The Wisconsin Supreme Court's 2016 decision upholding AI-assisted sentencing has become the leading authority on algorithmic due process — and a blueprint for the challenges still ahead in pretrial, parole, and civil contexts.
I. The Problem of the Black Box at Sentencing
The deployment of algorithmic risk assessment tools in American criminal courts has proceeded faster than legal doctrine has developed to govern it. Tools like COMPAS (Correctional Offender Management Profiling for Alternative Sanctions), PSI-7, LSI-R, and others are used in presentence reports, bail decisions, parole eligibility determinations, and reentry programs. They are developed by private companies, and their methodologies are typically protected as trade secrets. Courts have incorporated their outputs — recidivism risk scores, needs assessments — into consequential decisions affecting defendants' liberty without requiring, or in most cases permitting, disclosure of how those scores are calculated.
The constitutional question is acute: if a defendant has the right under Gardner v. Florida, 430 U.S. 349 (1977), to be sentenced only on information that the defendant has had an opportunity to rebut, how does that right coexist with a sentencing factor generated by an algorithm whose methodology is withheld as proprietary? State v. Loomis is the leading case, and its resolution is simultaneously instructive and unsatisfying.
II. State v. Loomis: The Wisconsin Supreme Court's Analysis
A. Facts
State v. Loomis, 881 N.W.2d 749 (Wis. 2016) arose from the prosecution of Eric Loomis on charges stemming from a drive-by shooting. Loomis pleaded guilty to two lesser charges — attempting to flee a traffic officer and operating a motor vehicle without the owner's consent. In preparation for sentencing, the Department of Corrections prepared a presentence investigation report that included a COMPAS risk assessment. COMPAS generated scores on three dimensions — pretrial recidivism risk, general recidivism risk, and violent recidivism risk — each on a scale of one to ten. Loomis scored high on all three. The circuit court referenced the COMPAS assessment in denying probation and sentencing Loomis to six years of imprisonment and five years of extended supervision.
Loomis moved for postconviction relief, arguing that the circuit court's consideration of COMPAS violated his due process rights on three grounds: (1) the proprietary nature of COMPAS prevented him from assessing its accuracy; (2) COMPAS provides only group-based predictions and cannot supply the individualized sentencing required by due process; and (3) COMPAS uses gender as a risk factor, which constituted unconstitutional gender discrimination in sentencing.
B. The Wisconsin Supreme Court's Holding
The Wisconsin Supreme Court affirmed in an opinion by Justice Ann Walsh Bradley. The court held that, if used properly, consideration of a COMPAS risk assessment at sentencing does not violate due process.
On the accuracy challenge, the court reasoned that COMPAS reports include the questions and answers on which the assessment is based — both from the defendant's criminal history (publicly available) and from the defendant's own responses to a questionnaire. Because the defendant has access to those inputs and an opportunity to deny or explain them, the accuracy-verification component of due process is satisfied. The court acknowledged that the proprietary algorithm prevents disclosure of how those inputs are weighted or how the score is calculated, but concluded that access to inputs was constitutionally sufficient.
On individualization, the court acknowledged that COMPAS generates predictions based on group statistics — it predicts the likelihood that persons similar to the defendant will reoffend, not the likelihood that this defendant will reoffend. But it concluded that because COMPAS is only one factor among many and courts retain full discretion to disagree with the assessment, the individualization requirement is not violated.
On gender, the court found that COMPAS uses gender because statistical evidence shows that men, on average, have higher recidivism rates than women — a nondiscriminatory actuarial purpose. Loomis provided insufficient evidence that the sentencing court had actually considered gender, as distinguished from the COMPAS composite score that incorporated it.
C. The Court's Safeguards
Though upholding COMPAS use, the court imposed requirements that reveal its unease. Any presentence investigation report containing a COMPAS assessment must include written advisements that:
- The proprietary nature of COMPAS prevents disclosure of how risk scores are calculated;
- Risk scores compare defendants to a national sample, but no cross-validation study for a Wisconsin population has been completed;
- Studies have raised questions about whether COMPAS disproportionately classifies minority offenders as higher risk;
- COMPAS scores are intended for the Department of Corrections, not for sentencing determinations;
- Risk scores cannot be used to determine whether an offender is incarcerated, the severity of the sentence, or as the determinative factor in community supervision eligibility.
These are not minor caveats. The court simultaneously upheld the practice and enumerated five reasons why it should be used with skepticism. The tension is unresolved.
III. Loomis v. Wisconsin: Cert Denied
Loomis sought certiorari in the United States Supreme Court. The Court denied cert in Loomis v. Wisconsin, 137 S. Ct. 2290 (2017) (mem.). The denial carried no precedential weight, but left the Wisconsin Supreme Court's ruling as the most prominent appellate authority on algorithmic sentencing. No federal circuit has issued a definitive ruling on the constitutional requirements for AI risk assessment tools in criminal sentencing.
IV. Doctrinal Weaknesses in Loomis
Loomis's constitutional analysis has drawn sustained academic and advocacy criticism that provides the roadmap for future challenges.
The "inputs aren't the algorithm" problem. The court's reasoning that a defendant can "verify" accuracy by confirming inputs assumes that an algorithm with correct inputs necessarily produces an accurate output. That assumption is demonstrably false: identical inputs processed through a different algorithmic structure produce different outputs. The right to verify inputs does not supply the ability to challenge whether those inputs are appropriately weighted or whether the algorithm's logic captures relevant risk factors. Gardner v. Florida's "opportunity to rebut" guarantee is hollow if a defendant cannot probe the translation from input to score.
The racial bias evidence. The ProPublica 2016 analysis (published months before Loomis was decided) documented that COMPAS classified Black defendants as higher risk at twice the rate of white defendants for crimes they did not ultimately commit, while white defendants were more likely to be misclassified as low-risk. The court's advisement acknowledging "questions" about racial disproportionality is a judicial acknowledgment that the tool may produce racially skewed results — which raises the question of why a tool with acknowledged racial bias satisfies due process.
The trade secret shield. COMPAS's methodology is shielded as a trade secret, which means no court has ever conducted adversarial examination of the algorithm's statistical foundations. The court cannot assess accuracy for what it cannot examine. Loomis effectively held that due process does not require access to the algorithm — but due process generally requires an opportunity to challenge the scientific basis of evidence used against a party.
V. Civil Applications: Pretrial Release and Parole
Loomis's holding is formally limited to sentencing, but algorithmic risk assessment has migrated to pretrial detention, bail-setting, and parole determinations — contexts with different constitutional and statutory frameworks.
Pretrial release. The Bail Reform Act, 18 U.S.C. § 3142, governs federal pretrial detention; state equivalents govern state proceedings. Pretrial detention raises Mathews v. Eldridge, 424 U.S. 319 (1976) balancing considerations: the private interest (liberty prior to conviction) is substantial, the risk of erroneous deprivation from algorithmic scoring is demonstrable, and the government's interest is in accurate risk prediction. The Loomis due process framework should not automatically transfer to pretrial settings where defendants have not yet been convicted of anything.
Parole. Morrissey v. Brewer, 408 U.S. 471 (1972) established minimum due process requirements for parole revocation; the parole-grant context is less doctrinally developed. When a parole board uses an algorithmic risk score as a substantial factor in denial of release, the same trade-secret and accuracy objections applicable in Loomis apply with equal force, and the liberty interest is substantial.
Civil commitment and sex offender registration. Actuarial risk tools like the Static-99 and STATIC-99R are routinely used in civil commitment proceedings and sex offender classification. These tools' reliability and validity are subject to expert challenge under Daubert and state equivalents. Unlike the sentencing context, civil commitment and registration proceedings typically allow adversarial expert testimony — and the literature on actuarial prediction instruments provides substantial fodder for Daubert motions.
VI. Practice Notes
In criminal cases: Preserve the due process challenge at sentencing. Object to COMPAS or other algorithmic risk scores on the grounds that: (1) the methodology is not disclosed; (2) the tool has not been validated for the local population; (3) the racial bias studies raise reasonable doubt about accuracy; and (4) use of the tool as anything more than a background consideration violates Gardner. Even if overruled, the preserved record supports appellate challenge.
Request the underlying data. Most risk assessment tools draw heavily on criminal history data maintained by state agencies. FOIA or state open-records requests for the defendant's official criminal history data — compared against the inputs COMPAS used — can reveal errors. The Loomis court's logic only works if the inputs are accurate.
Expert testimony. In civil commitment, parole, or sex offender registration contexts, retain a qualified psychologist or criminologist to challenge the actuarial instrument's validity, racial disproportionality, and predictive accuracy at the individual level. The base rate problem — that even a highly accurate actuarial instrument will misclassify most individuals in a low-base-rate population — is a powerful Daubert argument.
Equal protection. Loomis declined to address equal protection, noting that Loomis had not raised it. An equal protection challenge to a facially neutral risk tool that produces racially disparate outcomes is a distinct theory from Loomis's due process holding, and it is substantially stronger after Students for Fair Admissions v. Harvard, 600 U.S. 181 (2023), which reaffirmed that racial disparate treatment requires strict scrutiny.
VII. Where the Law Is Moving
Legislative reform is beginning: several states have enacted or considered legislation requiring validation studies, transparency disclosures, or prohibiting certain uses of risk scores in sentencing. The American Law Institute's Model Penal Code revisions have recommended caution about algorithmic scoring. Federal circuit courts have not definitively addressed the constitutional requirements. A case presenting a clean due process or equal protection record, in a context where the algorithmic tool's racial disproportionality can be specifically documented, is the vehicle most likely to generate significant appellate movement.
Talk to Yates Anderson
If you are litigating a matter in this area — or weighing whether to — the working analysis above only goes so far. Request a case evaluation and a Yates Anderson attorney will respond within one business day.
Informational only. Not legal advice. No attorney-client relationship is created by reading this post. Consult a licensed attorney in your jurisdiction.