Training machine learning to anticipate manipulation

August 01, 2023

Investigators: CDEP Affiliate Daniel Björkegren, Joshua Blumenstock, and Samsun Knight

An increasing number of important decisions are being made by machine learning algorithms. However, when algorithms are used to make consequential decisions, they create incentives for people to `game' the rule. When decision rules are gamed, they can produce decisions that are arbitrarily poor or unsafe. This problem is exacerbated when decision rules are disclosed, which inhibits efforts to make algorithms transparent.

This project develops and evaluates an approach to machine learning that anticipates manipulation, and produces more robust decisions even when decision rules are made transparent. It develops a framework for training machine learning models, and then demonstrates it in a field experiment with a custom app in Nairobi, Kenya. When implemented, decision rules estimated with this strategy-robust approach outperform those based on standard machine learning approaches. Additionally, the paper uses this framework to estimate the tradeoff between transparency and performance, which is small in this setting.