Artificial Intelligence in Criminal Justice Training Course

Participants can assess whether data proposed for an AI system is lawful, relevant, reliable and traceable.

• Data provenance mapped from collection source through transformation and model input
• Training, validation and operational datasets separated to prevent false assurance
• Representativeness assessed against population coverage, missingness and historical enforcement patterns
• Data quality dimensions: accuracy, completeness, timeliness, consistency and uniqueness
• Purpose limitation, minimisation, retention and access controls documented
• Sensitive personal data handling within criminal justice information-sharing arrangements
• Disclosure implications of data sources, labels, transformations and exclusion rules

Participants can interpret technical performance evidence and identify limits on operational claims.

• Confusion matrices explain true positives, false positives and missed cases
• Precision, recall, specificity and false discovery rate compared
• Threshold selection matched to harm, workload and legal consequences
• Calibration checks whether predicted probabilities match observed outcomes
• Base-rate effects examined before interpreting apparently high accuracy claims
• Validation datasets tested for independence, relevance and temporal drift
• Model cards record intended use, limitations, metrics and evaluation evidence

Participants can test AI proposals for disparate impact and define proportionate mitigation measures.

• Sources of bias traced through data, labels, features, thresholds and deployment
• Protected-characteristic analysis identifies unequal error patterns between relevant groups
• Disparate impact metrics interpreted alongside operational and legal context
• Equalised odds and calibration trade-offs examined for classification systems
• Proxy variables identified through correlation, domain knowledge and feature review
• Historical enforcement data assessed for feedback loops and sampling bias
• Mitigation options include data revision, threshold changes and use restrictions

Participants can apply enhanced safeguards to biometric and intelligence-led AI applications.

• Facial recognition workflows: enrolment, matching, candidate lists and human adjudication
• False match and false non-match rates interpreted by demographic group
• Watchlist construction controlled through lawful purpose, quality and review
• Biometric search results treated as investigative leads, not conclusive identity evidence
• Video analytics assessed for surveillance scope, retention and access restrictions
• Link analysis distinguishes correlation, association and evidential inference
• Real-time deployment authorisation records purpose, location, duration and safeguards

Participants can govern generative AI use in drafting, summarisation, translation and evidence review.

• Generative AI outputs checked for hallucination, omission, fabrication and attribution errors
• Prompt design separates authorised case information from restricted personal data
• Human verification requirements defined for summaries, translations and draft documents
• Digital evidence review maintains original files, metadata and chain of custody
• Output logging captures prompts, model version, reviewer and final disposition
• Confidentiality controls address third-party platforms, retention and cross-border processing
• Disclosure schedules identify AI-assisted analysis and relevant material generated

Participants can design human decision controls that prevent inappropriate reliance on AI outputs.

• Meaningful human oversight requires authority, competence, time and independent judgment
• Automation bias countered through structured review prompts and contradictory evidence checks
• Decision records distinguish AI output, human assessment and final rationale
• Escalation pathways address low confidence, data errors and exceptional cases
• Notice and explanation requirements considered for affected individuals
• Contestability procedures enable correction, review, appeal and complaint handling
• Frontline training covers permitted use, prohibited reliance and incident reporting

Participants can build an assurance cycle that monitors AI risks throughout system use.

• NIST AI RMF functions applied: Govern, Map, Measure and Manage
• Risk registers link harms, controls, owners, indicators and review frequency
• Pre-deployment assurance combines legal review, technical testing and operational acceptance
• Post-deployment monitoring detects drift, performance decline and changing user behaviour
• Independent audit evidence includes datasets, test results, logs and decision records
• Incident classification separates model failures, data breaches and misuse events
• Corrective actions define containment, notification, remediation and lessons learned

Participants can specify procurement and contract controls for accountable criminal justice AI.

• Requirements define intended purpose, excluded uses, users and decision authority
• Supplier due diligence tests technical claims, data practices and financial dependency
• Evaluation criteria weight accuracy, fairness, explainability, security and interoperability
• Contract clauses secure audit rights, change notification and documentation access
• Acceptance testing specifies datasets, thresholds, subgroup results and sign-off roles
• Change control governs model updates, retraining, feature additions and integrations
• Exit plans cover data return, record retention, continuity and supplier transition

Participants can produce a practical implementation plan for a controlled AI deployment or review.

• Use-case assessment consolidates operational need, rights impact and viable alternatives
• Governance operating model assigns accountable owner, risk owner and technical custodian
• Algorithmic impact assessment documents affected groups, harms and planned controls
• Implementation roadmap sequences policy, data preparation, testing, training and approval
• Key risk indicators trigger review when thresholds, errors or complaints rise
• Stakeholder engagement maps frontline users, legal advisers, communities and oversight bodies
• Ninety-day action plan assigns deliverables, owners, dates and evidence requirements