To limit potential bias, Mr Ghani says, avoid prejudice in the training data and set machines the right goals.
That could, say, help focus hip replacements on those 1966 chevrolet patch panel likely to live longest.
When a new defendant is tested against these patterns, the risk of racial skewing should be lower.
In hospitals, for instance, doctors try to predict heart attacks so they can act before it is too late.And in India, Microsoft and the state government of Andhra Pradesh are helping farmers choose the best time to sow their seeds.Getting rid of lead paint may be easy; even with clever algorithms, stopping traumatised policemen from drawing their guns is not.These are used in bail, parole and (most controversially) sentencing decisions.(Northpointe, the algorithm provider, disputes the finding.).Policing may be helped, too.Analytical skills, however, are scarce.
Get our daily newsletter, upgrade your inbox and get our Daily Dispatch and Editor's Picks.
But a system developed by Rayid Ghani at the University of Chicago and others increases the correctness of at-risk predictions by 12 and reduces the incorrect labelling of officers as being at risk by a third.
Once trained, it can study a different group and accurately pick those at risk.
Mr Mullainathan and his colleagues show that machine learning can help predict the risk of death.
Still, people want to know how decisions that affect them are made.Jens Ludwig of the University of Chicago and his colleagues claim that their algorithm, tested on a sample of past cases, would have yielded around 20 less crime (see chart while leaving the number of releases unchanged.Combining supposedly confidential data sets can heighten the risk of accidentally identifying individuals.Manual systems correctly predict around.Bail decisions, in which judges estimate the risk of a prisoner fleeing or offending before trial, seem particularly ripe for help.No one can be sure that machine learning would have prevented the Texas scare.Many areas of policy, he suggests, could do with a dose of machine learning.It is now being used by the Charlotte-Mecklenburg police department in North Carolina.Machines are trained to find patterns that predict future criminality from past data.Better bail decisions are a big priority of its Data-Driven Justice Initiative, which 67 states, cities and counties signed in June.The White House is taking notice.Now it tries to spot vulnerable youngsters before they are poisoned.An online streaming services software predicts what they might enjoy, based on the past choices of similar people.Prediction is anyway probabilistic, not perfect.According to Stephen Goldsmith, a professor at Harvard and a former mayor of Indianapolis, it could also transform almost every sector of public policy.