AI in the Courtroom: How Biased Algorithms Are Sending Innocent People to Prison

An updated ProPublica investigation in April 2026 found that AI sentencing tools incorrectly flag Black defendants as high-risk at 2.3 times the rate of white defendants with identical criminal histories. Latino defendants fare slightly better but are still flagged at 1.8 times the rate. These are not edge cases — they represent systematic, measurable discrimination baked into the mathematics of the algorithms.

When defendants challenge their AI-generated risk scores, they hit an impenetrable wall. The algorithms are proprietary — owned by private companies that refuse to disclose how they work, citing trade secrets. Judges who rely on these scores often do not understand the underlying methodology. Defendants are literally being sentenced based on evidence that no one in the courtroom — not the judge, not the defense, not the prosecution — can explain or verify.

The companies selling AI sentencing tools to courts have a perverse incentive structure. Their contracts depend on appearing effective, which means predicting recidivism accurately. But predicting recidivism using historical data means encoding historical biases — if Black communities were over-policed and over-prosecuted in the past, the algorithm will predict they will be in the future. The tool does not predict crime; it predicts policing.

Civil rights organizations are now pushing for a complete moratorium on AI in criminal sentencing until independent audits can verify these tools are not discriminatory. Three states have already introduced legislation to ban opaque AI sentencing tools. The question is whether the justice system will voluntarily give up a technology that makes its job easier — even when that technology is destroying lives.