When 'Bias' Is Strategic: The Limits of Debiasing

The Debiasing Playbook

Behavioral economics has a standard operating procedure. Step one: identify a cognitive bias — a systematic departure from the predictions of rational choice theory. Step two: document the bias in laboratory experiments, showing that it is robust, replicable, and prevalent. Step three: design an intervention — a nudge, a disclosure requirement, a reframing, a default — that corrects the bias and steers people toward choices that better serve their interests. Step four: evaluate the intervention with a randomized controlled trial. Step five: scale.

The playbook has produced real achievements. Auto-enrollment in retirement plans increased savings. Simplified disclosure improved financial decision-making. Commitment devices helped people quit smoking and lose weight. Default organ donation increased supply. The behavioral policy movement, launched by Thaler and Sunstein’s Nudge in 2008 and institutionalized by nudge units in governments around the world, has earned its place in the policy toolkit.

But the playbook rests on an assumption that is rarely examined: that the biases identified by behavioral economics are mistakes — deviations from optimal behavior that, if corrected, would make people unambiguously better off. This assumption is sometimes true. It is not always true. And when it is not true, debiasing interventions can backfire, produce unintended consequences, or simply fail — not because the intervention is poorly designed, but because the “bias” is not a malfunction.

Overconfidence and the Entrepreneur’s Gamble

Consider overconfidence — the tendency of people to overestimate their own abilities, knowledge, and prospects. In the standard behavioral account, overconfidence is a bias that leads to poor decisions: excessive trading in financial markets (Barber and Odean, 2001), overentry into competitive markets (Camerer and Lovallo, 1999), and underestimation of project costs and timelines (Buehler, Griffin, and Ross’s “planning fallacy”).

All of this is true. But overconfidence also has a function. Entrepreneurship, almost by definition, requires believing that your idea will succeed when most ideas like it fail. The base rate of startup failure is somewhere between 60% and 90%, depending on how you measure it. A perfectly calibrated person who knew the base rate would rarely start a company. But some of those companies — the ones that do succeed — produce enormous value: jobs, innovation, consumer surplus, economic growth.

If we could wave a wand and make all entrepreneurs perfectly calibrated about their odds of success, we would dramatically reduce the rate of new business formation. This would eliminate the waste of failed startups, but it would also eliminate the innovations that only emerge when someone is irrationally confident enough to try. The net effect on social welfare is ambiguous. Overconfidence is individually costly for the entrepreneurs who fail, but it may be socially beneficial if the winners create more value than the losers destroy.

This is not a hypothetical argument. Astebro, Herz, Nanda, and Weber (2014) surveyed the evidence and concluded that entrepreneurial overconfidence, while real and well-documented, serves an important economic function by overcoming the market failure that would result from too few entrants. In a world of rational agents, many profitable business opportunities would go unexploited because the expected return, after adjusting for risk, is negative for the individual entrepreneur — even though the option value to society of having someone try is positive.

The implication is uncomfortable for the debiasing agenda: correcting overconfidence in the entrepreneurial context might reduce social welfare. The bias is not a bug; it is a mechanism that partially solves a different problem.

Optimism and Health

A similar argument applies to optimism — the tendency to expect better outcomes than the evidence warrants. In the health domain, optimistic patients are more likely to adhere to treatment regimens, exercise, maintain social connections, and recover from surgery. A substantial body of evidence in health psychology, much of it summarized by Shelley Taylor and Jonathon Brown in their influential 1988 paper on “positive illusions,” suggests that moderately unrealistic positive expectations are associated with better mental and physical health outcomes.

The mechanisms are not mysterious. Optimistic people invest more effort in activities that contribute to good outcomes — precisely because they believe the effort will be rewarded. A cancer patient who believes (perhaps unrealistically) that treatment will succeed is more likely to complete a grueling chemotherapy regimen than a patient who has accurately assessed the survival probabilities. The unrealistic belief generates the behavior that makes recovery more likely.

This does not mean optimism is always beneficial. Unrealistic optimism about the risks of smoking, drinking, or unprotected sex can be deadly. The point is that the relationship between bias and welfare is not monotonic. A moderate degree of optimism may be better than perfect calibration, even as excessive optimism is worse. The optimal level of “bias” is not zero.

Loss Aversion as Insurance

Loss aversion — the tendency to weight losses more heavily than equivalent gains — is perhaps the most celebrated finding in behavioral economics. Kahneman and Tversky estimated that losses are roughly twice as painful as equivalent gains are pleasurable. In many contexts, loss aversion leads to suboptimal behavior: the endowment effect inflates selling prices, status quo bias prevents beneficial changes, and the disposition effect causes investors to hold losing stocks too long.

But loss aversion also prevents reckless risk-taking. A person who is equally sensitive to gains and losses — a person with no loss aversion — would be willing to accept any gamble with a positive expected value, no matter how volatile. In practice, this means accepting gambles that could wipe out their savings in exchange for a small expected gain. Loss aversion induces a demand for insurance, a preference for stable income over volatile income, and a reluctance to make bets that could be catastrophic even if they are favorable on average.

For most people, in most circumstances, this is prudent behavior. The marginal utility of wealth is not constant — losing your last $10,000 is much worse than gaining an additional $10,000 when you already have a comfortable cushion. Loss aversion approximately corrects for this asymmetry in a world where people do not carry around explicit utility functions. It is a heuristic that does the work of diminishing marginal utility of wealth without requiring the person to compute it.

Removing loss aversion, if it were possible, might increase expected returns (by inducing more risk-taking) but would also increase the incidence of financial catastrophe. Whether the tradeoff is favorable depends on the context, the individual’s circumstances, and the availability of other safety nets. The blanket claim that loss aversion is a “bias” that should be corrected obscures this complexity.

Gigerenzer’s Ecological Rationality

The most systematic intellectual challenge to the debiasing agenda comes from Gerd Gigerenzer and the Adaptive Behavior and Cognition (ABC) research group at the Max Planck Institute in Berlin. Gigerenzer’s program, developed over three decades, rests on a simple but powerful idea: the rationality of a decision procedure cannot be assessed in isolation. It must be assessed relative to the environment in which it operates.

Gigerenzer calls this ecological rationality. A heuristic — a simple decision rule that ignores some information — is ecologically rational if it performs well in the environment it was designed for (by evolution or by learning). The key insight is that heuristics are not always inferior to more complex strategies. In some environments, they are superior, because they exploit the structure of the environment while avoiding the costs of overfitting to noise.

The leading example is the “take-the-best” heuristic for paired comparisons. When choosing which of two options is larger on some criterion (which city has a larger population, which stock will earn a higher return), take-the-best looks at cues one at a time, in order of validity, and chooses the option favored by the first cue that discriminates. It ignores all remaining cues. This is a violation of the standard model, which says you should integrate all available information, weighting each cue by its reliability.

But Gigerenzer and his colleagues showed that in many real-world data sets, take-the-best performs as well as or better than multiple regression and other supposedly optimal statistical methods — especially when predicting out of sample. The reason is the bias-variance tradeoff: complex models fit the training data well but overfit to noise, producing poor predictions on new data. Simple heuristics fit the data less well but generalize better, because they are not distracted by irrelevant features of the training set.

This finding has profound implications for behavioral policy. If the heuristics people use are adapted to the environments they face, then “correcting” them by teaching people to use more complex strategies may actually worsen performance. A financial advisor who teaches a retiree to use mean-variance optimization for portfolio selection — the “rational” approach — may produce worse outcomes than the retiree’s simple rule of thumb (1/N allocation, spreading money equally across available options), because the optimization is sensitive to estimation errors in expected returns and covariances that the 1/N rule ignores.

Gigerenzer’s critique is not that biases do not exist. It is that the concept of “bias” is defined relative to a normative standard — usually expected utility theory or Bayesian updating — that may itself be inappropriate for the environments people face. If the normative standard is wrong, then the “bias” is not a bias. It is an adaptation.

When Nudges Backfire

Even when a bias is genuinely harmful and a debiasing intervention is well-designed, the intervention can fail or backfire through a variety of mechanisms.

Reactance. When people perceive that a nudge is an attempt to manipulate their behavior, they may do the opposite — a phenomenon psychologists call reactance. Telling a teenager that junk food is bad may make them eat more junk food, because the perceived external pressure triggers a desire to assert autonomy. In policy contexts, conspicuous nudges — calorie labels, graphic cigarette warnings, paternalistic framing — can provoke backlash, especially among individuals who value personal freedom highly.

Moral licensing. People who have done something virtuous often feel licensed to do something less virtuous afterward. A person who eats a salad for lunch may feel entitled to a large dessert. A person who buys carbon offsets may feel licensed to fly more. A person who donates to charity may feel less obligation to support redistributive taxation. If a debiasing intervention succeeds in one domain, it may be offset by worse behavior in a related domain. Mazar and Zhong (2010) showed that consumers who bought “green” products subsequently behaved less ethically in economic games — the green purchase provided moral license.

The “what the hell” effect. Dieters who break their diet with a single indulgence often abandon the diet entirely — “I’ve already blown it, so what the hell.” The same pattern appears in savings behavior: a person who fails to hit their savings target in one month may give up the savings plan entirely rather than resume it. Interventions that set bright-line targets (save $500 per month, exercise five days per week, eat fewer than 2,000 calories per day) can trigger the “what the hell” effect when the target is inevitably missed, producing worse outcomes than a vaguer, more forgiving approach.

Crowding out intrinsic motivation. When a nudge or incentive is introduced for a behavior that people already do for intrinsic reasons, the external intervention can crowd out the internal motivation. Gneezy and Rustichini (2000) showed that introducing a fine for late pickup at daycare centers increased lateness — parents treated the fine as a price (paying for the privilege of being late) rather than a moral sanction. Removing the fine did not restore the original norm. The incentive had permanently transformed the social meaning of the behavior.

Cultural Variation: Whose Bias?

The behavioral economics literature is overwhelmingly based on experiments conducted with Western, educated, industrialized, rich, and democratic (WEIRD) populations — primarily American and European university students. The assumption that the biases found in these populations are universal has been challenged by cross-cultural research.

Henrich, Heine, and Norenzayan (2010), in their landmark paper “The Weirdest People in the World?”, documented substantial cross-cultural variation in psychological phenomena that behavioral economists treat as universal, including risk aversion, fairness norms, time preferences, and responses to framing effects. The Muller-Lyer illusion — often cited as an example of a universal perceptual bias — varies dramatically across cultures. The endowment effect, robust among American students, is weaker or absent in some non-Western populations.

If biases vary across cultures, then so should debiasing interventions. A nudge designed for American workers may not work for Indian, Nigerian, or Japanese workers — not because the intervention is poorly implemented, but because the underlying psychology is different. The loss aversion that drives retirement savings behavior in the United States may have a different magnitude, or a different interaction with social norms, in a society with different family structures, different social insurance arrangements, and different cultural attitudes toward risk.

This cultural variation also raises a political question: who decides what counts as a “bias”? The behavioral policy framework assumes that there is an objective standard of rational behavior from which biases depart. But the standard is itself a cultural artifact — the axiomatic framework of expected utility theory, developed in mid-twentieth-century Western academic institutions. When a Kenyan smallholder farmer discounts the future steeply, is this “present bias” or a rational response to genuine uncertainty about whether the future will arrive? When an Indian family prioritizes social obligations over individual savings, is this “bias” or a different but coherent value system?

The Politics of Paternalism

The deepest objection to the debiasing agenda is not empirical but normative. Even when a bias is real, universal, and harmful, the question of who should correct it — and by what authority — is a political question that economics alone cannot answer.

Behavioral policy concentrates power in the hands of choice architects: the regulators, employers, and platform designers who set defaults, frame options, and structure choices. These architects are themselves human beings with biases, incentives, and political agendas. The assumption that they will use their power to maximize welfare, rather than to advance their own interests or their preferred political outcomes, is a strong one.

Rizzo and Whitman, in Escaping Paternalism (2020), argued that the slippery slope from nudges to harder forms of paternalism is real, not hypothetical. If the government can set defaults because people are biased, it can also restrict choices. If it can restrict choices because people are biased, it can also mandate choices. Each step is justified by the same logic — people do not know what is good for them — and each step is more intrusive than the last.

This does not mean all behavioral policy is illegitimate. It means that the case for any specific intervention must address not only whether the bias exists and whether the intervention corrects it, but also whether the intervention respects the autonomy of the people it targets, whether the choice architects are trustworthy and accountable, and whether the intervention creates risks of political abuse.

A Plea for Humility

The behavioral economics revolution has produced genuine insights about human decision-making. People are not the frictionless optimizers of textbook theory. Their judgments are shaped by heuristics, emotions, social context, and cognitive limitations. Understanding these patterns is valuable for both positive economics (predicting behavior) and normative economics (designing better institutions).

But the step from “people have biases” to “we should fix them” is larger than it appears. It requires knowing not only that a bias exists, but that it is harmful on net, that the proposed correction will work, that the correction will not produce worse side effects, and that the people designing and implementing the correction are competent and trustworthy. Each of these conditions can fail, and each has failed in practice.

The strongest version of the behavioral policy agenda — identify biases, design nudges, scale interventions, improve welfare — is too confident. A more modest version would acknowledge that biases sometimes serve functions, that heuristics can outperform optimization, that debiasing can backfire, that cultural context matters, and that the authority to reshape other people’s choices is not to be exercised lightly.

Gigerenzer’s ecological rationality program offers a framework for this humility. Before designing an intervention to correct a bias, ask: in what environment did this heuristic evolve? Does it perform well in that environment? Has the environment changed in ways that make it dysfunctional? Is the proposed correction adapted to the actual environment, or does it assume a world of frictionless optimization that does not exist?

These are hard questions, and answering them requires a deeper engagement with the specific contexts in which people make decisions than the typical lab experiment provides. It requires field research, cultural sensitivity, institutional knowledge, and a willingness to be wrong. It requires, above all, the recognition that the people whose behavior we seek to change may understand their own situations better than we do — and that our models of their “biases” may be as incomplete as their own models of the world.

The goal of behavioral economics should not be to make people rational according to the textbook definition. It should be to understand the strategies people actually use, the environments in which those strategies succeed or fail, and the institutional designs that help people navigate their world as it actually is — complex, uncertain, and irreducibly human.