Algorithmic reparation has been proposed as an alternative to algorithmic fairness.

Algorithmic fairness = refining or finetuning algorithms to reduce harm

Algorithmic reparation = broader, systemic approach that “displac[es] fairness in favor of redress.”

Left unresolved is the question of how algorithmic reparations ought to be implemented. The term “algorithmic reparation” invites at least two interpretations:

[fit] Two potential applications of “Algorithmic Reparations”

1. making reparations more effective by incorporating algorithms into the process

2. a particular form of reparations that targets harms caused by algorithms

^ The paper addresses #1, our paper addresses #2

Structure of the Paper

1. Develops a prototype for reparations that target harm caused by algorithms.

2. Tests this prototype against classic law-and-tech critiques

3. Proposes a framework for understanding the efficacy of algorithmic reparations in practice

Prototype for reparations that target harm caused by algorithms

Drawing upon international law and existing reparative frameworks, we can apply those principles to address algorithmic harm specifically.

Reparations principles

• Restitution
• Compensation
• Rehabilitation
• Satisfaction
• Guarantees of non-repetition

Critique #1:

Does algorithmic reparations — understood as a particular kind of reparations that targets harms caused by algorithms — require unecessarily specific legal rules when general rules should suffice?

In other words, is the existing legal framework for reparations in some way inadequate to handle algorithmic harm?

Critique #1:

Does algorithmic reparations — understood as a particular kind of reparations that targets harms caused by algorithms — require unecessarily specific legal rules when general rules should suffice?

Answer: No, because general rules and principles for reparations suffice. With algorithmic reparations, the process of reparations need not be made more particular, but the target of harm to be redressed should be made more particular.

Critique #2:

Why algorithmic reparations and not just reparations? Is it a solution in search of a problem?

Critique #2:

Why algorithmic reparations and not just reparations? Is it a solution in search of a problem?

Answer: Not always!

Reparations at large are not always possible. Algorithmic reparations may be politically more feasible.

But there’s a risk that algorithmic reparations will be insufficient or pinpoint the wrong target for redress. How do we address this?

Algorithmic reparations are an effective method of redress only for the upper left quadrant, when an algorithmic technology is both necessary and sufficient. Algorithmic reparations is less effective when algorithmic technology is not both necessary and sufficient for the harm to be redressed. This captures three out of the four squares on our grid: sufficient but not necessary, necessary but not sufficient, and not necessary and not sufficient.

Definitions:

Necessary: the harm only occurs when the technology is used

Sufficient: when the technology is used, the harm occurs

Examples

Necessary and sufficient The algorithmic technology of electronic monitorting producing the harms of stigmatization, false technical violations, and constant surveillance

Sufficient but not necessary Racially disproportionate enforcement of traffic laws through automated traffic systems

Necessary but not sufficient Someone receiving an inadequate defense in a criminal case because the local public defender service outsourced the work to a natural language processing model

Not necessary and not sufficient Police officers using an algorithmic GPS system to help navigate their car to a location where they subsequently commit the harm of unjustifiably attacking someone

Solid line = the appropriate target for redress Dotted line = what algorithmic reparations would target for redress