Authorities in the US state of Michigan plan to introduce an artificial intelligence (AI) system to assess the risk posed by prison inmates. The algorithm will classify inmates into four risk categories based on the likelihood of them committing acts of violence. The pilot project is scheduled to launch in December at several men’s correctional facilities, El.kz reports.
Previously, decisions on inmate classification were made by committees comprising prison staff. The authorities now plan to delegate this task to a system developed by the non-profit organisation Recidiviz, which originated as a Google project. According to official statements, the introduction of the algorithm is expected to improve the accuracy of predictions of future violent behaviour by approximately 30%.
The initiative has raised concerns, as an inmate’s assigned risk level affects their conditions of confinement, including access to educational and rehabilitation programmes and other resources. It also determines which other inmates they may be housed with within a correctional facility.
Problems associated with the use of similar algorithms have already emerged in other US states. Louisiana, for example, operates the TIGER system, which was originally developed as a tool to assess inmates and support their rehabilitation. However, the algorithm subsequently became the sole criterion for making decisions on the denial of early release.
The programme primarily takes into account an individual’s criminal history, including previous offences, employment history, age at first arrest and violations of parole conditions. However, it does not consider inmates’ conduct while in prison, their participation in educational courses, completion of therapy or other efforts towards rehabilitation.
An investigation by ProPublica highlights the case of a 70-year-old man who had almost completely lost his eyesight and relied on a wheelchair for mobility. While in prison, he completed anger management courses, received vocational training and enrolled in a substance abuse treatment programme.
However, two months before his scheduled early release hearing, the man was informed that he was no longer eligible for consideration for early release. The TIGER algorithm had assessed his likelihood of reoffending as moderate. Under the state’s new legislation, this assessment was sufficient to automatically disqualify him from a parole hearing.
Such cases raise broader questions about algorithmic transparency and accountability for decisions that directly affect people’s lives. In Michigan, it remains unclear how thoroughly the authorities will scrutinise AI-generated assessments and whether correctional facility staff will have the authority to review or override the system’s recommendations.
