A Board Meeting That Never Happened
In January 2024, an employee at the global engineering firm Arup joined what looked like an entirely ordinary video call with the company’s chief financial officer and several familiar colleagues. Everyone’s face was recognisable. The conversation was coherent. The instructions were specific enough that the employee authorised fifteen separate wire transfers totalling twenty-five million dollars before anyone realised the truth: every single person on that call, including the CFO, was an AI-generated fabrication. There had been no meeting. There was no CFO on the line. There was, very briefly, twenty-five million dollars that is now very hard to get back.
That story usually gets told as a fraud case, and it is one. But sit with what actually happened inside it for a moment, because it is the same trick a criminal court will eventually have to reckon with: a piece of audiovisual evidence, indistinguishable in the moment from the real thing, that never depicted anything that actually occurred. Fraud investigators call this a financial loss. A criminal court, faced with the same technology used the other way around, has to call it something much harder: reasonable doubt.
What India’s Evidence Law Actually Checks To see why deepfakes are such an unusually difficult problem for a courtroom, it helps to understand exactly what Indian law currently verifies when it admits a video or audio recording as evidence and, more importantly, what it does not.
Since July 2024, electronic evidence in India has been governed by Section 63 of the Bharatiya Sakshya Adhiniyam, 2023, which replaced the old Section 65B of the Indian Evidence Act with a stricter version of the same basic idea. Before a court will admit a video, audio clip, or any other electronic record, the party relying on it must comply with section 63 of the Bharatiya Sakshya Adhiniyam, 2023, including the certificate requirements prescribed under the provision and its Schedule. In Arjun Panditrao Khotkar v. Kailash Kushanrao Gorantyal (2020), the Supreme Court reaffirmed that the certificate required under section 65B of the Indian Evidence Act was a condition precedent to the admissibility of electronic evidence covered by that provision.
It is, by design, a genuinely rigorous system for answering one specific question: has this file been tampered with since it was created? What it was never built to answer, because this problem did not meaningfully exist when Section 65B was introduced in 2000, is a completely different question: did the events inside the file actually happen? A deepfake video can pass a hash-value integrity check with total ease, because nothing about the file has been altered since its creation. The fabrication happened at the moment of creation itself, inside the generative model, long before any certificate ever touched it. India’s statutory framework for electronic records is principally concerned with technical authenticity and the manner in which an electronic record was produced. It does not by itself establish that the events depicted in a recording actually occurred.
The Liar’s Dividend, in the Dock That gap cuts in two directions at once, and the second one is, if anything, more dangerous than the first. The first direction is the obvious one: a fabricated recording, properly hash-certified because nobody tampered with the fabricated file after the fact, gets treated by a court as though it were real. The second, subtler direction has a name in the academic literature: legal scholars Bobby Chesney and Danielle Citron called it the “liar’s dividend” in a 2018 paper that has turned out to be remarkably prescient. Their argument was simple: once everyone knows deepfakes exist, genuine evidence becomes deniable too, simply by asserting, with no proof required, that it might be fake. The accused does not need to prove conclusively that a recording is a deepfake. If the authenticity of crucial evidence is credibly called into question, the resulting uncertainty may contribute to reasonable doubt.
In a separate wrongful-death lawsuit against Tesla, the company’s own lawyers initially refused to authenticate a video of Elon Musk making public statements about Autopilot safety, suggesting it too could conceivably be an AI fabrication, a tactic the court explicitly rejected, warning that accepting it would let any sufficiently famous person disown anything they had ever said on camera. And the danger runs both ways at once: in September 2025, a civil case in Alameda County, California was dismissed after a judge concluded that videotaped witness testimony submitted in the case had been deliberately fabricated using AI a real instance of exactly the opposite failure, fake evidence nearly succeeding rather than real evidence being denied.
How Other Courts Are Responding: Courts elsewhere have started, tentatively, to build tools for exactly this problem, and it is worth understanding what they do and do not fix.
In the United States, the federal judiciary’s own rulemaking body opened a proposed Rule 707 that would introduce a framework governing the admissibility of certain AI-generated evidence by applying reliability standards in specified circumstances. Critics have already pointed out its central limitation: the rule only applies once a party admits their evidence is AI-generated. It does nothing at all for the harder case evidence to which nobody admits is fake, because either it genuinely isn’t, or because the party relying on it is not about to say so. Louisiana took a different approach by passing Act 250 in August 2025, requiring attorneys to exercise what the statute calls reasonable diligence to verify the authenticity of evidence before ever offering it to a court, with real professional consequences if they knew, or should have known, that evidence was fabricated and stayed silent about it. The issues of structure have been proposed once again, with more suggestions for more specific change proposals: the creation of new pretrial hearings specifically for contested authenticity, the shifting of the judge's role from a preparatory hearing to the substantive hearing of authenticity, and a higher evidentiary standard for video and audio, rather than the traditionally low bar of a witness merely testifying that a recording ‘looks accurate’.
India’s Rules Are Pointed at the Wrong Target India has moved too, but in a different place entirely and the gap between where it has moved and where the courtroom problem actually sits is the real story here. In February 2026, the government notified sweeping amendments to the IT Rules creating, for the first time, a formal legal category called Synthetically Generated Information, along with mandatory labelling, embedded provenance metadata, and takedown windows as short as two hours for deepfakes depicting real
people. It is a serious, fast-moving regulatory response, one of the most comprehensive regulatory responses to deepfakes adopted to date. But it is aimed entirely at stopping harmful synthetic content from spreading across platforms. Nothing in it touches the Bharatiya Sakshya Adhiniyam, and nothing in it tells a trial judge what to do when a defence lawyer stands up and says the video the prosecution just played might not be real. The provenance metadata the new rules require platforms to attach to AI content could, in principle, become exactly the kind of evidence a court needs to authenticate a recording’s origins but no bridge currently connects the regulatory regime built to keep deepfakes off social media to the evidentiary regime that decides what happens when one turns up in a case file.
What a Bridge Might Look Like None of this is unfixable, and the shape of a fix is already visible in pieces scattered across both regimes. Section 63 of the Bharatiya Sakshya Adhiniyam could, without much difficulty, grow a third limb to its certification requirement not just custody and technical authenticity, but provenance, drawing directly on the same watermarking and metadata infrastructure the 2026 IT Rules already force AI platforms to build for an entirely different reason. A pretrial hearing dedicated to contested authenticity, held before a judge rather than argued in front of a jury for the first time mid-trial, would borrow directly from what American scholars have been proposing under the banner of a gatekeeping role for the trial judge and Indian courts already run comparable hearings for other categories of contested expert evidence, so the institutional muscle for this already exists.
A Closing Thought
The entire architecture of a criminal trial rests on a division of labour that has held for as long as photography and recording have existed: courts decide what evidence means, and juries decide whether to believe it, but almost nobody has had to seriously litigate whether a video depicts something that actually happened at all. Deepfakes do not make evidence harder to interpret. They make it harder to trust in the first place, and they hand that same gift to the guilty as freely as they threaten the innocent. Until the law figures out how to answer the question a hash certificate was never designed to answer: not, ‘Was this file altered?’ but, ‘Did this ever happen?’ Beyond reasonable doubt may turn out to be the easiest thing AI has ever learned to manufacture.