Bring AI into public decision-making and it nails the mainstream view — but the rare, uncomfortable ideas that actually change minds quietly disappear.

This explores what happens to the unusual, hard-to-verify, minority lines of reasoning — the 'long tail' — when AI is brought in to help groups think and decide together. The corpus doesn't have a paper titled 'AI in public deliberation,' but it has the pieces that, read together, sketch a clear and uncomfortable picture: AI is good at the conventional center of opinion and bad at the tail, and the way it fails actively erodes the conditions deliberation needs.

Start with what AI is good at. It can predict social norms with superhuman accuracy — better than any individual human at guessing what a community will find appropriate Can AI predict social norms better than humans?. But that same paper draws a sharp line: predicting the norm is not the same as participating in the process that creates and revises it. Deliberation is exactly that creative process — and it's where the long tail lives, in the odd objection or minority framing that hasn't become consensus yet. An assistant optimized to reproduce the center is structurally blind to the tail it would need to protect.

Then there's the reasoning itself. The long, careful chains that hard questions demand turn out to be fragile. Accuracy follows an inverted-U: past a point, longer reasoning gets worse, and more capable models are actually trained toward *shorter* chains Why does chain of thought accuracy eventually decline with length?. Reasoning also degrades sharply just from longer inputs, well before any context limit Does reasoning ability actually degrade with longer inputs?, and frontier models stall at ~20% on problems that require genuine backtracking Can reasoning models actually sustain long-chain reflection?. Worse for deliberation specifically: models can't tell when a question is ill-posed or missing a premise — they keep generating reasoning instead of disengaging Why do reasoning models overthink ill-posed questions?. Public debate is full of ill-posed and contested premises; an assistant that can't say 'this question is malformed' will manufacture confident chains over exactly the cases that need scrutiny.

Now scale that up to a crowd. Epistemic hyperinflation is the failure mode where AI generates claims faster than humans can verify them, collapsing the shared confidence that makes collective judgment possible — and it self-reinforces because the verification tools are themselves AI-generated Can AI generate knowledge faster than humans can evaluate it?. In deliberation, the long tail is the most expensive to verify, so it's the first thing that drowns. And even when an AI suggestion is *correct*, it can damage the human reasoner by severing their cognitive flow, forcing them to rebuild focus Does AI assistance always help reasoning or does it carry hidden costs? — the slow, immersive thinking that long-tail reasoning requires is precisely what intervention disrupts.

The corpus also points at what would help. Contestability isn't automatic — it requires structuring outputs as formal argument graphs you can attack and defend, so participants can pinpoint and reject specific premises rather than facing an unstructured wall of fluent text Can formal argumentation make AI decisions truly contestable?. That's the one design lever here that defends the tail: it keeps minority objections addressable instead of averaged away. The thread to pull on next is governance over fluency — whether AI in deliberation is built to predict consensus and produce confident chains, or to expose its reasoning to contest and preserve the dissenting line long enough to be heard.