Admitted Without Permission
Imagine a civil dispute where one party produces a ChatGPT-generated chronology of events and tenders it as documentary evidence. Or a criminal trial where AI-enhanced CCTV footage is placed before the court to identify the accused. Does the court admit it? Under what standard? And who, exactly, vouches for its accuracy?
These are not hypothetical questions anymore. AI tools are embedded in legal practice, forensic investigation, and dispute documentation. Indian courts will increasingly face evidence whose author is not a human being. The Bharatiya Sakshya Adhiniyam, 2023 (‘BSA’) updated the language of evidence law but left its foundations untouched – carrying forward assumptions about authorship, origin, and verifiability that simply do not hold when a machine-learning model is the source. AI-generated evidence is neither clearly admissible nor clearly excluded, and courts have no framework to fall back on.
Not All Robots Are the Same
‘AI-generated evidence’ is not one thing – it covers a wide spectrum, and the legal difficulties differ depending on where a particular item sits. At one end is AI as a tool. Here, AI processes or enhances human-generated source material. Forensic image sharpening, audio noise reduction, AI-assisted transcription are instances where the underlying content exists, and the AI merely works on it. At the other end is AI as the source. Here, the model itself generates the output. An LLM-produced summary of documents, a generative reconstruction of a damaged image, a predictive analytics report inferring facts from data – there is no human author of the content itself.
The distinction matters enormously. A court familiar with AI as a tool may, without realizing it, apply the same logic to AI as a source – treating a generated output as though it were merely processed data. Legal systems were not built for this. They were built around human testimony and its functional equivalents. The BSA gives courts no vocabulary to tell these two categories apart, let alone handle them differently.
The Section 63 Problem: A Framework That Assumes a Human
The BSA carries forward the Indian Evidence Act’s approach to electronic records. Under Section 57, electronic records are admissible as evidence. But the provision assumes the electronic record comes from a deterministic system – one that produces the same output for the same input, every time. Generative AI models do not work like that. Their outputs are probabilistic and context-dependent. The same prompt, run twice, can produce different results. The thing that makes them powerful is precisely what makes them legally unreliable.
The certificate requirement under Section 63 of the BSA compounds the issue. Section 63 requires a certificate from a person in a responsible official position, attesting to how the electronic record was produced. The Supreme Court confirmed in Anvar P.V. v. P.K. Basheer (2014) 10 SCC 473 and Arjun Panditrao Khotkar v. Kailash Kushanrao Gorantyal (2020) 7 SCC 1, that this certificate is a condition precedent for admissibility. But who issues that certificate for an LLM output? A large language model has no responsible official. Its training data, weights, and inference settings are proprietary and opaque. The person who typed the prompt is a user of an interface, not the system’s operator. They cannot certify how the model produced its output because they do not know. Section 63 has no meaningful equivalent for AI-generated content.
Four Ways the Law Falls Apart
The framework’s inadequacy becomes even clearer when tested against what evidence law actually demands of any item placed before a court.
The first requirement is authenticity: evidence must be what it claims to be. AI models are well-documented to hallucinate – producing outputs that are fluent and confident but factually wrong, inconsistent, or fabricated, with no internal signal of error. A hallucinated answer looks exactly like a correct one. Indian law has no standardized mechanism to test whether an AI output corresponds to the reality it purports to describe. Traditional authenticity checks – signatures, cryptographic hashing, chain of custody – were built for human-generated content and do not cleanly apply here.
The second difficulty is the gap between admissibility and reliability. Indian evidence law does not draw a sharp structural line between them.[1] Once evidence clears the admissibility threshold, its weight is left to the trier of fact without a formal reliability gate. In contrast, in the United States, Daubert v. Merrell Dow Pharmaceuticals, Inc., 509 US 579 (1993) requires judges to act as gatekeepers before expert evidence is admitted at all, and the 2023 amendments to Federal Rule of Evidence 702 placed the burden of establishing reliability on the proponent. India has no equivalent mechanism.
The third difficulty – perhaps the deepest – is the cross-examination vacuum. When the AI itself is the source of evidence – not an expert’s tool but the evidence itself – cross-examination becomes structurally impossible. You cannot swear in a model. In State v. Loomis, 881 NW 2d 749 (Wis 2016), a defendant was sentenced partly on a proprietary algorithmic risk score. Its methodology was a trade secret; he had no means of challenging it. That is precisely the adversarial failure AI evidence produces.
The fourth issue is chain of custody. Any break in the chain of custody can render evidence inadmissible.[2] AI outputs often have no meaningful chain to speak of. Model versions update; temperature settings shift; the same prompt run a week later produces a different answer. The exact output may be genuinely irrecoverable – even by the person who first generated it.
What the Rest of the World Has Figured Out
India’s position becomes harder to defend when placed alongside other jurisdictions that have grappled with the same issues. In the United States, the Daubert framework requires courts to assess methodology, known error rates, and peer acceptance as conditions of admission – not merely factors going to weight. AI outputs can be excluded at the threshold when those criteria are not met. In Germany, machine evidence is routed through a court-appointed expert who must explain and validate the AI system’s methodology – structured judicial oversight that does not depend on partisan vetting. In the European Union, Regulation (EU) 2024/1689 – the EU AI Act classifies AI systems used in the administration of justice as high-risk under Annex III, triggering mandatory operational logging, human oversight, and full technical documentation of design logic and known limitations.
India has none of this – no reliability gate, no transparency mandate, no expert-validation structure – making it the only major common law jurisdiction with a recently enacted evidence statute and no framework for AI-generated evidence. The Digital Personal Data Protection Act, 2023 adds a further dimension: AI systems trained on personal data raise admissibility questions about consent and data provenance that the BSA is entirely silent on. The gap is not narrow – it is structural.
Building the Lock After the Door Has Opened
The most practical immediate step courts could adopt is a two-stage inquiry applied whenever AI-generated evidence is tendered. At the first stage of admissibility, the tendering party must show that the AI system’s methodology is documented and auditable, that it has been validated against known benchmarks, and that the output was generated under conditions that can be described and replicated. At the second stage of according weight, the court should require corroborating evidence from traditional or human sources before giving the output determinative importance. This is not a new doctrinal invention: it mirrors how courts already handle expert opinion under Section 45 of the BSA – admissible, but never conclusive, and always subject to scrutiny.
Legislatively, Section 63 of the BSA should be amended to create an ‘algorithmic certificate’ for AI-generated outputs – requiring disclosure of the model name and version, the exact prompt, known error rates, and confirmation that the output has not been edited post-generation. A false certificate should carry the same evidentiary consequences as producing a forged document, giving courts an enforcement mechanism rather than a mere formality. Where certification is not possible, the output should be routed through Section 45 BSA: a human expert adopts it, explains the methodology, and submits it to cross-examination. A Supreme Court practice direction or Law Commission reference would be the appropriate vehicle to formalize these principles before courts face high-stakes cases without guidance.
The Verdict on the Law Itself
The BSA 2023 was a real opportunity to rebuild Indian evidence law for the AI age. What came out of it was a digital facelift – the same foundational assumptions about human authorship, traceable origin, and cross-examinable sources, carried over into a new statute. Those assumptions do not hold for generative AI.
Without intervention, courts face a binary choice: blanket exclusion puts them out of step with the evidence parties rely on; uncritical admission allows hallucinated and unreproducible content to carry probative weight in proceedings that determine rights and liberties. Neither outcome is acceptable.
The tools already exist – in comparative jurisprudence, in the analogical extension of Sections 45 and 63 of the BSA, and in straightforward legislative drafting. What is required is not ingenuity. It is simply the institutional will to act on the recognition that a law which cannot answer ‘who vouches for this?’ has not kept pace with the evidence it is being asked to assess.
(This post has been authored by Sanchi Deshpande, 3rd Year student at National Law School of India University, Bengaluru)
CITE AS: Sanchi Deshpande, ‘The Witness Nobody Swore In’ (The Contemporary Law Forum, 08 July 2026) <https://tclf.in/2026/07/08/the-witness-nobody-swore-in/> date of access.