Wisconsin's Prison-Sentencing Algorithm Challenged in Court (engadget.com)

"Do you want a computer to help decide a convict's fate?" asks Engadget, telling the story of a Wisconsin convict who "claims that the justice system relied too heavily on its COMPAS algorithm to determine the likelihood of repeat offenses and sentenced him to six years in prison." Sentencing algorithms have apparently been in use for 10 years. His attorneys claim that the code is "full of holes," including secret criteria and generic decisions that aren't as individually tailored as they have to be. For instance, they'll skew predictions based on your gender or age -- how does that reflect the actual offender...?

[T]he court challenge could force Wisconsin and other states to think about the weight they give to algorithms. While they do hold the promise of both preventing repeat offenses and avoiding excessive sentences for low-threat criminals, the American Civil Liberties Union is worried that they can amplify biases or make mistakes based on imperfect law enforcement data.
The biggest issue seems to be a lack of transparency, which makes it impossible to determine whether convicts actually are receiving fair sentences.

Comments

Yes please

[penguinoid]

If you think a computer is biased and unfair and makes decisions using secret criteria, wait till you meet a human!

Re:Yes please

[Archfeld]

There is no AI involved, just some code written by humans. A computer cannot express compassion. IMHO Compassion is the highest virtue, and the first to be discarded. Hope can be rekindled, faith restored, love reunited but compassion once lost is rarely seen again. Even the guilty deserve compassion, it is that which elevates man from beast. We do that which is necessary but it should NOT stop us from being sorry it has to come to that.

Re:Yes please

[JoeMerchant]

The advantage of a computer algorithm is that it can be analyzed, refined, improved, and relied upon to function more or less as well as it did in the past (assuming the new inputs aren't radically different from the previous ones.)

Now, judges - they come to their position through politics and lawyering, two selection criteria that would seem to get you the last people you would want to determine what is fair or reasonable for "the people at large." Even 200+ years ago this was recognized and is the basis for "trial by jury of peers" laws, and everyone knows that amounts to a random lottery style decision, but it's still better than leaving it up to a judge.

Re:Yes please

[penguinoid]

Having the computer show compassion is easy.
if (criminal.gender == female & criminal.attractive == true)
      showCompassion = true;

Re:Yes please

[lucm]

Having the computer show compassion is easy.
if (criminal.gender == female & criminal.attractive == true)

      showCompassion = true;

that looks like the algorithm they use in custody battles.

Re:Yes please

[Anonymous Coward]

Most of the algo's used are "privately developed", and thus not released because that would involve revealing "business critical information" (which obviously trumps the interests of the person being sentenced on the basis of it). Because the primary aim is to provide private entities with contracts/work/govt money.

Re:Yes please

[cdrudge]

My compiler says that criminal.attractive is undefined. Can you post the code for that property so that the system knows what attractive is.

Also, I think criminal.gender is deprecated. These days, I think you're suppose to use criminal.birthGender or criminal.identifiedGender.

Re:Justice is blind and buggy

[JoeMerchant]

If the software is open source, transparent, and clear in how it functions, I'd rather have sentencing software than a judge in most cases.

The problem comes when you get politicians meddling in the software, writing in "zero tolerance" code and "mandatory minimum sentences" - they've done this already in the legislatures, tying judges hands in sentencing decisions. With the data available: prior convictions, credit scores, family ties, etc., software can determine probable outcomes of lenient sentences.

Yes, I'm saying credit score is a point for consideration in sentencing - not decision of guilt or innocence, but once guilt has been established, credit score tells about a person's history of making good on commitments, and should be a strong predictor of their likelihood of meeting terms of a suspended sentence, probation or parole.

The real issue is lack of transparency

[cowtamer]

While the idea of using an algorithm to sentence a human being is bone-chilling, you might be able to justify this as a "formula" for sentencing -- which, of course, merits its own debate.

What is unconscionable about this is the fact that it's a SECRET algorithm. As in closed source. Essentially a secret law.

This has no place in democracy.

(Also, any algorithm which ingests statistical and demographic data is bound to come up with unpalatable and/or spurious demographic correlations (since there is a causal link between poverty and crime and a historic link between race and poverty) which I wold rather have society refrain from codifying -- in law or in actual computer code).

Statistics

[Hognoxious]

For instance, they'll skew predictions based on your gender or age -- how does that reflect the actual offender...?

It doesn't and it's not intended to.

It's statistics and probabilities, just like how insurance premiums work.

Re:Statistics

[NotInHere]

like how insurance premiums work.

That *is* skewing predictions based on gender and age.

If a computer program deduces from the fact that you are male that you will live a shorter time, and this makes women receiving lower payments due to them being more likely to live longer.

As this discriminated against women, this practice was swiftly forbidden by the EU. It probably wouldn't have been forbidden if it discriminated against men, but that's another issue. It was actual gender based discrimination.

If you let a computer program take factors like skin color or religion into account, that computer program will arrive at similar conclusions that humans arrived:

* blacks do more crimes per thousand people than white people
* highly educated people do less crimes than less educated people
* More muslims commit terrorist attacks than non-muslims

This doesn't mean that black people are ape-like, or that less educated people are automatically criminals. It usually just is a correlation, which doesn't imply causation. If you took a black person, and remove them from all the racist and discriminatory influences, AND the lower social status (that has been caused by lots of discrimination before), then won't end up with a higher crime likeliness than if you took a "normal" white person. Its probably even more caused by the social status than discrimination.

As for muslims, its simply caused because there is no powerful christian group that radicalizes christians to commit terror attacks. If you take the radicals out, the average muslim is as peace-loving as the average christian or jew.

Denying the existence of these correlations to protect from discrimination is wrong. This is just fuel for people who claim that the existence is denied because people want to "cover up" something, or because people are too dumb, usual criticism by people like trump. It is also wrong though to do actual discrimination, like israel does it, or trump proposes.

Re:The 'real' software

[Firethorn]

Why don't they use the fundamentals used in developing the algorithm to identify and eliminate the root causes of crime (through rehabilitation/counseling) and thus reduce crime in highly prone areas?

Honestly enough they should, there are pushes to do so, it's just that funding is an issue, as always. Funny thing is, though, it's supposed to actually save money!

Okay, up here in Alaska low oil prices have resulted in a government crisis - the government's income has dropped drastically. One of the reforms being put in place, finally, is 'community corrections', which has been shown to save money by actually preventing repeat criminal occurrences. As they mentioned, locking people up for long periods actually increases the chance they'll reoffend, especially if you don't provide support after they get out.

So rather than locking somebody up for 12 years, you lock them up for, say, 4. You take the money for the next 4 years of prison and put it into rehabilitating the criminal, which is enough to cover extended services in prison, as well as at least 4 years outside, because, surprise, it's cheaper than keeping them in prison. The last 4 years of prison sentence avoided is pure savings, though they mentioned that they're putting half the money into what you mentioned - addressing the core situations causing criminality in the first place.

As for the op -

His attorneys claim that the code is "full of holes," including secret criteria and generic decisions that aren't as individually tailored as they have to be. For instance, they'll skew predictions based on your gender or age -- how does that reflect the actual offender...?

Secret criteria is a problem, but gender and age have clear differences on how likely you are to re-offend. If you're a first time offender at 40 odds are something very strange happened, and if you fix that problem the person is unlikely to offend again, while being a first time offender at 15 is a bad sign that it might become a habit.