Starting in December, Michigan will test AI-assisted housing classifications at a number of men’s prisons. The pilot puts software into a process that determines where incarcerated people live and what programs are available to them. The Michigan Department of Corrections (MDOC) says it will use nonprofit tech company Recidiviz’s software to “optimize placement of incarcerated individuals.”
Detroit channel Fox2 reports that the department is proceeding with an algorithm designed to classify incarcerated people by how dangerous they are and sort them into housing levels. Although Michigan plans to deploy the system across its prisons, the initial rollout will be a limited pilot rather than a statewide launch.
Security classifications divide Michigan’s prison housing into five levels. Screenings were previously handled by a staff committee chosen by the warden at each facility. In Futurism’s account, the state appears to be handing that decision to Recidiviz, or at least giving its software considerable influence over the outcome. The difference is important: the reporting doesn’t establish that software will independently decide final placements.
A risk score is only part of what’s at stake. Housing classifications are extremely consequential: they govern a person’s eligibility for programming and resources and determine whom they’ll live alongside for the duration of their sentence. The prediction can thus affect their actual prison conditions, not just add information to a file.
MDOC’s announcement says the system is expected to “improve MDOC’s ability to more accurately predict incarcerated individuals’ risk of future violence by approximately 30 percent,” although Futurism says the basis for that figure is unclear. The agency is describing an anticipated gain, not reporting a result demonstrated by the pilot, which has yet to begin.
Originally a project at Google for organizing data, Recidiviz set out to “uncover how technology could standardize criminal justice data and improve the system nationwide.” According to the organization, its platform, which integrates AI, can find people who qualify for state-offered programs, reclassification, or release, while coordinating access to transitional housing and employment resources. These are descriptions of the platform as a whole, not confirmation that Michigan’s housing pilot will use all of those functions.
For Futurism, the partnership’s unanswered questions center on the data the system will gather and use. Its accuracy will depend heavily on those inputs. As a comparison, the article points to a ProPublica investigation into a similar Louisiana risk-analysis program. That reporting found that the information used to classify incarcerated people carried inherent racial and socioeconomic bias.
ProPublica documented the case of a blind, wheelchair-bound 70-year-old with a clean disciplinary record. Despite that record, he received a “moderate risk” classification for reoffending, which resulted in an automatic loss of parole eligibility. Futurism draws on this reporting to argue that Louisiana’s carceral system became crueler and more harmful as a result to people earnestly working toward rehabilitation.
Because Michigan is using its tool for housing placements rather than parole, the Louisiana example doesn’t establish how Michigan’s rollout will turn out. Nevertheless, Futurism concludes that software by itself is very unlikely to fix Michigan’s ongoing corrections crisis. It warns that, if Louisiana’s adoption of AI is any example, Michigan’s system could make matters far worse. Futurism said it sought comment from the Michigan Department of Corrections.
Specific questions about the December trial are still open: which data will inform classifications, what effect the predictions will have on housing and access to programs, and how MDOC will assess its claimed approximately 30 percent gain in predicting violence.
