Facebook Ads Can Still Discriminate, Despite a Civil Rights Settlement

* housing, employment, and credit Ads

I think the easiest solution would be to disallow ads of those categories on their platform. I'd think the risk of "facebook/instagram is racist" damaging their brand and the cost of federal discrimination lawsuits would outweigh whatever revenue they project.

As an aside, I know it's faux pas to bring up any observed (and/or presumed) differences between the protected classes - but maybe (just maybe) Facebook's targeting is smart enough to correlate "most likely to care" about things that tend to have skewed demographics without looking at the demographic data itself. Like the example in of truck driver ads targeting men, what is Facebook using to determine who they target? And do those data points line up with demographics?

I don't know, but these kinds of systems are tough to introspect from the outside.

That "smart"ness you describe is almost definitely what the algorithm's doing. And this is all well and good, until you take into consideration the fact that these systems tend to amplify existing biases.

If it spots a correlation, it'll amplify the correlation, regardless of whether it's actually meaningful. There are probably some originally spurious correlations that Facebook has amplified into existence, given how big and all-encompassing it is. It's the same problems that lead to racist AI judges.

I can’t understand, in succinct plain terms, what the desired outcome is here.

What would it look like if the ads were not biased, or discriminating, in any bad way.

For stuff like "employment, housing and credit", it's illegal to discriminate in advertising. That's why they made a whole new portal to try to avoid discrimination for those kinds of ads. So after a while if you gather data and realize that your algorithm is sexist and ageist, then continuing to use that algorithm to place ads is knowingly using sexist and ageist techniques in advertising. I think the EEOC has the authority to determine how non-discriminating FB will have to be to avoid trouble.

In my opinion, incorrect bias can happen only if either the input or process is bad. So, if you consider the output undesirable then there is a problem with input or process.

Ideally, ads would not be targeted at all in any way and micro targeting should be a crime (for privacy reasons as well). Short if that, you target people because of their proven individual preferences and tendencies as opposed to using their membership of any larger group of any kind as an input variable. Regardless of how much accuracy it enables, associative generalization is the core problem here, even without protected classes like race for example, I don't think it should be legal for any ad content to target me because I am a member of HN,now if by my explicit consent the ad system discovers I like specific topics which are also common on HN then they can use those individual level inputs WITHOUT associating it with my HN membership.

Essentially, the enemy here is formation of social bubbles as a result of explicitly clustering individuals into easily targtable micro groups. In other words, it is easier to sell ads that target specific micro groups but not so much for specific individuals.