Maybe it is, maybe it isn't. But this is not relevant to what I'm saying. The goal of the AI is not to associate words in a hypothetical world where there is no discrimination or bias. The goal is to build a model that accurately reflects the real world.

Again, there may be justifications to deliberately introduce bias in the model to produce higher rates of association between women and technology than would normally occur if the creators of the AI so desire. But this is not elimination of bias, this is deliberate introduction of bias to make th model match the creators' worldviews.

> The goal is to build a model that accurately reflects the real world.

This is incredibly reductive - we use models for MANY purposes - prediction, inference, description, evaluation. Until you outline a specific use case for the model, understand where the data its trained on comes from and how it may or may not reflect reality, and think about whether the phenomena which you're measuring are important/sufficient for your application you CANNOT possibly justify your broad claims.

I'm well aware that models are used for a variety of applications. But regardless of the application, we usually want the model to reflect reality in some way.

Say you have an AI that predicts probability of re-offending for prospective parolees. If we observe that this system is flagging men as high risk more frequently than women, then this is not indicative of bias.

But!, one might say, men are socialized to be more violent and are more frequently recruited by gangs. These factors are being absorbed by the AI. Yes, they are, and this does not make the AI biased. The job of the AI to predict risk of re-offending in real world conditions. Not risk of re-offending in a hypothetical world where everyone has the same environment and experiences.

If we were to manually adjust the model to make it display equal risks for both men and women then that is not elimination of bias. It is deliberate introduction of bias to make the output fit our ideal worldview.

> Say you have an AI that predicts probability of re-offending for prospective parolees.

Already you've made one of the most common mistakes in this area: confusing the label for the latent truth. We don't measure "re-offending", we measure rearrest and re-incarceration, which as I've discussed elsewhere are biased (in the sense of "unjust").

> If we were to manually adjust the model to make it display equal risks for both men and women then that is not elimination of bias. It is deliberate introduction of bias to make the output fit our ideal worldview.

Again, you are using the word "bias" in a vague and meaningless way. In this context, our goals are constrained by the law - in the US we guarantee people certain rights, and prohibit certain kinds of discrimination. In this example, it would indeed be wrong (and illegal) to punish men more harshly just for being men, since gender is a protected class. In the same way, it would be wrong (and illegal) to punish black people more harshly just for being black. In my state, whites and blacks use cannabis at equivalent rates, and black people are arrested 8 times as often per capita - an algorithm looking at arrests will see that and make the problem worse.

You've staked out some sort of naiive and bizarre "AI Purity" stance which completely ignores how such models work, and which misunderstands how we use models to learn things and solve problems. You're also mixing up different definitions of "bias". Some facts which you might find interesting:

* Penalization/shrinkage, which is ubiquitous and incredibly useful in many predictive models, is the deliberate introduction of bias to improve prediction performance

* Adjusting for confounding in regression/classification type models is one way of accounting for or removing bias in a statistically rigorous way - as you (imprecisely and inaccurately) say, "modifying the model to make it display equal risks". This allows us to measure effects, conditional on known risks or existing patterns.

* All statistical and machine learning algorithms suffer from sampling bias, where the data we observe doesn't match the reality in important ways - a fact which you're completely ignoring. If a model reflects the bias in a bad sample, why on earth should we accept that?

> If we observe that this system is flagging men as high risk more frequently than women, then this is not indicative of bias.

If all it's doing is associating being a man with a higher risk factor without taking into account anything else in that man's life then, yes, it's being biased. "Being a man" in this situation is a correlation to, not casuation of, actual factors making a person more likely to commit crime.

This is a real problem that has been happening for years in Florida: the automated risk assessments used for a variety of decisions, including bail and sentencing factors, labels black people as inherently being more likely to reoffend than white people, even after controlling for actual crime rates and recidivism rates. https://www.propublica.org/article/machine-bias-risk-assessm...

That propublica article is the perfect example of the many aspects of the discussion here. It argues just one side of it, but the scenario it raises is one where an ML algorithm correctly found a solution to a badly framed problem.

Briefly, (and simplifying for clarity) it worked like this: the ML algorithm scored the risk of criminals reoffending. It gave a "high risk" score if someone was 80% likely to re-offend, and a "low risk" score if they had a 20% chance. Out of 100 blacks and 100 whites, here are the scores the ML algorithm gave, and how many ultimately did re-offend:

    100 black criminals
      * 50 "high risk", 40 re-offended
      * 50 "low risk", 10 re-offended
    100 white criminals
      * 10 "high risk", 8 re-offended
      * 90 "low risk", 18 re-offended

On the one hand, the algorithm was completely correct: 80% of high risk individuals re-offended, and 20% of low risk individuals did, and this is true even looking at just the black or the white criminals. It was unbiased according to its goals.

On the other hand, the failure mode disproportionately punished the black criminals: 10 were "high risk" who never re-offended, while only 2 white criminals were. Meanwhile, 18 "low risk" white criminals did re-offend while only 10 black criminals did. So the score was more strict for some unlucky blacks and more lenient for some lucky whites.

However, a key point is that because the underlying re-offense rates were different between the populations (50/100 for the black criminals, 26/100 for the white criminals), the algorithm could not have done otherwise. That is, given some n% re-offenders, if you have to fit them into 20% and 80% buckets, your "high risk" and "low risk" counts are mathematically fixed. In other words, it wasn't the ML by itself that was the problem, but the COMPAS score that it was trying to compute that had these issues inherent in it.

I think this is a good example where ML wasn't biased (at least, not anymore than reality), but where people were too eager to turn to it for a poorly considered project. By wrapping up important questions in high tech algorithms, it's too easy to fool yourself that what you're doing is the best thing you could be doing and miss problems in the fundamental framing.

1) Using skin color (or proxy) as basis for judgement or systematic discrimination, is illegal = punishable.

2) The whole point of systems is to impact the world, not replicate already recorded history or escalate problems.

Can the ML algo predict what the outcomes and consequences of its outputs are, and assess that? No?

The point is that the algorithm wasn't using skin color as an input, (directly, at least) and got those outputs anyways.

Would you prefer an outcome where half of the white men were also marked as high risk, but with 16 reoffending? (to give the low risk pools the same recidivism rate) Now instead of unjustly marking ten black men and two white men as high risk, you're doing so for thirty-four white men. Is that somehow less discriminatory?

Just because another option is worse doesn't make an unjust and repressive kind of social profiling better. The entire premise of using prediction may be found to have dire consequences, so the entire premise is dubious. There is no binary choice in this matter, but there needs to be accountability.

And if the decision isn't based on race, but on things that happen to correlate with it? Whether your father was present, for instance - since that's a major predictive factor for all sorts of life outcomes even when controlling for race. Should that not be used just because certain groups are more likely to come from single-parent households than others?

Because in the end, the choices are "accept that there will be a correlation between the outputs and race", "use no system that produces estimates or imperfect outputs", or "explicitly discriminate based on race to remove the race-result correlation". That's it.

It depends on the crime. But for crimes like, say, sexual assault being male is indeed a huge factor in likelihood to offend. Disparities between rates of women committing sexual assault and rates of men is immense - one or two orders of magnitude depending on the exact crimes counted.

I agree that an AI shouldn't use race or gender as an input. But we should not surprised when AIs have disparities when predicting risk of committing crimes, when there are disparities in the rates of actually committing crimes.

AI does not, in and of itself, have a goal. AI exists to be used for some human's goal. If the human's goal is to understand the level of association between women and technology without confounding factors like decades of marketing video games and computers to boys more than girls, the historical bias of such factors in the data fed to the AI becomes extremely important.

This is a common misunderstanding. When models impact the world, what matters more are their outcomes and consequences.

Right. And when you force a model to avoid disparities when disparities really do exist, the outcome is inaccurate data and the consequence is that decisions are made based on inaccurate data.

What people in this thread seem to be missing is that modeling and decision making do not have to be done simultaneously. You always want the most accurate model of reality as you can get, but what you do with that model is up to you.

Perhaps you think there are certain feedback loops that need to be broken (your model is usually a static representation of the reality), or perhaps you prefer low rate of false negatives / false positives, or perhaps if a model is uncertain you would like to defer to a human.

If a model predicts high probability of re-offence, you might decide to delay an action as an example, or re-examine the case in more detail.

I get a feeling that machine learning practitioners are not properly trained to recognize such subtleties, everything is a binary softmax output these days :)

EDIT: This seems to stem from the obsession with purely discriminative models that either directly model a binary function y = f(x) or model a function that returns some kind of score, which is later compared to a threshold y = f(x) > theta. Neither of these lend themselves nicely to this conceptual separation between modeling and decision making.

Sure, and there is also the assumption that good decisions are made on perfect/good data, which is also just an assumption. We can't even define good decisions well enough to be sure.

Just because bad data may tend (in hindsight) to yield inaccurate decisions, or decisions based on faulty data, doesn't promote data as the main driver behind sound decisions. That would just be another belief system, and we don't have all possible data or omniscience to prove it!

No, you can test and gauge the quality of that data. At this point, you're not really even making a point about whether or not we should introduce biases in our models to make them match our worldview - you're saying the entire endeavor of trying to build a model is worthless. If what you're saying is true, that we can't measure what decisions are successful or unsuccessful, then there's no point in trying to build a model because we no way of knowing whether the model is working or not.

It is the claim of vendors that their models works. It is irresponsible to just take their word for it, and such tests and critique need be ongoing.