Do Not Predict the Defendant
Risk scores can organise evidence and support system-level research, but they cannot turn group statistics into an individual fact capable of justifying guilt, sentence or liberty.
Human-centred technology—from everyday life to deep engineering.
Risk scores can organise evidence and support system-level research, but they cannot turn group statistics into an individual fact capable of justifying guilt, sentence or liberty.
Digital rules should face transparent synthetic cases, historical edge cases and community-designed scenarios before they change real entitlements or obligations.
Rules as code can make public decisions testable and consistent, but executable rules must remain traceable companions to law—not hidden substitutes for judgement.
A digital submission is not complete when a button turns green. A useful public-service receipt proves what was received, when it arrived and how the record can be corrected.
Structured legislation helps people and machines find the right law, version and source—without turning technical formats into legal authority or interpretation.
Human oversight fails when reviewers lack time, evidence or authority. Meaningful review requires independence, manageable workloads and measurable disagreement.
Automated triage distributes waiting time, attention and risk. Fair queues require defensible priorities, ageing, human overrides and monitoring beyond average speed.
Decision-support tools should show caseworkers evidence, provenance, gaps, uncertainty and lawful options—not compress a person into a recommendation score.
AI can assist judicial preparation, but a judge must still verify the record and law, resist anchoring, protect confidentiality and own every material reason.
Remote hearings can widen access, but cases involving liberty, family life, confidential advice, coercion or vulnerable participation may still require a physical courtroom.