Do algorithms dream of electric criminals ?

Not a terribly large amount of detail here, but interesting nonetheless.

On 73 occasions, over an eight-year-period, Ebrahimi had reported to the police that he'd been the victim of racially motivated crimes. His complaints went unheeded and a report into his murder concluded that both Bristol City Council and the police were guilty of institutional racism. "That was really a turning point from a data perspective," says Jonathan Dowey, who heads a small team of data analysts at Avon and Somerset Police. The question the force began to ask, he says, was: "Could we be smarter with our data?" 

Humans are susceptible to all manner of biases. And unlike computers, they're not good at spotting patterns - whether, for example, there have been multiple calls from the same address. When Avon and Somerset retrospectively ran a predictive model to see if the tragedy could have been averted, Ebrahimi's address popped up as one of the top 10 raising concern.

I think "not good" needs to be heavily qualified here. Under certain conditions they are absolutely astonishingly good at spotting patterns : try writing an algorithm to spot a tiger in a forest and then tell me humans aren't good at pattern recognition. True, there are plenty of cases where we screw up, but plenty where we do much, much better than an algorithm. Had the data been presented in an accessible format, it would have stood out like a sore thumb (because we're apparently very good at spotting those).

The public may be largely unaware of how algorithms are penetrating every aspect of the criminal justice system - including, for example, a role in sentencing and in determining whether prisoners get parole - but civil liberties groups are becoming increasingly alarmed. Hannah Couchman of Liberty says that "when it comes to predictive policing tools, we say that their use needs to cease".

OK, that is an absurd conclusion, but there's good reason to be skeptical :

Algorithms are informing vital decisions taken about peoples' lives. But if the computer suggests that someone is at high risk of re-offending, justice surely requires that the process by which this calculation is reached be not only accessible to humans but also open to challenge.

Shouldn't a prediction by an algorithm be treated with the same caution as a prediction by any other method ? A prediction based on correlations and extrapolations is not the same as an analytic deduction of what's going to happen next, which is what people may be confusing them with. Surely in general you respond to predictions from credible sources as though they were possible but not certain. If someone tells you that John Smith is going to murder Joe Bloggs, you keep watch, but you don't lock up Smith on those grounds alone. It would be foolish indeed to assume 100% reliability or completeness.

An even thornier issue is algorithmic bias. Algorithms are based on past data - data which has been gathered by possibly biased humans. As a result, the fear is that they might actually come to entrench bias. There are many ways in which this might occur. Suppose, for example, that one of the risk-factors weighed up in an algorithm is "gang membership". It's possible that police might interpret behaviour by white and black youths differently - and so be more likely to identify young black men as members of gangs. This discriminatory practice could then be embedded in the software.

Yes, but on the other hand, the best way of beating bias is through objective measurement. Using human judgement is hardly more likely to avoid bias than using an algorithm. An objective procedure is not the same as being objectively correct, but by defining the problem as clearly as possible, biases can be exposed much more readily than if using subjective judgement.

Which variables should go into an algorithm is hugely contentious territory. Most experts on police profiling want to exclude race. But what about sex, or age? What makes one variable inherently more biased than another? What about postcode? Durham initially included postcode in their Hart tool, but then removed it following opposition. Why should people be assessed based upon where they lived, the objection ran - would this not discriminate against people in less desirable neighbourhoods?

And there's the rub. Correlation doesn't equal causation, but statistical predictions don't care about that : they want as much data as possible. A statistically significant correlation can be absolutely meaningless. Without human oversight this is going to give ludicrous results, but equally, trying to figure out the causes of crime or predict it without data will also give nonsensical answers. And by entrenching bias against a particular demographic, there's a risk of driving that demographic toward further crime.

Perhaps a better approach would be to use this to identify areas where social policy needs to change, rather than finding areas to pre-emptively punish, and for displaying the data in different ways for humans to interpret rather than making direct predictions. We're still in the infancy of using big data to make statistical predictions, but to avoid using it, I think, would be foolish indeed. Only through rigorous examination of the data (and not our own feelings) do we have any chance of establishing which biases are justified, which have complex causes, and which are utterly false.

Could an algorithm help prevent murders?

Algorithms are increasingly used to make everyday decisions about our lives. Could they help the police reduce crime, asks David Edmonds. In July 2013, a 44-year-old man, Bijan Ebrahimi, was punched and kicked to death in south Bristol. His killer, a neighbour, then poured white spirit over his body and set it alight on grass 100 yards from his home.