Computer Informants: Coming to a Police Precinct Near You?

Privacy expert Jim Adler has written a computer program that predicts whether a person has ever committed a felony. Adler used the criminal records of defendants – felons and non-felons – in Kentucky's criminal records database to build his algorithm. The aim of the project was to provoke a debate about the limits of data profiling in the realm of crime prevention.

Published Aug. 15, 2013

Adler found several features that correlated highly with having a felony record in Kentucky: having light skin, hazel eyes, misdemeanor offenses, tattoos and being male.


Using a representative portion of the entire database, Adler extracted six traits from each defendant's record.

Classifying features

Each possible trait, such as blonde hair, is weighted based on its statistical prevalence among Kentucky felons. A numeric value is assigned to the features along the algorithm's decision tree (higher totals are more indicative of a likely felon). If the sum of the values exceeds the threshold, in this case 3.25, a likely felon is identified.

Profiling isn't perfect: balancing anarchy and tyranny

The accuracy depends on the threshold score set by the program. At its best balance – between anarchy and tyranny – the program is correct 73 percent of the time. The program labels non-felons as felons 4 percent of the time, and ignores a felon in 22 percent of its decisions.


At its most lenient, it correctly identifies only 9% of felons, misses 91% and misidentifies no one (0%).

Balance between anarchy and tyranny


At its most aggressive, it correctly identifies all felons (100%), while misidentifying 19%.

Sources: Jim Adler, Bloomberg Research