“Seeing Is No Longer Believing: How AI and Deepfakes Are Changing Evidence in Indian Criminal Courts”
The New Evidentiary Battle: When a Computer Becomes the Witness
There was a time when a photograph, a video recording or a voice recording carried an almost instinctive credibility.
If the camera showed it, people assumed it happened.
If the voice sounded like the accused, people assumed the accused spoke.
If a video appeared to place a person at a crime scene, the natural reaction was to ask what the person was doing there—not whether the video itself was genuine.
That assumption is no longer safe.
Artificial intelligence has changed the problem fundamentally.
A modern criminal trial may involve a video that never happened, a voice that was never spoken, a photograph of a person who was never present, a fabricated WhatsApp conversation, an AI-generated confession, or a facial-recognition system that gives the investigating agency a statistical match.
These are not all the same evidentiary problem.
That distinction, in my view, is going to become one of the most important questions in Indian criminal jurisprudence over the next decade.
The real question is no longer merely:
“Is this electronic evidence?”
It is:
“What exactly produced this evidence, what does the certificate prove, what does the technology prove, and what remains for the human witness and the court to establish?”
The distinction is critical because the authenticity of a digital file is not necessarily the authenticity of the event depicted in that file.
That is where the Bharatiya Sakshya Adhiniyam, 2023 (“BSA”), meets artificial intelligence.
“In the age of AI, a picture can be perfect, a voice can be flawless, and a video can be convincing—yet none of them can tell a court the truth. Technology can manufacture what we see; only evidence can prove what happened. When liberty is at stake, never let a machine’s confidence become a judge’s certainty.”
— Adv. Tarun Choudhury
Supreme Court Advocate | 25+ Years of Legal Experience
1. Two Completely Different Kinds of AI Evidence
The expression “AI evidence” is dangerously broad.
For legal purposes, I would divide it into at least two fundamentally different categories.
Also Read: AI-Generated Fake Nude Images: Minnesota vs. xAI and the Future of Deepfake Law
Category One: Synthetic Evidence Purporting to Show a Real Event
This includes:
- AI-generated or manipulated videos;
- deepfake videos;
- cloned voices;
- fabricated photographs;
- AI-created WhatsApp or Telegram conversations;
- fabricated emails;
- manipulated CCTV footage;
- synthetic confessions;
- AI-generated documents;
- images placing an accused at a location where he was never present.
Here, the principal danger is fabrication.
The exhibit itself purports to be evidence of a historical fact.
But the real question is whether that historical fact ever occurred.
A perfectly preserved deepfake remains a deepfake.
A hash value cannot transform a fictional event into a real event.
The Second Category: AI Used as an Investigative or Forensic Tool
The second category is entirely different.
Here, the underlying evidence may be genuine, but an AI system is used to analyse it.
Examples include:
- facial-recognition systems;
- speaker-identification systems;
- automated transcription;
- image enhancement;
- pattern-recognition software;
- predictive analytical systems;
- transaction anomaly detection;
- automated document analysis;
- reconstruction of events;
- forensic classification systems.
The danger here is not necessarily fabrication of the underlying file.
The danger is opacity and reliability.
A human expert can enter the witness box.
The expert can explain:
- what methodology was used;
- what material was examined;
- what assumptions were made;
- what error rate exists;
- what limitations apply;
- whether alternative explanations were considered.
A machine cannot itself enter the witness box.
Therefore, when the State places substantial reliance on an AI-generated conclusion, the Court must know what the system actually did and how reliable that process is.
This is a different evidentiary question from whether an electronic file has been properly certified.
2. What Does the Bharatiya Sakshya Adhiniyam Actually Require?
The BSA came into force on 1 July 2024 and replaced the Indian Evidence Act, 1872. The India Code places electronic or digital records specifically within Sections 61 to 63, while Section 39 deals with expert opinion.
The architecture is important.
| Provision | Subject |
|---|---|
| Section 61 | Electronic or Digital Record |
| Section 62 | Special Provisions |
| Section 63 | Admissibility of Electronic Records |
| Section 39 | Expert Opinion |
Section 61 — Electronic or Digital Record
Section 61 establishes that an electronic or digital record cannot be denied legal effect merely because it is electronic or digital, subject to the provisions governing admissibility.
Section 62 — Special Provisions
Section 62 directs attention to the statutory framework dealing with proof of electronic records.
Section 63 — Admissibility of Electronic Records
Section 63 is therefore central when a party seeks to rely upon electronic records through the statutory route for electronic evidence.
The modern Indian evidentiary question is consequently not:
“Does the Court recognise digital evidence?”
It plainly does.
The real question is:
“Has the statutory foundation required for this particular electronic record been established, and what exactly has been proved by that foundation?”
3. The Supreme Court’s Section 65B Jurisprudence Still Matters
Although the Evidence Act has been replaced, the Supreme Court’s jurisprudence under Section 65B remains enormously important for understanding the philosophy behind the new statutory framework.
In Anvar P.V. v. P.K. Basheer, (2014) 10 SCC 473, the Supreme Court treated the statutory certification requirement as a condition governing admission of secondary electronic evidence.
That principle was subsequently examined by the Constitution Bench in:
Arjun Panditrao Khotkar v. Kailash Kushanrao Gorantyal, (2020) 7 SCC 1.
The Constitution Bench reaffirmed the central importance of the statutory certificate and rejected the proposition that the certificate could simply be dispensed with as a matter of convenience.
The Supreme Court’s judgement in Arjun Panditrao is an important primary authority for the proposition that the statutory requirements concerning electronic evidence cannot casually be bypassed.
The significance of this history is obvious.
Indian law did not treat electronic evidence as ordinary paper evidence.
It recognised that digital material can be copied, transferred, altered and reproduced without leaving the sort of physical trace that traditional documentary evidence often leaves.
The BSA has carried that concern into the new statutory regime.
4. The 2026 Supreme Court Decision: Pune Bar Association v. Union of India
The most significant recent development is the Supreme Court’s decision in:
Pune Bar Association v. Union of India
W.P. (C) No. 599 of 2026, order dated 22 May 2026
The case directly concerned Section 63(4) of the BSA and the Schedule prescribing the certificate for electronic records.
The challenge questioned, among other things, the requirement concerning:
- disclosure of the hash value; and
- certification by an expert.
The Supreme Court rejected the constitutional challenge.
The Court explained that the hash value operates as an electronic fingerprint and serves the objective of verifying the authenticity and integrity of the electronic data.
The Court also considered the question of who may qualify as an expert for Part B of the certificate. Importantly, the Court did not hold that only persons notified under Section 79A of the Information Technology Act can necessarily provide the relevant expertise. It referred to the broader language of Section 39(1) of the BSA concerning persons possessing special skill.
The Supreme Court’s official cause-list material confirms that the matter was before a bench comprising the Chief Justice of India and Justices Joymalya Bagchi and Vipul M. Pancholi on 22 May 2026.
A Necessary Legal Caution
The supplied material describes the decision as though the Court definitively established a broad new evidentiary doctrine for AI and deepfakes.
That is too wide.
The decision principally concerned the constitutional challenge to Section 63(4) and the certification architecture.
It should not be read as a Supreme Court declaration that a Section 63 certificate itself establishes the truthfulness of everything depicted in an electronic recording.
That distinction is fundamental.
5. The Hash Proves Integrity — Not Necessarily Truth
This is perhaps the most important proposition for lawyers and judges dealing with AI-generated evidence.
Suppose the prosecution produces a video file.
The file has a valid hash.
The hash matches the value recorded when the evidence was acquired.
What does that establish?
It may establish that the file has not changed since the relevant hash was calculated.
That is extremely important.
But it does not necessarily establish:
- who created the original video;
- whether the original video recorded a genuine event;
- whether the scene was staged;
- whether the video was AI-generated;
- whether persons or objects were digitally inserted;
- whether the recording was manipulated before it came into the possession of the investigating agency.
In simple terms:
Integrity of the copy is not the same thing as authenticity of the event.
Imagine a perfectly preserved photograph of a nonexistent event.
The photograph may be forensically intact.
Its hash may be impeccable.
Its chain of custody may be perfect.
But the event depicted may still be fictional.
That is the central evidentiary challenge created by generative AI.
Also Read: Can America Reform the Supreme Court? Latest 5–4 Ruling Reignites Term Limits Debate
6. Admissibility Is Not Conviction
This distinction must be kept alive throughout every criminal trial involving AI-generated evidence.
There are at least three separate questions.
Question One: Can the Court Receive the Electronic Record?
This is an admissibility question.
Question Two: Is the Electronic Record Authentic?
This is an authentication question.
Question Three: Does the Evidence Prove the Prosecution’s Case Beyond Reasonable Doubt?
This is a weight and proof question.
They are not interchangeable.
A certificate may assist with the first question.
It may provide important evidence relevant to the second.
But it does not automatically answer the third.
A criminal court must ultimately determine whether the prosecution has proved the ingredients of the offence against the particular accused beyond reasonable doubt.
That constitutional and evidentiary burden cannot be outsourced to a certificate.
7. A Deepfake Should Not Become Admissible Merely Because It Is a Perfect Deepfake
This proposition deserves emphasis.
A sophisticated deepfake may satisfy the technical requirements applicable to an electronic record.
That does not make the underlying representation true.
The Court must distinguish between:
“This is an authentic copy of the file that was seized.”
and
“This file is an authentic representation of what actually happened.”
The two statements are entirely different.
The first is principally a question of digital integrity.
The second is a question of historical truth.
That difference may become decisive in criminal trials.
8. Section 39 BSA and the Role of Experts
The BSA separately recognises expert opinion.
The India Code identifies Section 39 as the provision concerning opinions of experts.
That provision creates an important avenue for dealing with AI.
Suppose the prosecution says:
“Our facial-recognition system has identified the accused.”
The Court should not automatically treat that sentence as equivalent to:
“An eyewitness has identified the accused.”
They are fundamentally different propositions.
The Court may need to ask:
- What algorithm was used?
- What version of the system was deployed?
- What input image was provided?
- Was the image of sufficient quality?
- What database was searched?
- What error rates exist?
- Was the system independently validated?
- How does it perform with Indian faces and comparable conditions?
- What false-positive rate exists?
- Was there human verification?
- Was the operator permitted to override the result?
- Were alternative candidates generated?
- Were investigators told the identity before examining the result?
- Could confirmation bias have affected the interpretation?
These questions are not technological decoration.
They go directly to reliability.
9. Facial Recognition Is Not an Eyewitness
A facial-recognition score should not be described casually as an “identification”.
A better legal description may be:
“The system generated a similarity or identification result under specified conditions.”
The difference matters.
The machine did not necessarily “see” the accused in the legal sense.
It processed data according to an algorithm.
The Court must decide what evidentiary significance that output deserves.
The greater the prosecution’s dependence on the algorithm, the greater the justification for disclosure sufficient to permit meaningful adversarial testing.
10. Article 21 Adds a Constitutional Dimension
This is not merely a debate about computer science.
A criminal prosecution ultimately concerns liberty.
Article 21 protects life and personal liberty except according to procedure established by law.
If an AI system materially contributes to the deprivation of liberty, the accused must have a meaningful opportunity to challenge the evidentiary basis of that system.
That does not necessarily mean that every proprietary algorithm must automatically be placed completely in the public domain.
But it does mean that “the computer says so” cannot become the legal equivalent of cross-examination.
The Court must remain the decision-maker.
The expert may assist the court.
The algorithm may assist the expert.
But neither should replace the judicial function.
11. The Delhi High Court and the Deepfake Problem
The Delhi High Court has already confronted the practical problem of deepfakes in Nirmaan Malhotra v. Tushita Kaul.
The case concerned photographs relied upon in matrimonial litigation. The Court took note of the increasing prevalence of deepfake technology and declined to treat the photographs as conclusively establishing the fact asserted through them without proper proof.
The significance of the decision is larger than the matrimonial dispute in which it arose.
A photograph is no longer necessarily a passive record.
It can be a manufactured representation.
That should change the courtroom instinct from:
“I can see it; therefore, it happened.”
to:
“I can see it; now prove to me that what I see corresponds to reality.”
12. India’s 2026 IT Rules: Important, But Different
India has also moved against synthetic media through the Information Technology (Intermediary Guidelines and Digital Media Ethics Code) Amendment Rules, 2026.
MeitY’s official materials confirm that the amendments were notified on 10 February 2026 and came into force on 20 February 2026. They introduce the concept of synthetically generated information (SGI) and impose additional obligations concerning unlawful synthetic content, labelling and provenance-related mechanisms.
The Government’s explanatory material expressly describes the new SGI framework and its obligations for intermediaries.
This is important.
But it must not be confused with the BSA.
The IT Rules primarily regulate the creation, dissemination and handling of synthetic content in the digital ecosystem.
The BSA governs the evidentiary treatment of electronic records in legal proceedings.
They address different stages of the problem.
One operates largely upstream.
The other operates at the courtroom door.
13. A Watermark Is Not a Certificate of Truth
Suppose an AI platform attaches an “AI-generated” label.
That is useful.
It provides provenance information.
It may alert the viewer that the material is synthetic.
But the label does not tell the Court whether the underlying proposition is true or false.
Similarly, the absence of a label does not automatically prove authenticity.
Technology can tell us something about the origin or processing history of a file.
It cannot, without more, answer every historical question contained within that file.
14. What the United States Is Considering
The United States provides an interesting comparison.
The Federal Rules of Evidence already contain general authentication and expert-evidence principles.
The U.S. Federal Advisory Committee on Evidence Rules has considered amendments specifically addressing deepfakes and machine-generated evidence.
The Committee’s May 2026 materials discuss a possible Rule 901(c) for potentially fabricated evidence created through generative AI.
The proposed approach would require the party challenging an item as a deepfake first to present sufficient evidence to warrant an inquiry; if that threshold is met, the proponent would then have to establish authenticity by a preponderance of the evidence.
This is important precisely because it is not yet settled federal law.
It is a proposal under consideration.
That distinction matters.
The American experience nevertheless illustrates an important procedural idea:
The Court should neither accept every deepfake allegation nor dismiss a credible deepfake challenge merely because the exhibit is accompanied by technical certification.
There must be a disciplined procedure.
15. Singapore Has Taken a Different Route
Singapore’s courts have issued a formal guide on the use of generative artificial intelligence tools by court users.
The Guide applies across the Supreme Court, state courts and family justice courts.
It permits appropriate use of generative AI but emphasises that court users remain responsible for ensuring that material placed before the Court is independently verified, accurate and appropriate.
The broader lesson is significant.
The solution is not necessarily to prohibit AI.
The solution is to preserve human responsibility for what is placed before a court.
That principle deserves serious consideration in India.
16. New South Wales Shows How Strict Courts Can Become
The Supreme Court of New South Wales has issued Practice Note SC Gen 23 dealing with generative AI.
The judicial guidance is particularly concerned with confidentiality, suppression orders, information produced under compulsion and the risks associated with feeding sensitive litigation material into AI systems.
The NSW experience demonstrates something important for Indian courts:
AI regulation inside the justice system does not have to be limited to evidence admissibility.
Courts can regulate:
- how AI is used in affidavits;
- how AI is used in submissions;
- how confidential information is handled;
- what lawyers must disclose;
- how experts use AI;
- how parties verify AI-generated material.
That is a procedural governance question, not merely an evidentiary one.
17. The European Union: Transparency Before Litigation
The EU AI Act adopts another approach.
Article 50 contains transparency obligations concerning AI-generated and manipulated content, including deepfakes.
The European framework requires certain AI-generated or manipulated content to be identifiable as artificially generated or manipulated, subject to the statutory qualifications and exceptions.
The European approach therefore seeks to create information about provenance before the material becomes courtroom evidence.
That is a valuable policy idea for India.
The strongest evidence system may ultimately be one in which provenance begins at creation rather than being reconstructed after a criminal investigation.
18. The Four-Layer Test India Should Consider
In my view, Indian courts should gradually move towards a four-layer approach when AI materially affects an electronic exhibit.
Layer One — Digital Integrity
Ask:
Has the file remained unchanged since acquisition?
This is where hash values, forensic imaging and chain-of-custody records become important.
Layer Two — Origin and Provenance
Ask:
- Where did the file actually come from?
- Who created it?
- On what device?
- When?
- Using what application?
- Was it directly recorded or generated?
- Was it subsequently exported, compressed or processed?
Layer Three — Technological Reliability
Ask:
If AI generated, reconstructed or analysed the material, how reliable was the system?
The Court should consider:
- methodology;
- validation;
- error rates;
- known limitations;
- data quality;
- human supervision;
- reproducibility;
- alternative explanations.
Layer Four — Evidentiary Corroboration
Finally ask:
Does independent evidence support the proposition for which the AI material is tendered?
For example:
- call records;
- location evidence;
- independent witnesses;
- financial transactions;
- CCTV from another source;
- contemporaneous documents;
- device metadata;
- admissions;
- forensic evidence.
The greater the AI dependence, the more important independent corroboration becomes.
20. A Practical Protocol for Prosecutors
If I were advising an investigating agency today, I would recommend a much more disciplined protocol.
Step 1: Preserve the Original Device
Do not begin by repeatedly opening, forwarding or editing the file.
Step 2: Forensically Acquire the Data
Create a proper forensic image wherever appropriate.
Step 3: Calculate and Record the Hash
Record the hash at acquisition.
Step 4: Maintain Chain of Custody
Every transfer should be documented.
Step 5: Preserve Metadata
Do not unnecessarily convert or “clean” the original file.
Step 6: Identify Whether AI Was Involved
The investigating officer should specifically investigate whether the material was:
- directly captured;
- edited;
- enhanced;
- reconstructed;
- generated;
- translated;
- transcribed;
- classified;
- matched by an AI system.
Step 7: Preserve Relevant AI Information
Where AI materially generated the evidentiary conclusion, preserve information concerning:
- system/version;
- input;
- output;
- relevant settings;
- operator;
- methodology;
- validation;
- error rates, where available.
Step 8: Establish the Human Facts
Who recorded the material?
Where?
When?
Under what circumstances?
Who possessed the device?
Who had access?
Was anyone capable of manipulating it?
Step 9: Corroborate
Do not build an entire prosecution around the visual attractiveness of an AI-generated exhibit.
21. The Defence Strategy Must Also Change
The defence lawyer should not simply say:
“This is a deepfake.”
That is too easy.
It should be converted into a structured evidentiary challenge.
Questions the Defence Should Ask
- Where is the original device?
- Who seized it?
- What was the acquisition process?
- What was the original hash?
- When was the hash calculated?
- Does the later hash match?
- Who handled the file?
- Was it exported?
- Was it transcoded?
- What metadata survives?
- What software was used?
- Was AI used?
- Was the material enhanced?
- Was the AI system independently validated?
- What is its known error rate?
- What alternative matches were produced?
- Was the operator aware of the suspect’s identity before analysis?
- Is there independent evidence corroborating the alleged event?
The objective is not to shout “deepfake”.
The objective is to expose the evidentiary gap.
22. The Prosecution Must Not Confuse Vividness With Reliability
AI-generated evidence has a psychological advantage.
Humans tend to trust vivid information.
A video appears more powerful than a paragraph.
A realistic voice appears more persuasive than a witness describing the conversation.
A photograph appears stronger than an allegation.
That psychological instinct can become dangerous in criminal trials.
The more realistic the synthetic material becomes, the less appropriate it is for the Court to rely upon human intuition alone.
The Court must become more forensic, not more impressionistic.
23. The “Liar’s Dividend” Is the Other Side of the Problem
There is a danger in going too far in the opposite direction.
If everyone knows that AI can fabricate almost anything, a genuine recording can be dismissed simply by saying:
“It may be AI-generated.”
That is also dangerous.
A genuine video should not become worthless merely because deepfake technology exists.
The answer, therefore, cannot be:
“Everything may be fake.”
The answer must be:
“A credible challenge to authenticity must be tested through evidence.”
This is why a disciplined threshold for raising a deepfake challenge may eventually become necessary.
The American proposals provide one possible model, although they remain proposals rather than enacted federal rules.
24. Should India Create a Special AI Evidence Hearing?
I believe it should eventually consider doing so.
Not necessarily through an entirely new statute.
The existing BSA already contains several building blocks:
- electronic-record provisions;
- expert opinion;
- rules governing proof;
- judicial evaluation of evidence.
But courts could develop a structured AI reliability hearing through procedural rules or authoritative judicial guidance.
What Could an AI Reliability Hearing Determine?
| Question | Issue to Be Determined |
|---|---|
| A | Whether the material is AI-generated or AI-assisted |
| B | Whether the technology has been reliably validated |
| C | Whether the output is reproducible |
| D | What known error rates exist |
| E | Whether the expert understands the limitations |
| F | Whether the defence has sufficient material to test the conclusion |
| G | Whether independent evidence supports the result |
This would not mean that every CCTV recording requires a scientific trial.
It would mean that machine-generated conclusions presented as proof of a disputed fact receive appropriate scrutiny.
25. What India Should Not Do
India should avoid two extremes.
Extreme One: Blind Technological Trust
“The computer generated it; therefore, it is objective.”
This is dangerous.
Algorithms can contain:
- biased data;
- flawed assumptions;
- inadequate training;
- implementation errors;
- false positives;
- false negatives.
Extreme Two: Technological Nihilism
“AI can create fakes; therefore, digital evidence cannot be trusted.”
That is equally wrong.
Digital evidence can be extraordinarily powerful.
The solution is not distrust of technology.
The solution is disciplined verification.
26. A Proposed Judicial Test for AI-Generated Evidence
I would propose that, when AI materially affects evidence, the Court ask seven questions:
| Test | Question |
|---|---|
| Test 1 | What is the original source? |
| Test 2 | Has the digital object remained intact? |
| Test 3 | Was AI involved in creation, modification or analysis? |
| Test 4 | If AI was involved, what exactly did it do? |
| Test 5 | Is the methodology scientifically and technically reliable for this use? |
| Test 6 | Can the defence meaningfully challenge the process? |
| Test 7 | Is there independent evidence supporting the conclusion? |
If these seven questions cannot be answered, the Court should exercise considerable caution before assigning substantial evidentiary weight to the material.
27. The Most Important Distinction: File Authenticity vs Event Authenticity
This deserves to become part of Indian courtroom vocabulary.
File Authenticity
Means:
This is the same digital object that was acquired or produced.
Event Authenticity
Means:
The digital representation accurately depicts something that actually occurred.
A hash can be powerful evidence concerning the first.
It cannot automatically establish the second.
That is why the phrase “electronic evidence is authentic” is sometimes legally incomplete.
The Court must ask:
Authentic in what sense?
28. Can AI-Generated Evidence Convict an Accused in India?
The answer is yes—but only as part of a legally sufficient body of evidence.
There is no sound basis for saying that AI-generated evidence is automatically inadmissible merely because AI was involved.
Equally, there is no sound basis for saying that an AI-generated exhibit is automatically reliable merely because it satisfies the statutory requirements applicable to electronic records.
The prosecution must still establish the facts necessary for conviction.
The Court must assess the reliability and weight of the evidence.
The accused must receive a fair opportunity to challenge material relied upon against him.
And where the AI output is central to the prosecution case, the Court should demand considerably more than visual persuasion.
29. The Future Criminal Trial May Have Three Witnesses
There is an interesting possibility ahead.
Traditionally, a criminal case may involve:
Witness → Evidence → Judge
With AI-assisted evidence, the structure may become:
Human witness → Digital evidence → AI system → Expert → Judge
That creates a new danger.
The further evidence travels from an observable human event, the greater the need for transparency.
A judge should therefore be cautious when the prosecution’s chain looks like this:
AI-generated material → AI analysis → expert interpretation → conviction.
Every additional technological layer introduces another possible point of failure.
30. The Supreme Court’s Pune Bar Association Decision Is a Beginning, Not the End
The 2026 Pune Bar Association decision is important because it confirms that the BSA’s stricter certification architecture is not, merely because it requires hash disclosure and expert certification, constitutionally arbitrary.
But it should not be misunderstood as answering every future question concerning AI evidence.
It does not eliminate questions of:
- authenticity;
- reliability;
- provenance;
- manipulation;
- scientific validity;
- expert methodology;
- constitutional fairness;
- evidentiary weight.
The court has strengthened the gate.
It has not removed the judge from the courtroom.
31. My Proposed Indian Framework
In my considered view, India should now move towards a three-part statutory and judicial framework.
Part I — Preserve the Existing BSA Certification Structure
Section 63 should continue to ensure integrity and proper certification of electronic records.
Part II — Introduce AI-Specific Disclosure
Where the prosecution or a party relies materially upon AI generation, alteration or analysis, the opposing side should ordinarily be informed.
Part III — Introduce Reliability Review
Where AI materially contributes to proof of a disputed fact, the Court should have power to require disclosure sufficient to permit meaningful testing of:
- methodology;
- validation;
- limitations;
- error rates;
- relevant inputs;
- human intervention.
This would not require the court to become a computer scientist.
It requires the court to remain a court.
32. The Real Constitutional Principle
The deepest principle is not technological.
It is constitutional.
A criminal court is not supposed to determine guilt by asking:
“Which image looks most convincing?”
It is supposed to determine guilt after evaluating admissible evidence, testing its reliability, hearing the accused and applying the criminal standard of proof.
Artificial intelligence changes the means by which evidence can be created.
It does not change the burden of proof.
It does not reverse the presumption of innocence.
It does not eliminate cross-examination.
It does not convert probability into certainty.
And it certainly does not transfer the judicial function from the judge to an algorithm.
Conclusion: In the Age of AI, Seeing Is No Longer Believing
The law is entering an evidentiary era in which the most convincing exhibit may be the least trustworthy.
The problem is not that machines lie.
The problem is that machines can now produce representations that look indistinguishable from truth.
That is why the Indian legal system must make one distinction absolutely clear:
A certificate may establish the integrity of an electronic record. It does not automatically establish the truth of the event represented by that record.
The BSA has taken an important step by strengthening the certification architecture for electronic evidence.
The Supreme Court’s decision in Pune Bar Association v. Union of India, W.P. (C) No. 599 of 2026, order dated 22 May 2026, reinforces the legitimacy of that architecture.
But the next battle will be harder.
The question will not simply be:
“Has this file been altered?”
It will be:
“Was the event ever real?”
And where artificial intelligence has been used to generate, reconstruct, identify or interpret evidence, the Court must ask another question:
“Why should I trust the machine?”
That question should never be treated as an attack on technology.
It is an affirmation of judicial responsibility.
The future of criminal justice cannot be:
“The algorithm says so.”
It must remain:
“Show me the evidence, explain the method, allow the other side to test it, and then persuade the Court.”
That is how the rule of law survives the age of deepfakes.
— Adv. Ameeta Sinha
Calcutta High Court Advocate
References
- BSA: The India Code confirms Sections 39, 61, 62 and 63 as the relevant provisions dealing respectively with expert opinions and electronic/digital records.
- Anvar / Arjun Panditrao: The Supreme Court’s Arjun Panditrao judgement confirms the importance of the statutory certification regime under the former Section 65B.
- Pune Bar Association: The Supreme Court’s 22 May 2026 order confirms the constitutional challenge to Section 63(4), the importance of the hash value, and the discussion of expertise under Section 39.
- Important correction: I have not treated the Pune Bar Association as creating a general AI-evidence admissibility rule. The primary order concerns Section 63(4) and its certificate architecture; the broader question of AI reliability remains distinct.
- 2026 IT Rules: MeitY’s official notification confirms the 10 February 2026 amendment and 20 February 2026 commencement, including the SGI framework.
- United States: The proposed deepfake Rule 901(c) remains a proposal under consideration, not enacted federal evidence law. The May 2026 Advisory Committee materials expressly describe it as a proposed amendment.
- Singapore: The official Singapore Courts guide confirms that court users may use generative AI subject to verification and continuing responsibility for accuracy and appropriateness.
- EU: Article 50 of the EU AI Act contains transparency requirements concerning deepfakes and AI-generated/manipulated content, subject to its statutory exceptions.
- New South Wales: Official NSW judicial material confirms the existence and operation of Practice Note SC Gen 23 concerning generative AI in court proceedings.
Frequently Asked Questions (FAQs)
Based directly on the article, here are 5 SEO-rich FAQs designed to capture searches around AI-generated evidence in Indian courts, deepfake evidence, BSA Section 63, and electronic evidence:
1. Can AI-generated evidence be admitted in Indian courts?
Yes. AI-generated evidence is not automatically inadmissible merely because artificial intelligence was used. Electronic records must satisfy the applicable requirements of the Bharatiya Sakshya Adhiniyam, 2023 (BSA), particularly the statutory framework under Sections 61–63. However, admissibility does not automatically establish that the contents of an AI-generated recording are truthful or reliable.
2. Is deepfake video evidence admissible under the Bharatiya Sakshya Adhiniyam, 2023?
A deepfake video is an electronic record, but its admissibility and evidentiary weight are separate questions. Compliance with Section 63 BSA, including the applicable certification requirements, may establish the necessary foundation for the electronic record. It does not by itself prove that the event depicted in the deepfake actually occurred. Authenticity, provenance and corroboration remain crucial.
3. Does a Section 63 BSA certificate prove that an electronic video is genuine?
No. A Section 63 BSA certificate primarily addresses the statutory requirements for proving an electronic record and its integrity. A hash value can help establish that a particular digital file has not been altered after acquisition or certification, but it does not automatically prove that the underlying event shown in the video was real. Digital integrity is not necessarily the same as event authenticity.
4. Can facial recognition or other AI forensic evidence be used to convict an accused in India?
AI forensic evidence, including facial-recognition results, speaker identification and automated analysis, may form part of the evidence in a criminal trial. However, an AI-generated score should not automatically be treated as equivalent to human eyewitness identification. Courts should consider the technology’s methodology, reliability, limitations, error rates, human oversight and independent corroborating evidence before assigning substantial evidentiary weight.
5. What should Indian courts consider when dealing with AI-generated evidence and deepfakes?
Indian courts should distinguish between AI-generated evidence and AI-assisted analysis of genuine evidence. Important considerations include the original source, chain of custody, hash value, metadata, Section 63 BSA certification, whether AI was involved in creating or modifying the material, the reliability of the AI system, expert evidence, the opportunity for the defence to challenge the technology, and independent corroboration. Ultimately, AI cannot replace the judicial assessment of proof beyond reasonable doubt.
Key Takeaways: AI-Generated Evidence in Indian Courts
- AI-generated evidence in Indian courts is not automatically inadmissible merely because artificial intelligence was used to create or analyse it.
- Under the Bharatiya Sakshya Adhiniyam, 2023 (BSA), electronic and digital records are governed principally by Sections 61–63, with Section 63 providing the statutory framework for proving electronic records.
- Section 63 BSA certification does not automatically prove that the event depicted in a video, photograph or audio recording actually occurred. Digital-file integrity and factual authenticity are two different questions.
- A hash value can help establish the integrity of an electronic file after acquisition, but it cannot by itself establish that the original recording was genuine or that the event shown actually happened.
- Deepfake evidence creates a unique legal challenge because AI can manufacture realistic videos, photographs, voices, chats and other material that may appear genuine to the human eye or ear.
- Courts should distinguish between AI-generated evidence and AI-assisted forensic evidence, such as facial recognition, speaker identification, automated transcription and algorithmic analysis.
- AI forensic results should not automatically be treated as equivalent to eyewitness evidence. Courts should examine methodology, validation, error rates, limitations, human oversight and corroborating evidence.
- Expert evidence under Section 39 BSA can become increasingly important where specialised AI or forensic technology has materially influenced the prosecution’s case.
- Admissibility is not the same as proof. Even when an electronic record satisfies the statutory requirements for admission, the Court must independently determine its authenticity, reliability and evidentiary weight.
- Article 21 and the right to a fair trial become particularly important where an AI-generated or AI-assisted system materially contributes to the deprivation of an accused person’s liberty.
- Indian courts should increasingly examine provenance, metadata, chain of custody, original devices, forensic imaging and hash values when the authenticity of digital evidence is seriously challenged.
- A mere assertion that “this could be a deepfake” should not be enough to destroy genuine evidence. A credible challenge should be supported by circumstances requiring proper forensic examination.
- The emerging international approach shows different solutions: reliability scrutiny in the United States, transparency requirements in the European Union, judicial AI guidance in Singapore and Australia, and upstream synthetic-content regulation in several jurisdictions.
- India’s 2026 IT Rules concerning synthetically generated information address the broader digital ecosystem, but they should not be confused with the evidentiary requirements of the BSA.
- The central legal principle is simple: a perfectly preserved digital file can still contain a completely fabricated event.
