Michigan prisons to pilot Recidiviz AI risk tool

That is the bull case for a technology increasingly being pushed into public-sector decision-making: better data could mean better outcomes. But the downside is equally material. The security label assigned to a prisoner determines who they are housed with and what programs or support they can access, so errors can have immediate consequences. If the model reflects flawed data or built-in bias, it could reinforce racial or socioeconomic disparities rather than remove them.

The concern is not theoretical. A ProPublica investigation into a similar Louisiana risk tool found that biased data distorted assessments and pushed some inmates into higher risk categories. That history is likely to intensify scrutiny of Michigan’s pilot and of Recidiviz, which emerged from a Google-linked data project and now positions itself as a digital infrastructure provider for the correctional system.

For investors, the story is less about prisons alone than about a wider theme: governments are increasingly willing to adopt AI for high-stakes operational tasks, but only if vendors can prove the systems are auditable, accurate and defensible. That creates a commercial opening for correctional-tech providers, but it also raises regulatory and reputational risk. Any widely publicized failure could slow procurement, tighten oversight and make adoption slower and more expensive.

The broader implication is that AI in public safety is moving from testing to governance. Michigan’s pilot will be watched as a case study in whether algorithmic tools can improve prison management without importing the biases they are meant to reduce. If the system performs well, other states may follow. If it fails, the backlash could shape how far AI is allowed to go inside U.S. prisons.