RatedWithAI

RatedWithAI

Accessibility scanner

Algorithmic DiscriminationJuly 19, 2026

AI Pretrial Risk Assessment Tools and Bail/Sentencing Discrimination 2026

Algorithmic risk scores now inform bail and, in some jurisdictions, sentencing decisions for hundreds of thousands of defendants a year. The racial-disparity findings behind the leading legal challenges have not gone away — they've become the baseline every county, vendor, and defense counsel now has to account for.

Error Asymmetry
Independent analyses found Black defendants more likely misflagged high-risk, white defendants more likely misflagged low-risk
Loomis (2016)
Wisconsin Supreme Court allowed risk-score use in sentencing but required cautionary disclosures to defendants
Proprietary Scoring
Most deployed tools keep the underlying formula undisclosed, complicating individual challenges

Why Pretrial Risk Scores Became Standard Practice

Bail reform advocates and court administrators pushed algorithmic risk assessment as a fix for two problems at once: judges setting inconsistent bail amounts for similar defendants, and cash bail keeping low-risk, low-income defendants detained pretrial simply because they could not pay. A risk score, the argument went, would replace inconsistent human judgment with a standardized, data-driven number.

That standardization is exactly what has drawn scrutiny. A tool trained on historical arrest and conviction data reflects the same racial disparities present in decades of policing and charging decisions. When that history becomes the training signal for a score judges rely on today, the tool can reproduce — and give a veneer of statistical objectivity to — the disparities it was meant to remove.

What the Disparity Research Actually Found

A landmark statistical analysis of one widely used tool's scores compared outcomes by race and found a specific asymmetry: among defendants who did not go on to reoffend, Black defendants were roughly twice as likely to have been labeled high risk as white defendants in the same category. Among defendants who did reoffend, white defendants were more likely to have been labeled low risk.

The tool's vendor pushed back using a different fairness definition — calibration, meaning that among defendants who received the same risk score, the actual reoffense rate was similar across race — and argued the tool passed that test. Both findings can be true at once; academic work on this specific tool helped establish that different, mathematically valid definitions of "fair" can produce opposite conclusions about the same tool, and that a vendor's chosen metric is not automatically the right one for a given legal question.

That unresolved tension — error-rate parity versus calibration — is not a settled question in court, and it is the core reason risk-assessment litigation has been slower to produce clear precedent than more straightforward discrimination claims.

Where Legal Challenges Have Landed So Far

Due Process — Proprietary Scoring

HIGH RISK

Defendants have argued they cannot meaningfully contest a score whose weighting formula the vendor treats as a trade secret

Judicial Reliance Without Disclosure

MEDIUM RISK

Courts using a score as a factor in sentencing without explaining its limitations or margin of error to the defendant

Equal Protection — Disparate Impact

MEDIUM RISK

Claims that facially race-neutral inputs (e.g. prior arrests, zip code proxies) produce racially disparate outcomes

Vendor Validation Claims

MEDIUM RISK

Whether a tool validated on one jurisdiction's population was adequately re-validated before deployment elsewhere

What Courts and Vendors Are Doing Differently in 2026

The response to a decade of disparity findings has mostly been procedural rather than a wholesale abandonment of these tools. Jurisdictions that continue using risk assessment are increasingly expected to document:

  • Local validation studies showing the tool performs consistently across demographic groups in that specific jurisdiction, not just the vendor's original training population
  • A standing cautionary instruction to judges and defendants explaining the score's known limitations, following the model set by Loomis
  • A published policy on how much weight the score carries relative to other factors, so it functions as one input rather than a de facto decision
  • A periodic re-validation schedule, since population and charging patterns shift over time and a tool validated years ago may no longer reflect current outcomes

For vendors, the emerging baseline is disclosure of the general scoring methodology and input categories — even where the exact weighting remains proprietary — plus jurisdiction- specific validation data made available to courts and defense counsel on request.

Compliance Checklist for Jurisdictions and Vendors

Relevant to court systems, pretrial services agencies, and risk-assessment tool vendors.

Commission or obtain a validation study specific to the local defendant population, not just the vendor's original training dataRequired
Publish the categories of inputs the tool uses, even if exact weighting remains proprietaryVendor
Adopt a standing cautionary instruction for judges on the score's known limitations, modeled on the Loomis requirementRequired
Document how much weight the score carries relative to other bail or sentencing factorsPolicy
Set a periodic re-validation schedule tied to a fixed interval, not an indefinite deploymentOperations
Track outcomes by demographic group post-deployment to catch drift from the original validation studyMonitoring
Give defense counsel a documented process to request the tool's general methodology for a specific caseDue Process

Audit your AI product's compliance exposure

RatedWithAI helps tech teams understand their compliance posture across accessibility, privacy, and AI regulation requirements. Start with a free scan.

Scan Your Product for Free →

Frequently Asked Questions

Are pretrial risk assessment tools banned anywhere?

Some jurisdictions have discontinued specific tools or reverted to judicial discretion after disparity findings or local advocacy, but there is no nationwide ban. The more common trend is added validation, disclosure, and monitoring requirements rather than outright prohibition.

Does removing race as an input field fix the discrimination problem?

Not on its own. Inputs like prior arrest counts, neighborhood-linked variables, or employment history can correlate strongly with race even when race itself is excluded, a well-documented proxy problem in algorithmic fairness research. Removing the explicit field does not remove a statistical relationship baked into the other inputs.

Can a company that builds risk-assessment software for courts be sued directly by an affected defendant?

It's a developing area. Judicial immunity generally shields the court's own decision, but vendors have faced scrutiny over marketing and validation claims, and some litigation has targeted the vendor's representations about accuracy rather than the judge's reliance on the score itself.

Related Guides