Plate viii in print

The Ecology

Gigerenzer; ecological rationality and the great rationality debate, whether the biases are defects against logic or fast-and-frugal heuristics fitted to the world.

A mind is not convicted of folly merely for breaking the rules of logic in the empty air; reason is cut to fit a world — of scarce time, thin information, and outcomes that matter — and in that world the swift rule that ignores most of what it might consider will often outrun the patient reckoning that tries to consider everything. What the logician, from his chair, marks down as a defect may prove, out in the field, the very form of good judgment.

— the argument of this plate, after Gerd Gigerenzer, 1996

By the end of the fourth plate the case against the human chooser looks overwhelming. The rational model of decision under risk has been shown descriptively false at its foundation; the mind values not final fortunes but gains and losses from a shifting reference point, with losses looming larger and risk-attitudes flipping across the line; it judges likelihood by resemblance and recall and whatever number lies near to hand, erring in lawful directions; and it cannot keep faith with its own future selves, discounting the present until its resolutions break. The natural verdict, and the one the field largely drew, is that human beings are systematically irrational — not randomly mistaken, which experience could correct, but reliably, predictably, structurally flawed, a species of bungling Bayesians forever falling short of the rational ideal. This last plate sets against that verdict the most serious challenge it has faced, mounted over decades by the German psychologist Gerd Gigerenzer, and the challenge is not that the departures are unreal. They are real; this volume has not been describing a mirage. The challenge is that calling them defects rests on a buried and disputable assumption about what rationality is — and that once the assumption is dragged into the light, a great many of the so-called biases turn out to be not the failures of reason but the workings of a different and arguably better kind.

The challenge begins with a question about the measuring stick. The heuristics-and-biases program scores human judgment against a fixed standard — the laws of logic, the axioms of probability, the rule that forbids a conjunction from outranking its part, the demand that preferences obey independence and that discounting hold steady through time — and every deviation from that standard is entered in the ledger as an error. But why, Gigerenzer asks, is that the right standard? The classical norm of rationality was built for an imagined reasoner of unlimited resources: infinite time to deliberate, complete information to draw on, unbounded power to compute — a Laplacean intelligence that could hold the whole problem in view and grind out the optimal answer. No actual mind is remotely like that. Real choosers decide under the pressure of the clock, on fragments of information, with a brain that cannot perform the integrations the optimal model demands. To convict such a mind of irrationality because it fails to compute what no real mind could compute, in conditions no real mind ever enjoys, is to fault a sprinter for falling short of the speed of light. The standard is not wrong because it is demanding; it is suspect because it is the wrong KIND of standard — a measure drawn from logic in a vacuum, applied to a creature that must act in a world.1

1. The challenge begins with the measuring stick. The heuristics-and-biases program scores judgment against a fixed standard — logic, probability, the conjunction rule, independence — and enters every deviation as error. But why is that the right standard? The classical norm was built for an imagined reasoner of unlimited time, information, and computation, a Laplacean intelligence grinding out the optimal answer. No real mind is like that. To convict a mind of irrationality for failing to compute what no real mind could, in conditions none ever enjoys, is to fault a sprinter for falling short of the speed of light.

Against the standard drawn from logic Gigerenzer sets one drawn from the world, and calls it ECOLOGICAL RATIONALITY: a rule of thought is rational not if it satisfies the logician but if it succeeds in the actual environment it is used in, given that environment's real structure and the chooser's real limits. The idea has a distinguished parent in Herbert Simon, whose bounded rationality had insisted a generation earlier that real choosers do not optimize but SATISFICE — search until they find an option good enough, then stop — and who left behind a metaphor Gigerenzer made the cornerstone of his program: that rational behavior is cut by two blades of a single pair of scissors, the capacities of the mind and the structure of the environment, so that you can no more understand cognition by studying the mind alone than you can understand how scissors cut by examining one blade. The heuristics-and-biases program, on this view, had examined only the one blade — the mind, and its departures from logic — and had ignored the other, the environment, against which those departures must be judged. Gigerenzer's project was to take up the neglected blade: to study the heuristics of real judgment not as regrettable shortcuts but as instruments, and to ask, of each, the question the logician never thought to ask — not is it valid, but does it WORK, and in which environments, and why.2

2. Against the standard drawn from logic, Gigerenzer sets one drawn from the world: ECOLOGICAL RATIONALITY — a rule is rational if it succeeds in the actual environment, given its structure and the chooser's limits. The parent is Herbert Simon, whose bounded rationality held that real choosers SATISFICE rather than optimize, and whose metaphor Gigerenzer made central: rational behavior is cut by two blades of a scissors, the mind and the environment, and you cannot understand cognition by studying one blade alone. The biases program examined only the mind; Gigerenzer took up the neglected blade.
⁂

The program's findings are genuinely startling, because they overturn the assumption that a simple rule can only ever be a cheap approximation to the correct, complex computation — a corner cut, paid for in accuracy. Often it is the reverse: the simple rule is MORE accurate than the elaborate one. Consider the recognition heuristic, the most frugal rule imaginable: faced with two options and asked which scores higher on some criterion, if you recognize one and not the other, choose the one you recognize. Asked which of two cities is larger, pick the one you have heard of. The rule consults a single bit of information — bare recognition — and throws away everything else, and yet across many real domains it is strikingly accurate, because recognition is not random: the larger city, the stronger team, the more successful firm is the one more likely to have reached your ears in the first place, so that the very pattern of your ignorance carries information about the world. This yields an effect that ought to be impossible on the standard view — the less-is-more effect, in which a person who recognizes FEWER of the items can score HIGHER than one who recognizes them all, because the one who knows every city can no longer use recognition to discriminate between them and must fall back on shakier knowledge, while the one who knows only some can let his partial ignorance do the sorting. Or consider take-the-best, a rule for choosing between two options described by many cues: do not weigh and sum all the cues as the optimal linear model prescribes; find the single most reliable cue that distinguishes them and decide on that one alone, ignoring all the rest. In a wide range of real environments — predicting which of two firms will prove more profitable, which patient is at higher risk, which of two outcomes will obtain — take-the-best matches or beats the full weighting of all the evidence. The mind that ignores most of what it knows does better than the mind that uses it all.3

3. The findings overturn the assumption that a simple rule is only a cheap approximation. Often the simple rule is MORE accurate. The recognition heuristic: faced with two options, choose the one you recognize — asked which city is larger, pick the one you have heard of. It consults a single bit and discards the rest, yet is strikingly accurate, because recognition tracks size; the pattern of your ignorance carries information. Hence the less-is-more effect: one who recognizes fewer items can score higher than one who knows them all. And take-the-best: decide on the single most reliable cue, ignoring the rest, often beats weighing all the evidence.

Why a rule that discards information should beat one that uses it is not a paradox but a theorem, and naming it supplies the rigorous heart of the whole ecological case. It is the bias-variance tradeoff, known to every statistician and every builder of predictive machines. Any rule that learns from a finite, noisy sample of the world and must then perform on fresh cases faces two distinct sources of error. One is bias: the rule may be too simple to capture the true pattern, and so be systematically off. The other is variance: the rule may be so flexible, so richly parameterized, that it bends to fit not only the real pattern in the sample but the noise — the accidents and flukes peculiar to that sample — and so swings wildly from one sample to the next, fitting the past beautifully and the future badly. The elaborate optimal model, with its many free parameters, drives bias to nothing but pays in variance: it overfits, mistaking noise for signal, and generalizes poorly. The fast-and-frugal heuristic, with its few parameters, carries some bias — it cannot capture every nuance — but its very rigidity protects it from chasing noise, holding variance low, so that in a world of small samples and large noise it predicts the unseen case more accurately than its sophisticated rival. Here is the deep vindication the program needed, and the sharpest possible rebuke to the word that titles this volume: bias, in the precise technical sense, is not the enemy of accuracy but sometimes its guardian. A measured dose of bias — a willful simplicity, a deliberate refusal to use all the information — is exactly what keeps a judgment robust. The heuristic's "irrational" neglect of information is, in an uncertain world, a defense against the overfitting that destroys the rational ideal.4

4. Why discarding information can beat using it is a theorem: the bias-variance tradeoff. Any rule learning from a noisy sample faces two errors — bias (too simple to catch the pattern) and variance (so flexible it fits the noise and swings from sample to sample). The elaborate model drives bias to nothing but overfits, paying in variance. The frugal heuristic carries some bias but, being rigid, holds variance low, and so predicts unseen cases better where samples are small and noise large. Bias, precisely understood, is not the enemy of accuracy but sometimes its guardian.

From this vantage Gigerenzer mounts his most pointed assault on the biases themselves, and it is an assault by reframing. Many of the celebrated errors, he argues, are not failures of the mind but artifacts of the unnatural form in which the experimenter posed the question — and they shrink or vanish when the same problem is cast in the shape the mind was built to handle. The keenest example is the one that most embarrassed human judgment: the neglect of base rates in problems of diagnosis. Tell people, in the language of single-event probabilities, that a disease afflicts one percent of women, that a test catches ninety percent of those who have it, and that it falsely alarms on some small percentage of those who do not, and then ask the chance that a woman who tests positive is truly ill, and most people — including, notoriously, most physicians — answer wildly too high, forgetting how the rarity of the disease swamps the test's accuracy. A grievous bias, on the standard reading. But pose the identical information in NATURAL FREQUENCIES — out of every thousand women, ten have the disease; of those ten, nine test positive; of the nine hundred and ninety who are well, some ninety also test positive; so of roughly ninety-nine positive tests, only nine mark real illness — and the fog lifts. People, doctors among them, now see almost at once that a positive result means real disease only about one time in eleven. The competence was there all along; what defeated it was the demand to reason in abstract single-event probabilities, a recent and artificial cultural invention foreign to the way experience actually accumulates, which is by counting cases as they come. The mind, Gigerenzer argues, is well tuned to frequencies, the natural currency of a creature that learns from a stream of encounters, and ill tuned to the percentages of the statistics class; pose the problem ecologically and the bias is revealed as a creature of the laboratory, not of the head.5

5. Gigerenzer's sharpest move is reframing: many celebrated errors are artifacts of unnatural framing and vanish when the problem takes the shape the mind handles. Told in single-event percentages that a disease hits one percent, a test catches ninety percent, and alarms falsely on a few percent, most people — most doctors — wildly overstate the chance a positive means illness. Recast in NATURAL FREQUENCIES — of a thousand women, ten are ill and nine test positive, ninety of the well also test positive, so nine of ninety-nine positives are real — and the fog lifts. The competence was always there; the abstract framing manufactured the error.
⁂

It would be a poor reckoning, though, that let Gigerenzer have the last word as easily as the field once gave it to his opponents, and the dispute between them — the great rationality debate, as it came to be called — is not a rout but a genuine and unfinished quarrel, with real force on both sides. The defenders of the heuristics-and-biases program grant much of the ecological case and then press hard on what it leaves out. That a heuristic can be accurate in some environments, they observe, does not show that the documented biases are not errors in the environments where they were found; that natural frequencies dissolve base-rate neglect does not dissolve the conjunction fallacy, the anchoring on numbers known to be random, the framing effects that swing choices between life and death, the overconfidence that survives every reformulation — and many of these do not melt away when the problem is posed kindly. They insist, too, that the adaptiveness of a rule on average is cold comfort to the individual it misleads on the occasion that matters, and that a bias which is systematic is, by that very fact, exploitable: the marketer who sets a high anchor, the swindler who frames a sure loss as a gamble worth taking, the lender who buries the long-run cost behind a salient short-run ease, all make their living from departures that are perfectly real and perfectly costly, ecological defenses notwithstanding. The most influential settlement, associated with Kahneman's later work, grants that intuitive judgment is fast, frugal, and usually serviceable, exactly as Gigerenzer says — and maintains, all the same, that it is prone to systematic and consequential error that a slower, more deliberate reasoning must sometimes be summoned to catch. And the hardest case for the ecological view is the one the previous plate set out: present bias, the under-saver and the addict, who are not misjudging an abstract probability but acting, year upon year, against their own considered and repeatedly affirmed interest. There the talk of a wisely weighted bird in hand grows thin; some of what we do across time looks less like a rule fitted to uncertainty than like a genuine failure to be the person we ourselves endorse.6

6. The debate is no rout but a genuine quarrel. The biases program grants much and presses on the rest: that a heuristic can be accurate somewhere does not show the documented biases are not errors where found; natural frequencies dissolve base-rate neglect but not the conjunction fallacy, arbitrary anchoring, life-and-death framing, or overconfidence. A systematic bias is exploitable — the marketer's anchor, the swindler's frame, the lender's hidden cost. Kahneman's later settlement: intuition is usually serviceable yet prone to consequential error a slower reasoning must catch. The hardest case for the ecological view is present bias, the addict acting against his own endorsed interest.

What the quarrel finally exposes, beneath the trading of experiments, is that the word over which it is fought has more than one meaning, and that the combatants have been measuring with different units. There are, at the least, two conceptions of rationality at large in the argument, and they can come apart. The first is COHERENCE: a mind is rational if its beliefs and choices hang together by the internal rules of logic and probability — if it never ranks a conjunction above its part, never lets a frame reverse its preference, never violates independence or lets the calendar overturn its plans. Coherence looks inward, at the consistency of the system, and asks nothing of the world. The second is CORRESPONDENCE, or ecological success: a mind is rational if its rules actually deliver — accurate predictions, sound decisions, survival and flourishing — in the real environment it inhabits, whatever the logician makes of their form. Correspondence looks outward, at the fit between rule and world, and asks nothing of internal elegance. The crucial discovery of the whole debate is that these two need not coincide and frequently do not: a perfectly coherent agent can be an ecological fool, computing the optimal model and overfitting himself into ruin, while an ecologically triumphant agent can be, by the logician's lights, incoherent — wielding a fast-and-frugal rule that flouts a rule of logic and predicts the world better for it. Which conception is the RIGHT one — which deserves the honorific "rational" — is not a question any experiment can decide, for it is not a question about the mind or the world but about the meaning of a word, about what we are asking for when we ask whether a creature is rational. And economics has an uncommon stake in the answer, having raised its entire edifice on the first conception, the coherent maximizer, while the second quietly suggests that the agent who built the wealth of nations may have been succeeding all along by lights the theory never recognized.7

7. The quarrel exposes two meanings of rationality. COHERENCE: a mind is rational if its beliefs and choices hang together by logic and probability — never ranking a conjunction above its part, never letting a frame reverse a preference. It looks inward, asking nothing of the world. CORRESPONDENCE, or ecological success: a mind is rational if its rules deliver accurate predictions and good outcomes in the real environment, whatever their form. The two come apart — a coherent agent can overfit into ruin; an incoherent one can predict the world better. Which deserves the name is not an empirical question but one about the meaning of the word, and economics built its edifice on the first.

So the volume can at last return its answer to the question the table set it — that the mind departs from the model in regular ways, and the question was which, and at what cost. The which it has catalogued: the certainty effect and the violations of expected utility, the reference-dependence and loss aversion that bend the value of every gain and loss, the heuristics of resemblance and recall and anchoring that govern judgment, the present bias that divides the self across time, and in the plates left for later the aversion to ambiguity, the clinging to what one owns, the overconfidence, the dread that distorts risk. The cost is the harder half, and the volume's real finding is that there is no single answer to it, because the cost is not a property of the bias alone. It is a joint property of the departure, the environment it occurs in, and — decisively — the standard of rationality against which it is scored. Measured against coherence, in the abstract air of the logician's problem, every departure is a cost, an error to be tallied and deplored. Measured against correspondence, in the concrete and uncertain world the mind evolved to act in, a great many of the same departures are costless or better than costless — the modest price of a robustness that keeps judgment from shattering on noise. The departures are real, every one of them; the volume has not explained them away. What it has unsettled is their STATUS, the easy verdict of defect, by showing that the verdict was smuggled in with a choice of yardstick that the science of bias had taken for granted and never argued for. The deepest yield of that science, in the end, is not an inventory of human folly but a question turned back upon its own foundation: before we can say how far the mind falls short of rational, we must say what rational is — and the mind's reliable "failures," read the other way around, are among the best evidence we have of what reason, fitted to a real and uncertain world, actually requires of a creature that must decide and cannot wait.8

8. The volume's answer to which, and at what cost: the which is catalogued — the certainty effect, loss aversion, the heuristics, present bias, and the rest. The cost has no single value, for it is a joint property of the departure, the environment, and the standard scored against. Against coherence, every departure is an error; against correspondence, in the real uncertain world, many are costless or better. The departures are real; what is unsettled is their STATUS as defects — a verdict smuggled in with an unargued yardstick. The yield is not an inventory of folly but a question: before saying how far the mind falls short of rational, we must say what rational is.

With that doubt the volume closes, and with it the family of the behaviors, and with it the whole of this table. The earlier groups had built the architecture of an economy conceived as a vast machine for fitting scarce means to competing ends — the prices that carry its signals, the quantities at its margins, the failures that bedevil it, the institutions that hold it together, the aggregates by which it is measured — and through all of it ran a single figure, taken mostly on faith: the agent who knows what he wants and chooses the best means to it. The behaviors turned the lens at last upon that figure. The Choice anatomized the theory of rational choosing and found it straining at its seams; The Game set the chooser among other choosers, each anticipating the rest; The Norm bound him in the unwritten rules that bear more of the world's weight than any contract; and The Bias has asked, last of the twenty-six, whether the maximizer at the foundation of the entire structure is a faithful likeness of the human creature or a serviceable fiction we have mistaken for one. The answer is neither the agent's vindication nor his demolition, but something more demanding than either: that he is an indispensable idealization, false in regular and knowable ways, and that the lawful shape of his failure is not an embarrassment to be excused but the opening of a truer account — of a mind that is neither a flawless engine of calculation nor a bungler, but an instrument fitted, imperfectly and with a craft we are only beginning to read, to the uncertain world it must act in. The series took its name from the numéraire, the one good against which all the others are priced, the unit that makes the values commensurable; and it ends by turning that question upon the one who does the pricing, asking what the measure of the measurer might be — and finding that even reason has more than a single unit, and that which unit to reach for is the open question on which, quietly, all the others have depended. Here the periodic table of economic life is complete. It is fitting that it should close not on an answer but on a doubt, for the doubt is the most honest thing the subject has to offer, and the truest measure of how much remains to be understood.