1. Introduction

Artificial Intelligence (AI) is increasingly used in criminal and civil justice systems for facial recognition, predictive policing, forensic analysis, risk assessment, evidence classification, and decision-support tools. Although these technologies may improve efficiency, their use in trials raises fundamental questions concerning transparency, procedural fairness, the right to a fair trial, and judicial accountability.

The framework for explainability of AI decisions in trials refers to the legal and procedural safeguards that enable parties, lawyers, experts, and judges to understand how an AI system produced a result, what evidence it relied upon, what limitations affected its output, and whether the result can be independently challenged.

Explainability does not necessarily require disclosure of every line of source code. It requires sufficient information to evaluate the reliability, relevance, and fairness of an AI-generated result. An AI output should not replace judicial reasoning or become conclusive evidence merely because it was generated by sophisticated technology.

2. Legal Foundations of AI Explainability

The framework rests on several established legal principles.

A. Right to a fair trial: An accused person must have a meaningful opportunity to understand and challenge evidence used against them. If the prosecution relies on an opaque AI system, the defence may be unable to test the accuracy of the result or identify errors.

B. Right to equality and non-discrimination: AI systems trained on biased or unrepresentative datasets may produce discriminatory outcomes. Courts must consider whether the technology treats individuals fairly and whether its use creates unjustified disparities.

C. Right to know and challenge evidence: The parties should receive sufficient information about the AI system's methodology, input data, error rates, validation, and limitations to challenge its evidentiary value.

D. Judicial independence: Judges must independently evaluate AI-generated information rather than mechanically accepting algorithmic recommendations. The responsibility for the final decision remains with the court.

E. Reasoned judicial decisions: Where AI materially influences a judgment, the court should explain whether it accepted or rejected the AI output and why. The system's conclusion cannot substitute for the court's own legal reasoning.

In India, these principles are connected with Articles 14 and 21 of the Constitution, the requirements of procedural fairness, and applicable rules governing admissibility and proof of electronic and expert evidence.

3. Essential Components of the Explainability Framework

A. Disclosure of the AI system

The party relying on AI evidence should disclose the system's name, developer, version, intended purpose, input data, and relevant operating procedures. Where necessary, disclosure should include source code, technical documentation, training-data information, or access to an independent audit.

Trade-secret claims should be balanced against the accused person's right to challenge evidence. Confidentiality may justify protective arrangements, but it should not automatically prevent meaningful examination of evidence central to the prosecution.

B. Scientific reliability and validation

The court should examine whether the AI system has been independently tested, validated for its intended purpose, and evaluated for false-positive and false-negative results. Testing should reflect the population, conditions, and circumstances in which the system was used.

For example, a facial-recognition system's identification of a suspect should not be treated as proof of identity without examining image quality, similarity thresholds, error rates, and possible alternative matches.

C. Human interpretability

An AI-generated conclusion should be accompanied by a comprehensible explanation of the principal factors that influenced it. The explanation should identify relevant inputs, assumptions, limitations, confidence measures where meaningful, and the possibility of alternative outcomes.

Importantly, a confidence score is not necessarily the probability that the accused committed the offence. Its meaning depends on how the system was designed and validated.

D. Adversarial testing and cross-examination

The defence must have a practical opportunity to question the expert or operator responsible for the AI evidence. Relevant questions include whether the system was properly configured, whether its data were accurate, whether the model was updated, and whether human operators influenced its output.

E. Independent judicial assessment

The judge must determine whether the AI evidence is admissible, relevant, and sufficiently reliable under the applicable law. The court should not delegate the determination of guilt, liability, or credibility to an algorithm.

4. Important Case Laws

1. State v. Loomis, 881 N.W.2d 749 (Wisconsin Supreme Court, 2016)

This is a leading decision concerning algorithmic risk assessment in criminal sentencing. The court considered the COMPAS risk-assessment system, which estimated the defendant's risk of reoffending.

The court permitted its use subject to limitations and cautions, including warnings about its application, gender-related concerns, and the proprietary nature of the algorithm. It did not authorize COMPAS to determine a sentence independently or to be treated as conclusive evidence.

Legal significance: The decision demonstrates why courts must understand an algorithm's limitations and prevent opaque risk scores from becoming decisive factors in sentencing.

2. Daubert v. Merrell Dow Pharmaceuticals, Inc., 509 U.S. 579 (1993)

The United States Supreme Court established a judicial gatekeeping approach for scientific expert evidence under the Federal Rules of Evidence. Relevant considerations include testability, peer review, known or potential error rates, operating standards, and general acceptance within the scientific community.

Justia U.S. Supreme Court Center

+1

Legal significance: Although not an AI-specific decision, Daubert provides a useful framework for evaluating AI-generated expert evidence. Courts should assess the reliability of the underlying methodology rather than simply accept an expert's conclusion.

3. Anvar P.V. v. P.K. Basheer, (2014) 10 SCC 473

The Supreme Court of India examined the statutory requirements governing the admissibility of electronic records under the former Indian Evidence Act, 1872, particularly Section 65B.

Legal significance: AI-generated reports, machine logs, and electronic outputs may raise questions about authentication and statutory compliance. The case illustrates that technological sophistication does not remove the need to satisfy applicable evidentiary requirements. The current framework must be read alongside the Bharatiya Sakshya Adhiniyam, 2023, which came into force on 1 July 2024. Sections 62 and 63 address electronic records.

IndiaCode

+1

4. Arjun Panditrao Khotkar v. Kailash Kushanrao Gorantyal, (2020) 7 SCC 1

The Supreme Court clarified the importance of the certificate requirement under Section 65B for electronic records when that provision applies, while explaining the treatment of electronic evidence and the circumstances in which certification is required.

Legal significance: The decision supports careful scrutiny of the authenticity and legal admissibility of electronically generated material. It does not establish a separate, comprehensive AI-explainability test, and its statutory analysis must be applied in light of the current evidence legislation.

5. Application Under Indian Law

The Indian framework can be developed through constitutional protections and evidentiary safeguards.

Article 14: AI-assisted decisions should not result in arbitrary or discriminatory treatment.

Article 21: The right to life and personal liberty requires legally fair procedures, particularly where AI evidence influences criminal conviction or sentencing.

Bharatiya Sakshya Adhiniyam, 2023: Sections 39 and 40 address expert opinions and grounds of opinion, while Sections 62 and 63 govern electronic records. Their application depends on the nature of the evidence and the proceedings.

Cross-examination: The defence should be able to challenge the expert's reasoning, data quality, methodology, and conclusions under applicable procedural and evidentiary rules.

These provisions do not automatically make every AI system inadmissible unless its source code is disclosed. Rather, the court must assess the applicable legal requirements and whether the parties have a meaningful opportunity to contest the evidence.

6. Conclusion

A sound framework for explainability of AI decisions in trials should combine transparency, scientific validation, disclosure, independent expert examination, effective cross-examination, protection against discrimination, and reasoned judicial decision-making.

The principles reflected in State v. Loomis, Daubert, Anvar P.V., and Arjun Panditrao demonstrate the importance of scrutinising algorithmic outputs and electronic evidence through established legal safeguards. However, these decisions do not collectively establish a universal AI-explainability rule.

Ultimately, AI should assist the administration of justice, not replace judicial reasoning. A court should give an AI-generated conclusion weight only to the extent justified by its reliability, relevance, lawful admissibility, and the opportunity available to the parties to challenge it.