AI-Generated Evidence in Indian Courts: Can a Machine Be a Witness? : Author: Khushi Goswami
In September 2025, a case before a Delhi trial and appellate court took an unusual turn when it was discovered that the petitioner had relied on Supreme Court judgments that had never actually been given. The citations looked authentic — case names, party titles, paragraph numbers but on verification, they simply did not exist. They had been generated by an artificial intelligence tool and filed without a second glance
ARTICLE
Introduction
In September 2025, a case before a Delhi trial and appellate court took an unusual turn when it was discovered that the petitioner had relied on Supreme Court judgments that had never actually been given. The citations looked authentic — case names, party titles, paragraph numbers but on verification, they simply did not exist. They had been generated by an artificial intelligence tool and filed without a second glance. The episode was more than an embarrassing lapse; it was a preview of a question Indian law is only beginning to confront. As technology becomes more involved in the legal system, courts are increasingly seeing information produced directly by machines, such as forensic reports based on machine-learning models, deepfake detection results, and predictive risk scores. This raises an important question: how much importance should courts give to information generated by a machine? Should a machine-generated result be treated like a witness's statement, an expert opinion, or simply as a document that still needs a human expert to explain and support it?
Documents, Not Witnesses
Indian evidence law has never recognised a machine as a person capable of testifying. Under the Indian Evidence Act, 1872, now largely replaced by the Bharatiya Sakshya Adhiniyam, 2023 (BSA), testimony, cross-examination, and the oath are based on the idea that a real person gives evidence. That person can be questioned, challenged if they are mistaken, and held responsible if they deliberately give false evidence.
A machine cannot swear an oath or explain, in terms a judge can test, why it reached a conclusion. Whatever an AI system produces is generally treated as a document or electronic record. It is not treated as a witness that can speak or testify on its own.
This means that AI-generated material must be considered under the legal rules that apply to electronic evidence. Previously, the admissibility of electronic records was mainly dealt with under Section 65B of the Indian Evidence Act, 1872. With the enactment of the Bharatiya Sakshya Adhiniyam, 2023 (BSA), these provisions have now been incorporated into Sections 61 to 63. Under the earlier Section 65B(1), information contained in an electronic record could be accepted as evidence when the conditions prescribed under Section 65B(2) were properly satisfied. The BSA carries this forward, with Section 63(4) also requiring the record's hash value, reflecting that authenticity now rests on tamper-evidence, not signature.
The Case Law Foundation
Two Supreme Court cases help us understand how AI-generated material may be treated as evidence.
In Anvar P.V. v. P.K. Basheer, the Supreme Court said that secondary electronic evidence cannot be accepted in court without the certificate required under Section 65B. The same rule was later confirmed by the Constitution Bench in Arjun Panditrao Khotkar v. Kailash Kushanrao Gorantyal. The Court also said that if a person cannot obtain the certificate for a reason, the person may ask the court for help.
Neither of these cases was directly about AI evidence, but the same principle can be applied to AI-generated material. An AI-generated report should not be accepted just because it was created by an AI system. The court should also check how the information was created, stored and kept, and whether the information was properly checked and verified by a person.
The Delhi High Court's 2025 experience with fabricated AI-generated citations illustrates what happens when that verification step is skipped. On discovering that the cited paragraphs and case titles corresponded to no real judgment, the bench permitted the petition's withdrawal and flagged the episode as a caution to the profession - reportedly India's first instance of a filing undone by AI hallucination. Courts examining deepfakes similarly rely on forensic examiners under Section 45A to authenticate what the eye cannot judge.
Illustrative Contexts
Facial recognition and forensic analytics
Police increasingly use facial recognition software to identify suspects from CCTV footage. A probability match is, at most, an investigative lead, not proof of identity - the model cannot be cross-examined on its error rate or training data. Similarly, forensic laboratories using algorithmic tools to interpret DNA mixtures produce reports where the human analyst who operated the software, not the machine, becomes the witness, echoing the older position on computer-generated records taken in State (NCT of Delhi) v. Mohd. Afzal.
AI-drafted legal submissions
The Delhi episode shows that lawyers using AI to draft pleadings risk invented, plausible sounding citations. Courts are telling lawyers that counsel must fully verify work before filing.
Can a Machine Be an Expert Witness?
Section 45 of the Evidence Act (Section 39 of the BSA) allows a specially skilled person's opinion as expert evidence. An AI system is not a 'person' and cannot hold this status. Courts can accept the opinion of a human expert who used AI tools, provided that expert can be cross-examined on the methodology and the model's limitations. The 'black box' problem — where even designers cannot fully explain an output — is a genuine obstacle, since cross-examination is how Indian courts test reliability, and a black box offers nothing to interrogate.
The Road Ahead
Indian law has absorbed AI-generated material into the existing electronic-record framework rather than creating a bespoke regime. This works for static outputs a human custodian can certify, but less well for opaque, self-updating systems with no auditable chain of reasoning. Courts may eventually need supplementary safeguards — mandatory disclosure of the model and version used, independent audit trails, and certification tailored to algorithmic outputs. Until legislation catches up, the safest principle remains the one courts have applied since Anvar P.V.: an algorithmic output is only as trustworthy as the human chain of custody and explanation standing behind it.
Conclusion
On the present state of Indian law, a machine cannot be a witness. The more pressing question is how much unexamined trust courts should place in what a machine produces. The 2025 fabricated-citation episode and the continuing caution around facial recognition and deepfakes suggest Indian courts are moving in the right direction — treating AI output as a document requiring certification and a human sponsor, not as an autonomous source of truth.
References
1. The Indian Evidence Act, 1872, ss. 45, 45A, 65, 65A, 65B.
2. The Bharatiya Sakshya Adhiniyam, 2023, ss. 39, 61-63.
3. Anvar P.V. v. P.K. Basheer, (2014) 10 SCC 473.
4. Arjun Panditrao Khotkar v. Kailash Kushanrao Gorantyal, (2020) 7 SCC 1.
5. State (NCT of Delhi) v. Mohd. Afzal, 107 (2003) DLT 385 (Del.).
6. Greenopolis Welfare Association matter (Delhi High Court, September 2025) — reported in legal news coverage on AI-generated fabricated citations and withdrawal of the petition.
7. LiveLaw, "AI In Digital Forensics: Are Indian Evidence Laws Equipped To Handle Machine-Generated Proof?", available at livelaw.in.
8. LiveLaw, "When Artificial Intelligence Becomes Witness: Who Bears Responsibility For AI-Generated Evidence?", available at livelaw.in.
9. iPleaders Blog, "Admissibility of E-Evidence in Indian Courts" and "Supreme Court's Judgment on Requirements of Certificate under Section 65B, Indian Evidence Act, 1872", available at blog.ipleaders.in.