A Coda About Equal Treatment vs. Equal Outcomes

Separate from my general observations about "algorithmic bias", I wanted to dwell for a bit on something that Narayanan says around 32:06 in his presentation. Paraphrasing slightly:

If we want to harmonize, to balance outcomes between different groups, we have to treat different people from different groups differently, even if they are similar in all the ways we think that matter to the decision making task. That seems like a very uncomfortable notion to deal with.

He's absolutely right; you can't have both equal treatment and equal outcomes if there's any sort of difference in prevalence between groups. The tension between the two should make us uncomfortable, because no matter how you slice it it seems that someone is being treated unfairly.

Honestly, I don't find this problem to be nearly as vexing as he does. Here's my recapitulation of the underlying reasoning:

1. Treating different groups differently, when all the relevant facts about them are the same, is presumptively bad; "failure to treat like groups alike" is actually a pretty good definition of "unjustified discrimination".
   
2. This presumption can be overcome if such treatment serves to rectify injustices at the group level.
   
3. Inequality in outcome X at the group level is indicative of just such injustice.
   
4. From 1, 2, 3: Differential treatment of like groups is justified.
   

The problem here, though, is that it's easy to demonstrate that 3 doesn't hold for all X.

Consider Ibram X. Kendi's recent statement regarding racial disparities:

As an anti-racist, when I see racial disparities, I see racism.

This is a concrete example of the reasoning in step 3. If we take Kendi's statement at face value we should, for example, treat the overwhelming prevalance of African American employees in instituations catering to African Americans as a sign of anti-white bias. After all, racial disparities in workforce composition are, per Kendi, a clear sign of racism. But no reasonable person (including, presumably, Kendi himself) actually believes this to be the case, which demonstrates two things:

• Kendi's statement has unvoiced caveats.
  
• Group disparities in outcome can, in some cases, be explained by innocuous causes.
  

Or, put more plainly, it doesn't take an assumption of invidious motives to explain why Ebony's staff is mostly African-American.

Having demonstrated that unequal outcomes can occur for morally blameless reasons it follows that step 3 above needs to be rewritten:

3. Inequality in outcome X at the group level is indicative of just such injustice, provided a showing can be made that the disparity results from unjustified differential treatment between those groups.

"But", you may say, "you've set your standard of proof too high. It's quite difficult, in practice, to prove that differences in outcome are due to unequal treatment". My rebuttal is that's a feature, not a bug; it should be difficult.

I'm going to go all Rawlsian for a bit, because that seems to be a good framework for talking about this issue. It's plausible that an arbitrary individual, looking at this issue from behind the veil of ignorance, might agree to "take one for the team" and cede eir right to equal treatment in order to further a more just society, provided that it's clear that there is acctually an injustice at the group level. However, it's a much harder sell if the injustice is merely speculative; why should anyone give up eir claim to equal treatment to correct an injustice that is stricly conjectural? Assuming that disparities in group outcomes must be rectified is bad policy because it fails the test of public justification.

Before I sign off I should also point out that I've said nothing about step 2 so far. I don't think it holds either, but it wasn't necessary to go that far in this post. I'll have more to say about that next.