Written by Adarsh Parashar, a fourth-year B.A., LL.B. (Hons.) student at Chanakya National Law University, Patna

Two Kinds of AI Error, One Regulatory Response

The Supreme Court on 2 July 2026 quashed an insolvency order, on the ground that the National Company Law Tribunal relied on six precedents that were not in existence. In the case of Pooja Ramesh Singh vs Jammu and Kashmir Bank Ltd, a bench of P.S. Narasimha and Alok Aradhe held that no reliance can be placed on unverified content created by AI and called for the Bar Council of India to draft guiding principles. Though it was a grave transgression, it was undoubtedly the simplest kind of an AI mistake to detect: there’s a precedent somewhere in the law reports or there isn’t one, and any competent clerk with database access can easily check in minutes.

Effective from 1 November 2025, witness depositions in every trial court in Kerala were being recorded using a speech-to-text application called Adalat.AI, developed on a legal speech model in Malayalam language, following a pilot program that began on 1 February 2025 in the Ernakulam courts, hearing cases of atrocities and sexual violence against Women and Children. This is a unique type of issue. Unlike the independent register of reported judgments, there is no independent one concerning what a witness actually said. If a transcription model takes one word and replaces it with another that changes meaning without affecting the syntax, there may never be anything in the case file to catch it.

On 3 June 2026, a month before Pooja Ramesh Singh and seven months after Kerala began its implementation, the Supreme Court’s AI Committee published its draft Regulations for Use of Artificial Intelligence in Courts. The Draft Regulations were designed primarily to correct identifiable fabrications, and their design does not deal with undiscoverable transcription drift nearly as effectively. The Draft Regulations classify courtroom transcription as a merely permissible use requiring only internal certification, and confine their own grievance mechanism to harms arising from prohibited uses. A defectively certified transcript does not come under the remedy the Regulations provide for themselves, precisely the risk Kerala’s courts have been running for a full nine months in the country’s most fair-trial sensitive class of criminal proceedings.

Why the Fidelity of the Record, Not Just its Existence, is a Fair Trial Guarantee

In the Indian criminal procedure system, the role of the written transcript is significant as it forms the basis for subsequent actions in the case. According to Section 313 of Bharatiya Nagarik Suraksha Sanhita 2023 (replacing the Code of Criminal Procedure 1973 from 1 July 2024), as soon as the evidence is recorded from a witness, it must be read back to the accused and corrected if they disagree; if the witness objects to any part, the presiding officer shall record a memorandum of the objection without changing the text. Section 308 requires that evidence be taken in the accused’s physical presence, with an exception for child rape victims that allows them to testify via video link without the need for direct confrontation, while still maintaining cross-examination. 315 necessities that the judge record the demeanour of the witness, which cannot be preserved by any transcript.

This apparatus rests upon an assumption that a judge is personally responsible for each word that may be put into the record, since an appellate court does not re-hear witnesses, it reads the deposition. A mistake at the time of transcription would not then be a ‘clerical inconvenience’. It is the sole account of testimony that a higher court will ever have in exactly the offences under the Protection of Children from Sexual Offences Act 2012, where the specificity of the language of a child or survivor witness has its own evidentiary value. As opposed to a fabricated citation, a transcription error does not have an external register to which it can be compared. The record and the only practical way to check it, are produced by the same machine.

Kerala’s Mandate and the Certification it Never Had

The Kerala High Court’s own July 2025 policy on AI tools in the district judiciary anticipated that such tools would make mistakes, and placed the burden of catching them on the presiding officer’s unaided diligence. The Registrar’s memorandum of 27 September 2025, which made Adalat.AI mandatory from 1 November, requires only that the deposition be endorsed and signed by the witness and judicial officer before upload to the District Court Case Management System. There is no published, tool-specific error rate for Malayalam legal and testimonial speech, despite Indic languages being comparatively low-resource for AI training and errors being likelier where training data lacks the acoustic variation of a distressed or child witness’s voice. There is no indication in the memorandum that tells the reviewing trial judge, viewing a screen for the first time with a traumatised witness in attendance, what constitutes a reasonable error-rate, or how a substitution of words, which changes the meaning of the rule without altering the structure of the sentence, would be identified.

Permissible, Not Prohibited: Where Regulation 19 and Regulation 52 Miss Each Other

When read in full, the Draft Regulations include automated transcription of court proceedings in the scope of what is allowed as a permissible use of AI under Regulation 19(1)(b), provided it is ‘mandatorily reviewed and certified as accurate by a Designated Officer”, defined under Regulation 3(1)(u) as a judicial or court officer ‘nominated to supervise and take responsibility’ for a particular AI tool. This is already more challenging than Kerala’s protocol as certification assumes a standard and an accountable, named officer, rather than a reading-back. However, it is still an internal, one-time institutional certification, rather than an independent, published, language-specific accuracy audit.

When read in context with the Regulations’ own remedy, the consequences of their classification as permissible and not prohibited becomes evident. Regulation 52, the sole grievance mechanism the Draft Regulations create, applies only ‘where any harm is caused to any party to a proceeding as a direct or indirect effect of prohibited use of AI.’ Transcription is not prohibited use; it is a permissible use that is subject to certification. The exact situation which Kerala’s design invites, a litigant getting hurt by a certified but incorrect transcript, has no textual presence in Regulation 52. Indeed, in principle, there is no bar on the litigant exercising ordinary vindications under other law, as provided in Regulation 53, but the specific grievance provision of the Regulations, backed with the AI Secretariat and AI Committee with knowledge of the technology, does not make the same provision. It reaches risk-scoring, undisclosed AI evidence and surveillance, all listed as prohibited under Regulation 20, but not a poorly certified transcript.

Regulation 41: Fixes the Tool’s Future, not the Testimony’s Past

Regulation 41 requires that AI systems already in use at the commencement of the Regulations be reviewed by the AI Secretariat for compliance within one year, with the Appropriate Authority determining the course of action for any system found non-compliant. This is a genuine mechanism, and if applied to Adalat.AI it could in principle force Kerala’s courts to adopt a proper certification standard going forward. But Regulation 41 governs continued deployment of the tool. It does not address the evidentiary use of existing depositions that may have been taken between the date of 1 February 2025 and the date the Regulations come into force, which in the Regulations’ own commencement clause is merely “as considered appropriate by the Chief justice of India or the Chief justice of the respective High Court” to be notified at some point by those Chief Justices. A finding of non-compliance under Regulation 41 will have the effect of halting the use of the future output of Adalat Ai without affecting any word of testimony that it has already produced in POCSO and atrocity trials, as there is no mention of the reliability of its past output.

The European Union’s AI Act provides a useful contrast in architecture not outcome: High-risk AI systems used to support judicial bodies are required to keep automatic logs of events and appoint human oversight to monitor and rectify anomalous operations, requirements that are inherent to the operation of the system not just a single sign-off. Even this framework, however, does not address what should be done with output produced prior to the existence of the certification system, a concern India’s Draft Regulations also shares.

Closing the Gap Before Notification

The consultation window has been closed but the Draft Regulations have not yet been notified; and the AI Committee has, on its own, extended its deliberation beyond the original consultation day set on 20 June. Three specific amendments would close the gap identified here without upsetting the skeleton of the Regulations. Most of all, the current text of Regulation 52 should be expanded to include harm caused by defective certification under Regulation 19, as well as prohibited uses under Regulation 20, because that is the category of harm most likely to be caused by a transcription tool. Second, the certification requirement of Regulation 19(1)(b) should be read as requiring a published language specific error-rate audit existing prior to certification of a transcription tool, especially for Indic languages, to make the Designated Officer’s accountability a verifiable standard, and not a formality. Third, an express provision should be included in the Regulations that provides for the evidentiary value of testimony recorded prior to a system’s Regulation 41 compliance finding; otherwise, a compliance finding will only prevent future deposition testimony, meaning that every prior deposition will remain exactly where it was. In fact, Kerala’s courts have already produced over nine months of testimony in trials against women and children using a tool which the Regulations’ own certification logic would not yet approve. A Regulation 41 review which merely corrects the tool without a regulation that gets to the record is a correction that comes one deposition too late.