# Bayes' Theorem **The model:** A formal rule for how to update what you believe in light of new evidence — and, more fundamentally, a theory of *what it means to reason well*. Bayes' theorem states that your revised belief (posterior) should depend on three things: what you believed before the evidence arrived (prior), how likely the evidence would be if your belief were true (likelihood), and how likely the evidence is regardless (base rate). The theorem states this as a precise mathematical relationship, but its analytical power is mostly conceptual: it forces you to keep three numbers honest, two of which most people forget. The formal statement: **P(H|E) = P(E|H) × P(H) / P(E)**. Read aloud: the probability of hypothesis H given evidence E equals the likelihood of E under H, times the prior probability of H, divided by the overall probability of E. Every cognitive bias in the Kahneman-Tversky catalog, every reasoning failure in the Munger Biases folder, and every propaganda exploit in the Information Warfare cluster can be redescribed as a *specific violation of Bayes' theorem*. Base-rate neglect violates the denominator. Confirmation bias violates the likelihood by only counting evidence *for* and ignoring evidence *against*. Anchoring violates the prior by fixing it at an arbitrary starting point and under-adjusting. The theorem is the normative standard against which all of Decision Science measures human failure — and against which all of Information Warfare measures human exploitability. ## The Mechanism ### The Three Numbers Most People Forget **1. The prior matters.** What you believed before the evidence arrived is part of what you should believe after. The same piece of evidence is more or less compelling depending on how plausible the hypothesis was to begin with. A positive drug test from a random person is weak evidence (most positives from low-prevalence populations are false positives). A positive drug test from someone caught with paraphernalia is strong evidence (the prior is already high). Strong evidence for a wildly implausible claim still leaves the posterior at "probably not, but updated slightly." This is where most conspiratorial reasoning fails its first test. The claim "the election was stolen" requires a prior: how often have American presidential elections been stolen through coordinated multi-state fraud? The base rate is zero in modern history. Any evidence — statistical anomalies, affidavits, video clips — must overcome that prior. Vivid evidence against a near-zero prior produces a posterior that is higher than zero but still very low. Conspiracy reasoning treats the evidence as though the prior were 50/50 (it's either stolen or not, so it could go either way) — which is the mathematical error that makes everything downstream go wrong. **2. The likelihood ratio is what does the work.** The right question is not "could this evidence appear if the hypothesis is true?" — almost any evidence *could* appear under almost any hypothesis. The right question is: **how much more likely is this evidence under my hypothesis than under the alternative?** The likelihood ratio P(E|H) / P(E|¬H) is what actually moves the posterior. Evidence that is equally likely under both hypotheses — a statistical anomaly that would appear in a fraud scenario *and* in a no-fraud scenario — has a likelihood ratio near 1 and does essentially no updating work, no matter how vivid it is. This is the diagnostic tool for evaluating political claims: not "is this consistent with my theory?" (almost everything is consistent with almost every theory) but "is this *more* consistent with my theory than with the alternative?" If the evidence is equally expected under both hypotheses, it's noise, not signal. **3. Base rates dominate intuition.** The most robust finding in the heuristics-and-biases program: people consistently underweight base rates and overweight vivid, salient, case-specific evidence. Kahneman and Tversky documented this across dozens of experiments. The most famous demonstration: the "cab problem" (a witness identifies a hit-and-run cab as Blue; Blue cabs are 15% of the fleet; the witness is 80% accurate — most people say the cab is probably Blue, but the correct Bayesian answer is only ~41% Blue, because the base rate of Green cabs is so much higher that Green cabs generate more false identifications than Blue cabs generate true ones). The political version: "immigrants commit terrible crimes" → check the base rate (immigrants commit crime at lower rates than native-born citizens). The vivid anecdote (System 1 evidence) overwhelms the base rate (System 2 evidence) unless you discipline yourself to look for it. ### Multiple Updates and the Independence Problem Bayes lets you combine evidence sequentially. Each piece updates the prior, and the new posterior becomes the prior for the next piece. Multiple independent pieces of evidence can drive a posterior from improbable to near-certain. But — and this is the critical qualification for information-warfare analysis — **the updates compound only if the evidence is independent**. If your twenty data points all trace back to one source, or are correlated through shared methodology, shared bias, or shared coordination, they function as a single piece of evidence, not twenty. The right-wing media ecosystem's amplification model systematically destroys independence. A claim originates on one platform, is amplified by twenty outlets, is cited by thirty pundits, and appears in a hundred social-media posts. To System 1, this feels like a hundred independent confirmations. To Bayes, it's one data point with a very large megaphone. The mathematical structure of the information environment — its correlation structure — determines how much evidence you actually have. The media ecosystem is designed to make the correlation structure invisible. ## Bayes as Epistemology ### Beyond a Formula: A Theory of Reasoning Bayes' theorem is not just a tool for probability calculations — it's a *normative theory of belief change*. The Bayesian epistemological claim, developed most forcefully by E.T. Jaynes and the subjective-probability tradition, is that *rational belief updating just is Bayesian updating*. There is no alternative normative standard. Every other approach to changing your mind in response to evidence is either a special case of Bayes (when done right) or a violation of Bayes (when done wrong). This is a strong claim, and it's contested (see Critical Reception below), but its implications for this vault are profound: if Bayes is the normative standard, then every documented cognitive bias is a *specific, characterizable departure from Bayes*, and every propaganda technique works by *inducing a specific Bayesian violation* in the target population. ### The Bayesian vs. Frequentist Debate Two traditions in probability theory compete for foundational status: **Frequentist.** Probability is the long-run frequency of events. A coin has a 50% probability of heads because, in infinite flips, half will be heads. Frequentists reject the idea of assigning probabilities to one-time events or hypotheses — "what's the probability Trump wins the 2024 election?" is meaningless in the frequentist framework because the election happens once. **Bayesian.** Probability is a measure of uncertainty — a degree of belief. You can assign probabilities to one-time events, to hypotheses, and to any proposition about which you have incomplete information. Probabilities update as evidence arrives. This is the framework that permits Tetlock's superforecasting, intelligence analysis, medical diagnosis, and every other domain where you need to reason about unique events under uncertainty. For the vault's purposes, the Bayesian framework is overwhelmingly more useful. Political events are one-time occurrences; intelligence assessments concern unique situations; propaganda analysis requires evaluating belief change in individuals. The frequentist framework has nothing to say about these domains. The Bayesian framework says everything. ### Superforecasting Is Applied Bayesianism [[Philip Tetlock]]'s superforecasting research is the empirical demonstration that Bayesian reasoning works in practice. Superforecasters — the top performers in the Good Judgment Project — share a cognitive profile that is essentially "good Bayesian updating": they start with well-calibrated priors (base rates), update incrementally in response to new evidence, weight evidence by its diagnostic value (likelihood ratios, not vividness), aggregate multiple independent sources, and maintain calibrated uncertainty (they know what they don't know). The superforecaster profile is also a System 2 profile: deliberate, effortful, uncomfortable with premature closure, willing to hold multiple hypotheses simultaneously. The anti-profile — confident, fast, narrative-driven, uncomfortable with uncertainty — is the System 1 profile that dominates punditry, partisan media, and political reasoning. The accuracy gap between superforecasters and pundits is the gap between Bayesian and non-Bayesian reasoning applied to the same evidence. ## KB Applications ### Conspiracy Theories Run on Bayes-Violations The structural failure mode of conspiracy theories ([[QAnon]], [[Stop the Steal]], birtherism, anti-vax, climate denial) is systematic violation of all three Bayesian components simultaneously: **Prior violation.** Conspiracy theories implicitly set the prior at 50/50 ("it's either a conspiracy or it isn't") when the actual prior for large-scale coordinated conspiracies is extremely low. The base rate of successful, secret, multi-institutional conspiracies in open democracies is near zero — not because powerful people don't try, but because the coordination and secrecy requirements are almost impossible to sustain at scale. **Likelihood violation.** Conspiracy theories evaluate evidence by asking "is this consistent with the conspiracy?" rather than "is this *more likely* under the conspiracy than under normal explanations?" Since almost any anomaly is *consistent with* a conspiracy theory, this test admits everything and excludes nothing. The likelihood ratio does no work because the denominator (probability of this evidence under non-conspiracy explanations) is never calculated. **Independence violation.** Conspiracy-theory evidence is heavily correlated — the same anomalies recirculated through the same networks, each repetition feeling like a new data point. The QAnon ecosystem is the extreme case: "drops" are interpreted, reinterpreted, and cross-referenced in a process that generates enormous *apparent* evidence from tiny *actual* evidence, because each interpretation is treated as independent when it's derived from the same source. ### Election-Fraud Claims: A Worked Bayesian Example A concrete application: the claim "Biden received more votes than registered voters in this precinct." **Prior:** How often do American precincts report vote totals exceeding registered-voter counts due to fraud? Near zero in documented history. How often due to clerical errors, same-day registration, data-reporting timing, or registration-database lag? Routinely — in every election cycle, across thousands of precincts. **Likelihood ratio:** P(this anomaly | fraud) is moderate — fraud could produce this pattern. P(this anomaly | clerical error in a system processing millions of records) is high — this pattern is expected without fraud. The likelihood ratio P(E|fraud) / P(E|no fraud) is close to 1, possibly below 1. The evidence barely moves the posterior. **Base rate:** In 100,000+ precincts, the base rate of at least *some* anomalies appearing somewhere is effectively 1.0. The Texas Sharpshooter fallacy (paint the target after seeing where the bullets hit) compounds the problem: anomalies are cherry-picked from the full dataset, and the number of precincts examined is never reported alongside the number of anomalies found. **Posterior:** Very low probability of fraud, even after observing the anomaly. Most "Stop the Steal" arguments collapse under this analysis because they present evidence without priors, likelihoods without alternatives, and correlated data without independence checks. ### Medical Testing and COVID-Era Discourse The false-positive problem is Bayes' theorem's most intuitive teaching example — and it was directly relevant to COVID-era public discourse. A COVID test with 99% sensitivity (catches 99% of true positives) and 99% specificity (only 1% false positive rate) sounds nearly perfect. But in a population where 1% is actually infected: out of 10,000 people, 100 are infected and 9,900 are not. The test correctly identifies 99 of the 100 infected (true positives). It also falsely identifies 99 of the 9,900 uninfected (false positives). Total positives: 198. True positives: 99. **A positive test result means you're only 50% likely to be infected** — not 99%. The political dimension: anti-vax and COVID-skeptic discourse weaponized testing statistics (both false-positive rates and PCR cycle-threshold arguments) without Bayesian context. "The tests produce false positives!" is true but misleading without the base-rate context. Conversely, public-health messaging that failed to explain false-positive rates in low-prevalence populations undermined trust by overstating testing reliability. Both sides of the discourse failed Bayesian discipline — and the failure was consequential for policy credibility. ### Racial Profiling and "Suspicious Behavior" The base-rate problem is the mathematical core of the racial-profiling debate. Suppose a certain criminal behavior has a base rate of 0.1% in the general population and 0.3% among a specific demographic group — three times the base rate. A "profile" that targets the demographic group is detecting a *relative* difference that is large (3x) but an *absolute* difference that is tiny (0.2 percentage points). The vast majority of people in the profiled group (99.7%) are innocent. Profiling them imposes enormous costs (civil liberties, community trust, resource misallocation) to capture a marginal increase in detection probability. The Bayesian frame makes the trade-off explicit in a way that narrative framing ("they commit crimes at three times the rate!") obscures. The 3x multiplier sounds alarming; the 99.7% innocence rate sounds reassuring. Both are true. Bayes forces you to hold both numbers simultaneously, which is exactly what System 1 refuses to do. ### Adversarial Bayesian Environments Standard Bayesian reasoning assumes the evidence you observe is generated by nature — randomly, without strategic intent. In adversarial environments — intelligence analysis, information warfare, political disinformation — the evidence is generated by an agent who is *trying to manipulate your posterior*. This changes everything. **Dezinformatsiya as adversarial evidence injection.** Russian disinformation operations generate evidence (forged documents, staged events, planted stories) designed to shift the target's Bayesian posterior toward the operator's preferred conclusion. The Bayesian defense: treat evidence from adversarial sources with a lower likelihood ratio — not because it's necessarily false, but because the source's incentive to mislead reduces the evidence's diagnostic value. **Reflexive control as prior manipulation.** [[Reflexive Control]] works by shaping the target's information environment so that the target's Bayesian updating — even when done correctly on the *available* evidence — produces the operator's desired posterior. The manipulation is in the *evidence selection*, not the reasoning process. This is the deepest form of Bayesian attack: the target reasons correctly from manipulated inputs and reaches the wrong conclusion while believing they reasoned independently. **Flood-the-zone as likelihood-ratio destruction.** Bannon's strategy works Bayesianly by saturating the information environment with so many claims (true, false, half-true, misleading) that the target's ability to estimate likelihood ratios collapses. When you can't distinguish high-diagnostic from low-diagnostic evidence, all evidence receives similar weight, and the posterior is dominated by the prior — which is exactly what the flooding party wants if their goal is to prevent belief change (maintain priors) or to make all claims seem equally uncertain ("who knows what's true?"). **Preference falsification as prior distortion.** [[Preference Falsification]] distorts the evidence environment for Bayesian social reasoning. When everyone publicly supports the regime, your Bayesian estimate of regime support is high — because you're updating on public behavior, which is the only evidence available. The falsification makes your evidence systematically misleading, so your correctly-executed Bayesian update produces a posterior that is systematically wrong. The preference cascade happens when a shock reveals the true distribution, forcing a massive simultaneous update across the entire population. ### Intelligence Analysis Intelligence assessment is Bayesian reasoning under adversarial conditions — the domain where Bayes matters most and fails most consequentially. **The Iraq WMD failure** is a Bayesian case study. Analysts' priors were high (Saddam had previously possessed WMD and had used them). The evidence they observed (Saddam's obstructive behavior, claims from sources like Curveball, satellite imagery of suspicious sites) had moderate likelihood ratios. The posterior — "Saddam probably has WMD" — was a reasonable Bayesian conclusion *from the evidence available*. The failure was twofold: the evidence was not independent (multiple sources traced to the same unreliable channels), and the likelihood ratios were miscalculated (Saddam's obstructive behavior was equally likely under "has WMD" and "is bluffing Iran about having WMD" — a likelihood ratio near 1 that should have done no updating work). **The Bayesian lesson for intelligence:** the quality of the posterior is bounded by the quality of the inputs. Bayesian reasoning applied to contaminated evidence produces contaminated conclusions. The defense is *source criticism* (how independent is this evidence?), *alternative-hypothesis discipline* (what's the likelihood ratio relative to the best alternative explanation?), and *calibrated humility* (how confident should I be given the quality of my inputs?). ## The Dual-Process Connection Bayesian updating is a System 2 operation. It requires holding multiple numbers in working memory, comparing hypotheses, calculating ratios, and overriding System 1's confident-but-wrong intuitive answer. System 1 provides the vivid evidence, the emotional weight, and the narrative frame; System 2's job is to run the Bayesian check before accepting System 1's conclusion. The entire heuristics-and-biases program can be restated as: *System 1 is a bad Bayesian*. Specifically: - **Availability heuristic** — System 1 estimates probability from ease of recall (a proxy) rather than from base rates (the correct input). Bayesian violation: wrong prior. - **Representativeness heuristic** — System 1 estimates category membership from similarity rather than from base rates. Bayesian violation: prior ignored, likelihood assessed by resemblance rather than diagnostic value. - **Anchoring** — System 1 sets the prior at an arbitrary anchor and under-adjusts. Bayesian violation: corrupted prior, insufficient updating. - **Confirmation bias** — System 1 searches for evidence consistent with the hypothesis and ignores evidence against it. Bayesian violation: asymmetric likelihood estimation (only the numerator, never the denominator). - **Base-rate neglect** — System 1 ignores P(E) entirely. Bayesian violation: missing denominator. The implication: Bayesian reasoning is not a natural cognitive mode — it's a trained discipline that requires *activating System 2 against System 1's defaults*. This is why Bayesian thinking feels effortful, counterintuitive, and unsatisfying: it is. It's also why it works. ## Critical Reception ### Taleb's Objection: Fat Tails Break Bayes [[Black Swan]]'s critique: Bayesian reasoning assumes you can estimate meaningful priors and likelihood ratios. In fat-tailed domains (financial markets, geopolitics, pandemic risk), the priors are unknowable and the likelihood ratios are dominated by events that have never occurred. Bayes works for "Mediocristan" (normal-distribution domains) and fails for "Extremistan" (power-law domains). The practical lesson: in fat-tailed domains, Bayesian confidence is itself overconfident. Build robustness (margin of safety, antifragility) rather than relying on Bayesian predictions. The synthesis for the vault: Bayes is the best available tool for updating beliefs under uncertainty *within* a known distribution. It's insufficient for events that fall outside any known distribution. Both truths matter. Most political reasoning failures are Bayesian failures (fixable by better base rates and likelihood ratios); some are black-swan failures (fixable only by robustness). ### Gigerenzer's Critique: Natural Frequencies Beat Formulas [[Gerd Gigerenzer]] argues that presenting Bayesian problems as formulas (P(H|E) = ...) guarantees failure because human cognition doesn't process abstract probabilities. Presenting the same problems as *natural frequencies* ("out of 1,000 people, 10 have the disease; of those 10, the test catches 9; of the 990 healthy, the test falsely flags 99") produces dramatically better reasoning — even from people who cannot solve the formula. The practical lesson: Bayesian reasoning is learnable, but the pedagogy matters. Frequency formats activate the right cognitive machinery; probability formats don't. ### The Problem of Prior Selection The deepest philosophical challenge: where do priors come from? If Bayesian updating is the normative standard, the prior is the one input the theorem doesn't supply — it must come from somewhere. In practice, priors come from base rates, from domain knowledge, from previous updating, and (inevitably) from cultural assumptions, ideological commitments, and identity. The recognition that priors are culturally shaped means that two Bayesian reasoners processing the same evidence can reach different posteriors — and both can be "correct" given their priors. This is the mathematical formalization of why political disagreements are so hard to resolve: if you start with different priors (different life experience, different trust in institutions, different media diet), the same evidence updates you to different places, and *both updates are Bayesian*. The practical implication for the vault: disagreement is not always error. Sometimes it's prior divergence. The diagnostic question is: are the priors responsive to evidence (good Bayesian) or fixed regardless of evidence (bad Bayesian / motivated reasoning)? Prior rigidity — refusing to update when evidence accumulates — is the Bayesian definition of irrationality. ## Practicing Bayes 1. **Always ask for the base rate.** Whenever someone presents vivid evidence, ask: under what null hypothesis would this evidence be unsurprising? If the null hypothesis predicts the evidence too, the evidence isn't doing the work it's claimed to do. 2. **Calculate the likelihood ratio, not just the likelihood.** "Could this evidence appear if my theory is true?" is the wrong question. "Is this evidence *more likely* under my theory than under the alternative?" is the right one. 3. **Check independence of sources.** Multiple confirming sources matter only if they're independent. If they share data, methods, or coordination, treat them as a single source. The media ecosystem systematically destroys independence in ways you must actively track. 4. **Update incrementally.** Don't go from prior to posterior in one big leap. Each piece of evidence is one update. Strong opinions formed from a single piece of evidence are usually Bayes-violations. 5. **Be willing to have weak posteriors.** Bayesian thinking often leaves you mildly unsure rather than confident. The discomfort of mild uncertainty is the discipline working as designed. [[Doubt-Avoidance Tendency]] (Munger #4) is the cognitive mechanism that makes this difficult. 6. **Distinguish adversarial from natural evidence.** In adversarial environments (politics, intelligence, information warfare), the evidence is generated by someone trying to manipulate your posterior. Downweight the likelihood ratio accordingly and search for independent confirmation. 7. **Express beliefs in probabilities, not certainties.** "I think there's a 30% chance X is true" is more honest, more useful, and more updatable than "X is true" or "X is false." Calibrated uncertainty is the Bayesian's signature. ## See Also - [[Dual-Process Theory]] — Bayesian reasoning is System 2 work; every cognitive bias is System 1 failing to be Bayesian - Bayesian Game — game theory with Bayesian players; strategic interaction under incomplete information - [[Preference Falsification]] — distorts the evidence environment, making Bayesian social reasoning systematically wrong - [[Signaling Game]] — signals are evidence for Bayesian updating; the separating condition is a likelihood-ratio story - [[Black Swan]] — Taleb's argument that Bayes breaks in fat-tailed domains - [[Confirmation Bias]] — the Bayes-violating cognitive default: asymmetric evidence weighting - Neglecting Probability — direct base-rate failure - Anecdotal — the fallacy form of base-rate neglect - Burden of Proof — implicitly sets the prior at 50/50 when it shouldn't be - [[Reflexive Control]] — adversarial prior manipulation; the deepest Bayesian attack - [[Game Theory Hub]] — hub page - [[Daniel Kahneman]] — the heuristics-and-biases tradition that documented systematic Bayes-violations - [[Philip Tetlock]] — superforecasting as applied Bayesianism