The power of learning

IN “Minority Report”, a policeman, played by Tom Cruise, gleans tip-offs from three psychics and nabs future criminals before they break the law. In the real world, prediction is more difficult. But it may no longer be science fiction, thanks to the growing prognosticatory power of computers. That prospect scares some, but it could be a force for good—if it is done right.

Machine learning, a branch of artificial intelligence, can generate remarkably accurate predictions. It works by crunching vast quantities of data in search of patterns. Take, for example, restaurant hygiene. The system learns which combinations of sometimes obscure factors are most suggestive of a problem. Once trained, it can assess the risk that a restaurant is dirty. The Boston mayor’s office is testing just such an approach, using data from Yelp reviews. This has led to a 25% rise in the number of spot inspections that uncover violations.

Governments are taking notice. A London borough is developing an algorithm to predict who might become homeless. In India Microsoft is helping schools predict which students are at risk of dropping out. Machine-learning predictions can mean government services arrive earlier and are better targeted (see article). Researchers behind an algorithm designed to help judges make bail decisions claim it can predict recidivism so effectively that the same number of people could be bailed as are at present by judges, but with 20% less crime. To get a similar reduction in crime across America, they say, would require an extra 20,000 police officers at a cost of $2.6 billion.

But computer-generated predictions are sometimes controversial. ProPublica, an investigative-journalism outfit, claims that a risk assessment in Broward County, Florida, wrongly labelled black people as future criminals nearly twice as often as it wrongly labelled whites. Citizens complain that decisions which affect them are taken on impenetrable grounds.

These problems are real, but they should not spell the end for machine learning as a policy tool. Instead, the priority should be to establish some ground rules and to win public confidence. The first step is to focus machine learning on applications where people stand to gain—extra help at school, say, rather than extra time in jail.

More can be done to assuage concerns about transparency. Algorithms can be modified to reveal which components of their inputs had the most influence on their decisions, for example. But full transparency has risks. If restaurants know that five-star reviews will guarantee fewer inspections, they may make them up. Even so, regulators should insist that government users know the factors behind predictions, and that these are explained to affected citizens upon request. Above all, algorithms should help people make decisions, not make decisions for them—as can be the case with credit-scoring.

Colour-blind computing

The trickiest issues lie in criminal justice, but here too machine learning could still do much good. The threat of racial bias can be minimised by paying close attention to the distribution of false-positive results while the system is being trained. With or without programs to help them, judges have to make plenty of predictions, for instance about whether a person will commit a crime or flee before trial. They can display lifelong bias (they are, after all, only human). The right machine could make their decisions fairer.

In the end Mr Cruise’s psychics were banished to an isolated island. Machine learning deserves no such fate. But to avoid rejection, it needs to be used in the right situations with the right caveats; and it must remain a tool in human hands. Do that, and the benefits promise to be vast.