The Punjab and Haryana High Court has dismissed a third bail petition in which counsel asserted that a phone video of the deceased victim's statement was fabricated or AI-generated, without deciding the point and leaving authenticity to be tested at trial

What the petitioner argued

This was the third bail petition by the same petitioner under section 483 of the Bharatiya Nagarik Suraksha Sanhita, 2023, in a case arising out of FIR No. 18 dated 18 January 2025 under sections 110 and 3(5) of the Bharatiya Nyaya Sanhita, 2023, with sections 103 and 238 added later.

On the prosecution case as the order records it, the victim was set ablaze on 16 January 2025 after a quarrel over her not giving birth to a son, and died during treatment on 15 February 2025. The complainant, her elder sister, recorded a statement from her on a mobile phone, and that recording forms part of the challan report.

Counsel for the petitioner advanced several points together, and the AI point was one item in a list, not the centrepiece. He was falsely implicated. No specific overt act was attributed to him beyond coming forward to assault the victim before being pushed aside. He had undergone heart surgery in 2024 and was in poor health. The recording was not believable because it appeared to have been prepared without the assistance or presence of any medical staff, and because the deceased had been given heavy medication including sedatives, making it highly improbable that she was in a fit mental state to record anything. Then the sentence that brought this order into our window: "The video recording is fabricated or AI-generated one."

The order records no expert report, no forensic examination, no application for one, and no material of any kind offered in support of that sentence.

What the court did with the assertion, which was very little

The court's reasoning on the recording occupies one sentence, and the sentence is about procedure rather than about the video. Having noted that the respondent relied on the recording of the victim's statement taken by the complainant on her phone, the judge wrote: "Even though the authenticity of the same has to be tested during trial, nonetheless, the allegations leveled against the petitioner prima facie establish his complicity in the crime."

That is a deferral, not a ruling. The court did not find that the video is genuine. It did not find that the video is AI-generated. It did not say what would be needed to make good such an assertion, did not direct any forensic examination, and did not address the sedation argument or the absence of medical staff at all.

What carried the refusal was elsewhere. The court held that although the act of igniting the fire might not be attributed to the petitioner, both his presence at the spot and the fact that he assaulted the victim stand prima facie established from the allegations levelled in the FIR. It added that the allegations are serious, that there is nothing on record to suggest undue delay in concluding the trial, and that taking into account the gravity of the allegations, the quantum of sentence a conviction may entail and the attendant facts and circumstances, the petitioner does not deserve bail.

Read in sequence, the AI assertion did not fail on its merits. It was simply routed to the forum that decides authenticity, and the bail question was answered on other material.

Two things this order is not

It is not a holding about AI-generated evidence, and it is not a judicial finding that the video is authentic. Nothing in it states a test, allocates a burden, or sets a standard for challenging a recording as synthetic. Anyone citing it for a proposition about AI evidence would be relying on a sentence whose only decision is that the question is not decided here but at trial, and the order preserves that challenge expressly. If the recording is contested at trial, this order does not foreclose it.

It is not, on its own face, marked for reporting. The concluding lines of the order carry the court's standard template entries, "Whether speaking/reasoned:- Yes/No" and "Whether reportable:- Yes/No", with both options still present in the text we read and neither struck out. We therefore do not describe it as a reportable judgment.

A vocabulary point worth getting right

The order does not use the phrase dying declaration anywhere. It describes a statement recorded by the complainant on her own mobile phone, forming part of the challan report, made by a victim who later died. Whether that recording is treated as a dying declaration is exactly the sort of question the trial court will reach, and this order does not reach it.

The distinction matters for anyone reading across jurisdictions. A statement recorded by a magistrate or by a doctor sits on a different evidentiary footing from a family member's phone video, and the safeguards argument counsel made, that no medical staff were present, is aimed squarely at that gap.

For defence practitioners the practical lesson is about pairing. An assertion that a recording is synthetic, standing alone in a list of submissions, did nothing here. Whether the same assertion supported by an examiner's report would move a bail court is untested in this order, and we do not predict it.

For prosecutors and investigating officers the lesson points the other way. Provenance material for phone-recorded statements, the device, the transfer chain, the original file and its metadata, is what makes an authenticity challenge answerable instead of merely deferred.

What we did not verify

What we opened: the full text of the five-page order as reproduced on Indian Kanoon, read end to end, including the cover table giving the reserved, pronounced and upload dates, paragraphs 1 to 9, the signature block reading (MANISHA BATRA) JUDGE 07.09.2026, and the digital attestation footer timestamped 2026.09.07 19:01.

A sourcing limitation we are disclosing rather than hiding. We read the judgment through Indian Kanoon's reproduction and did not retrieve the High Court's own copy from its website. The text layer we read carries visible extraction artefacts, in which fragments of words and docket numbers appear doubled. We selected quotations from passages free of those artefacts, and the sentence quoted from paragraph 7 runs across a page break in the underlying document, with the page-foot attestation stamp interposed by the extraction, not by the court.

What we did not open: the FIR, the challan report, the recording itself, the two earlier bail petitions by the same petitioner, and any order made by the trial court. We describe the prosecution case only as this order records it, and we make no statement about what the evidence shows.

What we refuse to claim: we do not say the court found the video to be genuine, or to be AI-generated, because it made neither finding. We do not say the court rejected the AI argument, because it did not rule on it. We do not call the recording a dying declaration, because the order does not. We do not say this is the first Indian order in which such an assertion has been made, because the order does not say so and we have not tested it. We name the petitioner and the judge as the order names them, and we do not describe anyone as guilty of anything; the petitioner stands accused, the trial is pending, and the order says its observations have no bearing on the merits.

Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.

Related briefings

• India's Supreme Court draft rules on AI in courts
• A new US federal rule of evidence for AI-generated evidence
• Brazil's STJ on AI evidence and lawyer hallucination
• Asia Pacific AI regulation tracker
• AI Regulation News hub