Prediction, Pre-Crime, and the Presumption of Innocence

This workstream deliberately grants the strongest technical assumption: future machine prediction may become very good. The constitutional issue survives. A person can be in a statistically elevated group and still not have committed the forecast event. Individual adjudication therefore needs a temporal boundary between future risk and past fact.

BOP’s PATTERN is a current, bounded example of formal recidivism risk assessment in the federal prison context. It demonstrates that person-level risk scores are not purely hypothetical. It does not establish that prediction has become general-purpose evidence across society, and Concresca does not treat it that way.

The EU AI Act provides important counterpressure by prohibiting specified individual criminal-risk assessment based solely on profiling or personality while preserving risk analytics tied to objective verifiable facts and human assessment. The architecture matters: forecasting can support where to look without becoming proof of what happened.

The Prediction–Evidence Wall therefore asks what new individualized facts exist before a risk output leads to detention, asset freezes, travel restrictions or other serious consequences. It also audits automation deference: a human reviewer with no practical power or incentive to override the model is not meaningful independence.