Prior Belief: The Adversarial Model
The dominant account in cognitive psychology treats heuristics and Bayesian inference as incompatible strategies. The Bayesian ideal requires an agent to maintain coherent probability distributions over hypotheses and update them exactly on each observation. Heuristics — availability, representativeness, anchoring, Take-the-Best — are characterised as departures from this standard: computationally cheaper, systematically biased.
Two research programs structured the debate. The heuristics-and-biases program (Kahneman and Tversky[2]) catalogued conditions under which human judgment deviates from normative standards. The fast-and-frugal program (Gigerenzer and colleagues[1]) documented conditions under which the same heuristics outperform elaborate strategies in ecologically valid environments. Both programs accepted that exact Bayesian computation does not describe human behaviour. The disagreement was over what that fact implies.
The adversarial framing assumed a fixed, narrow definition of Bayesian optimality — one that ignores the cost of computation. Three independent lines of evidence undermine that assumption and support a revised position: many heuristics are not approximations to Bayesian inference. Under specifiable conditions, they are the Bayesian-optimal strategy.
Making many quick decisions based on very few samples may be the globally optimal strategy over long periods.
Core finding — Vul et al. (2014)
Formal Equivalence
Parpart et al. (2019)[3] proved that Take-the-Best is mathematically equivalent to a Bayesian model under a Laplacian prior over cue orderings. The equivalence is exact, not approximate. Given that prior, Take-the-Best does not deviate from Bayesian inference — it implements it.
This result reframes the evaluative question. The relevant question is not whether a heuristic deviates from the Bayesian standard, but what prior renders it the correct Bayesian strategy. For Take-the-Best, that prior is well-motivated in environments with sparse, ordered cue structures — a common natural condition.
Sampling Approximation
Exact Bayesian computation over realistic hypothesis spaces is intractable. Marginalising the posterior requires integrating over exponentially many states. No biological system does this. The second line of evidence concerns what biological systems do instead.
Vul et al. (2014)[4] showed that one-sample Monte Carlo estimation of the posterior — selecting a single hypothesis and acting on it — is globally optimal over long decision sequences when sampling is cheap and decisions are frequent. Evaluated on individual decisions, single-sample reasoning appears irrational. Evaluated across sequences, it is not. The apparent bias is an artefact of the evaluation frame.
Chater et al. (2020)[5] reviewed the broader pattern, finding that behaviour characterised as heuristic shortcutting is largely consistent with approximate Bayesian computation — stochastic estimation of the posterior, not departure from probabilistic reasoning.
Resource Rationality
Lieder and Griffiths (2020)[6] formalise the unifying principle as resource rationality: behaviour is evaluated against the optimum achievable given actual computational costs, not against an unconstrained Bayesian ideal. Cognitive effort is treated as a limited resource; optimal behaviour maximises expected accuracy per unit of cognitive cost.
Under this framing, heuristics are not approximations to the correct computation — they are the correct computation for an agent with bounded resources. Anchoring, for example, minimises the number of costly adjustments required from a starting estimate. The heuristic optimises the right objective; the prior analysis was optimising the wrong one.
Bruckner et al. (2020)[7] provide supporting evidence from development: children and older adults — populations whose cognitive constraints differ from young adults in different ways — show stronger reliance on prior beliefs when learning under uncertainty. This pattern is consistent with resource rationality's prediction that tighter computational constraints shift inference toward prior-heavy strategies.
The Evidence Map
| Study | Claim | Method | Citations |
|---|---|---|---|
| Parpart et al. (2019)[3] | Take-the-Best ≡ Bayes under Laplacian prior | Formal proof + simulation | 312 |
| Vul et al. (2014)[4] | One-sample Monte Carlo is globally optimal over sequences | Mathematical + behavioural | 453 |
| Lieder & Griffiths (2020)[6] | Resource rationality unifies Bayesian and heuristic accounts | Theoretical + empirical review | 891 |
| Chater et al. (2020)[5] | Human inference consistent with approximate Bayesian computation | Meta-analysis + modelling | 278 |
| Bruckner et al. (2020)[7] | Age-related differences in prior reliance consistent with resource constraints | Developmental + behavioural | ~145 |
| Xu & Tenenbaum (2007)[8] | Infants' word learning follows Bayesian generalisation | Developmental + computational | 1,204 |
| Gopnik et al. (2017)[9] | Children acquire Bayesian reasoning before explicit reasoning | Developmental | 387 |
| Mangalam (2025)[10] | Bayesian brain hypothesis remains falsifiable and unfalsified | Theoretical critique | 8 |
Ecological Rationality
Gigerenzer's ecological rationality account holds that heuristics are accurate in the environments that selected them. A heuristic performs well in a given task structure because that structure shaped the heuristic, through evolution or development. This is equivalent to saying the heuristic embeds a prior that is well-calibrated to the ecological distribution of that environment.
The ecological account and the Bayesian account are therefore not competing explanations. Ecological rationality describes how a prior becomes calibrated. Resource rationality describes why that prior is used rather than a more expensive computation. Formal equivalence describes the relationship between the resulting behaviour and the Bayesian optimum.
The pattern Kahneman and Tversky documented — systematic errors in judgment — is consistent with this account. Errors arise when heuristics with priors fitted to one environment are applied to another. This is the predicted consequence of deploying a well-calibrated prior outside its calibration domain, not evidence of non-Bayesian reasoning.
The heuristic is not deviating from the Bayesian ideal. It is the Bayesian ideal, instantiated under a particular prior.
Synthesis from Parpart et al. (2019)
Developmental Evidence
Any account of inference must explain how the relevant computations are acquired. Xu and Tenenbaum (2007)[8] showed that infants learning word meanings prefer hypotheses consistent with a size principle — the generalisation pattern a Bayesian agent with that prior would produce. Bayesian-consistent behaviour precedes explicit linguistic or numerical reasoning.
Gopnik et al. (2017)[9] found that children explore hypothesis spaces more broadly than adults or adolescents — a pattern consistent with Bayesian inference under flatter priors about environmental structure, which is an appropriate prior when the environment is not yet modelled. The developmental trajectory from broad exploration to selective use of heuristics is consistent with prior calibration over time.
Falsifiability
The unfalsifiability objection — that any behaviour can be rationalised by selecting an appropriate prior — is methodological, not principled. Bayesian models generate testable predictions when priors and likelihoods are specified and fixed in advance of data collection. Mangalam (2025)[10] argues this point directly: the difficulty is in constraining models sufficiently, not a property of the framework.
Properly specified Bayesian models fail when their priors are wrong. That falsifiability condition holds — the literature contains numerous cases of Bayesian models that were tested against alternatives and rejected. The framework is not immune to disconfirmation.
Posterior
The prior — that heuristics and Bayesian inference are adversarial — does not survive this evidence. The updated position:
- Formal equivalence is established for key heuristics. Take-the-Best is mathematically identical to Bayesian inference under a Laplacian prior. The equivalence is not approximate.
- Low-sample reasoning is globally optimal under realistic conditions. One-sample Monte Carlo is the correct Bayesian strategy when sampling is cheap and decisions are frequent. Evaluating it on single decisions misidentifies the objective function.
- Computational cost resolves the apparent conflict. Resource rationality closes the gap between the heuristic account and the Bayesian account. Given accurate cost modelling, the two accounts converge.
- Bayesian-consistent behaviour precedes explicit reasoning developmentally.Infants and children exhibit Bayesian generalisation patterns before they can articulate probabilistic reasoning, suggesting the underlying computation is not a sophisticated learned overlay.
- Systematic errors are miscalibrated priors, not framework failures.Biases arise when a heuristic — calibrated to one environment — is applied to another. This is the predicted output of Bayesian inference under a mismatched prior.
The revised position: heuristics are the output of a Bayesian inference process operating under real computational constraints. The biases documented by Kahneman and Tversky are real. Their explanation is Bayesian.
BayesCore extracts the evaluation criteria implicit in any structured document and scores it against them — Bayesian confidence per predicate. Free to start.
Evaluate free →Citation counts approximate as of early 2025.