How to Challenge AI-Generated Risk Assessment Scores (COMPAS) at Sentencing in California Courts
California judges must “consider” risk assessments at sentencing, but under People v. Dueñas (2019) and People v. Hernandez (2022), defense counsel can challenge reliability, notice, and due process—often winning limits or exclusion. COMPAS-style scores raise accuracy, bias, and transparency issues that matter under California’s evidence and sentencing rules. This article explains California-specific strategies to object, litigate discovery, request hearings, and build an appellate record.
What “COMPAS” Means in California Sentencing (and Why It’s Contestable)
Risk assessment tools such as COMPAS (Correctional Offender Management Profiling for Alternative Sanctions) and similar actuarial instruments are designed to predict recidivism risk using demographic, criminal history, and questionnaire data. In practice, these tools can enter a California sentencing proceeding through probation reports, pretrial reports, or attachments that summarize a “low/medium/high” risk category and sometimes a numerical score.
Even when a court says it is not “bound” by the score, the score can still anchor the sentencing narrative—framing the defendant as inherently dangerous or unlikely to rehabilitate. That anchoring effect makes the score legally and strategically contestable, particularly where the defense has not been given transparent access to inputs, validation studies, or error rates for the local population.
California’s Legal Landscape: Due Process, Evidence, and Sentencing Discretion
Risk tools are not immune from challenge simply because sentencing is “informal”
California sentencing hearings are not full trials, but they are still constrained by due process. A defendant must have a meaningful opportunity to know and contest information used in sentencing. When an algorithmic score is presented as scientific or data-driven, the defense can argue that fairness requires disclosure of the basis for the conclusion, the reliability of the method, and an opportunity to rebut.
Key California themes to frame your challenge
While your exact motion practice will depend on the case posture and the way the score appears in the record, successful California challenges typically rely on these themes:
(1) Notice and ability to rebut: The defense must be able to review the underlying inputs and methodology, not merely the label (“high risk”).
(2) Reliability and relevance: The tool must be relevant to the sentencing factors actually before the court, and its predictive claims must be reliable for the population and context in question.
(3) Bias and unfair prejudice: Many tools have been criticized for disparate impact and for using proxies correlated with protected characteristics (e.g., neighborhood, prior contacts, socioeconomic factors). Even if not intentionally discriminatory, they may skew outcomes.
(4) Overstated scientific certainty: If the score is presented without error rates, base rates, false positive rates, or limitations, it can mislead the court.
Authorities to cite early
Two California Supreme Court decisions often appear in AI-sentencing litigation strategies:
People v. Dueñas (2019) 30 Cal.App.5th 1157, while primarily focused on ability-to-pay and due process, is frequently used as a broader due process framework for sentencing practices that impose consequences without meaningful consideration of individualized circumstances.
People v. Hernandez (2022) 74 Cal.App.5th 332 addresses use of actuarial risk assessment instruments in sentencing contexts and is commonly cited for the proposition that risk tools can raise due process concerns and that courts must be careful about how such instruments are used and disclosed. When a risk score functions like “expert” or “scientific” evidence, the defense should be prepared to demand the foundational information that would ordinarily accompany such claims.
Practice note: Always confirm how your appellate district has treated these cases and whether there are local rules or standing orders affecting probation report attachments and disclosures.
Step 1: Identify Exactly How the Risk Score Entered the Case
Your first task is not legal—it’s factual. Risk scores show up in multiple ways, and each pathway changes your objections and remedies.
Common entry points
Probation report: The most common route. A probation officer includes a risk score summary and recommends a sentence based on “risk/needs.”
Pretrial services report: Although more common at bail, these reports can be referenced later for “pattern” arguments.
Departmental assessment: A jail, treatment program, or contractor provides a risk/needs assessment.
Prosecutor’s sentencing memo: The DA cites a score as support for prison time or upper term arguments.
Why it matters
If the score is embedded in a probation report, courts often treat it as part of the informational basis for sentencing, but that does not eliminate your right to contest it. If the score is presented as “scientific” support for a harsher sentence, you have stronger arguments to require foundational reliability and to prevent misleading characterizations.
Step 2: Demand Discovery—Inputs, Training/Validation, and Error Rates
A “high risk” label is not rebuttable unless you can see how it was produced. Your discovery request should be specific and tied to due process and reliability.
What to request (a checklist)
(1) Raw inputs: All answers to questionnaires, all criminal history fields used, and any administrative data pulled into the score.
(2) Scoring rules: The rubric that converts inputs into scores, including weighting and thresholds for categories.
(3) Versioning: Tool version number, date administered, and any updates between versions.
(4) Local validation: Studies or audits showing performance on a comparable California population (or your county), including calibration.
(5) Error metrics: False positive/negative rates and AUC/ROC or similar metrics, ideally broken down by relevant groups.
(6) Administrator training: Manuals, training records, and whether the assessor was certified.
(7) Quality control: Procedures for verifying data accuracy and correcting errors.
(8) Any contracts or vendor materials: Especially where a private vendor claims trade secret protections.
How to handle “trade secret” pushback
Vendors often argue proprietary protection. In California criminal cases, courts can use protective orders, in camera review, or limited disclosure to protect sensitive details while still preserving the defendant’s right to challenge a score that may affect liberty. Your theme: you cannot constitutionally rely on a black-box score while denying the defense a meaningful chance to rebut it.
Step 3: File Targeted Sentencing Objections and Motions to Strike/Exclude
In many cases, the fastest path is to ask the court to strike the risk score from consideration or, at minimum, limit its use.
Core objections to raise in writing
Due process / lack of meaningful opportunity to rebut: If you do not have the inputs and methodology, the score is functionally unchallengeable.
Relevance: If the tool predicts rearrest (not reconviction), or predicts a broad population outcome rather than the defendant’s individual sentencing factors, argue mismatch.
Unreliable foundation: No showing of validation for your county, your defendant’s demographics, or the sentencing context.
Misleading “scientific” aura: Without limitations and error rates, the score risks being treated as objective truth rather than a probabilistic estimate.
Unfair prejudice: A numerical “risk” label can be more prejudicial than probative, especially when the defense cannot test it.
Request specific relief
Don’t ask generically for “appropriate relief.” Ask for one or more of the following:
(1) Exclusion: Order that the court will not consider the score.
(2) Redaction: Remove the score and narrative references from the probation report.
(3) Limiting instruction (sentencing context): Court states on the record it will not treat the score as scientific proof and will not use it to aggravate.
(4) Continuance: If disclosure arrives late, request time to consult an expert.
Step 4: Request an Evidentiary Hearing When the Score Is Doing Real Work
If the prosecution or probation department relies on the risk score to argue for prison, an upper term, or denial of probation, consider requesting an evidentiary hearing to test foundation and reliability.
When a hearing is most justified
Score conflicts with the record: e.g., labeled “high risk” despite long offense-free periods, stable employment, or minimal history.
Input errors suspected: arrests treated as convictions, mis-coded priors, wrong age, wrong residence history.
Tool not validated locally: no evidence it works for your county or the subpopulation.
Disparate impact concerns: score appears to reflect policing intensity rather than actual risk.
What to examine
At the hearing, focus on the administrator and the foundation:
(1) Data integrity: Who entered data? What was verified? Any audit trail?
(2) Administration protocol: Was it administered as required? Any deviations?
(3) Interpretation: What does “high risk” mean numerically? Risk of what outcome, over what time horizon?
(4) Limitations: Ask for known limitations, confidence intervals (if any), and known misclassification patterns.
Step 5: Use an Expert Strategically (Not Automatically)
An expert can be decisive, but only if you match the expert’s testimony to your legal theory. In many California sentencing disputes, the most effective expert testimony is not “AI is bad,” but rather: this instrument’s output is not reliable for this decision, and the court is being asked to overread it.
Helpful expert profiles
Forensic psychologist/criminologist:
