Beyond Admissibility: Rethinking AI-Generated Documentary Evidence under the Sakshya Adhiniyam

This article examines the admissibility and authentication of AI-generated documentary evidence under the Bharatiya Sakshya Adhiniyam, 2023, and proposes a structured framework based on provenance disclosure, reliability assessment and independent corroboration.

The Sakshya Adhiniyam, 2023 creates no categorical bar to AI-generated documentary evidence

Much of the emerging debate surrounding artificial intelligence (AI)-generated evidence has focused on whether such material is admissible. That focus, however, is doctrinally misplaced. The statutory framework governing electronic records creates no categorical exclusion for AI-generated documentary evidence.

The Sakshya Adhiniyam, 2023 (BSA), defines “document” in terms broad enough to absorb any output capable of conveying information through letters, figures, or marks, irrespective of the method of production. Section 61 goes further and states, without qualification, that electronic or digital records shall have the same legal effect, validity, and enforceability as paper records.¹ The statutory scheme does not distinguish between documentary evidence produced with conventional software and that produced with generative AI. Its concern is with the legal status and proof of electronic records, rather than the technology used to produce them.

The certificate regime under Section 63 BSA — the successor to Section 65-B, Evidence Act, 1872 — supplies the mechanism. In Anvar P.V. v. P.K. Basheer², the Supreme Court held that secondary electronic evidence is inadmissible without a certificate identifying the device and the manner of production, but it did not hold that electronic evidence itself is suspect; rather, it held that uncertified electronic evidence is procedurally deficient. A three-Judge Bench of the Supreme Court in Arjun Panditrao Khotkar v. Kailash Kushanrao Gorantyal³ reaffirmed the certificate as mandatory while softening its procedural rigidity, permitting production at a later stage and empowering courts to issue a certificate where a party is unable to produce it. Taken as a whole, these rulings do not constitute a barrier to new types of digital evidence. They set up a process. AI-generated documentary evidence often enters the evidential framework through the same legal procedures that govern the admissibility of electronic data.

In summary, there is no categorical restriction to the acceptance of AI-generated evidence imposed by the legislative structure. In terms of admissibility, the same rules controlling electronic records would typically allow such evidence to enter the evidentiary framework. At that point, the theological door opens. What comes next is a more difficult issue.

The real evidentiary challenge lies beyond admissibility

Authentication is the problem — and it is a different problem, wearing the same clothes.

Admissibility asks a threshold question: Does the law permit this category of material to enter the record at all? Authentication asks a continuing question: Can this particular item be trusted to mean what it purports to mean? Indian evidentiary reform since the Anvar case has answered the first question with increasing generosity, while largely leaving the second question untouched. That asymmetry is not incidental. It is the central fact a tribunal must confront when AI-generated material is placed before it, because the two questions do not scale together. A photocopier and a large language model both produce documents. In contrast to a photocopy, a generative AI system may create information identical to that of authentic documents, since it has no physical counterpart. By treating them as posing the same evidential inquiry, this oversimplification risks a categorical mistake.

Three failure points deserve separate treatment, because collapsing them into a single inquiry — “is it AI-generated?” — produces neither useful law nor useful practice.

Provenance asks where the record came from: which system generated it, on what input, at whose instance, and for what purpose. Conventional electronic evidence answers this through a chain of custody anchored in a device and an operator — the Section 63 certificate exists precisely to fix that anchor.⁴ AI-generated material typically resists this anchor in a particular way: the same or very similar prompts can yield plausible documentary outputs without any trustworthy origin indication, and there is no device-level signature that can distinguish a real reconstruction from a fake one. Because a CCTV recording’s origin is physically limited, it has a rich provenance. Because its source is not provenance-poor, an AI output is.

There is more to the distinction than just semantics. When admissibility and authentication are confused, one runs the danger of introducing procedural procedures intended to prove the provenance of electronic documents into a situation where the more important concern is whether the content itself is accurate. That ideological mismatch is shown by generative AI.

Integrity asks whether the record, once produced, has remained unaltered — the question hash values and metadata⁵ were designed to answer for conventional digital files. AI-generated material introduces a prior integrity question that conventional doctrine never had to ask: was the record ever anchored to a real event in the first place, or does it represent the model’s statistically plausible construction of what such a record would look like? A tampered photograph has a genuine original somewhere against which tampering can be measured. A synthetic image generated from a text prompt has no original to tamper with. Integrity analysis presupposes a baseline; AI-generated content often supplies none.

Reliability, the third and most demanding inquiry, asks whether the process that produced the record is one whose outputs can be trusted as a class — the inquiry Indian Courts have historically reserved for expert and opinion evidence. Generative AI systems are, on any honest account, probabilistic text and image predictors optimised for plausibility rather than truth.⁶ This is the design specification, not a flaw that can be litigated away. Therefore, a closer example would be an expert report rather than a witness who made a mistake on a specific occasion: Courts do not accept an expert opinion as self-proving merely because the expert followed a recognised protocol. They inquire as to whether confidence is inspired by the underlying process. The long-term, adversarial analysis that has enabled courts to determine the appropriate weight to assign to expert techniques has not yet been applied to generative AI systems.

The BSA‘s certificate mechanism, built for provenance and integrity in the narrow sense of unaltered transmission, was never built to interrogate reliability in this deeper sense. The BSA does not expressly address evidentiary issues arising from generative AI systems, and it is unsurprising that a certificate mechanism designed for the transmission of existing records is silent on the trustworthiness of records generated by a system itself. But this means a court which stops at “was there a Section 63 certificate” has answered a provenance question while leaving the reliability question entirely unasked. That is the gap this article is written to name.

When an evidentiary system conflates admissibility with authentication, it risks making two opposite mistakes: excluding trustworthy AI-assisted evidence that could aid fact-finding, and giving synthetic material undue weight simply because it complies with procedural requirements intended for electronic records. Therefore, the goal is to create a more accurate law of evidence rather than a more restrictive one.

Towards a structured framework for evaluating AI-generated evidence

Not by treating AI-generated material as inadmissible — the statute forecloses that path — and not by treating it as ordinary electronic evidence either. The correct response is a graduated authentication framework, applied at the point of evaluation rather than the point of admission.

First, rather than requiring access, tribunals ought to demand generative provenance disclosure as a prerequisite for weight. The system utilised, the type of prompt or input, and whether the output was validated against an outside source should all be disclosed by a party submitting AI-generated content. In the same manner as unexplained gaps in the chain of custody already allow the Tribunal to make an unfavourable inference on weight, failure to disclose should not dismiss the evidence.

Second, much as expert techniques gain established court standing over time, the reliability inquiry suggested above should be formalised as a threshold determination, carried out once for a class of instruments rather than relitigated in every instance. Tribunals should demand further confirmation before giving AI-generated material major evidentiary weight when it is presented as substantive proof of disputed facts until such standing is proven.⁷

Third, corroboration should be treated as the operative safeguard, not certification. A Section 63 certificate confirms transmission integrity; it says nothing about whether the underlying content was ever true. Tribunals should therefore ask not, “Was this properly certified?” but “Is there evidence independent of the AI system that this content corresponds to reality?” Where such corroboration exists, admissibility and authentication converge, and the evidence should be treated as any other document. Where it does not, the Tribunal should say so — and assign weight accordingly, rather than deciding the question at the threshold where the statute has already spoken.

The law was never the obstacle. The discipline of asking the right question is.