When Artificial Intelligence Becomes Witness: Who Bears Responsibility For AI-Generated Evidence?

Testifying from memory, documents, or expert opinion have long been accepted as types of evidence in court. The third category is generated by AI — a product that resembles testimony, reads like a document and is confident of being an authority, but isn't in the law. A facial recognition match, a predictive policing risk score, a detection report of the ability to clone a voice, AI-generated CCTV footage and a generative reconstruction of a crime scene are increasingly being arresting items in the Indian courtroom. The problem is no one, not even the machine, its manufacturer or the investigators who operated it, can be subjected to the kind of cross-examination that can be allowed for a human witness. This presents a real accountability void that the profession should be addressing head-on, not simply by ignoring AI-generated records as another variant of electronic record.

The Changing Face of Evidence

The traditional evidential theory is based on an assumption that a source can be interviewed. The witness is sworn in and is subject to cross-examination; Under Section 45 of the old Evidence Act (now Section 39 of the Bharatiya Sakshya Adhiniyam 2023), the document is examined and proved by the party who created it or holds it; Under Section 45 of the Evidence Act, 1875 (now Section 39 of the Bharatiya Sakshya Adhiniyam 2023), an expert opinion is obtained by asking the expert to explain the method. This is where AI-generated evidence comes in handy. A neural network isn't about to tell you why it marked a face as a match or why an individual got a high score when it called them a high risk of recidivism; it is only going to spit out a statistic made of weights that are even as obscure as they may be to their own creators. In granting such output it is effectively asking to accept an assertion of a source who is unable to take an oath, who can't be cross-examined, and who often can't explain its written opinion in a way a judge can evaluate.

The Indian Statutory Framework and Its Silence

The evidentiary framework in India for any document created by a computer remains electronic-record driven, both under Indian Evidence Act 1872 s 65B as well as under its successor, the Bharatiya Sakshya Adhiniyam, 2023 (BSA) ss 61 to 63. As explained in Anvar P.V. v. P.K. Basheer and confirmed by the Constitution Bench in Arjun Panditrao Khotkar v. Kailash Kushanrao Gorantyal, the following are to be included in the certificate: This is one that was created for records that a computer stores and can pass on – a call log, an email, a CCTV clip. It was never meant for records made by a computer by making inferences. Section 63 does not presume that the analytical process used to arrive at an AI's conclusion is sound, that there was sufficient representative training data, or that the error rate for the analytical model has been shared. Certification is used to establish both custody and integrity of output file. It does not (or cannot) guarantee the veracity of the argument that produced the contents of that file. This is the divide that still has to be bridged in Indian evidence law.

Neither Witness, Nor Document, Nor Expert

The obstacle for the doctrines is that there is no clear-cut category to which AI-generated output belongs. It cannot be, "It can't be a witness within the meaning of §§118-134 of the BSA, because the requirement for a human declarant of the kind envisaged in §§118-134 of the BSA assumes that a declarant can take an oath, which is the same thing as asking that a human can take an oath, and is when the person can be examined cross-examination, which is the same as asking that a human may be examined cross-examination. It appears to be easily classified as a just a document—it's different from a photo or a ledger, where the information is captured directly, but in this case it's created from an inferential model. For example, in some cases, the machinery is introduced into evidence via an expert witness, who not only gives the jury information about the machine, but also about its general reliability which the expert is not trying to illustrate for the specific case before him — something that defence counsel seldom presses. Put another way, AI-generated content can belong to a new kind of hybrid category of material -- admissible if admitted (when appropriate to do so) but subject to the limitations it has as something corroborative but not necessarily substantive evidence, and only when it is clear enough that it could be understood as such to be meaningfully challenged.

Locating Responsibility: A Four-Point Matrix

When the machine is unable to answer it should be spread out between the human and institutional actors it is behind. There are four areas of accountability that require discussion on their own merits: a) The report's focus on new technologies and their potential impact on state law enforcement and how to manage them, b) The specific type and purpose of the application, c) The appropriate deployment model, and d) The AI's specific function within that application.

The first place that a failure occurs relates to design-time failures: the underlying model developer may have used biased or unrepresentative training data, the developer failed to provide users with confidence bounds, or the confidence bounds built into the underlying model were not designed to be “basically explained.” If there is no way to validate a forensic tool for the specific population and context that it will be applied to and is marketed for court room purposes, then it is not possible to later disclaim any liability because it is just a tool.

Second, the agency from which it comes (most often, the police or a forensic lab), gets to decide how it is used: does it allow the authorities to proceed within the limits of the validated range? Was the operator trained to use it? etc.; and how a lead it provides is treated: as a mere trail of fellow investigators or as definitive proof of guilt. Predictive policing scores and facial recognition matches are investigative tools and should not be mistaken as a finding of fact; it's a problem with how they're used not how they work.

Thirdly, certifying officer's burden under Section 63 of the BSA is distinct. The certificate is a legal statement that the record is as described and an officer who signs it without revealing that it has been generated by an inferential model (not a passive recording device) implicitly makes a factual statement that it has been generated in a different fashion.

Last but not least, there is the gatekeeping role that the court plays. Fourth, and last, there is the gatekeeping function which has not yet been adequately expressed in Indian jurisprudence. The application of Daubert-style scrutiny in the United States calls on judges to evaluate the scientific soundness of an expert's method prior to placing weight on the discovery of AI-generated evidence, Indian courts hearing such cases ought to demand artificial intelligence systems disclose their known error rates, the makeup of their training data, and validation studies conducted in the Indian context before attributing any evidentiary weight to the systems output. Failing to explore this output after acceptance by a judge without explanation is not neutrality. It's shutting your eyes to the unaccountable system you've outsourced a judicial determination to.

A Comparative Glance

Some other governments are responding (less successfully) to the lack. Forensic and law-enforcement organizations rely on AI systems, which the European Union's AI Act has identified as "high-risk," for deployment in the administration of justice, and for which the use of such systems requires documentation and monitoring by humans, with accuracy tests specified. To avoid introducing evidence unchallenged in the record, some courts overseas have introduced a requirement that claims that generative AI has been utilized to make submissions and/or evidence. India's own set of IT Rules of 2026 is a welcome but partial measure: While it will have a requirement on tracing back to the source of AI-generated content deployed publicly, the rules do not touch upon the evidentiary value that AI content has in a court of law. At the moment, there is no such thing in India as compulsory test of forensic AI in litigation.

Towards a Workable Framework

A coherent approach does not have to wait for new legislation – much can be done by reading existing provisions purposively, and developing procedural practice. Strategies that could fill the existing gap in a meaningful way include three steps. One, in the case of AI-generated content, a technical disclosure, including parameters validated for the AI, its documented error rate, and whether it was applied within parameters above, should be attached to the certificate (and not optional context). Two, AI content in evidence as evidence in support of a claimed fact should, normally, be admitted as a corroborative piece of evidence, rather than substantive evidence, and thus necessarily require some independent human verification before it can be used to secure a conviction or a substantive finding in civil proceedings – an outcome several of the analyses reviewed above suggest would be possible in criminal cases, anyway. Three, courts, over the course of practice directions (or eventually amendment) should establish a structured 'threshold” inquiry akin to a 'reliability” hearing before allowing case determinative evidence arising from AI use to be adduced, with the burden of establishing reliability falling on the party desiring its use.

The answer to the question raised in the title of this piece – who is the witness when artificial intelligence becomes a witness? – is not simple and is actually just as complicated as the problem itself. The responsibility should be spread upon the developer who created the model, the agency who put the model in place, the officer who certified the model and the court which determined the weight to be accorded to the model. That spread-out responsibility can not turn into no responsibility at all because the immediate source of the assertion is a machine, that the law simply does not allow. The Bharatiya Sakshya Adhiniyam, 2023 did make available to India a modern system of electronic records and records that never imagined that some of them will think for themselves – albeit in a rather coarse way. It is the need of the hour to fill the vacuum created by those standards (disclosure, corroboration, and effective judicial gatekeeping), in Indian evidence law, before it begins the next iterative cycle.