Journal of African Development

Algorithmic evidence is moving from the police workstation to the criminal courtroom. Facial recognition outputs, probabilistic genotyping likelihood ratios, risk scores, and tool-generated investigative leads now shape arrest, charging, bail, plea bargaining, and, increasingly, proof at trial. Yet criminal procedure still assumes that evidence is either human testimony or a readable artefact whose reliability can be tested through disclosure and cross-examination. This paper argues that courts must examine not only the algorithm’s final result, but also the full process that produced it, because that process decides whether the result is trustworthy. Through doctrinal and comparative analysis of India, the European human rights and data protection framework, the United Kingdom’s disclosure and expert evidence regime, and the United States’ Daubert and Frye reliability gatekeeping, the paper develops a three-part account. First, admissibility must demand demonstrated validity for the claimed use, plus error characterisation that is usable in Court. Second, disclosure must be structured around meaningful defence challenge rather than vendor secrecy, using calibrated protective orders where necessary. Third, defence rights must include effective access to technical assistance, preservation of run artefacts, and credible remedies for non-disclosure. Recent judicial responses, including exclusion of facial recognition outputs under Frye Mack, appellate scrutiny of undisclosed facial recognition disclaimers in warrant practice, and appellate approval of probabilistic genotyping where code access issues are addressed on the record, show both the promise and the limits of existing tools. For India, the immediate task is doctrinal adaptation, electronic record admissibility establishes authenticity, but algorithmic inference requires a reliability and contestability layer anchored in Article 21 fair trial, Article 14 non-arbitrariness, and surveillance legality.