Algorithms were going to save us, make our system more fair, more precise, more . . . real. Even bad algorithms were better than no algorithms, leaving it up to biased, if minimally competent, judges to make decisions they were incapable of making. Algorithms. Yeah, that’s the ticket.

There were, of course, some Luddites who didn’t appreciate the power of numbers, complained that algorithms would never be capable of taking into account the particulars of every defendant’s world. They would homogenize decision-making, reduce people to caricatures and, ultimately, end up replacing human bias with machine bias. We would feel more science-y about the law, but it would do nothing more than provide algorithmic cover for the same shallow bias it replaced.

For example, when ProPublica examined computer-generated risk scores in Broward County, Fla., in 2016, it found that black defendants were substantially more likely than whites to be rated a high risk of committing a violent crime if released. Even among defendants who ultimately were not re-arrested, blacks were more likely than whites to be deemed risky. These results elicited a visceral sense of injustice and prompted a chorus of warnings about the dangers of artificial intelligence.

And as it turned out, the algorithms used effective proxies for race and poverty, so they gave the appearance of being very high-tech while employing the same quantitative factors that humans applied instinctively. When judges did it on their own, it was called bias. Then machines did it, it was called science. And it turned out to be pretty much the same.

The next step in the process was taking the algorithm’s outcome and then introducing a post-hoc human tweak to rid it of the biased outcome. After all, if the basic premise was that algorithms would eliminate racial prejudice from the process, then produced disparate outcomes anyway, the algorithms had to be biased. Of course, the irony of human tweaking of algorithmic outcomes to align with ideology to remove bias from the system by re-introducing human bias into the system didn’t seem to bother people too much.

But now, a study shows that we’ve done no more than engage in a combination of mental and techno masturbation, albeit with the best of intentions.

The academics used a database of more than 7,000 pretrial defendants from Broward County, Florida, which included individual demographic information, age, sex, criminal history and arrest record in the two year period following the Compas scoring.

The online workers were given short descriptions that included a defendant’s sex, age, and previous criminal history and asked whether they thought they would reoffend. Using far less information than Compas (seven variables versus 137), when the results were pooled the humans were accurate in 67% of cases, compared to the 65% accuracy of Compas.

So random people did better than COMPAS? Even worse, while COMPAS used 137 measures, it turned out that only two mattered.

In a second analysis, the paper found that Compas’s accuracy at predicting recidivism could also be matched using a simple calculation involving only an offender’s age and the number of prior convictions.

“When you boil down what the software is actually doing, it comes down to two things: your age and number of prior convictions,” said Farid. “If you are young and have a lot of prior convictions you are high risk.”

Try as we might, the numbers refuse to play along. As it turns out, the same factors that judges intuitively applied all along were the best predictors of recidivism. No, they were hardly conclusive predictors, and no, they didn’t address root causes of the problem, such as why poor black defendants had a rap sheet as long as their arm in the first place, but that’s not the question for the legal system or the judge in deciding bail or sentence.

The paper also highlights the potential for racial asymmetries in the outputs of such software that can be difficult to avoid – even if the software itself is unbiased.

The analysis showed that while the accuracy of the software was the same for black and white defendants, the so-called false positive rate (when someone who does not go on to offend is classified as high risk) was higher for black than for white defendants. This kind of asymmetry is mathematically inevitable in the case where two populations have a different underlying rate of reoffending – in the Florida data set the black defendants were more likely to reoffend – but such disparities nonetheless raise thorny questions about how the fairness of an algorithm should be defined.

The reveal is that science doesn’t necessarily align with ideology. If the algorithm is unbiased, and algorithms are only as good or bad as their input, and still don’t yield the results one expects of them, is the fault in the algorithm or expectations?

Farid said the results also highlight the potential for software to magnify existing biases within the criminal justice system. For instance, if black suspects are more likely to be convicted when arrested for a crime, and if criminal history is a predictor of reoffending, then software could act to reinforce existing racial biases.

Or, as argued from the outset, the problem isn’t on the back end, where algorithms are employed in lieu of bespoke human involvement, but on the front end, with cops tossing poor black kids for kicks, because that’s where the action happens. Once a defendant has a conviction, it all goes downhill from there. Jobs, school, housing, all of it. And they end up being recidivists because what else are they going to do?

There’s nothing new about the input causes of recidivism, about how poverty means the inability to make bail, imposed for no particularly good reason, which produces pleas of convenience to get out of jail as soon as possible. How cops use ghetto kids to make their numbers, because that’s where they believe all the bad seeds can be found.

But none of this is going to be cured by algorithms. It never was. Garbage in, garbage out, and we’ve been putting garbage into the criminal justice system forever. What did we expect to be the outcome? The Sentence-O-Matic 1000 isn’t broken. It never worked in the first place, and it never will.