From Signals to Propositions: Matching Validation to the Level of Inference in AI-Assisted Forensic Genetics

Artificial intelligence (AI) is being explored across forensic genetics, from electropherogram review to support for activity-level evaluation. Treating these applications as a single technology class obscures a problem: an error close to the analytical signal and an error that shapes propositions have different forensic consequences. This narrative framework review, informed by structured literature identification and evidence mapping, proposes validation proportionality, whereby the evidence, auditability, oversight and reporting required of a tool reflect the inferential level at which its output may influence a conclusion. The framework distinguishes AI and machine learning from probabilistic genotyping, Bayesian networks and likelihood-ratio modelling, except where an AI/ML component is involved. It also distinguishes source and biological-source questions from propositions concerning transfer and activity. Validation requirements depend on intended use: analytical review aids and tools that influence mixture assumptions, biological-source classification, transfer, persistence, prevalence, recovery or reporting require different, task-specific tests of representativeness, failure modes, version control and human interaction. At the activity level, AI may retrieve and structure evidence, preserve provenance and support expert-led sensitivity analysis, but it should not formulate propositions, assign definitive probabilities, generate a casework activity-level likelihood ratio or infer the historical event. Validation must follow the forensic question, not the technology label.