How to Challenge AI-Generated Risk Assessment Scores in California Sentencing Hearings (2026)
In California sentencing hearings in 2026, attorneys can challenge AI-generated risk assessment scores through discovery, evidentiary objections, and due process arguments grounded in California’s determinate sentencing framework. Courts are increasingly asked to rely on “risk” tools for custody, probation, and supervision decisions, even when the underlying model is opaque. This article explains practical motions, hearing tactics, and appellate issue preservation to contest AI risk scores in California.
Why AI-Generated Risk Scores Matter in California Sentencing (2026)
“Risk assessment” scores—often generated by algorithmic or AI-assisted tools—can influence whether a client receives probation, what conditions are imposed, eligibility for treatment programs, recommended custody level, or how a probation department frames “public safety” concerns. In California, even when a judge states the score is “just one factor,” a risk label (e.g., low/medium/high) can anchor the court’s perception and shift the burden to the defense to “disprove” danger.
In 2026, the practical defense question is not whether courts should use these tools, but how to keep unreliable, biased, or nontransparent scoring from shaping the sentence. The best challenges combine (1) targeted discovery, (2) evidentiary and procedural objections, (3) expert-driven rebuttal, and (4) meticulous issue preservation for appeal or writ relief.
Where the “AI Risk Score” Typically Appears in a California Case
Most defense attorneys encounter algorithmic risk scoring in one of these pathways:
1) Probation reports and addenda. The report may cite a tool’s “risk to reoffend” score, violence risk flag, or supervision recommendation. Sometimes the score is embedded indirectly through “validated instrument” language or “assessment suggests high risk.”
2) Pretrial services and release decisions. While your hearing may be sentencing, pretrial risk scoring can bleed into later proceedings through judicial familiarity or probation file references.
3) Specialty courts and program eligibility. Treatment, mental health diversion-adjacent screening, or custody programming may reference an algorithmic “needs” and “risk” profile.
4) Departmental decision-making that becomes part of the record. Even if the court does not admit the score as evidence, it may influence recommendations the court heavily relies on.
Core Legal Theories to Challenge AI Risk Assessment Use
1) Due Process: Reliability, Notice, and a Meaningful Opportunity to Respond
California sentencing is broad, but it is not lawless. If the prosecution or probation department urges reliance on an AI score, the defense can frame the issue as a basic due process problem: a defendant cannot meaningfully rebut a conclusion when the methodology, inputs, and error rates are concealed or when the score is presented as scientific without adequate foundation.
Practice point: Demand that the court require transparency sufficient for a “meaningful opportunity to be heard,” including the data sources used, the variables considered, and how the score was generated and validated for the relevant population.
2) Evidence Code: Foundation, Hearsay, and “Black Box” Assertions
Even in sentencing, where courts may consider a wide range of information, you can still attack the risk score as unreliable hearsay or lacking foundation—particularly if the score is imported from a vendor tool and presented through a probation officer who cannot explain the model. If the score is framed as expert-like opinion (“validated,” “predictive”), consider whether it triggers foundation requirements akin to expert testimony.
What to highlight:
- Unidentified declarant and methodology. Who “said” the score is high risk? A tool? A vendor? A database?
- Inability to cross-examine. If no competent witness can explain the scoring process, the defense cannot test it.
- Overstatement of scientific certainty. “Validated” does not mean accurate for your client’s demographic, county, or current year.
3) Equal Protection and Bias: Disparate Impact and Proxy Variables
Risk instruments may incorporate variables that function as proxies for race, poverty, disability, neighborhood policing patterns, or prior justice-system contact. A tool can be facially neutral yet operationally discriminatory. Your challenge should be specific: identify which inputs likely encode bias (e.g., arrest history versus conviction history, prior supervision failures, zip-code correlates, employment stability measures) and demand validation evidence addressing disparate impact.
4) Statutory/Rule-Based Arguments: Sentencing Must Be Individualized
California sentencing requires individualized consideration of the defendant and the offense. A generic risk label can conflict with individualized sentencing if the court treats it as a shortcut. Your theme: the tool is a population-level statistical instrument that cannot substitute for individualized findings on mitigation, culpability, and the specific facts of the case.
Step-by-Step: A Defense Playbook to Challenge the Risk Score
Step 1: Identify the Tool and Get the “Provenance” Into the Record
Start by forcing clarity. At or before sentencing, ask the probation officer (or the court) to identify:
- The tool’s name and version (tools change; versions matter).
- Whether the score was AI/ML-driven or rules-based.
- Who administered it and when.
- What data sources were used (CJIS extracts, county records, self-report, arrest databases, school/work histories).
- Whether any “override” or human adjustment occurred.
Practical tip: If the report is vague (“assessment indicates…”), file a short ex parte or noticed request for clarification before sentencing so the hearing is not dominated by uncertainty.
Step 2: File a Targeted Discovery Motion or Subpoena Strategy
Many risk tools are vendor products. Vendors often resist disclosure citing trade secrets. But the defense can narrow requests to what is necessary to test reliability and bias without demanding full source code in the first instance.
Discovery targets that are often defensible and proportional:
- All inputs used for your client’s score (the “feature list” for this scoring event).
- Scoring rubric or explanation output (reason codes, top factors, confidence bands).
- Validation studies relied upon by the county/vendor (including local validation, if any).
- Error rates and performance metrics (false positives/false negatives; calibration).
- Policies on overrides, audits, and drift monitoring.
- Training data geography/timeframe (e.g., pre-2020 vs post-2020, local vs national).
Trade secret workaround: Propose a protective order, in camera review, or limited disclosure to defense counsel and a defense expert. Offer a confidentiality stipulation that still preserves your client’s right to challenge the score.
Step 3: Object Early—Before the Score Becomes the “Default Truth”
If a probation report contains the score, do not wait for oral argument. File written objections to the probation report and request an addendum striking or clarifying the AI-risk language. Ask the court to rule on the objection on the record.
Key request: “The defense requests the court state on the record whether it is considering the risk score and, if so, what weight it is giving it.” This is critical for appeal.
Step 4: Demand a Foundation Hearing (or Equivalent) for Algorithmic Claims
When the score is presented as predictive science, ask for a mini-foundation hearing. Your goal is not necessarily to exclude all risk information, but to prevent the court from treating an opaque metric as validated, individualized truth.
Questions to ask the probation officer or proponent witness:
- What is the model’s documented error rate for people like the defendant?
- How does the tool handle missing or incorrect data?
- Was the defendant allowed to review and correct inputs?
- Does the tool rely on arrest data (which may reflect policing patterns)?
- When was it last validated for this county?
Step 5: Use an Expert to Translate “AI” Into Plain Reliability Problems
An expert in statistics, ML auditing, or psychometrics can be decisive—especially if the prosecution leans on “validated tool” rhetoric. The defense expert can explain, in understandable terms:
- Base rate and calibration: a “high risk” label can still be wrong most of the time depending on prevalence.
- False positives: risk tools often over-classify certain groups as high risk.
- Dataset shift: a model trained on older data may not predict current outcomes.
- Proxy discrimination: neutral variables can encode protected traits.
Cost-sensitive strategy: If budgets are tight, consider a consulting expert to help craft cross-examination and motions, even if the expert does not testify.
Step 6: Offer a Competing, Human-Centered Sentencing Narrative
Judges may welcome a risk score because it appears objective and efficient. Replace that vacuum with a credible, individualized alternative:
- Documented treatment plan (provider letter, intake date, funding plan).
- Employment/education proof and supervision support.
- Character letters that address risk factors directly (substance use, stability, compliance).
- A proposed probation plan with measurable milestones.
Make the contrast explicit: “This plan is based on the defendant’s actual circumstances—not a population-level estimate whose inputs we cannot verify.”
Specific Examples of Effective Challenges (Phrasing You Can Use)
Example A: Unknown Inputs and Data Errors
Scenario: The probation report says “assessment indicates high risk,” but the defense discovers the tool used an old arrest record that was dismissed or
