7: Hellman on Giving Algorithms Access to Race

What is disparate treatment? What are the two ways in which a practice can constitute disparate treatment (explicit racial classification and invidious intent)?

What is an example of each?

Disparate treatment occurs when a decision procedure engages equal protection restrictions by making distinctions based on protected traits. The two ways it occurs are through explicit racial classification (such as considering race directly in college admissions) and invidious intent (such as using zip codes in loan applications as an intentional proxy for race).

What is the first option for using racial classifications to limit error disparities: setting different “cut points” (thresholds) for labeling defendants “high risk,” depending on the defendant’s race? What does Hellman think about this option?

The first option sets different risk score thresholds or "cut points" for labeling defendants as high risk depending on their race, such as classifying White defendants as high risk at a score of six but Black defendants at a score of eight. Hellman rejects this option, concluding that an algorithm may not deploy different cut points for Black defendants than for White defendants.

What is the second option for using racial classifications to limit error disparities: using different “tracks” within the algorithm for different racial groups? Give an example of how this might work.

The second option uses different "tracks" within an algorithm so that non-racial predictive traits are evaluated differently for each demographic group based on their specific predictive power. For example, an algorithm might include housing stability as a predictive factor for White defendants while excluding it for Black defendants because housing stability is less predictive of recidivism for Black defendants.

What are the two requirements Hellman identifies for the use of racial classifications to count as disparate treatment? Be able to describe them as well as name them.

What examples does Hellman give to illustrate these two requirements?

In what way do the examples support the requirements?

Hellman identifies proximate effect (the requirement that the use of racial classification produces a direct, proximate effect rather than remote downstream consequences) and racial generalization (the requirement that the classification relies on a statistical generalization about a racial group).

Examples and How They Support the Requirements:

• Hellman uses the U.S. Census (requiring self-reported race) to show that downstream consequences of collecting race data are too remote to count as proximate effects, and police suspect descriptions (using eyewitness accounts of a perpetrator's race) to show that identifying a specific individual does not rely on a statistical generalization about a racial group.

These examples support the requirements by demonstrating that using racial classifications without direct harm or racial generalizations does not constitute legally prohibited disparate treatment.