Belief Updating and the Question of Cumulative Knowledge in Behavioral Economics

In economics, the normative model of how an agent comes to know the world is Bayesian updating: a prior belief is combined with the likelihood of newly observed data to yield a posterior, and that posterior becomes the prior for the next observation (Savage, 1954). The Bayesian apparatus is attractive precisely because it describes knowledge as cumulative: under a correctly specified model, an agent who keeps observing informative data will, in the limit, concentrate belief on the truth. This convergence property is the implicit benchmark against which any claim about "cumulative knowledge" must be measured (Benjamin, 2019).

Behavioral economics takes this benchmark seriously not because people satisfy it, but because departures from it are systematic, measurable, and consequential. The research program therefore has two halves: a positive account of how real belief revision differs from Bayes' rule, and an account of what those differences imply for whether knowledge accumulates at all - in the head of a single decision-maker and in the literature of the field.

Two canonical departures anchor the literature. In base-rate neglect, agents underweight the prior relative to new evidence and therefore over-update; in conservatism, they underweight the likelihood and therefore under-update (Benjamin, 2019). The two coexist and can even appear in the same person depending on how the problem is framed, which already signals that the errors are not a fixed "bias parameter" but a response to the structure and difficulty of the inference (Benjamin, 2019).

Around this core sit a family of related regularities: confirmation bias, in which ambiguous evidence is read as supporting a maintained hypothesis (Rabin & Schrag, 1999); correlation neglect and sample-size insensitivity, in which agents misjudge how much independent information a body of data contains (Benjamin, 2019); and motivated beliefs, in which people update asymmetrically toward conclusions they find pleasant or ego-protective (Benabou & Tirole, 2016). Benjamin (2019) argues that much of this can be traced to a single heuristic - representativeness - whereby people judge probabilities by similarity rather than by Bayes' rule.

A more recent literature proposes that many of these phenomena share a common mechanism rooted in imperfect information processing rather than in distinct preferences or heuristics. Enke and Graeber (2023) introduce and measure "cognitive uncertainty" - a person's subjective uncertainty about what the optimal action or belief actually is - and show that when cognitive uncertainty is high, people compress their responses toward an intermediate "mental default," such as a 50:50 probability. This single force ties together the probability-weighting function in choice under risk, base-rate insensitivity and conservatism in belief updating, and predictable over- and under-optimism in forecasts (Enke & Graeber, 2023).

Enke, Graeber, Oprea, and Yang (2024) generalize the idea into "behavioral attenuation": across roughly thirty experiments spanning choice, valuation, belief formation, strategic games, forecasting, and information acquisition, people's decisions respond too weakly to the underlying fundamentals, and the degree of attenuation rises with the complexity of the problem and with measured cognitive uncertainty. In about 93 percent of their experiments the elasticity of decisions to fundamentals decreases in cognitive uncertainty, and in problems that have an objectively correct answer the elasticities are universally smaller than is optimal (Enke et al., 2024). Crucially for the present discussion, the authors emphasize that attenuation is not confined to matters of taste:

"An important indication that behavioral attenuation primarily reflects imperfect information processing is that it is pervasive both in choice contexts that involve people's preferences and in situations in which an objectively correct decision exists, such as belief updating, forecasting, and information acquisition problems." (Enke, Graeber, Oprea & Yang, 2024)

This observation matters because it relocates belief-updating errors from a list of separate "biases" to a general property of bounded cognition: when the mapping from evidence to the correct belief is hard to compute, the computed belief is shrunk toward a default. Attenuation thus predicts chronic under-reaction to genuinely informative signals - a direct brake on the rate at which knowledge can accumulate.

The forecasting evidence adds an important wrinkle: the direction of the error depends on the level of aggregation. Individual forecasters tend to overreact to recent news, whereas consensus forecasts tend to under-react relative to full-information rational expectations (Coibion & Gorodnichenko, 2015; Bordalo, Gennaioli, Ma & Shleifer, 2020). The overreaction is well captured by diagnostic expectations, in which the representativeness heuristic leads agents to overweight recent news and exaggerate the probability of states that have just become more likely, generating the excess volatility and boom-bust dynamics seen in credit and asset markets (Bordalo, Gennaioli & Shleifer, 2018).

Whether knowledge accumulates also depends on whether people gather information in the first place. Both rational-inattention models, in which limited attention is allocated optimally subject to a cost (Sims, 2003), and behavioral-inattention models, in which attention is boundedly and imperfectly deployed (Gabaix, 2019), imply that agents leave informative signals unprocessed. Attenuation can therefore arise either as a near-optimal response to real processing costs or as a hard cognitive limit; in both cases the practical consequence is the same - new information moves beliefs less than its objective content warrants.

It should be said plainly that behavioral economics has no established construct literally called "cumulative knowledge." The closest formal object is the theory of learning - the study of where beliefs go as data accrue - and its verdict is mixed. Under a correctly specified model, a Bayesian learner's beliefs form a martingale that converges to the truth, so knowledge does cumulate; this is the benchmark sense in which "cumulative knowledge" exists in the framework (Savage, 1954).

That guarantee, however, is fragile. When agents hold even slightly misspecified models of the world, learning generically converges not to the truth but to the belief that best fits the data among the wrong models the agent entertains - the content of Berk-Nash equilibrium (Esponda & Pouzo, 2016; Fudenberg, Lanzani & Strack, 2021). Confirmation bias can entrench a false but confident belief and drive polarization between people who see the same evidence (Rabin & Schrag, 1999). Selective attention can create stable, self-confirming errors - "learning traps" - that a purely Bayesian agent inside the misspecification can never detect (Gagnon-Bartsch, Rabin & Schwartzstein, 2021). And in social settings, early signals can trigger informational cascades in which individuals rationally ignore their own information, so that the group stops aggregating knowledge altogether (Bikhchandani, Hirshleifer & Welch, 1992).

Layering behavioral attenuation on top of these results sharpens the picture. Persistent under-reaction to fundamentals slows the march toward truth even when the model is correct, while episodic overreaction injects noise and instability (Enke et al., 2024; Bordalo et al., 2020). The honest summary is therefore that, at the individual level, knowledge accumulation is partial, path-dependent, and not guaranteed to converge - closer to a noisy, sometimes self-trapping process than to the clean cumulative ideal of the Bayesian benchmark.

The phrase "cumulative knowledge" also invites a second, meta-scientific reading: does behavioral economics, as a discipline, build a reliable and growing stock of findings? The replication movement provides the most direct evidence. The Open Science Collaboration (2015) successfully replicated only about a third of a large sample of psychology results, whereas Camerer et al. (2016) replicated roughly two-thirds of laboratory experiments in economics - better, but still far from complete cumulation, and a reminder that individual published estimates are noisy draws rather than settled facts.

Generalizability is a second hurdle. Interventions that work in academic trials often shrink when deployed at scale: nudge effects estimated inside two large government "nudge units" are markedly smaller than those in the published literature (DellaVigna & Linos, 2022), and the average effect of choice-architecture interventions is contested once publication bias is taken into account (Mertens et al., 2022; Maier et al., 2022). Knowledge that does not travel from the lab to the field, or from a small trial to a population, has not really cumulated.

Against these fragilities, the field has developed machinery that makes knowledge genuinely additive. Structural behavioral economics estimates portable parameters - a present-bias coefficient, a degree of inattention - that can be carried from one setting to another and tested out of sample (DellaVigna, 2018). Theory disciplines field experiments so that each study speaks to a common framework rather than standing alone (Card, DellaVigna & Malmendier, 2011). And the systematic elicitation of expert forecasts before results are known both records what the field collectively believes and exposes where that belief is wrong, turning accumulated intuition into a testable object (DellaVigna & Pope, 2018). The verdict for this second sense mirrors the first: behavioral economics is becoming cumulative - through replication, meta-analysis, pre-registration, and structural integration - but unevenly and with well-documented soft spots.

Behavioral attenuation is the thread that connects the two senses of cumulative knowledge. The same imperfect information processing that shrinks an individual's response to evidence also cautions the field that any single estimate is an attenuated, noisy signal of the truth, recoverable only by replication and aggregation. Reliable knowledge, in other words, cumulates slowly for the same cognitive reasons at both levels.

Two implications follow. For policy, because people systematically under-respond to information, interventions that merely supply facts - disclosure, warnings, financial-literacy campaigns - tend to have limited reach, while defaults, salience, and simplification, which do not require correct updating, tend to do more work (Gabaix, 2019; DellaVigna & Linos, 2022). For the study of trust specifically, the belief-dependent preferences that drive cooperation are not exempt: the first-order beliefs about others' trustworthiness that raise contributions to public goods (Kim, Putterman & Zhang, 2022) are themselves formed by attenuated, biased updating. Trust-building interventions must therefore reckon not only with what people believe about one another, but with how slowly and imperfectly those beliefs move in response to new evidence.