The Law of Using AI — India First

In this session

1. 01

   The opening experiment re-framed in one slide: the room's real-or-AI score, reliably worse than its confidence, is the factual premise of modern evidence law. If trained eyes cannot tell, authenticity must rest on provenance rather than plausibility: where the file came from, who can speak to it, what the hash and metadata show. Every doctrine in this hour is the legal system converting that premise into procedure.

2. 02

   Authentication, the current Indian frame: electronic records come in under s.63 of the Bharatiya Sakshya Adhiniyam 2023, with the s.63(4) certificate in the Schedule's form (Part A disclosing the record's hash value, Part B signed by an expert, alongside the person in charge of the device). The case law: Anvar (certificate mandatory for secondary electronic evidence, overruling Navjot Sandhu) and Arjun Panditrao (mandatory for secondary; not required where the original is produced), both decided under the old s.65B, their carry-over to the BSA a working assumption of practice rather than a settled holding, and flagged as such.

3. 03

   The 2026 authority Session 4's bundle-builders have met: Pune Bar Association v. Union of India (SC, 22 May 2026): a three-judge order rejecting a constitutional challenge to s.63(4), holding the hash-value requirement rationally connected to integrity ('an electronic fingerprint'), and clarifying that Part B's 'expert' is not confined to examiners notified under s.79A of the IT Act: qualified cyber-forensic professionals may sign, subject to the court's satisfaction. The order names artificial intelligence and deepfake technology as why the safeguards matter. It is an order disposing of a petition without notice, not a judgment after contest, the kind of distinction this course trains students to state.

4. 04

   Synthetic media's two-front pressure, named precisely: fabricated evidence that must be kept out, and the liar's dividend (genuine evidence dismissed as fake) that must be defended against; both are provenance problems, and the second is the more corrosive because it taxes every authentic record. India's answer at the platform layer arrived in force on 20 February 2026: the amended IT Rules define 'synthetically generated information', require prominent labels and embedded provenance metadata on it, oblige platforms to prevent unlawful synthetic content, and compress takedown clocks to hours.

5. 05

   Two boundaries matter to lawyers: the definition covers audio-visual material and not text (an AI-drafted pleading is not 'SGI'), and a platform label is a new species of provenance evidence whose absence is now itself a signal. No Indian court has yet ruled on the admissibility of a contested deepfake; the personality-rights docket (chatbot impersonation in Sonakshi Sinha, voice cloning in Arijit Singh) is a stream of interim injunctions, remedies rather than evidence law, and the course says so.

6. 06

   The White Paper, by register: it recommends. Its firm lines are Guidelines 12 (no confidential or privileged information into any AI tool), 14 (independent verification before reliance), 15 (no AI verifying AI) and 18 (responsibility stays with the human). It contains no cloud restriction and no approved-tools regime, whatever summaries claim.

7. 07

   The draft Regulations for Use of AI in Courts (June 2026), by register: they would bind if notified. Fifty-seven regulations, including permitted judicial uses, an absolute bar on AI deciding, risk-scoring or credibility assessment (reg. 20(d), India's structural answer to the Loomis problem), reg. 43(3)'s requirement that lawyers certify disclosure of AI assistance in prepared material, and reg. 49's mandatory AI training for advocates; a sixteen-hour course like this one is reg. 49 arriving early. Publicly opposed in part by the advocates-on-record association, judicially doubted in speeches, and, as of the date on this slide, a draft.

8. 08

   The judiciary's own stack, and the first binding instrument: SUPACE, SUVAS, TERES, LegRAA, e-filing AI; Adalat AI transcription mandatory in one state's district courts; and the Kerala High Court's 2025 policy for its district judiciary, the first binding Indian AI-use instrument for judges. What is notified changes; the ladder's registers do not.

9. 09

   The profession's frame, with the pinpoints this course checked: the BCI Rules (Part VI, Chapter II) duties: Rules 3 and 4 on conduct toward the court, Rules 15, 17, 19 and 24 on the client, including the confidentiality rule that still cites the repealed Evidence Act section whose successor is BSA s.132, enforced through s.35 of the Advocates Act's misconduct jurisdiction. No BCI rule yet mentions technology or competence, which is why Pooja Ramesh Singh matters structurally: it converts 'cite it unverified' into misconduct within the existing rules and directs the BCI to frame AI-specific guiding principles (no committee reported constituted as of the date checked).

10. 10

    The BCI's July 2026 circular on digital ethics (reproduced but not readable at the original) reaches students directly: undertakings on deepfake and digital conduct at admission and before internships. And the DPDP Act, dated honestly: enacted, Rules notified, but the substantive obligations commence 13 May 2027. Today the binding grounds for the never-paste rule are privilege and the professional rules; the fiduciary-and-processor structure with its ₹250-crore penalties is the frame you are training for. Re-checked each cohort.

11. 11

    Privilege, as construed now: ss.132–134 BSA carry the advocate-client privilege forward; In Re: Summoning Advocates (2025 INSC 1275) holds the privilege is the client's, invocable by the advocate on the client's behalf; sharply limits summoning the accused's advocate; protects the advocate's devices with court-supervised procedures mindful of other clients' confidences; and places in-house counsel outside s.132, with only s.134's narrower protection. The AI consequences: a consumer chatbot is nobody's advocate; a paste is a disclosure the client did not authorise; and the in-house gap means a company's legal team has less privilege cover than external counsel for the material likeliest to end up in productivity tools.

12. 12

    Comparative, as contrast and in three sentences each. The EU AI Act classifies AI for judicial decision-support as high-risk, its obligations phasing in with dates that have themselves moved; the direction is architecture and conformity rather than case-by-case duty. The United States has fragmented into judge-by-judge standing orders, over a hundred requiring disclosure or certification, with no national rule; India's draft reg. 43 would be the opposite design, one rule for every court.

13. 13

    England and Wales has judicial guidance whose updated text says lawyers need not disclose responsible AI use (the opposite of draft reg. 43), alongside a High Court decision referring fake citations to the regulators and a fresh solicitors'-regulator warning notice; it is the common-law cousin whose case law the Indian Supreme Court itself cited in Pooja. The pattern across all three: verification duties converge; disclosure and institutional design diverge. A student who can state which jurisdiction requires what, as of when, is ahead of most of the profession.

14. 14

    The student's own layer, today, and the warning this course owes you: the institutional landscape is uneven. One NLU has a published AI-content policy; another's examination regulations make 'use of any means of artificial intelligence' in assessed projects serious academic malpractice; the flagship student journal permits disclosed assistive use but bans generative drafting and AI-supplied citations; the leading international moot's rules are silent, leaving its work-product rule to do the work; and the BCI's digital-ethics push reaches students directly, with undertakings at admission and before internships.

15. 15

    So the rule of rules: this course requires AI use under its own assessment scheme; your other subjects are governed by your university's regulations, and in some of them the same conduct is a disciplinary offence. Know your institution's rule before you rely on a tool for assessed work; disclose per the rule, and where none exists, disclose anyway in the form Session 8 drills; keep the prompt log and verification trail as your record; and never let a detection tool's verdict, in either direction, substitute for that record.

The mirror

Run on the class, then on the machine.

An experiment on the class, the same effect explained in the model, a live demonstration on a real tool, and a named takeaway skill.