INTRODUCTION
Indian courts have always had a complicated relationship with digital evidence. Disputes over intercepted messages, call records, and CCTV footage have kept lawyers and judges busy for years. But nothing in recent memory has created quite the same headache as deepfakes. These are videos, audio clips, or images generated by machine learning tools that can put words in someone’s mouth, put their face on someone else’s body, or recreate their voice saying something they never said.
As deepfakes start turning up in defamation cases, criminal proceedings, election disputes, and domestic violence trials, courts are going to have to answer a question the law has not properly addressed: can this material be admitted, and if so, what does it take to get it in?
This blog looks at what the current legal framework says, where it falls short, and what courts ought to do in the gap.
WHAT A DEEPFAKE ACTUALLY IS?
A deepfake is produced using a type of machine learning architecture called a Generative Adversarial Network, or GAN. The basic idea is that two neural networks are set against each other: one tries to generate convincing fake content, the other tries to spot the fakes. Over thousands of training cycles, the generator gets better and better at fooling the discriminator. The output can be a video of a named person saying something they never said, or audio in someone’s voice, or a photograph that never existed.[1]
Older fake photos had obvious mistakes, like weird shadows or visible edits. Deepfakes don’t, to the naked eye, look completely real. This creates a massive legal problem because courts have always relied on spotting those visible “seams” to prove evidence was faked. With deepfakes, the seams are perfectly hidden.
WHAT DOES THE BHARATIYA SAKSHYA ADHINIYAM, 2023 SAYS?
The Bharatiya Sakshya Adhiniyam (BSA), which replaced the Indian Evidence Act of 1872, brought Indian evidentiary law into the digital age in several respects. Section 57 consolidates the rules on electronic records and makes them admissible once certain conditions are met. Section 63 defines documents broadly enough to include digital and electronic files, which means a deepfake video or audio clip is, in principle, a document that can be tendered in court.[2]
The four conditions for electronic records:
- Regular computer use: The record must have been produced by a computer that was regularly used for lawful activities by the person relying on it. It cannot be a one-off output generated for litigation.
- Proper operation: The computer must have been functioning correctly at the relevant time, or any errors must not have affected the accuracy of the output.
- Accurate input: The information in the record must have come from information fed into the computer in the ordinary course of its lawful use.
- A certificate: A responsible official must sign a certificate confirming how the record was produced and from which device it was extracted.
A deepfake, by its very nature, fails conditions one and three. It is not produced during the ordinary course of any lawful activity. It is generated, usually secretly, to simulate something that did not happen. The “information” inside it does not come from any authentic underlying event. On a strict reading of the statute, a deepfake offered as genuine footage should not clear the admissibility bar at all.[3]
WHAT THE COURTS HAVE SAID SO FAR?
Indian courts have, over the past two decades, built up a reasonably clear line of authority on electronic evidence. The most important ruling came in Arjun Panditrao Khotkar v Kailash Kushanrao Gorantyal,[4] where the Supreme Court held that compliance with the certificate requirement is mandatory, not optional. Without the Section 65B certificate (now Section 57 BSA), electronic evidence simply cannot go in.
The courts have also said, in proceedings involving intercepted digital communications, that formal compliance with certification is not enough on its own where authenticity is specifically challenged. If a party puts genuineness squarely in issue, the proponent has to demonstrate that the content is actually reliable, not just that it came with the right paperwork.[5]
The actual challenges, one by one
Authentication: When we talk about authentication in a courtroom, like under the Federal Rules of Evidence, it just means proving that a piece of evidence is exactly what you say it is. With deepfakes, proving this is a huge headache. The whole reason these fake videos are so dangerous is that regular people simply cannot tell them apart from the real deal. If a witness takes the stand and says, “Yes, that sounds like me, but I never said those words,” the court does not have much to go on. They need hard, technical proof to back up that claim.
Right now, the forensic tools used to catch deepfakes look for things we cannot see with our own eyes. They scan for weird edges around a face, unnatural blinking, or tiny glitches at the pixel level. The catch, as pointed out by research groups like DARPA’s Semantic Forensics program, is that these tools are not perfect. They cannot just give a simple “yes” or “no” answer. The AI making the deepfakes is always learning how to fix the exact mistakes that the detection tools are looking for.
Relying on probabilities totally changes the game in court. For a judge to even allow this kind of technical analysis as evidence, they have to run it through strict legal tests, like the Daubert standard. They have to ask if the detection software has a known error rate, if it is based on solid, peer-reviewed science, and if the broader forensic community actually trusts it. Because of this, juries and judges are not just looking at a video and deciding if it looks real anymore. Instead, they have to sit through complex expert testimony and try to make sense of statistical probabilities.
The Hearsay Angle: A deepfake purporting to show someone making a statement raises hearsay issues that are separate from the question of whether it is synthetic. Even a genuine video of a person speaking is only admissible if the statement captured falls within a recognised hearsay exception: a dying declaration, an admission, a confession, a res gestae utterance.[6] A fabricated video is essentially a false attribution of speech. The BSA’s provisions on admissions and confessions were written with real statements in mind. They need to be thought through again in light of content that was never uttered at all.
The Expert Gap: This is the most immediately practical problem. Detecting a deepfake requires forensic expertise that India’s government laboratories are not currently equipped to provide at scale. The Central Forensic Science Laboratories have no standardised protocols for deepfake analysis. That means courts would be entirely dependent on privately retained experts whose credentials and methodologies would themselves become battlegrounds in the litigation.
A PARALLEL WORTH DRAWING
There is an interesting structural similarity between the deepfake admissibility problem and the challenges courts face in arbitration when asked to bind parties who never signed an arbitration agreement. Just as the Group of Companies doctrine[7] requires a careful look at the connections between entities to determine whether an arbitration clause should reach further than its signatories, deepfake admissibility requires courts to look past the surface of the evidence and ask harder questions about what it actually represents.
In both cases, the formal appearance tells you less than you think. A signed contract can obscure a complex web of corporate relationships. A video recording can obscure the fact that nothing in it ever happened. Doctrine has to evolve to catch up with both.
The Supreme Court’s willingness, in Cox and Kings Ltd v SAP India (P) Ltd,[8] to convene a larger bench to examine the foundations and limits of the Group of Companies doctrine, which offers a useful model here. A similar exercise applied to AI-generated digital evidence would give trial courts across the country something to actually work with.
A WORKABLE FRAMEWORK
In the absence of specific legislation or binding precedent, courts can adopt a consistent approach by following a straightforward sequence when deepfake evidence is tendered.
Threshold Scrutiny: If any party raises a genuine challenge to the authenticity of audio-visual material, the court should require the proponent to produce forensic expert analysis before the material is placed before the fact-finder. The certificate alone is not sufficient.
Chain of Custody: Beyond the BSA certification requirement, courts should ask for documented provenance: where did this file come from, who handled it, and how did it get to court? Gaps in the chain should be treated seriously.
Methodology Disclosure: Any expert offering deepfake detection analysis must disclose the tool and version used, the dataset it was trained on, and the known error rate of the method. These go to the weight of the evidence and must be open to cross-examination.
Conditional Admission: Where forensic analysis cannot give a definitive answer, courts should consider admitting the evidence conditionally while giving the fact-finder a clear direction on the uncertainty and what weight, if any, it can reasonably be given.
CONCLUSION
Ultimately, the Indian legal system is currently sailing in uncharted waters when it comes to deepfakes. The Bharatiya Sakshya Adhiniyam provides a starting point for handling electronic records, but it is not a complete map. Right now, the law’s instinct is to strictly reject synthetic media. That makes perfect sense when a party is trying to pass off a fake video as the truth. However, this rigid approach creates a massive hurdle for victims who need to submit the very deepfakes used against them to prove defamation, fraud, or harassment.
Moving forward, the solution will not come from rewriting the rulebook overnight. Instead, it will require Indian courts to develop a new kind of technical literacy. Judges will need to move beyond relying on visual authentication and start working comfortably with digital forensics, probability scores, and expert testimonies. Just as the Supreme Court has carefully untangled complex modern issues in other areas of law, it will eventually need to set clear, predictable standards for authenticating and evaluating synthetic media. Until then, the focus must be on creating a framework where the law is rigorous enough to catch digital lies without shutting the courtroom door on the people harmed by them.
Author: Anchita C Patil (PES University, RR Campus, Bengaluru)
References:
[1] Ian Goodfellow and others, ‘Generative Adversarial Nets’ (NeurIPS, 2014)
[2] Bharatiya Sakshya Adhiniyam 2023, ss 57, 63; Indian Evidence Act 1872, s 65B (repealed)
[3] Tomaso Degasperi, ‘Admissibility of AI-Generated Evidence’ (2025) 29 International Journal of Evidence and Proof 112
[4] Arjun Panditrao Khotkar v Kailash Kushanrao Gorantyal (2020) 7 SCC 1
[5] State (NCT of Delhi) v Navjot Sandhu (2005) 11 SCC 600
[6] Bharatiya Sakshya Adhiniyam 2023, ss 17-30 (admissions and confessions)
[7] Chloro Controls India (P) Ltd v Severn Trent Water Purification Inc (2013) 1 SCC 641
[8] Cox and Kings Ltd v SAP India (P) Ltd (2022) SCC OnLine SC 570
