a link here to the full proposal once I publish it; a demo will be ready by ATmosphereConf 2026.