Artificial Intelligence and Administrative Decisions Under Uncertainty

Abstract

How should artificial intelligence guide administrative decisions under risk and uncertainty? I argue that artificial intelligence, specifically machine learning, lifts the veil covering many of the biases and cognitive errors engrained in administrative decisions. Machine learning has the potential to make administrative agencies smarter, fairer and more effective. However, this potential can only be exploited if administrative law addresses the implicit normative choices made in the design of machine learning algorithms. These choices pertain to the generalizability of machine-based outcomes, counterfactual reasoning, error weighting, the proportionality principle, the risk of gaming and decisions under complex constraints.

Appendix

• 1. An ordinary least squares regression takes the functional form:

  $$ {y}_i={\beta}_0+{\beta}_1{x}_{1i}+\dots +{\beta}_n{x}_{ni}+{\varepsilon}_i $$

• 2. The objective function usually used for regression algorithms is to minimize the mean squared error (MSE) between the vector of observed values Y_i and the vector of predicted values ${\hat{Y}}_i$ given a set of n observations:

  $$ MSE=\frac{1}{n}\sum \limits_{i=1}^n{\left({Y}_i-{\hat{Y}}_i\right)}^2 $$

The closer MSE is to zero, the higher the accuracy of the predictor.

• 3. Technically, variance describes the difference between estimates of a value in one sample from the expected average estimate of the value if the algorithm were retrained on other data sets:

  $$ V=E\left[{\left(\hat{f}(x)-E\left[\hat{f}(x)\right]\right)}^2\right] $$

• 4. Technically, bias is the difference between the expected average prediction of the machine learning and the true value that the model is intended to predict:

  $$ B=E\left[\hat{f}(x)-f(x)\right] $$