How to Challenge AI-Generated Risk Assessment Scores in Sentencing Hearings: A Step-by-Step Defense Strategy
AI-generated risk assessment scores can add 6–12 months or more to a recommended sentence when courts treat “high risk” labels as aggravating. These tools—often based on proprietary algorithms—are now used in many jurisdictions at bail, probation, and sentencing, raising due process, confrontation, and reliability concerns. This article gives defense counsel a step-by-step strategy to identify the tool, obtain discovery, litigate admissibility, and build a record for appeal.
Why AI Risk Scores Matter at Sentencing
“Risk assessment” instruments—sometimes marketed as AI—convert a defendant’s history and demographics into a numerical score or category (e.g., low/medium/high risk of recidivism or failure to appear). Judges and probation departments may incorporate those scores into pre-sentence investigation (PSI) reports, guideline recommendations, supervision intensity, or program eligibility.
Defense counsel should treat a risk score like any other aggravating evidence: if it is inaccurate, biased, opaque, or methodologically unsound, it can improperly inflate a sentence. Even when the court says the score is “only one factor,” it can anchor decision-making and shift the burden onto the defense to prove “low risk.”
Step 1: Identify the Exact Tool and the Decision Point
Start by pinning down what the score is and how it entered the case. “Risk score” is not a single product. Common examples include COMPAS-type tools, PSA-style pretrial instruments, state-specific actuarial tools, and proprietary vendor models embedded in probation software.
Defense checklist: confirm the basics
Request or confirm in writing:
1) Tool name and version. Risk tools change over time; validation for Version 2.1 may not apply to Version 3.0.
2) Stage of use. Was the score generated for pretrial release, a PSI, probation classification, or sentencing?
3) Output relied upon. Numerical score, risk category, narrative “flags,” recommended conditions, or supervision level.
4) Human overrides. Did the officer or judge override the score? If so, what rationale and documentation?
5) Legal consequence. Identify the concrete sentencing effect (guideline range, upward variance argument, program exclusion, custody level).
Step 2: Demand Discovery Early—Before the PSI Hardens
Many risk scores reach the judge through the PSI. Once a PSI is filed, courts often treat it as presumptively reliable unless the defense makes specific objections. Make discovery requests as soon as you learn a tool is involved—ideally before the probation interview or before the PSI is finalized.
Key discovery targets
Raw inputs and worksheets. Every data field used to generate the score (criminal history entries, age, employment, education, residence stability, prior FTA, substance history, etc.), including the officer’s notes and source documents.
Scoring rubric and weighting. The factor weights, cutoffs, and how the score is calculated. If proprietary, request at least a scoring explanation sufficient to test accuracy and error rates.
Validation studies. Tool-wide and jurisdiction-specific validation, including base rates, false positive/negative rates, calibration by subgroup, and any revalidation schedule.
Training materials. Probation/vendor training on data entry, interpretation, and limitations—often revealing that the tool is not meant for sentencing severity.
Audit logs. Who entered data, when, and any edits. Input errors are common and highly litigable.
Contracts and policies. Vendor contracts, policy manuals, and agency directives governing use, prohibitions, and “do not use for sentencing” limitations.
Source code/model documentation (when feasible). For some courts, full source code is a long shot; but model cards, technical documentation, and independent audit reports are more attainable.
Practical motion language to tee up the issues
In discovery motions, frame the score as evidence offered to increase punishment. Argue that fundamental fairness requires access to (1) inputs to verify accuracy, (2) methodology to test reliability, and (3) validation/error-rate data to rebut probative value. If the prosecution claims it is not “their” evidence, point out that sentencing evidence still must be reliable and that the state is seeking to benefit from the score’s persuasive effect.
Step 3: Object to Inaccurate Inputs—The Most Winnable Attack
The fastest path to discrediting a risk score is showing that it rests on wrong data. Many instruments are sensitive to a small number of fields (prior violent offense classification, number of prior arrests/convictions, prior supervision failures). A single mis-coded event can move a defendant into a higher category.
Common input errors to hunt
Arrest vs. conviction confusion. Some records list arrests that never led to conviction; incorporating them can raise risk and raise due process concerns.
Out-of-state equivalency errors. Mislabeling non-violent convictions as violent, or counting expunged/sealed matters.
Juvenile history misuse. Juvenile adjudications or contacts may be improperly included or over-weighted.
FTAs and warrants. Administrative warrants or failures caused by homelessness, hospitalization, or lack of notice can be mischaracterized.
Subjective fields. “Antisocial peers,” “attitude,” or “compliance” fields invite bias and should be scrutinized for factual support.
File written PSI objections with citations to docket entries, certified dispositions, and affidavits where needed. Ask the court to strike the score or order recalculation with corrected inputs, and request a continuance if recalculation cannot be completed before sentencing.
Step 4: Litigate Reliability and Admissibility (Even in “Relaxed” Sentencing Rules)
Sentencing often allows broader information than trial evidence, but most jurisdictions still require that information used to increase punishment be reliable and that the defendant have a meaningful opportunity to rebut it. Your motion should focus on reliability, transparency, and the inability to test the tool.
Core arguments to develop
1) Unreliable hearsay / lack of foundation. If the state cannot show who generated the score, how inputs were verified, and what the tool measures, the score lacks foundation.
2) Methodological validity. Ask whether the tool was validated for the purpose used here. A tool designed for supervision intensity may not validly support an upward variance in incarceration.
3) Error rates and calibration. A score is less probative if the false-positive rate is high or if the tool overpredicts risk for certain groups. Demand subgroup performance metrics.
4) “Black box” opacity as a rebuttal impairment. If the defense cannot access the scoring logic or sufficient documentation, the defendant cannot meaningfully challenge it. That is a fairness problem even when the model is proprietary.
5) Overreliance and anchoring. Argue that the score’s scientific veneer risks undue weight, particularly when displayed as a single number or category without uncertainty intervals.
Use state constitutional and local-rule hooks
Many successful challenges are built on local PSI procedures, state due process protections, and sentencing statutes requiring accurate information. Cite any rule that permits correction of PSI inaccuracies, requires disclosure of information relied upon, or mandates an evidentiary hearing on disputed sentencing facts.
Step 5: Seek an Evidentiary Hearing and Put the Tool on the Stand
Don’t settle for argument alone if the score is being used to justify a longer sentence. Move for an evidentiary hearing to examine the probation officer, a records custodian, and—when possible—a vendor representative or technical witness.
Direct goals for the hearing
Pin down purpose and limits. Establish whether the tool’s documentation warns against using the score to increase incarceration length.
Expose data-entry discretion. Show which fields are subjective and how officers are trained to complete them.
Reveal missing validation. Ask whether the tool was validated in your county/state population and how often revalidation occurs.
Quantify error. Get admissions about known error rates, confidence intervals (if any), and what “high risk” actually means in probability terms.
Show disparate impact risks. Elicit testimony on whether the agency monitors outcomes by race, ethnicity, sex, or socioeconomic status.
Sample cross-examination themes
“What does the score measure?” Recidivism over what period? Arrest or conviction? Any new offense or only violent offenses?
“How do you verify inputs?” What sources are checked? Are dismissed cases excluded? Are expungements filtered?
“Can you reproduce the calculation?” If not, argue that neither the witness nor the court can test the score’s correctness.
“Is there a confidence interval?” If the score lacks uncertainty, argue it is presented with false precision.
Step 6: Retain the Right Expert—Statistics, Data Science, or Forensic Social Science
An expert can translate technical critiques into courtroom-relevant reliability points. Depending on the tool and your theory, consider:
A statistician/psychometrician to critique validation, calibration, and base-rate neglect.
A data scientist to evaluate model documentation, feature selection, and transparency gaps.
A forensic psychologist or criminologist to address actuarial limits, ecological fallacy, and misuse for sentencing severity.
What your expert should produce
1) A rebuttal report identifying missing documentation and performance metrics.
2) A sensitivity analysis showing how small input changes move the defendant across thresholds.
3) A “fit-for-purpose” critique explaining why the tool does (or does not) answer the legal question at sentencing—individual
