How to Challenge AI-Generated Risk Assessment Scores in California Criminal Sentencing Hearings
AI risk scores can affect bail and sentencing outcomes, and California courts must protect a defendant’s due process rights when such tools are used. Across California criminal sentencing hearings, judges increasingly encounter algorithmic “risk assessment” inputs from probation or pretrial services. This article explains practical, California-focused ways to challenge AI-generated risk scores through discovery, evidentiary objections, expert proof, and appellate preservation.
Why AI-Generated Risk Scores Matter in California Sentencing
Risk assessment scores—often presented as “low,” “moderate,” or “high” risk—can influence how probation departments frame a defendant, how the court views public safety, and whether the court leans toward incarceration, jail alternatives, supervision conditions, or treatment. Even when a score is formally “advisory,” it can anchor the court’s perception of dangerousness or recidivism in a way that is difficult to unwind.
In California, these scores may appear in probation reports, pretrial services summaries, specialty court screening tools, or vendor-generated dashboards integrated into case management systems. Regardless of the source, the defense should treat an AI-generated risk score like any other influential but potentially unreliable piece of information: challenge its foundation, demand transparency, expose limitations, and build a record that allows meaningful review.
Where Risk Scores Show Up (and How They Typically Get Introduced)
In practice, AI-derived or algorithmic risk assessments most commonly enter the record through:
Probation reports at sentencing. A probation officer may include a numerical score, a risk tier, or a narrative recommendation “supported” by a tool.
Pretrial assessments referenced later. A pretrial risk label can follow a defendant into later stages, implicitly informing sentencing arguments about community safety.
Treatment/alternative-to-incarceration screening. Eligibility for programs may be tied to “risk/need” tools; a “high risk” label can be used to justify exclusion or more restrictive conditions.
Prosecution argument. The People may cite the score in aggravation or as justification for custody, even when the underlying methodology is not disclosed.
Core Legal Theories for Challenging AI Risk Scores in California
1) Due Process: The Right to a Fair Sentencing Based on Reliable Information
California sentencing must be fundamentally fair. If the court relies on materially false, unreliable, or untested information, the defense should frame the issue as a due process problem: the defendant is being sentenced based on an opaque, potentially erroneous risk label that cannot be meaningfully confronted.
Practical argument: a score derived from undisclosed features, unknown error rates, and untestable assumptions is not “reliable information” for individualized sentencing. If the probation report uses the score to support aggravating inferences—future dangerousness, likelihood of reoffense, supervision failure—due process concerns intensify.
2) Confrontation and Cross-Examination: Limited at Sentencing, Still Powerful in Practice
Formal confrontation rights are narrower at sentencing than at trial, but California courts still expect a fair process and permit challenges to the accuracy of probation report content. A defense strategy is to use the procedural tools that do exist—cross-examination of the probation officer (where available), subpoenas for records, and expert declarations—to show the score is not fit for sentencing reliance.
Even when live testimony is not typical, you can request an evidentiary hearing when the score is disputed and materially impacts the recommendation or the court’s intended sentence.
3) Discovery and Transparency: You Cannot Rebut What You Cannot See
Many AI tools are proprietary. Vendors often assert trade secret protections for model architecture, feature weights, training data, or source code. Your response should be to narrow and prioritize what is necessary to test reliability and bias—without conceding that trade secret claims defeat disclosure altogether.
Key principle: if the state wants the court to rely on an AI-generated score, the defense is entitled to sufficient information to challenge it. At minimum, that typically includes the inputs used for the defendant, the scoring rules as applied, validation studies, known limitations, and policies governing use.
4) Equal Protection and Anti-Discrimination: Bias Can Be a Sentencing Fact
Risk tools can encode historical policing, charging, and conviction patterns. Even if a tool does not use race explicitly, correlated variables (e.g., ZIP code, employment, prior contact patterns) can produce disparate impacts. In sentencing, bias is not an academic critique; it is a concrete reason the output may be unreliable or unfair as applied.
Defense counsel can argue that an AI score with demonstrated disparities cannot be used as an aggravating proxy for dangerousness without violating basic fairness norms and California’s commitment to individualized sentencing.
Step-by-Step Defense Playbook: How to Attack the Score
Step 1: Identify the Tool and Its Path Into the Record
Start with simple, record-building questions:
What is the tool’s name? (Vendor, version number, acronym.)
Who generated the score? (Probation, pretrial services, third-party contractor.)
When and where was it generated? (Intake, interview, automated pull from databases.)
What decision did it influence? (Probation recommendation, custody level, program eligibility.)
Ask the probation officer to identify the specific document or screen print, and request it as an attachment to the probation report or as a supplemental disclosure.
Step 2: Demand the “Inputs” and the “Explainability” for Your Client’s Score
At minimum, you need the factors used for your client and how they were coded. Common issues include:
Incorrect priors. Overcounting arrests, mislabeling dismissed cases, or including juvenile contacts improperly.
Data freshness problems. Old warrants cleared, charges reduced, or convictions vacated not reflected.
Proxy variables. Employment history, housing instability, or neighborhood metrics used in ways that penalize poverty.
Interview coding. Subjective intake answers converted into “risk points” based on unclear rubrics.
Write your request in a way that forces specificity: “All variables, fields, or questionnaire responses used to compute Defendant’s score; the value assigned to each; and all rules mapping those values to the final score.”
Step 3: Attack Foundation and Reliability Like Any Other Expert Method
Even if the prosecution calls the score a “tool” and not “expert testimony,” the defense can still argue the court should not rely on it absent a foundation showing reliability. Focus on:
Validation in comparable populations. Was the tool validated on California counties, similar demographics, and similar offense types?
Error rates and calibration. How often does “high risk” actually reoffend? Over what timeframe? For what definition of recidivism (arrest vs. conviction)?
Drift and updates. Has the model changed since the cited validation study? Was your client scored with a different version?
Human override and workflow. Does probation adjust the score? If so, where is the audit trail?
If the state cannot provide these basics, argue the score is at best a generalized administrative label and at worst misleading scientific-sounding evidence.
Step 4: Challenge Hearsay and Unsworn Summaries in the Probation Report
Probation reports often compile hearsay. While sentencing courts may consider a broad range of information, your job is to highlight what is untested and unreliable. If the “risk score” is based on database entries, third-party questionnaires, or vendor analytics, argue that the report is layering hearsay upon hearsay—then presenting it as objective.
Practical remedy requests include: (1) striking the risk score, (2) ordering probation to file a revised report without it, or (3) holding an evidentiary hearing where the scoring methodology and inputs are produced and examined.
Step 5: Use an Expert Strategically (Data Science + Sentencing Mitigation)
A credible expert can turn abstract “AI concerns” into concrete courtroom points. Consider retaining:
A statistician/data scientist to critique validation, disparate impact, calibration, and variable selection.
A forensic social scientist or mitigation specialist to reframe “risk factors” (housing instability, substance use disorder) as treatable needs rather than reasons for punishment.
Have the expert prepare a declaration that: (1) explains why the score should not be used for individual sentencing, (2) identifies what information must be disclosed for meaningful testing, and (3) flags potential bias and measurement error.
Step 6: Offer a Competing Narrative With Concrete Alternatives
Courts are more receptive when the defense provides a workable alternative. Pair your challenge with a sentencing plan:
Documented treatment placement (bed date, provider letter, payment plan).
Structured supervision proposal (reporting schedule, testing, counseling).
Employment and housing verification that counters “instability” assumptions.
Restitution plan showing accountability.
The message: the AI score is not a trustworthy proxy for public safety, and the defense has a more individualized, evidence-based plan.
Example Defense Arguments That Work in Real Sentencing Rooms
Example 1: The “Wrong Data In, Wrong Score Out” Challenge
Your client is labeled “high risk” based partly on a prior “conviction” that was actually dismissed after diversion. You file a sentencing brief attaching minute orders and DOJ rap sheet corrections, and you argue the score is objectively unreliable because it is built on inaccurate inputs. You request the court disregard the score and order a corrected probation report.
Example 2: The “Arrest-Based Recidivism” Critique
The tool defines recidivism as “any arrest within two years,” not conviction. You argue arrest-based endpoints amplify policing patterns and are not a valid measure of criminal behavior for sentencing purposes. You request the court treat the score as non-probative and instead rely on proven conduct and individualized mitigation.
