Rebuilding Epistemic Trust in an AI-Mediated Internet

A video goes viral of a health official “admitting” that vaccines weren’t tested. It spreads fast—first as shock, then outrage, then a cascade of sadness, sarcasm, confusion, even jokes. People send it to group chats with “???” or “can this be real??” It looks authentic enough to believe in the moment: perfect lighting, familiar voice, even the right inflection. Only later do we learn it’s a deepfake. But by then, the damage isn’t the lie itself; it’s the lingering uncertainty it seeds. When anything can look true, everything starts to feel unstable.

That erosion of confidence isn’t limited to public health. It’s crept into climate communication, election reporting, and even education.¹ People aren’t dismissing evidence; they’re struggling to weigh it amid competing claims that all sound equally sure. The result isn’t disbelief so much as fatigue: an uncertainty about what (or whom) to trust.

As AI systems increasingly filter how information reaches us, the challenge isn’t only to correct falsehoods, but to rebuild the mechanisms of trust that those systems have replaced. Fact-checks and content warnings help, but they won’t rebuild confidence on their own. To restore epistemic trust, we’ll need systems that earn it. That means making AI-generated information traceable to its sources, teaching models to indicate when they’re unsure, and surfacing multiple credible perspectives instead of a single polished answer. By designing environments where transparency feels natural, uncertainty feels honest, and credibility finally feels visible again, behavioral science can help make those changes stick.