No! Because, for some reason that has no basis in reality, people think computers are less biased than humans! They trust computers OVER PEOPLE, even if the computer is racist, or just wrong.

Humans are accountable, changeable, cross-examinable. Computer systems are not. Please do not give enforcement systems over to computer systems, and have the bias calcified forever.

I'm not arguing for or against computers vs humans, but your argument is puzzling to me.

Computer systems are accountable, changeable, and cross-examinable. Arguably more so than humans.

I do stipulate that computer systems can be racist or otherwise biased and can be "just wrong".

I fully agree; a neural network can be continuously trained. Software can continuously perform A/B testing and change its mind when its current procedure proves inefficient.

The human mind tends to be set in its ways.

But it doesn't. Current systems do not do this, but people trust in them because machine learning is "more accurate", somehow. It's a huge problem managing expectations around a black box. Online training does not happen in any form at scale. These "fair" systems are, by and large, nonsense.

It really sounds preferable to me, even if the initial versions of these algorithms are biased. They will provide a concrete thing against which to test for bias and improve against. It is easily verifiable if your changes to an algorithm have removed the bias you are worried about. Conversely it is basically impossible to tell if your training on a human has changed their biases in the field, sort of just having them do their job again and see if they are biased.

> It is easily verifiable if your changes to an algorithm have removed the bias you are worried about.

Please never enter politics.

Yeah, god forbid we have audit capabilities in law enforcement.

Until we can enforce use of/non-deletion of footage from body cams, I think policing and governing with AI is a tick out of our depth, as a society.

What, would your worldview survive if we came to find that bias didn't drive the outcomes disparities that vex you?

Quite the opposite - I know how easily those administrating a system can bias it.

> They will provide a concrete thing against which to test for bias and improve against.

The history of junk forensic science has pretty clearly demonstrated that nobody in law enforcement cares about testing their biases, or improving the processes.

If it gets them results, they'll use it.

Law enforcement is general is not good at self-accountability. That accountability has to be rigorously enforced from the outside.

With algorithms, I think it can be done... but given how subtle the negative effects can be, the process must be stringent and err on the side of caution. In other words, any algorithm should be assumed to be biased until proven otherwise, and with a relatively high bar at that. There should be clear technical standards for what constitutes a valid proof. The testing process should be reproducible by third parties. And there should be a straightforward way to challenge either the process itself, or its results in any particular case.

In practice, given how low our bar for law enforcement accountability is today in other cases - including those where people die or are badly hurt - I'd say that we need to fix those issues first before talking about face recognition and its regulation in this context. And until then, it should be off-limits, because the potential for abuse is so high.

> Law enforcement is general is not good at self-accountability. That accountability has to be rigorously enforced from the outside.

Absolutely. One major positive outcome of the OJ Simpson trial was how Simpson's attorneys completely trashed the police handling of forensic evidence on a national stage, forcing law enforcement across the country to raise their game.

> That accountability has to be rigorously enforced from the outside.

I agree. But the reason that we don't have that right now isn't related to whether or not computers are involved, and introducing algorithms won't solve that situation.

The reason that police accountability is weak is political.

Sure. But until we resolve the accountability, we shouldn't increase the power of law enforcement. The more unaccountable power they have, the more it will be abused, and the graver the abuses.

When you're a hammer, your metric is how many nails you...nail.

> Risk scores are computed for every traffic stop, wouldn't you want to take that process away from the office and give it to a computer?

What I want is this process to be cross-examinable. You can't cross-examine an algorithm in any meaningful way at a jury trial. You can't cross-examine a neural network at all.

You know all the bullshit about how police dogs are used to conduct illegal searches, and how you can't cross-examine a dog? It'll be like that, but much, much worse.

No, you can challenge the validity of the risk assessment. Bayes is used in jury trials all the time.

https://en.m.wikipedia.org/wiki/Bayesian_inference#Applicati...

What we have today is institutional racism based on people’s gut reactions. An NN can be tested and shown that given all other factors it risks race, or any other protected class higher than it should.

> What we have today is institutional racism based on people’s gut reactions.

Institutional racism is not based on people's gut reactions. Institutional racism is based on flawed algorithms. Not computer algorithms in this case, but bureaucratic ones.

I don't see how a flawed computer algorithm would be any easier to correct than a flawed bureaucratic algorithm.

Because we can black box test the “risk assessment” that whoever sells to police departments. We can’t black box test bureaucracies.

Testing the results of the algorithms doesn't address the problem of institutional racism, though. Any bureaucracy will simply use the results that the algorithm supplies in a way that maintains the behavior that the bureaucracy is already programmed to maintain.

To address things like institutional racism, you have to change the institution. That's a human problem for which there is no technological solution (although technology can certainly be one of the tools used to create a larger change).

We can black box test bureaucracies just as poorly as we can software. For both, all we can do is measure inputs, and outcomes.

Since we've not made much traction on blackbox testing bureaucracies, I'm going to go ahead and say that blackbox testing algorithms will be even worse. (For some dumb reason, people are inclined to trust algorithms.)

You are aware that AI programs reflect the prejudices of the organizations that develop them?

https://science.sciencemag.org/content/356/6334/183.full

Yes, and it is well known that Tesla's self-driving program reproduces Elon Musk's reckless driving habits... /s

I'm just intervening to point out that what you're saying is not what the paper you linked implies. Its title is "Semantics derived automatically from language corpora contain human-like biases", and the general conclusion you could derive from it is that AI programs reflect the stereotypes ... of the data they are trained with. This is why they use the right word: stereotype, instead of the charged word you used: prejudice.

And who selects the training set?

Do you not mind that your point isn't reflected in the paper you linked?

I"m not aware of that result, can you help connect your paper to that claim?