In 1983, Amos Tversky and Daniel Kahneman described Linda, a 31-year-old philosophy graduate deeply concerned with social justice. They asked participants which was more probable: that Linda is a bank teller, or that she is a bank teller active in the feminist movement. Roughly 85 percent chose the second option. That violates a basic rule of probability, because a conjunction can never be more probable than either of its parts.
Findings like this launched the heuristics-and-biases programme. It has since catalogued well over a hundred systematic departures from the normative standards of logic, probability, and decision theory. The word systematic matters. These are not random slips but patterned errors, and patterns are evidence about the machinery that produces them.
That is what makes bias research philosophically interesting. If we want to know how the mind computes, studying where it predictably fails works much like studying visual illusions to reverse-engineer vision. But the inference cuts both ways. Before concluding that human reasoning is flawed, we have to ask which standard of rationality we are measuring against, and whether the mind was ever built to meet it.
Sorting Biases by Mechanism, Not by Name
Lists of cognitive biases tend to read like a bestiary: anchoring, framing, availability, the sunk cost fallacy, each with its own name and its own founding experiment. This proliferation is a problem for a science of cognition. Naming an effect is not the same as explaining it. A catalogue of a hundred phenomena might reflect a handful of underlying processes, or a hundred unrelated ones.
One influential unification is Kahneman and Frederick's notion of attribute substitution. Faced with a hard question, such as how probable an outcome is, the system quietly answers an easier one, such as how representative it seems or how easily examples come to mind. On this view, representativeness, availability, and affect-based judgment are not separate biases. They are a single strategy applied to different target attributes. Keith Stanovich adds a complementary axis. Some errors come from failing to override a fast default response. Others come from missing mindware, the learned rules of probability and logic needed to compute the correct answer at all.
A third family fits neither description. Confirmation bias and motivated reasoning seem driven less by computational economy than by goals. We search for, weight, and remember evidence in ways that protect beliefs we already hold. A fourth explanation is strikingly austere. Martin Hilbert has argued that noise alone, meaning random error in how information is encoded and retrieved, can generate several classic effects without any dedicated bias mechanism. These include conservatism in belief updating and illusory correlation.
Jerry Fodor's theory of modularity is useful here. Some biases behave like the outputs of encapsulated systems. The Müller-Lyer lines still look unequal after you measure them, and anchoring persists even when people know the anchor was generated at random. Other biases shift with what the reasoner believes and wants, which suggests central, non-modular processing. A mechanism-based taxonomy therefore does more than tidy the list. It predicts which biases should yield to knowledge and which should not.
TakeawayA bias's name describes a pattern of error, but its mechanism determines how that error can be changed. The more useful question is not which bias this is, but which process produced it.
Errors, or the Wrong Test?
Gerd Gigerenzer's challenge to the heuristics-and-biases programme begins with a methodological observation. Many classic tasks present information in formats the mind rarely encounters outside the laboratory. When the Linda problem is rephrased in natural frequencies (of 100 people like Linda, how many are bank tellers?), conjunction violations drop dramatically. Base-rate neglect in medical diagnosis problems shrinks in the same way. On this view, some apparent biases are artefacts of a mismatch between task format and the representations our reasoning evolved to process.
Content matters too. In the abstract version of the Wason selection task, a test of conditional logic, typically only around one in ten participants answers correctly. Yet when Leda Cosmides recast the same logical structure as a matter of detecting people who break a rule of social exchange, performance rose sharply. Researchers still dispute whether this reveals a dedicated cheater-detection module or simply the benefit of familiar rules about obligation and permission. Either way, reasoning competence appears domain-sensitive rather than uniformly poor.
Error management theory pushes the adaptive argument further. Haselton and Buss argued that when two kinds of error carry asymmetric costs, a well-designed system should be biased toward the cheaper mistake. Mistaking a stick for a snake costs a moment of alarm, while the reverse can cost a life. A detector that minimises expected cost will make more errors in total than an unbiased one, and still be the better design. Even the hot hand fallacy has been revisited. Miller and Sanjurjo showed that the original statistical analysis contained a subtle bias of its own, and corrected estimates suggest streak shooting may be real.
The adaptive reading is powerful but limited. It works best for biases that would have been cheap or beneficial in ancestral or everyday environments. It struggles with errors that persist under favourable conditions and high stakes. Its deeper contribution is conceptual. It replaces a single, context-free norm of rationality with ecological rationality, the fit between a strategy and the structure of its environment. A heuristic is not good or bad in itself, only good or bad in a particular setting.
TakeawayRationality may be a relation rather than a property. The same heuristic can be a flaw in one environment and an elegant solution in another.
Can Reasoning Be Retrained?
If biases arise from fixed cognitive architecture, debiasing should be hard, and early results seemed to confirm this. Warnings about a bias rarely eliminate it. Emily Pronin's work on the bias blind spot shows that people readily detect biases in others while rating themselves as relatively immune. Awareness is not the same as control, just as knowing the Müller-Lyer lines are equal does not make them look equal.
Yet some interventions work. Prompting people to consider the opposite, by generating reasons their initial judgment might be wrong, reduces anchoring and overconfidence more reliably than simply telling them to be unbiased. Carey Morewedge and colleagues found that a single training session reduced several biases, and that the improvement persisted for months. The training took the form of either an instructional video or an interactive game with personalised feedback. A later field study found that trained students were less likely to choose a confirmation-biased solution in a disguised business case presented in a different context.
A second strategy leaves the reasoner alone and redesigns the environment. Presenting risks as natural frequencies improves physicians' diagnostic inferences. Default settings work with inertia rather than against it, which is why opt-out organ donation and automatic pension enrolment produce far higher participation. Checklists, structured interviews, and pre-mortems move the corrective step outside the head, so it no longer depends on vigilance in the moment.
The pattern across this evidence is instructive. Interventions that supply missing mindware or change the input format tend to succeed. Interventions that simply urge people to try harder tend to fail. This is exactly what a mechanism-based taxonomy predicts. Errors rooted in encapsulated or highly automatic processes are best handled by redesigning the task, while errors rooted in absent knowledge can be trained away. In Andy Clark's terms, the most robust debiasing may be less about fixing the brain than about extending the mind into well-structured tools and environments.
TakeawayThe most dependable cure for a bias is often not a better reasoner but a better-designed situation, one that makes the right inference the easy one.
Bias research began as a demonstration of human irrationality. It has become something more useful: a set of probes for mapping the architecture of reasoning.
The emerging picture is not of a uniformly flawed reasoner, nor of a secretly perfect one. It is of a mind made up of processes with different properties. Some are fast and encapsulated, some depend on knowledge, some are goal-driven, and some are simply noisy. Each produces characteristic errors, and each responds to different remedies.
That has a philosophical consequence. Questions about rationality can no longer be settled from the armchair by comparing behaviour to a single normative model. They require knowing which mechanism is operating, in which environment, and at what cost. The errors turned out to be some of the best evidence we have about how the mind works.