Should a judge rely on AI risk-assessment and sentencing tools, and who owns the decision?

The shift

Predicting a defendant's likelihood of reoffending — and benchmarking a sentence against thousands of comparable cases — moves from scarce (a probation officer's interview, a clinician's assessment, a judge's accumulated intuition over years on the bench) to abundant, cheap, and instant: a score computed from historical data in seconds. What stays scarce is the accountable human who owns the sentence, the individualized moral judgment the law demands, and any ground truth about what this person will actually do.

The axioms

1. Sentencing is an individualized moral judgment about a specific person, not a lookup against a population base rate.
2. The judge is personally accountable for the sentence and owns the decision — it is issued in their name and reviewable on their reasons.
3. A defendant has a due-process right to know the basis of their sentence and to contest it.
4. The reasons for a sentence must be stated and are reviewable on appeal.
5. Predicting recidivism risk was expensive and slow — it took clinical assessment, probation interviews, or a judge's hard-won intuition.
6. Consistency across similar cases is a goal of justice, but achieving it manually is hard because no judge can hold thousands of prior cases in mind.
7. A sentence's legitimacy rests partly on it being seen as a considered human act, not a computation.
8. The information a judge acts on (priors, presentence report, the person in front of them) is curated by accountable humans who can be questioned about it.

Invalid axioms

1. Predicting recidivism risk is expensive and slow. Scoring a defendant against historical reoffending data is now near-instant and near-free, and does it at a scale no clinician or probation officer matches. The habit-trap: jurisdictions treat the arrival of a cheap score as if it added a scarce new expert to the room, and build it into bail, parole, and sentencing workflows as a load-bearing input — when what actually got cheap is plausible pattern-matching against the past, not knowledge of this person's future.
2. No judge can hold thousands of prior cases in mind, so manual consistency is out of reach. Benchmarking a proposed sentence against a large body of comparable dispositions is now trivial. The habit-trap: treating the analytic that surfaces the "typical" sentence as a neutral consistency aid, when it also quietly encodes and reproduces the disparities already in that historical distribution — consistency with a biased past is not fairness.

Unchanged axioms

1. Sentencing is an individualized moral judgment about a specific person. A risk score is a claim about a group someone resembles; the sentence is a decision about them. Collapsing the individual into the base rate is precisely what the tool cannot be allowed to do, and no gain in model accuracy changes that — it is a question of what justice is for, not what can be predicted.
2. The judge is personally accountable for the sentence and owns the decision. Accountability attaches to a person with standing, not a tool. A model cannot be appealed, reversed, disciplined, or held to have erred — the judge signs the order and answers for it. Cheaper prediction does not make accountability cheaper or transferable.
3. A defendant has a due-process right to know and contest the basis of their sentence. This rests on a legal guarantee, not on scarce analytic capacity. A right to contest the reasons doesn't weaken because the reasons now include a machine output — if anything it bites harder.
4. The reasons for a sentence must be stated and be reviewable. Appellate review runs on articulable human reasoning. A number a judge cannot explain is not a reason, however predictive it is.

New axioms

1. A score built on arrest and conviction data launders historical bias into something that reads as objective. When the training data reflects who got policed, charged, and convicted, the output inherits that skew — but arrives wearing the authority of a number, which is harder to challenge than a human hunch precisely because it looks neutral. Whether a given tool does this, and by how much, is contested and studied case by case; the structural risk is not.
2. The judge defers to a usually-plausible score while still owning the outcome (automation bias). A score that is right often enough to feel reliable is the most dangerous kind: it anchors the human decision, gets adopted as the safe default, and quietly shifts the real call to the tool — while the accountability stays, correctly, with the judge who can no longer fully reconstruct why they landed where they did.
3. Opacity collides with the right to contest. If the model's logic is proprietary, or simply too complex to state as reasons, the defendant is being sentenced partly on a basis they cannot examine or rebut — and the judge cannot fully explain. Courts have already had to rule on whether relying on a closed scoring tool is compatible with due process, and the boundary is unsettled.
4. The score anchors human judgment even when used only "as one input." Once a number is on the page, it is hard to un-see. The move from "the judge decides, informed by a report" to "the judge adjusts away from a score" reframes the whole act — and nobody has agreed how much a human should be allowed to lean on a prediction they cannot verify for the individual in front of them.

Where it breaks

Jurisdictions adopt risk tools for consistency and objectivity — treating cheap prediction as a neutral upgrade over fallible intuition (invalid habit) — while the same tools reproduce the historical distribution's bias under the cover of a number and anchor the judge who is supposed to be checking them (new). The efficiency case and the fairness case point in opposite directions on the same output, and the more accurate and routine the score becomes, the stronger the pull to defer and the weaker the judge's ability to say, on the record, why they didn't.

A second collision: the law still demands stated, contestable reasons and a human who owns the call (still holds), while the practical basis of the sentence increasingly includes a score that is opaque to the defendant and not fully reconstructable by the judge (new). Accountability stays fixed on the judge; the reasoning it is supposed to attach to is partly migrating into a tool nobody in the courtroom can fully open.

ai-axioms.vercel.app/axioms/legal/judge-sentencing-tools-ai-axioms