Six Points on the Draft AI Regulations for Courts That Deserve Attention in the Final Text
The consultation on the draft AI regulations for courts closed in July and the regulations have not been notified. That makes this a useful moment to set out six questions the draft leaves open, each of which will be harder to fix once the text is settled.
Where the Draft Stands
The Supreme Court published the draft Regulations for Use of Artificial Intelligence in Courts, 2026 under a notice dated June 3, 2026, inviting comments by June 20. A further notice dated June 16 recorded that the deadline stood extended till July 15, 2026. The regulations remain a draft and have not been notified.
Sansa Kanoon Pranali Partners filed a submission dated June 10, 2026. The points below are drawn from it. Whether any of them is taken up is entirely a matter for the Committee, and nothing here suggests that any has been.
One: The Draft Treats All AI as One Thing
A system that retrieves and assembles text a qualified person already wrote behaves very differently from one that generates new text by predicting the next word. The first cannot invent a clause that nobody drafted. The second can, and that risk cannot be engineered away.
The draft defines artificial intelligence and large language models in Regulation 3 but does not separate these two behaviours for the purpose of calibrating obligations. A definition of a structured system, meaning one that retrieves or assembles pre-authored and human-verified content without generating new content by predictive modelling, would let the heavier controls fall where the generative risk actually sits.
Two: Training, Fine-Tuning and Training Data Are Not Defined
The draft's definitional framework is extensive, but it does not define training, fine-tuning or training data. Those three terms carry most of the weight in any argument about what a vendor may do with material a court gives it.
The practical difficulty is that training embeds data in model parameters. Data that has been trained into a model is not stored in a way that can simply be deleted on request, which is why a deletion obligation drafted without a definition of training may not do the work it appears to do.
Three: What Happens to Court Data After a Vendor Has Processed It
Three consequences of court material reaching a vendor's training pipeline are worth separating out, because each engages a different body of law.
Privilege: Filings and written arguments contain privileged communications. Once that material is embedded in a vendor's model, the confidentiality on which the privilege rests is difficult to assert in any practical sense.
Pre-judgment information: A vendor processing draft orders holds material non-public information about listed companies before the market has it, which raises questions under the insider dealing framework that the draft does not address.
Consent: A litigant files documents because the law requires it, not because they have agreed to a private company improving a commercial product with their data.
On the last point, the exemption in clause (b) of sub-section (1) of section 17 of the Digital Personal Data Protection Act, 2023 covers processing by a court or tribunal where it is necessary for performing a judicial function. It is not obvious that it reaches a private vendor retaining the same material for commercial training. The exemption is also partial: sub-section (1) leaves sub-sections (1) and (5) of section 8 intact, so the obligation to take reasonable security safeguards survives it.
Four: Five Frameworks the Draft Intersects With
The draft does not operate in isolation, and five intersections would benefit from being addressed on the face of the text rather than left to argument later.
Consumer protection on interface design: Litigant-facing conversational tools permitted under the draft would sit within the dark patterns guidelines of 2023, which matter most where a user is unrepresented.
Intermediary status: Whether a court-deployed chatbot makes the court or its vendor an intermediary under the Information Technology Act, 2000, and so attracts the rules made under it, is unresolved.
Electronic evidence: Section 63 of the Bharatiya Sakshya Adhiniyam, 2023 is built around computer output produced in the ordinary course of a regular activity, with a certificate under sub-section (4). It was not designed for output generated on demand by a model.
Competition: Approving a small number of vendors without a procurement framework concentrates a market, which is a competition question before it is a technology one.
Product liability: Placing responsibility on the judicial officer when a tool malfunctions sits awkwardly with the product liability regime of the Consumer Protection Act, 2019.
Five: The Effect on Lawyers Who Have Not Yet Qualified
India has roughly twenty lakh enrolled advocates, and the All India Bar Examination held in November 2025 recorded a pass rate of 69.21 per cent. Deloitte has estimated what it calls a fifty per cent shock to the profession's workforce structure, which is an estimate rather than a measurement and is used here as one.
The exposure is not to senior practitioners. It is to the research, drafting and diligence work through which junior lawyers acquire judgment. There is also a narrower effect worth naming: the firms building these systems recruit young lawyers to label and structure legal data, and when a dataset is finished those lawyers have spent formative years neither practising law nor acquiring technical skills.
Six: A Penalty Schedule Instead of Predicting the Technology
Rules written to anticipate particular capabilities tend to be either too vague to enforce or too specific to survive the next release. Parliament took a different approach in the data protection statute: it set penalties high enough that non-compliance is not commercially rational, and left organisations to work out their own controls.
The Schedule to that Act puts failure to take reasonable security safeguards under sub-section (5) of section 8 at up to two hundred and fifty crore rupees, with a residual entry of up to fifty crore rupees. A comparable schedule attached to these regulations, covering unauthorised use of court data for training, failure to delete it, and undisclosed AI-generated content, would place the risk on the party best able to assess it without requiring anyone to forecast the technology.
Key Takeaways
The draft is not law: Published under a notice of June 3, 2026, comments extended to July 15, 2026, and not notified.
Retrieval and generation are different risks: A definition separating systems that assemble pre-authored content from those that generate new content would let obligations fall where the risk is.
Three terms do the heavy lifting and are undefined: Training, fine-tuning and training data.
Training is not reversible in the way deletion assumes: Data embedded in model parameters cannot simply be deleted on request.
The judicial exemption is narrow and partial: Clause (b) of sub-section (1) of section 17 covers the court's own processing, and sub-sections (1) and (5) of section 8 survive it.
Five existing frameworks intersect with the draft: Dark patterns, intermediary status, electronic evidence, competition and product liability.
The workforce effect falls on entrants: On the research and drafting work through which junior lawyers acquire judgment.
A penalty schedule avoids forecasting: The data protection statute sets up to two hundred and fifty crore rupees for a security safeguards failure and leaves controls to the regulated party.
Key Statutory Provisions
Section 17 of the Digital Personal Data Protection Act, 2023: Exemptions. Clause (b) of sub-section (1) covers processing by a court or tribunal, or other body entrusted by law with a judicial, quasi-judicial, regulatory or supervisory function, where necessary for performing that function. The exemption is partial, leaving sub-sections (1) and (5) of section 8 intact.
Section 8 of the Act, with the Schedule: Sub-section (5) requires reasonable security safeguards to prevent a personal data breach. Entry (1) of the Schedule sets up to two hundred and fifty crore rupees for its breach, and the residual entry up to fifty crore rupees.
Section 63 of the Bharatiya Sakshya Adhiniyam, 2023: Admissibility of electronic records. Sub-section (2) sets conditions turning on regular use of the device for an ongoing activity, and sub-section (4) requires a certificate accompanying the record.
Sources and References
Section 17 of the Digital Personal Data Protection Act, 2023
Digital Personal Data Protection Act, 2023, sections 8 and 17 and the Schedule; Bharatiya Sakshya Adhiniyam, 2023, section 63; and the draft Regulations for Use of Artificial Intelligence in Courts, 2026
Disclaimer: This article is for informational purposes only and does not constitute legal advice. Readers should consult a qualified legal professional for advice specific to their circumstances.
Comments