Although I have to congratulate you for sticking to your guns and not RTFA, TFA does explain the situation. How the system is inaccurate varies by race.
The system errs on the side of predicting more risk than there actually is for blacks, and less than there actually is for whites. The study picked out 7K predictions and compared them with actual results.
The actual recidivism rates by race are irrelevant to this discussion about the bias of the system.
There is no debt, legally speaking, in an exchange of things, so vendors don't have to accept cash.
Way back when, paper money was sliver or gold certificates, which were IOUs, in that you could demand an appropriate silver or gold coin in exchange. Nowadays, it's Federal Reserve Notes, which are currency by themselves, not IOUs. Coins always were currency by themselves.
The "legal tender" issue means that you can always pay off a debt with cash. If you owe someone $20, you can satisfy the debt with a $20 bill, and your lender has no legal choice other than to accept it or forgive the debt.
I can buy something on Amazon on my own account, and pay shipping. Or my wife can buy on her Prime account, and she pays the same price except she doesn't pay shipping. That's what "free" means here.
Then there was the time my wife said she was going to the Apple Store and just looking. I got a call asking if this large transaction at the Apple Store was legit. I teased her about it when she came home.
Except that the predictions are inaccurate, and the inaccuracy depends on race. The algorithm used is racist in that it predicts that more blacks will re-offend than actually do, and predicts that fewer whites will re-offend than actually do. Regardless of any reality-based bias, the AI is racially biased somehow.
You seem to have just assumed that the algorithm is basically correct, and that attitude is really frickin' dangerous in real life. We have real people here who received worse treatment just because the algorithm was incorrectly biased against blacks.
Properly trained AI systems make rational decisions; they don't "discriminate" in any meaningful sense.
In which case this AI system was improperly trained, since it predicts blacks will re-offend more than they actually do, and predicts white will re-offend less than they actually do. The discrimination is very real and meaningful when you consider that the justice system uses these scores to determine what to do with an offender.
I realize this won't be a popular opinion, but perhaps the bias is warranted?
It isn't. That was determined by comparing the predictions to the actual outcomes. The system is biased toward expecting blacks to re-offend at a higher rate than reality, and whites to re-offend at a lower rate than reality.
Correction: People build a tool in a proprietary manner, so we don't know what's in it. Race is not an explicit input, but data related to race is.
The tool shows results that some people don't want to admit.
Correction: On later study, the tool shows results that are racially biased. The tool predicts blacks to re-offend more often than they do in real life, and does the opposite for whites.
The tool has to be racist and sexist.
The study either didn't cover sex, or it didn't find significant disparities. The tool is observed to be racist.
Now people will BUILD IN race and sex rules to counteract unbiased decisions.
Adding race as an input, and using it to predict that blacks will re-offend less and whites will re-offend more, would improve the accuracy of the tool.
So now the tool is racist and sexist.
There's no evidence given that the tool is sexist. After taking race into account, in a way that looks "politically correct", the tool would be more accurate.
The AI is, in fact racist. A study of 7000 individuals showed that it seriously overpredicted repeat offenses for blacks, and seriously underpredicted repeat offenses for whites. If it had a dose of political correctness added, it would be more accurate when compared to actual real-life data. Altering the AI so that it's more often correct is the right way to go about it.
That an AC doesn't understand is no surprise, actually.
The study compared the actual data to the algorithm predictions, and found that the algorithm predicted more crime for blacks than actually happened, and less for whites that actually happened.
I don't know what the rates of offending again were for blacks and whites. However, the algorithm predicted higher than the actual rate for blacks, and lower than the actual rate for whites. Obviously the algorithm has errors, but the errors are racially biased.
The narrative in this case is that the algorithm is racist. The actual crime rates were used in the comparison, and the algorithm predicted much higher re-offense rates for blacks than they actually had, and the reverse for whites.
You seem to be talking about cases where predictors have racially disparate but reasonably accurate outcomes. In this case, the algorithm is racially biased in predicting outcomes, as determined by a comparison of its output to real-life outcomes. How the bias got into the system is worth studying, but it doesn't affect the fact that it is biased.
The study compared the algorithm's predictions against real outcomes gathered two years later, and found that the errors were racially distributed. Blacks re-offended at a lower rate than the algorithm predicted, and whites at a higher rate. The algorithm is showing bias in a way that isn't reflected in the real data.
Ideally, we'd like an algorithm that would give a good prediction, but we don't have one. This one overestimates recidivism rates for blacks and underestimates them for whites. Therefore, if we gave a certain reduction to scores of blacks, the algorithm would be more accurate, and similarly if we gave whites a certain increase. It still would be unfair to some people, but it would be overall more fair.
Coming up with a new and better algorithm is difficult. Tweaking one we've already got is easier and less likely to introduce further error.
The study showed that the algorithm seriously overestimated the chances that blacks would commit another crime, and underestimated the chance for whites. It did that by looking at risk scores and criminal records. The program doesn't use race as an input, but it manages to be unfair to blacks anyway.
This isn't an accusation. This is a statistical study of known facts. If it's harsh, it's because the algorithm is seriously racist.
The current difference between black and white IQs in the US is, IIRC, less than the difference between IQs in the 1930s and IQs today. Genetics have not had time to change in less than a century, so obviously environmental factors were involved.
GMOs are like Dick Cheney deciding he owns your house because his out of control dog shat on your lawn.
And nobody's been able to find me a case supporting that. Monsanto has apparently only sued farmers who deliberately attempted to violate Monsanto patents.
Slashdot Mirror
Although I have to congratulate you for sticking to your guns and not RTFA, TFA does explain the situation. How the system is inaccurate varies by race.
The system errs on the side of predicting more risk than there actually is for blacks, and less than there actually is for whites. The study picked out 7K predictions and compared them with actual results.
The actual recidivism rates by race are irrelevant to this discussion about the bias of the system.
If "free" is to mean anything, it means free of additional cost to the decision maker.
There is no debt, legally speaking, in an exchange of things, so vendors don't have to accept cash.
Way back when, paper money was sliver or gold certificates, which were IOUs, in that you could demand an appropriate silver or gold coin in exchange. Nowadays, it's Federal Reserve Notes, which are currency by themselves, not IOUs. Coins always were currency by themselves.
The "legal tender" issue means that you can always pay off a debt with cash. If you owe someone $20, you can satisfy the debt with a $20 bill, and your lender has no legal choice other than to accept it or forgive the debt.
I can buy something on Amazon on my own account, and pay shipping. Or my wife can buy on her Prime account, and she pays the same price except she doesn't pay shipping. That's what "free" means here.
Then there was the time my wife said she was going to the Apple Store and just looking. I got a call asking if this large transaction at the Apple Store was legit. I teased her about it when she came home.
Another reason for an IPO is to let investors cash out.
Which is not what is happening here. The predictions are racist based on comparison to the actual data.
Except that the predictions are inaccurate, and the inaccuracy depends on race. The algorithm used is racist in that it predicts that more blacks will re-offend than actually do, and predicts that fewer whites will re-offend than actually do. Regardless of any reality-based bias, the AI is racially biased somehow.
You seem to have just assumed that the algorithm is basically correct, and that attitude is really frickin' dangerous in real life. We have real people here who received worse treatment just because the algorithm was incorrectly biased against blacks.
In other words, if an algorithm is wrong, and discriminates against a group, the fault is that that group hasn't learned to give correct input data?
The algorithm is not from ProPublica. ProPublica published the study showing that the algorithm is racially biased.
This one did. Some equalization by race would produce more accurate results.
In which case this AI system was improperly trained, since it predicts blacks will re-offend more than they actually do, and predicts white will re-offend less than they actually do. The discrimination is very real and meaningful when you consider that the justice system uses these scores to determine what to do with an offender.
It isn't. That was determined by comparing the predictions to the actual outcomes. The system is biased toward expecting blacks to re-offend at a higher rate than reality, and whites to re-offend at a lower rate than reality.
Correction: People build a tool in a proprietary manner, so we don't know what's in it. Race is not an explicit input, but data related to race is.
Correction: On later study, the tool shows results that are racially biased. The tool predicts blacks to re-offend more often than they do in real life, and does the opposite for whites.
The study either didn't cover sex, or it didn't find significant disparities. The tool is observed to be racist.
Adding race as an input, and using it to predict that blacks will re-offend less and whites will re-offend more, would improve the accuracy of the tool.
There's no evidence given that the tool is sexist. After taking race into account, in a way that looks "politically correct", the tool would be more accurate.
I rest my case.
The AI is, in fact racist. A study of 7000 individuals showed that it seriously overpredicted repeat offenses for blacks, and seriously underpredicted repeat offenses for whites. If it had a dose of political correctness added, it would be more accurate when compared to actual real-life data. Altering the AI so that it's more often correct is the right way to go about it.
That an AC doesn't understand is no surprise, actually.
The study compared the actual data to the algorithm predictions, and found that the algorithm predicted more crime for blacks than actually happened, and less for whites that actually happened.
I don't know what the rates of offending again were for blacks and whites. However, the algorithm predicted higher than the actual rate for blacks, and lower than the actual rate for whites. Obviously the algorithm has errors, but the errors are racially biased.
The narrative in this case is that the algorithm is racist. The actual crime rates were used in the comparison, and the algorithm predicted much higher re-offense rates for blacks than they actually had, and the reverse for whites.
You seem to be talking about cases where predictors have racially disparate but reasonably accurate outcomes. In this case, the algorithm is racially biased in predicting outcomes, as determined by a comparison of its output to real-life outcomes. How the bias got into the system is worth studying, but it doesn't affect the fact that it is biased.
Except that this isn't what's happening.
The study compared the algorithm's predictions against real outcomes gathered two years later, and found that the errors were racially distributed. Blacks re-offended at a lower rate than the algorithm predicted, and whites at a higher rate. The algorithm is showing bias in a way that isn't reflected in the real data.
Ideally, we'd like an algorithm that would give a good prediction, but we don't have one. This one overestimates recidivism rates for blacks and underestimates them for whites. Therefore, if we gave a certain reduction to scores of blacks, the algorithm would be more accurate, and similarly if we gave whites a certain increase. It still would be unfair to some people, but it would be overall more fair.
Coming up with a new and better algorithm is difficult. Tweaking one we've already got is easier and less likely to introduce further error.
The study showed that the algorithm seriously overestimated the chances that blacks would commit another crime, and underestimated the chance for whites. It did that by looking at risk scores and criminal records. The program doesn't use race as an input, but it manages to be unfair to blacks anyway.
This isn't an accusation. This is a statistical study of known facts. If it's harsh, it's because the algorithm is seriously racist.
The current difference between black and white IQs in the US is, IIRC, less than the difference between IQs in the 1930s and IQs today. Genetics have not had time to change in less than a century, so obviously environmental factors were involved.
And nobody's been able to find me a case supporting that. Monsanto has apparently only sued farmers who deliberately attempted to violate Monsanto patents.
There are scientists at the FDA, tasked with reviewing applications. I knew one.