As we learned before, hate facts “are true statements about reality that our elites demand remain occult and unuttered.”
The problem is that hate facts will routinely pop up in statistical (a.k.a. machine learning, a.k.a. artificial intelligence) algorithms, and when they do the algorithms are said to be “biased”. The paradigmatic example are algorithms which estimate the chance of persons paying back loans. Race was found to be highly informative in these algorithms, but race is also unwelcome, so modelers were forbidden to use it.
The blame for the “bias” is put on the algorithm itself, but, of course, the algorithm is not alive, not aware, and so does not know the numbers it manipulates are anything but numbers. The meaning to numbers is found only in our eyes.
Which brings us to the Nature article “Bias detectives: the researchers striving to make algorithms fair: As machine learning infiltrates society, scientists are trying to help ward off injustice.”
It begins with a sob story, as is, we guess, mandatory in pieces like this.
In 2015, a worried father asked Rhema Vaithianathan a question that still weighs on her mind. A small crowd had gathered in a basement room in Pittsburgh, Pennsylvania, to hear her explain how software might tackle child abuse…the system does not catch all cases of abuse. Vaithianathan and her colleagues had just won a half-million-dollar contract to build an algorithm to help…
After Vaithianathan invited questions from her audience, the father stood up to speak. He had struggled with drug addiction, he said, and social workers had removed a child from his home in the past. But he had been clean for some time. With a computer assessing his records, would the effort he’d made to turn his life around count for nothing? In other words: would algorithms judge him unfairly?
In other words, this father guessed the algorithm might use the indicator “past druggie”, and use it to up the chances he’d abuse a kid. Which certainly sounds reasonable. Druggies are not known to be as reliable with kids as non-druggies, on average. You dear reader, would for instance use the information were you deciding on baby sitters.
However, past drug use is a hate fact in the eyes of the Nature author. How to ensure it’s not used?
I changed the colors from “blue” and “purple” to the more accurate, but hate fact, “white” and “black” in the following passage:
Researchers studying bias in algorithms say there are many ways of defining fairness, which are sometimes contradictory.
Imagine that an algorithm for use in the criminal-justice system assigns scores to two groups ([white] and [black]) for their risk of being rearrested. Historical data indicate that the [black] group has a higher rate of arrest, so the model would classify more people in the [black] group as high risk (see figure, top). This could occur even if the model’s developers try to avoid bias by not directly telling their model whether a person is [white] or [black]. That is because other data used as training inputs might correlate with being [white] or [black].
Knowing a person’s race is useful information in predicting recidivism. Note, again, the algorithm does not, and is incapable, of saying why race is useful information. It is entirely neutral, and cannot be made non-neutral. It cannot be biased, it cannot be unbiased. It cannot be equitable, and it cannot be unequitable. The interpretation, I insist, is in the eyes’ of the users.
“A high-risk status cannot perfectly predict rearrest, but the algorithm’s developers try to make the prediction equitable”. What is in the world can that possibly mean? Since the algorithm cannot be equitable or biased, it must be that the modelers insist that model does not make use of hate facts, or create them.
Now the author prates on about false positives and negatives, which are, of course, undesirable. But the better a model gets, in the sense of accuracy, then the fewer false positive and negatives there will be. If the model is hamstrung by denying hate facts as input, or it is butchered because it produced hate facts, then model inaccuracy must necessarily increase.
What makes the whole thing laughable, is that algorithm builders are being denied even access to hate facts, so they can’t check whether their models will be judged as “biased.” For instance, race cannot be input, or even viewed, except by authorities who are free to use race to see whether the models’ outputs correlate with race. If it does, it’s “biased.”
The best way to test whether an algorithm is biased along certain lines — for example, whether it favours one ethnicity over another — requires knowing the relevant attributes about the people who go into the system. But the [Europe’s General Data Protection Regulation]’s restrictions on the use of such sensitive data are so severe and the penalties so high, Mittelstadt says, that companies in a position to evaluate algorithms might have little incentive to handle the information. “It seems like that will be a limitation on our ability to assess fairness,’ he says.
Diversity is our weakness.