Bayesian Theory of Truth: Prediction Accuracy as the Only Test of Understanding

Forum topic · 小凯 · 2026-02-01

Summary

This forum post presents a poster-style infographic outlining a 'Bayesian theory of truth'—the idea that predictive power is the only standard for testing whether you understand the truth, and the more accurate your predictions, the closer you are to reality. The core thesis: truth is not merely 'seeing is believing,' but the ability to make accurate predictions based on evidence. The poster contrasts the 'trap of kindness'—where moral preferences suppress the prior probability of disturbing news—with the 'Bayesian antidote': if a shocking piece of news has extremely high likelihood (explaining previously unexplained events), belief in it must rise even if the prior is low. It also presents a four-step method for judging truth: (1) set an honest prior acknowledging bias, (2) evaluate evidence by comparing likelihoods under competing hypotheses, (3) perform a Bayesian update regardless of how surprising the result is, and (4) validate by prediction—if forecasts come true, you have touched the truth. The post closes with the aphorism: 'If you cannot predict, you do not understand. If you refuse to update, you have drifted from truth.'

A zhichai.net forum post shares an infographic-style poster titled "Bayesian Theory of Truth" (贝叶斯真相理论), built around the claim: *"Predictive ability is the only standard for testing whether you understand the truth. The more accurate your predictions, the closer you are to the truth."*

Core Thesis

Truth is not simply "seeing is believing" — it is the ability to make accurate predictions about the future based on evidence, expressed through the Bayesian framing:

> P(Truth | Evidence)

with the standard form: *posterior probability = (likelihood × prior probability) / normalization factor*.

Kindness vs. Truth

The trap of kindness: When facing unbelievable news, our priors are constrained by social morality and wishful thinking. Because we don't want to believe the world is so dark and complicated, we assign an extremely low subjective probability to "cruel truths."
The Bayesian antidote: If a piece of news, however shocking, has extremely high *likelihood* — i.e., it perfectly explains a series of previously unexplainable events — then even with a very low prior, we must substantially raise our confidence in it.

The Four-Step Method for Judging Truth

1. Set a prior — acknowledge subjective bias and honestly state your initial probability judgment. 2. Evaluate the evidence — compare how likely the evidence is if the hypothesis is true versus false. 3. Bayesian update — compute the posterior and adjust your beliefs, no matter how surprising the result. 4. Validate by prediction — make predictions based on your new beliefs. If they come true, you have touched the truth.

Closing Quote

> "If you cannot predict, you do not understand. If you refuse to update, you have drifted away from the truth."

The poster credits its design to an AI assistant, based on Bayesian statistics and cognitive psychology.