Interfaces with Decision Making

Topics

• The gap between model output and decisions; amplification of bias and feedback loops

Reading (Required)

This blog post looks at how judges use the COMPAS risk score in practice, and how it differs from the assumptions we make about idealized algorithmic decisions:

• If You Give a Judge a Risk Score

This article examines the problem of ‘selective label’ bias:

• The Selective Labels Problem: Evaluating Algorithmic Predictions in the Presence of Unobservables

This article defines a theoretical model to illustrate how, beginning with very slight existing bias, algorithmic systems will usually converge to an extremely biased steady-state. This model explains the reinforcement of existing inequities in police interactions.

• Runaway Feedback Loops in Predictive Policing

Reading (Optional)

This law-review article provides a comprehensive background to the discussion in the If You Give a Judge a Risk Score reading.

• Asseessing Risk Assessment in Action

This article surveys approaches to batch model updates of healthcare models that are retrained on prior decisions:

Hidden Risks of Machine Learning Applied to Healthcare

Reading Responses

For any two of the required readings:

1. Summarize the main argument of the reading
2. Give a supporting argument the author uses
3. Discuss any objects you may have or subtleties the author must deal with.