Redstart: IELTS Academic Reading practice test

A bird's calls and a bird's song are different things. Calls are short, innate and functional: an alarm, a contact note, a begging signal from a chick. Song is longer, more elaborate, produced mainly by males in the breeding season, and used to hold a territory and attract a mate. The important difference is that song, in the group of birds we call songbirds, has to be learned. Most birds do not learn their vocalisations at all; among the ten thousand or so species, only songbirds, parrots and hummingbirds acquire their sounds by listening and imitating, which is a capacity they share with a small number of mammals and with us.

The learning happens in stages and the stages have been mapped in detail. A young bird passes through a sensitive period, often before it sings at all, in which it memorises the song of a tutor, usually a neighbouring adult of its own species. Some weeks or months later it begins to produce subsong, which is quiet, rambling and without clear structure, rather like the babbling of an infant. This develops into plastic song, in which recognisable elements appear and the bird practises them, comparing what it hears itself produce against the model it stored earlier. Finally the song crystallises: it becomes fixed, and in many species changes very little for the rest of the bird's life. A bird raised in isolation produces an abnormal and simplified song. More tellingly, a bird that memorises a tutor normally but is made deaf before the practice stage never matches the model at all, which shows that it must be able to hear itself in order to learn.

Because song is learned from neighbours, it varies from place to place. The classic study was conducted on a sparrow in the area around San Francisco Bay, where populations only a few kilometres apart sing recognisably different versions of the same song, with boundaries between them that are surprisingly sharp and that have remained in roughly the same places for decades. A recording made at one site can be assigned to its dialect by ear. These are not different subspecies, and birds from either side of a boundary are physically indistinguishable.

Why the dialects exist is less settled than that they do. One explanation is acoustic: sound of different frequencies travels differently through different vegetation, and songs recorded in dense forest do tend to use lower frequencies and slower repetitions than songs of the same species in open country, which is what the physics of sound transmission would predict. A second and more ambitious explanation holds that dialects act as a barrier, so that females prefer local males and the populations diverge. This has been tested by comparing genetic differences across dialect boundaries, and the differences found are small. The barrier, if it exists, is a leaky one.

What is not in doubt is that song traditions change over time in the way languages do. The best-documented case concerns a North American sparrow whose song had ended, for as long as anybody had recorded it, in a series of three-note phrases. In the 1960s birds in western Canada began ending with two-note phrases instead, and over the following decades the new version spread eastward across the continent, tracked by amateur recordists, until it had almost entirely replaced the old ending across a range of several thousand kilometres. Nothing about the birds changed. A variant appeared, was copied, and was copied again.

Cities have introduced a new pressure. Traffic noise is concentrated at low frequencies, exactly where many bird songs sit, and birds in noisy places sing louder and at higher pitches than the same species in quiet ones. Whether this should be called an adaptation is disputed, since singing louder raises pitch as a matter of physiology and part of the shift may be a side effect rather than a solution. What is clear is that it is not free. In several species, experiments in which females are played higher and lower versions of the same song find that they respond less strongly to the higher one, so a male who makes himself audible in traffic is making himself less attractive at the same time.

There remains the question of what song is for, from the bird's point of view rather than the observer's. The most persuasive account is that it works as an honest advertisement. Learning a complex song requires the development of particular structures in the brain during a narrow window in early life, and that development is expensive and easily disrupted by poor nutrition or illness. A male with a large and accurate repertoire is therefore providing evidence about the nest he came from, which no amount of later effort can fake. On this account a female listening to a song is not appreciating it. She is reading a record of his first few weeks.

It is worth resisting the temptation to call all of this language, which it is not, while noticing how many of the interesting properties it does have. Song is transmitted socially rather than genetically, varies geographically in ways that map onto social groups rather than onto biology, changes over time through copying error and innovation, and can be lost when a population is small enough that no tutor is available. Whatever else those features add up to, they are the features that make a tradition, and they are running in an animal whose entire brain would fit inside a walnut.

In 1954 a clinical psychologist published a short book reviewing about twenty studies which compared the predictions of trained professionals with those of simple statistical formulas. In almost every case the formula did at least as well, and often better. The book was expected to provoke a decisive rebuttal. Instead the comparison has been repeated several hundred times in the intervening decades, across medical diagnosis, academic admissions, staff selection, the prediction of violence, the assessment of credit and the forecasting of business outcomes, and a large meta-analysis covering well over a hundred such studies reached the same conclusion as the original. On average the formula wins; where the professional wins, the margin is small; where the formula wins, the margin is often considerable.

The reasons are not mysterious and none of them involves the professional being stupid. First, human judgement is inconsistent: presented with the same case twice, separated by a few weeks and disguised, experienced practitioners disagree with themselves at a rate that surprises them when it is pointed out. A formula given the same inputs returns the same answer every time, and that consistency alone accounts for a good part of its advantage. Second, people weight vivid information too heavily and dull information too lightly, so a striking detail in a file displaces three duller ones that between them carry more of the signal. Third, experts add variables. A practitioner who has noticed something once tends to keep watching for it, and most of what accumulates in this way is noise.

The most surprising result in this literature concerns how little the formula needs to know. Models built by working out the optimal weight for each variable do well, as one would expect. Models that simply take the same variables, point them in the right direction and give them all equal weight do very nearly as well, and better than the professionals. What this implies is that almost all of the useful work is done by choosing which variables matter and which way each of them runs, and hardly any by the arithmetic of combining them. That is a humbling result for statisticians as much as for clinicians.

It does not follow that the professional has nothing to contribute, and the boundary is reasonably well marked. Somebody has to decide which variables enter the model, and no procedure can do that from the data alone. There is also the case, given a memorable name in the early literature, in which the practitioner knows something the model cannot: the formula predicts that a man will go to the cinema tonight as he does most Tuesdays, and the practitioner knows he has broken his leg. Such information is genuinely decisive and genuinely outside the model. The difficulty is that practitioners believe they encounter it far more often than they do, and a licence to override the model whenever one feels one has spotted something is a licence to reintroduce every inconsistency the model removed.

Opposition to all of this is persistent and comes in two forms, of which only one is serious. Noemi Abrantes, who trains clinical psychologists, argues that the studies systematically choose the kind of question a formula can answer: a single defined outcome, measurable within a fixed period, with a known base rate. Much of professional practice is not like that. Deciding what a person's difficulty actually is, in a conversation that generates the information as it goes, is not a prediction problem at all, and a literature about prediction has nothing to say about it. She is right, and the finding is narrower than its advocates generally admit.

The second objection is that a formula cannot explain itself, and this matters when the decision concerns a parole hearing, a mortgage or a child. A person is entitled to know why. This argument seems to me to point the opposite way from the direction it is usually aimed. A scoring rule with eight variables and stated weights can be written on one page, checked by the person it concerns, challenged on its facts and audited across thousands of cases for systematic bias. A professional's impression cannot be any of those things, and the confident reconstruction a practitioner offers when asked to justify a judgement is, on the evidence, frequently not the reason for it. If accountability is the standard, it favours the rule.

What actually works, where it has been tried, is neither pure judgement nor pure formula. The model produces a recommendation which stands as the default; the professional may depart from it and must record, in advance, the specific reason; and the departures are reviewed periodically against outcomes. This last part is the one that is always omitted and the one that matters, because unmonitored discretion to override reliably destroys the model's advantage: the overrides are made in exactly the cases where the vivid detail is most compelling, which are exactly the cases where the professional is least reliable.

It is worth asking why a result this well established has changed so little. Part of the answer is that the question is heard as an insult, since being told that a page of arithmetic outperforms one's professional intuition is not an invitation to collaborate. But the finding is not that practitioners are unnecessary. It is that a particular sub-task, the combining of several pieces of information into a single prediction, is one that human beings perform badly and a rule performs well, and that the professional's real work lies on either side of it: deciding what to measure, gathering it, and acting on the answer. That is a description of a division of labour, and it has been presented, for seventy years, as a verdict.