Admissibility of AI-generated digital evidence in criminal proceedings: A framework for enhancing judicial reliability in Sri Lanka

Review

M.N.D. Jayasumana examines how Sri Lankan criminal procedure should treat evidence generated, transformed, enhanced, classified, or analysed by AI systems. The paper’s central argument is deliberately incremental: Sri Lanka does not need an entirely new philosophy of evidence or a comprehensive AI-evidence statute. Existing principles of authenticity, safe custody, anti-tampering safeguards, disclosure, expert evidence, and fair trial remain usable, but they need procedural mechanisms capable of making AI-mediated transformations visible and challengeable. Through doctrinal analysis of Sri Lankan law and functional comparison with England and Wales, the European Union, and India, the paper proposes an AI Evidence Certificate, provenance and audit requirements, independent verification where necessary, and a nine-question judicial admissibility checklist.

The paper’s strongest contribution is to treat admissibility as a governance problem across the evidentiary lifecycle rather than as a binary judgment about whether an AI system is accurate. The relevant object of scrutiny is the process through which a particular exhibit was captured, preserved, transferred, processed, transformed, disclosed, reviewed, and presented. This shifts authority away from the apparent plausibility of an output and toward an inspectable chain of evidence. The proposed certificate operationalizes that shift by requiring source identifiers, hashes, model and version information, prompts or settings where material, processing logs, human interventions, validation material, known limitations, and provenance credentials where available.

This is also where the paper makes an important distinction between provenance and legitimacy. A hash can show that a file has not changed after a particular point; a C2PA credential can strengthen a claim about origin or transformation; a model card can describe intended use and limitations. None of these artifacts establishes that the evidence is true, unbiased, legally obtained, or sufficient to convict. The paper correctly places them inside a broader judicial inquiry into legality, relevance, authenticity, integrity, transparency, validity, bias, fairness, and weight. Its statement that a checklist should discipline judicial discretion rather than replace it is institutionally sound.

The proposed framework is practical, but its implementation risks deserve more attention. An AI Evidence Certificate is initially completed by the party producing the evidence, which means the court remains dependent on the completeness and accuracy of information supplied by investigators, experts, laboratories, or vendors. Proprietary tools may not expose training data, internal logic, or complete logs. The paper answers this partly through a standard of practical sufficiency and by treating non-disclosure or missing audit trails as relevant to weight or admissibility, but the harder governance question is how an accused person obtains meaningful equality of arms when the prosecution has access to technical systems, vendor documentation, or expert resources that the defence cannot independently reproduce.

A second implementation gap concerns institutional capacity. Independent expert verification, immutable audit trails, secure repositories, version-aware forensic practice, and technically informed judicial review all require sustained funding, accredited expertise, procurement controls, and procedural time. The framework identifies training and institutional reform, but it does not specify who bears verification costs, when independent review becomes mandatory, how conflicts between vendor confidentiality and disclosure are resolved, or what minimum technical capacity courts should require before admitting high-impact AI-mediated evidence. Those questions will determine whether the framework becomes a genuine contestability mechanism or a documentation layer that sophisticated parties can satisfy more easily than resource-constrained defendants.

The paper nevertheless establishes a credible reform path. Its staged approach, beginning with practice directions and targeted amendments rather than comprehensive new legislation, is proportionate to the evidentiary problem it identifies. The durable governance proposition is that AI evidence should not become authoritative because a system produced it or because it looks persuasive. Judicial reliability depends on whether the process that generated the exhibit can be reconstructed, tested, challenged, and bounded by fair-trial safeguards.

Key Insight

AI-generated evidence should acquire courtroom authority only through a verifiable evidentiary process that preserves provenance, transformation history, disclosure, independent challenge, and judicial control.