Shield Public Safety Training · Police Technology Case Law Center

State v. Loomis

2016 WI 68, 371 Wis. 2d 235, 881 N.W.2d 749

The leading state supreme court decision defining constitutional limits on the use of a proprietary algorithmic risk assessment at criminal sentencing.

CourtWisconsin Supreme Court
DecisionJuly 13, 2016
Majority OpinionJustice Ann Walsh Bradley
DefendantEric L. Loomis
TechnologyCOMPAS risk and needs assessment
IssueDue process at sentencing
Core RuleConstrained, nondeterminative use permitted
Prohibited UseIncarceration or sentence length
Required SafeguardWritten limitations and cautions
DispositionJudgment and order affirmed
Modern RelevanceAI, explainability, validation, bias, and human review
Last ReviewedAugust 29, 2026

Executive Summary

The Case in One Paragraph

Eric Loomis challenged a Wisconsin sentence after the presentence investigation report included scores from COMPAS, a proprietary risk and needs assessment. He argued that reliance on a model whose internal weighting he could not inspect violated due process and that its use of gender denied him an individualized sentence. The Wisconsin Supreme Court affirmed. It concluded that COMPAS could be considered when accompanied by specified cautions and used only to supplement—not replace—the sentencing court's independent judgment. The court emphasized that a COMPAS score may not determine whether a defendant is incarcerated, set the severity or length of a sentence, or become the determinative answer to whether community supervision is appropriate.

Core RuleA sentencing court may consider a COMPAS risk assessment only within a circumscribed role, with disclosure of its material limitations and with independent reasons supporting the sentence. Loomis did not authorize an algorithm to decide punishment.

Loomis is binding in Wisconsin and influential elsewhere. Its importance in 2026 lies as much in its warnings as in its result: prediction is probabilistic, group-based data does not identify what one person will do, proprietary design limits scrutiny, performance may vary across populations, and a model can drift as conditions change.

Key Holdings at a Glance

Limited Use Was ConstitutionalOn this record, considering COMPAS with safeguards did not violate due process.
The Score Was Not DeterminativeThe sentencing judge identified independent factors supporting the sentence.
No Incarceration DecisionCOMPAS may not decide whether a person is sent to prison.
No Sentence-Length DecisionCOMPAS may not determine the severity or duration of punishment.
Warnings Are RequiredA presentence report containing COMPAS must explain specified limitations and cautions.
Inputs Could Be ChallengedLoomis could review the report and dispute the information used to produce his scores.

Facts and Sentencing Record

Loomis was charged in connection with a drive-by shooting. He denied participating in the shooting but admitted driving the vehicle involved. He pleaded guilty to attempting to flee or elude a traffic officer and operating a motor vehicle without the owner's consent. Other charges were dismissed and read in for sentencing.

The presentence investigation report included a COMPAS assessment that classified Loomis as high risk for pretrial misconduct, general recidivism, and violent recidivism. The circuit court imposed consecutive sentences totaling six years of initial confinement followed by five years of extended supervision.

Record-Specific DecisionThe supreme court reviewed what the sentencing judge actually said and did. The result depended on the court's independent reliance on the seriousness of the offense, criminal history, prior supervision, character, rehabilitative needs, and protection of the public.

What COMPAS Is—and Is Not

COMPAS stands for Correctional Offender Management Profiling for Alternative Sanctions. The assessment uses information about a person and comparison data to produce classifications intended to assist correctional management, supervision, treatment, and risk-related decisions.

ConceptFunctionLegal Caution
Risk assessmentEstimates the likelihood of a defined outcome for members of a comparison groupIt is not a factual finding that a particular person will reoffend
Needs assessmentIdentifies factors potentially relevant to services, treatment, or supervisionA need classification should not be converted automatically into punishment
Risk classificationPlaces a score into a category such as low, medium, or highCut points and labels can obscure uncertainty and error rates
Professional judgmentApplies law and case-specific facts to an individual decisionThe accountable official—not the software—must make and explain the decision
Group-to-Individual LimitationStatistical performance for a group does not establish what a specific person will do. A risk estimate should never be described as certainty, guilt, dangerousness, or a clinical diagnosis unless the tool was validated for that distinct purpose.

Procedural History and Result

Loomis sought postconviction relief and requested resentencing, arguing that the circuit court had relied on inaccurate or impermissible information. The circuit court denied relief. The Wisconsin Court of Appeals certified the case to the Wisconsin Supreme Court because use of algorithmic risk assessment at sentencing presented an issue of statewide importance.

DispositionThe Wisconsin Supreme Court affirmed the conviction and the order denying postconviction relief. It concluded that the circuit court would have imposed the same sentence without COMPAS and that the assessment was not the determinative factor in the sentencing decision.

The Due Process Framework

A defendant has a due process right not to be sentenced on materially inaccurate information. The defendant generally must show both that information was inaccurate and that the sentencing court actually relied on it.

Loomis focused on the proprietary nature of COMPAS, the inability to inspect its internal weighting, the use of group data, and gender-based scoring. The supreme court evaluated those objections against the limited role the score played in the complete sentencing record.

Decision ReviewSeparate three questions: Was the information accurate? Was the tool valid for this population, setting, and outcome? Did the decision-maker actually rely on the output—and, if so, for what legal purpose?

Proprietary Algorithm and Ability to Respond

COMPAS was proprietary, and Loomis could not inspect how the developer weighted variables or calculated risk. The court nevertheless found no due process violation because he could review and challenge the factual inputs and the resulting scores, and because the sentencing court did not use the output to determine incarceration or sentence length.

That holding is narrow. It does not create a general rule that trade-secret status defeats disclosure, that source code is never relevant, or that a vendor's output is presumptively reliable. A different system, decision, record, error, or degree of reliance may present materially different due process, discovery, evidentiary, statutory, or state constitutional questions.

Do Not Overread LoomisThe case approved a restricted use of one assessment on one sentencing record. It did not approve secret algorithms as judges, validate every COMPAS application, or authorize officials to rely on outputs they cannot meaningfully explain.

Individualized Sentencing and Independent Reasons

Wisconsin sentencing requires consideration of the individual defendant and offense. The supreme court concluded that Loomis received an individualized sentence because the judge discussed case-specific facts independent of COMPAS.

A defensible decision record should identify which legally permissible factors mattered, how the evidence supports each factor, what role—if any—the model played, and why the same legal conclusion does not rest on the score alone. Merely stating that a human made the final decision is insufficient if the record shows practical automation bias.

Gender and Potential Bias

Loomis argued that the assessment's use of gender violated due process. The court rejected that claim on the record before it, reasoning that gender-specific data was used to improve statistical accuracy and that the sentencing judge did not actually rely on gender in selecting the sentence.

The court did not decide an equal protection challenge because one was not presented. The opinion therefore should not be cited for the broad proposition that using sex or another protected characteristic in an automated decision is always lawful.

Current ReviewTest both direct and proxy effects. Evaluate protected attributes, correlated variables, subgroup error rates, calibration, false positives, accessibility, data quality, and whether the legal context permits the distinction at all.

Uses Loomis Identified as Permissible

Subject to the opinion's cautions, the court described a limited role for risk and needs information:

  • Identifying low-risk individuals who otherwise might be incarcerated and could be considered for alternatives.
  • Helping assess whether risk may be managed safely in the community, so long as the score is not determinative.
  • Informing the level and conditions of probation or supervision.
  • Identifying treatment, intervention, and service needs.

Each use still requires lawful authority, a tool validated for the actual purpose and population, accurate data, trained reviewers, an opportunity to correct material errors, and a documented exercise of independent judgment.

Uses Loomis Prohibited

Prohibited RoleWhy It Is ImproperRequired Alternative
Deciding whether to incarcerateThe tool was not developed to determine punishment and cannot replace individualized sentencingApply governing law to verified case-specific facts
Determining sentence severityA group prediction cannot supply the legal measure of blame or punishmentExplain the lawful sentencing factors independently
Determining sentence lengthThe score does not establish the proper duration of confinementBase duration on statute, precedent, and individualized reasons
Determinatively denying community supervisionRisk output is supplementary information, not the final answerConduct a documented, human review of feasibility and conditions

Required Advisement and Cautions

The court directed that any presentence investigation report containing a COMPAS assessment include a written advisement explaining the tool's limitations. The warning should remain current as research and the system change.

Proprietary DesignThe developer did not disclose how factors were weighted or how scores were calculated.
Group PredictionComparison data identifies group tendencies, not what a particular person will do.
Disparate ClassificationResearch raised concerns about differential classification of minority defendants.
Local ValidationAt the time, the tool relied on a national sample and had not been cross-validated for Wisconsin.
Ongoing MonitoringAssessment systems require monitoring, updating, and periodic renorming.
Original PurposeCOMPAS was developed for correctional management, treatment, supervision, and parole—not sentencing.
Living WarningA historical disclaimer is not enough. Document the current model version, intended use, validation population, performance date, known limitations, error measures, update history, and conditions that make results unreliable.

Human Judgment, Explanation, and Accountability

Loomis requires more than a nominal human in the loop. The accountable decision-maker must understand the output's limited meaning, verify material data, consider contrary evidence, exercise genuine discretion, and give reasons independent of the model.

ControlOperational Requirement
AuthorityDefine who may use the tool and for which decisions
CompetenceTrain users on intended use, error, uncertainty, and prohibited inferences
ContestabilityProvide a practical method to inspect and correct material inputs
ExplanationRecord lawful, case-specific reasons that stand without the score
OverridePermit and document informed disagreement with the output
AuditPreserve version, inputs, output, user actions, and downstream reliance

Algorithmic Decision Systems in 2026

Public agencies now encounter predictive models, machine-learning classifiers, generative AI summaries, biometric scores, automated alerts, and vendor decision-support platforms. These systems differ technically from COMPAS, but Loomis supplies durable questions about purpose, transparency, validation, bias, data quality, and the boundary between advice and decision.

Generative systems add distinct risks: fabricated facts, unstable answers, prompt sensitivity, hidden model changes, and persuasive explanations that may not reveal the actual basis of an output. A COMPAS-style warning does not by itself make a generative output fit for a liberty-affecting decision.

Technology-Neutral PrincipleThe greater the consequence for liberty, safety, employment, benefits, or legal status, the stronger the need for validated purpose, reliable evidence, meaningful notice, human review, contestability, and an auditable explanation.

Practical Guidance for Courts and Public Agencies

Define the DecisionState precisely what the system may inform and what it may never determine.
Validate the Actual UseRequire evidence for the relevant outcome, population, jurisdiction, and operating environment.
Know the VersionRecord the model, data source, configuration, thresholds, and date used.
Disclose LimitationsGive reviewers and affected persons understandable, current cautions.
Enable CorrectionCreate a timely process to challenge inaccurate inputs and material outputs.
Audit ReliancePreserve who viewed the output, what they did, and the independent basis for the decision.

Algorithmic Assessment Review Checklist

QuestionEvidence to Preserve
What exact decision and legal authority govern?Policy, statute, rule, order, and approved-use statement
Was the system designed and validated for this use?Technical documentation, validation studies, dates, and populations
Which version produced the output?Model identifier, software version, configuration, and threshold
Are the inputs accurate and legally obtained?Source records, corrections, provenance, and access history
How does performance vary across groups?Subgroup error, calibration, false-positive, and false-negative results
Can the affected person meaningfully respond?Notice, understandable explanation, correction process, and timing
Did a qualified person exercise independent judgment?Reviewer identity, analysis, override, and case-specific reasons
Can the complete event be reconstructed?Inputs, output, logs, user actions, exports, and retention controls

Litigation and Discovery Checklist

  1. Identify the model, vendor, version, configuration, and stated intended use.
  2. Determine exactly how the output entered the decision and who saw it.
  3. Obtain the input data, source records, corrections, and score report.
  4. Compare the deployed use with validation studies and contractual representations.
  5. Examine subgroup performance, calibration, error rates, and known limitations.
  6. Preserve audit logs, prompts, outputs, model updates, user notes, and overrides.
  7. Review warnings given to the decision-maker and affected person.
  8. Separate risk prediction, needs identification, factual findings, and legal judgment.
  9. Identify independent reasons and test whether the result stands without the output.
  10. Evaluate due process, equal protection, evidence, discovery, statutory, contractual, and state-law claims separately.

Frequently Asked Questions

What did State v. Loomis hold?

It held that the limited consideration of COMPAS in Loomis's sentencing did not violate due process because the score was not determinative, specified warnings applied, and the judge gave independent reasons for the sentence.

Did the court approve algorithmic sentencing?

No. It allowed one risk assessment to play a restricted supplementary role. The court prohibited using the score to decide incarceration or determine the severity or length of a sentence.

Did Loomis have access to the source code?

No. The model was proprietary. The court found no due process violation on this record because Loomis could challenge the inputs and scores and because the assessment's role was circumscribed.

Does Loomis mean proprietary algorithms never require disclosure?

No. The opinion does not establish that rule. Disclosure obligations may depend on the tool, legal claim, jurisdiction, consequence, contested issue, and degree of actual reliance.

What warning must accompany COMPAS?

The presentence report must explain proprietary limitations, group-versus-individual prediction, bias concerns, local-validation limitations, the need for monitoring and renorming, and that COMPAS was not developed to determine sentences.

Is Loomis binding nationwide?

No. It is a Wisconsin Supreme Court decision. It is binding on Wisconsin courts and persuasive elsewhere, subject to federal law and each jurisdiction's constitution, statutes, rules, and precedent.

Does the decision apply directly to generative AI?

No. Loomis addressed COMPAS at sentencing. Its concerns are instructive, but generative AI presents additional reliability, traceability, and fabrication risks requiring separate analysis.

Primary Authorities

State v. Loomis, 2016 WI 68, 371 Wis. 2d 235, 881 N.W.2d 749
Wisconsin Supreme Court opinion addressing COMPAS, due process, proprietary scoring, gender, individualized sentencing, required warnings, and prohibited uses.
Read the complete State v. Loomis opinion
Wisconsin Statutes § 973.01 and Wisconsin Sentencing Law
Applicable sentencing law should be reviewed in its current form together with controlling decisions and local procedure.
Review Wisconsin Statutes § 973.01

Final Assessment

State v. Loomis is neither a blanket approval nor a categorical rejection of algorithmic tools. It is a warning-bound decision allowing a proprietary risk assessment to supplement an individualized sentencing record while sharply limiting what the score may do.

The practical lesson is governance. Before an algorithm affects a high-consequence public decision, officials should define its permitted purpose, validate it for the actual context, disclose material limits, provide a method to correct errors, measure disparate performance, preserve an audit trail, and require a qualified person to give independent reasons grounded in law and verified facts.

Shield Practice RuleNever allow a risk score or AI output to determine incarceration, punishment, or another high-consequence legal result. Verify the data and intended use; disclose current limitations; test performance and bias; give the affected person a meaningful opportunity to respond; preserve the complete decision record; and require the accountable official to reach and explain an independent judgment.

Shield Public Safety Training · Police Technology Case Law Center

This monograph is provided for training and general informational purposes. It is not legal advice and does not replace review of the complete opinion, subsequent history, controlling jurisdictional authority, current statutes, state constitutional law, agency policy, technical documentation, validation evidence, or consultation with agency counsel.

© 2026 Shield Public Safety Training. All rights reserved. Reviewed August 29, 2026.