Machine Learning as a Tool for Hypothesis Generation
Problem
Hypothesis generation in social science remains informal; specifically, understanding what facial features drive judges' jailing decisions.
Why it matters
Judicial detention decisions affect human liberty, so uncovering hidden drivers of those decisions has high stakes.
What the AI did
An algorithm trained only on mug shot pixels predicts jailing decisions; humans interact with the black box to extract novel, interpretable hypotheses.
Results
Pixel-only algorithm accounts for up to half of the predictable variation in jailing decisions; generated hypotheses are novel and interpretable.
Impact
Demonstrates a systematic ML-assisted procedure for generating novel hypotheses about human behavior.
Record
| Venue | The Quarterly Journal of Economics |
|---|---|
| Published | 2024 |
| Work type | article |
| Authors | Jens Ludwig, Sendhil Mullainathan |
| Citations | 85 (OpenAlex, at import time) |
| References | 117 |
| Model or method | not stated in source |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1093/qje/qjad055 |
| Open access | closed |
| Keywords | machine learning, hypothesis generation, judicial decisions, interpretability, black-box models |
Importance rating
significant — thousands to millions affected
Judicial jailing decisions affect liberty; abstract shows face alone accounts for up to half of predictable variation and claims a general procedure for novel hypothesis generation from any high-dimensional data.