Filter by Category

How to Challenge AI-Generated Risk Assessment Scores in Sentencing Hearings Under Daubert and Due Process

Courts in at least 20 states now use algorithmic risk assessment tools at some stage of criminal sentencing or supervision. These AI-adjacent scores can materially affect incarceration length, probation conditions, and release decisions. This article explains how defense counsel can challenge AI-generated risk scores under Daubert/Frye, procedural and substantive due process, confrontation principles, and practical evidentiary motions.

Why AI-Generated Risk Scores Matter at Sentencing

Risk assessment scores—often described as “algorithmic,” “actuarial,” or “AI-assisted”—are used to predict a defendant’s likelihood of recidivism, failure to appear, or supervision violations. In practice, these scores can shift where a defendant falls on a custody continuum, influence whether a judge selects incarceration versus community-based sanctions, and shape probation conditions and length.

Even when a court treats the score as “advisory,” the number can carry outsized psychological weight. It can also become a shortcut for disputed facts—substituting a proprietary model’s output for individualized judicial reasoning. That makes risk scores fertile ground for evidentiary and constitutional challenges, especially when the score is generated by an opaque tool and embedded in a presentence report (PSR).

Know the Tool: “AI” Risk Assessment Is Often a Mix of Statistics and Policy Choices

Many tools marketed as AI are not large language models; they are statistical models (logistic regression, random forests, gradient boosting, etc.) trained on historical criminal justice data. The legal vulnerability is often not the math but the inputs and assumptions:

Common input categories: prior arrests/convictions, age at first contact, employment, education, housing stability, substance use history, and “antisocial attitudes” proxies. Some tools use neighborhood or peer variables that can correlate with protected characteristics.

Common legal problems: hidden weighting, unvalidated local performance, data errors in criminal history, “label leakage” from policing bias, and confusion between correlation and causation. These issues inform both Daubert/Frye admissibility and due process reliability arguments.

Threshold Question: Is the Risk Score “Evidence” Subject to Daubert/Frye?

Before litigating scientific reliability, frame the procedural posture. In many jurisdictions, sentencing courts can consider a broad range of information. Prosecutors may argue that relaxed evidentiary rules make Daubert inapplicable. Defense counsel should respond in two steps:

1) Argue the score is expert-like opinion, not mere background

A risk score is not simply a recitation of records; it is an inferential prediction generated by a specialized methodology. When the prosecution asks the court to rely on that prediction, it functions like expert testimony—particularly if a probation officer “interprets” the score or testifies about its meaning.

2) Even in relaxed sentencing evidentiary regimes, due process requires reliability

Many courts hold that sentencing information must have “sufficient indicia of reliability” and that defendants must have a meaningful opportunity to rebut adverse information. That reliability floor provides a pathway to Daubert-like scrutiny even when formal rules of evidence are relaxed.

Daubert (and Frye) Framework Applied to AI Risk Tools

Where Daubert applies (or where courts import its logic), structure your challenge around the familiar factors: testability, known error rate, peer review/publication, standards controlling operation, and general acceptance. Under Frye, focus on general acceptance in the relevant scientific community, but still develop the same record.

Testability and reproducibility

Demand to know whether the model can be independently tested. If the vendor’s model is proprietary and cannot be reproduced from disclosed materials, argue it fails basic scientific norms. Reproducibility is especially important where minor input variations materially change the score.

Practice point: Ask whether the score can be re-run using the same inputs and produce the same output (determinism), whether updates occur (model drift), and whether the jurisdiction is using the same version validated in studies.

Known or potential error rate

Force the proponent to identify error rates in metrics judges can understand: false positives (predicting “high risk” when the person does not reoffend) and false negatives. Many tools report AUC or correlation statistics that are not intuitive and can obscure high false-positive rates for the subgroup most relevant to sentencing (e.g., people without recent convictions).

Cross-examination hook: “For defendants in my client’s demographic and offense category, what is the false-positive rate? How many ‘high-risk’ defendants are actually not rearrested within the prediction window?”

Peer review and publication vs. marketing claims

Distinguish independent peer-reviewed validation from white papers commissioned by the vendor. If the “studies” are non-public, lack raw data access, or are not specific to the jurisdiction’s population, argue they do not satisfy Daubert reliability.

Standards and controls: how inputs are defined and collected

Risk tools often depend on questionnaire responses, probation officer coding, and criminal history databases. Challenge the lack of standardized administration and inter-rater reliability: different officers may score the same person differently, affecting the output.

Example: If “employment stability” is coded subjectively, show how the scoring rubric permits inconsistent classification for gig work, caregiving, or informal employment.

General acceptance and the relevant community

Under Frye, define the “relevant scientific community” narrowly: criminometrics/psychometrics experts, not probation departments or vendor sales teams. “Widespread use” by agencies is not the same as scientific acceptance.

Due Process: The Core Argument Is Meaningful Notice and Opportunity to Rebut

Even where a court permits wide-ranging sentencing information, due process principles still require basic fairness. AI-generated risk scores raise recurring due process concerns:

Opacity and the right to contest adverse information

If the defense cannot see how the score was generated—inputs, weights, training data, calibration methods, and validation results—then the defendant cannot effectively rebut it. Argue that “trust us” is not an adequate substitute when the score may increase punishment or restrict liberty.

Accuracy of inputs: garbage in, garbage out

Many scores rely on criminal history and arrest data that contain errors or reflect enforcement disparities. Due process supports a robust challenge to the underlying records and assumptions:

Common input disputes: misclassified dispositions, sealed/expunged matters, old arrests without convictions treated as risk-elevating, juvenile contacts improperly included, and duplicative entries.

Procedural regularity: undisclosed policy choices disguised as science

Some tools embed normative judgments—for example, treating unemployment as a major risk factor. That may penalize poverty rather than predict crime in a constitutionally acceptable way. Frame this as a due process concern: punishment should be individualized and tied to legally relevant factors, not socioeconomic proxies dressed up as neutral computation.

Confrontation Clause: Use It Strategically (Even If It’s an Uphill Fight)

Traditional Confrontation Clause doctrine is strongest at trial, and many courts limit confrontation rights at sentencing. Still, confrontation principles can be persuasive when the score is introduced through testimonial statements (e.g., a report prepared for litigation) or when a probation officer relays the vendor’s assertions as truth.

Strategic use: Even if the court rejects a categorical confrontation claim, it may order production of underlying materials or permit robust cross-examination to satisfy due process reliability.

Discovery and Motion Practice: Build a Record That Forces Transparency

Effective challenges depend on obtaining the right materials. Consider a coordinated set of motions:

1) Motion to compel production of the tool’s technical and validation documentation

Request: the model version used; input variables and definitions; scoring rubric; training and validation datasets (or at least summary statistics); calibration and local validation studies; error rates; subgroup performance; update history; and any internal audits.

If the state claims trade-secret privilege: propose a protective order, in camera review, limited expert access, or disclosure of sufficient information to test reliability without public release. Emphasize that liberty interests outweigh generalized proprietary concerns when the state uses the tool to seek a harsher sentence.

2) Motion to exclude or limit under Daubert/Frye (and “reliability at sentencing” standards)

Ask the court to (a) exclude the score; (b) prohibit “high risk” labels; (c) allow only a narrative discussion of verified, contestable facts; or (d) limit use to supervision planning, not incarceration length.

3) Motion to correct PSR and strike unsupported conclusions

Where the score appears in the PSR, object in writing and request findings. Preserve issues by forcing the judge to rule on disputed factual predicates and to state whether the score influenced the sentence.

4) Request for an evidentiary hearing

Ask for a focused hearing with testimony from: the probation officer who administered the tool, the agency’s risk assessment coordinator, and (if possible) a vendor representative or custodian of records. If the court refuses, make a proffer of what your expert would show.

Expert Strategy: The Right Expert Can Reframe the Fight

An effective defense expert is often a statistician, psychometrician, criminologist, or computer scientist with experience in validation and fairness metrics. The expert’s goals are practical:

Explain calibration and base rates: A tool can look “accurate” overall while being poorly calibrated for your client’s subgroup or local population.

Translate performance into courtroom terms: false positives, false negatives, and “net benefit” compared to simpler baselines.

Audit the inputs: show how subjective items and data quality issues create instability in the score.

Demonstrate disparate impact mechanisms: without overstating conclusions; connect to policing intensity, arrest-as-proxy variables, and socioeconomic factors.