From Misconduct to Method.

When I published my Field Note on the frontier model releases of April, I wrote that the threat surface was no longer theoretical. The months since have converted that claim from argument into record.

The clearest way to see the change is to place the judicial findings in sequence. In December 2024, the Bengaluru bench of the Income Tax Appellate Tribunal recalled an order in a dispute worth several hundred crores after it emerged that the order rested on four non-existent citations, reportedly sourced through a chatbot and copied without verification. In March 2025, the Karnataka High Court directed an inquiry against a trial judge in Bengaluru who had rejected an application by relying on two Supreme Court decisions that were never delivered. In September 2025, a petition before the Delhi High Court quoted paragraphs 73 and 74 of the Raj Narain judgment; the judgment runs to 27 paragraphs. In October 2025, the Bombay High Court in KMG Wires Pvt. Ltd. v. National Faceless Assessment Centre quashed a faceless tax assessment because the assessing officer had justified a multi-crore addition with three decisions that do not exist, holding the reliance a breach of natural justice. In January 2026, the Bombay High Court in Deepak Bahry v. Heart and Soul Entertainment, 2026:BHC-AS:828, order dated 7 January 2026, imposed costs of fifty thousand rupees for AI-generated submissions containing an untraceable authority, recording the failure of verification as a professional conduct failure. In February 2026, a bench headed by the Chief Justice of India flagged, in open court, a fictitious case titled Mercy v. Mankind that had been cited in a petition. And in December 2025, the Kerala High Court in Blue Star Aluminium v. Federal Bank, WP(C) 43123/2025, order dated 10 December 2025, confronted writ petitions that carried formal legal structure but no material facts, drafted so mechanically that the advocates on record could not answer the court's questions about their own pleadings.

Then came the escalation. On 2 July 2026, in the appeal arising from the Essel Infraprojects insolvency, Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd., 2026 INSC 668, a bench of Justices P. S. Narasimha and Alok Aradhe set aside orders of the National Company Law Tribunal and the Appellate Tribunal after finding that of six judgments the Tribunal relied upon, three did not exist and three were genuine citations carrying invented paragraphs or a wrong title. No counsel had cited them; the Tribunal had sourced the material through its own research, and it passed unnoticed through an entire tier of appeal. The Court held that a decision resting on such material is no decision in the eyes of the law, prescribed zero tolerance for producing, citing, or using AI-generated precedents without verification, described the conduct as misconduct for an advocate and a serious lapse for a judge, directed the Bar Council of India, as the apex statutory body of the profession, to constitute a committee, prescribe guiding principles, and specify the disciplinary action that follows their violation, and likened the entry of fabricated case law into the system to the release of methyl isocyanate in the province of law and justice.

The same failure is being recorded across the common law world, and expertise offers no immunity anywhere. In England, the Divisional Court in R (Ayinde) v. London Borough of Haringey, [2025] EWHC 1383 (Admin), judgment of 6 June 2025, dealt with five fabricated cases in judicial review grounds, imposed wasted costs, and referred the lawyers to their regulators, holding that a general-purpose language model is not capable of conducting reliable legal research. In the United States, courts imposed over a hundred and forty-five thousand dollars in AI-related sanctions in the first quarter of 2026 alone, and one datum from that record is worth every statistic: in Lacey v. State Farm, a retired federal magistrate judge serving as special master, reviewing a brief with the specific intention of scrutinising it, recorded that the fabricated citations "affirmatively misled" him and that he had nearly incorporated them into an order. Expertise did not save him. Attention did not save him. The intention to scrutinise did not save him.

Around these judgments, the institutional machinery has begun to move. In May 2026, the Supreme Court asked the Bar Council of India to constitute an expert panel on artificial intelligence. In June 2026, the Court's AI Committee released draft Regulations for the Use of Artificial Intelligence in Courts, built on the principle that AI may assist but can never replace judicial decision-making, with absolute prohibitions on algorithmic adjudication and risk scoring, and mandatory disclosure by lawyers who use AI in preparing any pleading. The draft is not uniformly strict: it expressly permits AI-assisted legal research including citation verification, and it allows the responsible officer to waive verification of AI output for reasons recorded in writing, a waiver that sits uneasily beside the zero-tolerance standard the Court itself laid down three weeks later. Meanwhile, the High Courts have split: Punjab and Haryana and Gujarat have prohibited their judicial officers from using these tools outright, while the Supreme Court operates research, translation, and transcription systems of its own. And beyond the courts, the wider environment has hardened. The CERT-In advisory of 26 April 2026 on AI supply-chain compromise put every advocate who advises a regulated entity on notice that client confidentiality now runs through vendor infrastructure the advocate does not control, and the India AI Impact Summit closed in New Delhi with a Declaration joined by ninety-two countries and international organisations. Governance of this technology is no longer a specialist's subject. It is the weather.

Read the sequence as a whole and its direction is unmistakable. In January 2026, the Andhra Pradesh High Court could still hold, in the very Gummadi litigation, that non-existent citations do not vitiate an order so long as the legal principle applied is correct. By July, the Supreme Court had replaced that tolerance with a categorical standard. The system has moved, in eighteen months, from treating fabricated authority as an embarrassment to treating it as a contaminant. The vacuum I described in my book as total has begun, visibly and unevenly, to fill.

My work on this subject did not begin as scholarship. It began as self-defence. I practise constitutional, civil, revenue, criminal, and family matters before the Allahabad High Court, and I adopted AI tools in my own chamber early enough to feel, personally, the pull they exert on professional judgment: the fluent draft that invites acceptance, the confident citation that discourages the trip to the reporter. The frameworks I have published are the discipline I first had to impose on myself.

The public sequence began in February 2026, when I presented a paper on the imprints of AI on Indian litigation at the International Conference on New Age Legal Dynamics at the UPES School of Law, Dehradun, and was conferred the Award of Excellence as the session's best presenter. The following month I published the doctrinal foundation: AI for Indian Advocates: The Practitioner's Standard for Supervised Intelligence, a sixteen-chapter volume built around three instruments. The Five Doctrines of Professional AI Use state the obligations: Cognitive Sovereignty, Epistemic Diligence, Fiduciary Confidentiality, Adversarial Anticipation, and Intellectual Amplification. The Supervised Intelligence Method converts them into a five-stage workflow: Legal Framing, Pattern Expansion, Doctrinal Reconstruction, Verification, Strategic Judgment. The AI Responsibility Test supplies the accountability instrument: four questions on framing, supervision, verification, and independent judgment that an advocate must be able to answer before filing AI-assisted work, and may equally deploy against an opponent's.

In March, I published the cognitive grounding as an SSRN working paper, System 1 Lawyering and the Governance of AI in Legal Practice, applying dual-process theory to explain why fluent machine output suppresses precisely the verification that professional duty demands. In the same month I distilled the position into a five-page Declaration on Responsible AI Use in Legal Practice and dispatched it physically to the Registry of the Supreme Court of India and to the Bar Council of India. Two journal papers followed: System 1 Lawyering as Governance Failure, submitted in April to the AI, Law, Politics journal in Warsaw and revised on editorial review in May, now under peer review; and Cognitive Sovereignty as a Digital Human Right, submitted on 30 April to the RMLNLU-UPHRC Journal on Human Rights, now under review. In November, I will carry the work to the Solidarity AI Conference at Chulalongkorn University, Bangkok, where the conversation turns to law, regulation, and the cooperative governance of AI.

I want to be precise about one thing. I make no claim that my dispatches or publications moved any court or council to act. I claim something more modest and, I think, more useful: that the direction of the institutional response was predictable from inside practice, and predicted. The duties now crystallising in judicial directions and draft rules, disclosure of use, preservation of human judgment over outcomes, and accountability for the adoption of machine output, are the duties the Five Doctrines, the SIM, and the ART articulated from the practitioner's side months before the institutions reached for them. Where the instruments diverge, as the draft rules' written-reasons waiver diverges from the Court's own zero-tolerance language, the divergence itself marks the work still to be done. When the Bar Council frames the norms it has now been directed to frame, it will find that a working method already exists, built and tested where the problem actually lives.

Five lessons. Some were earned in practice. One was a correction I did not enjoy and now would not surrender.

First, the failure is cognitive, not technological, and it will not be cured by better models. Every documented Indian case follows the same anatomy: fluent output, correct format, plausible citation, and a professional whose deliberate verification never engaged because nothing on the surface signalled the need. The empirical record now bears this out with numbers. Peer-reviewed testing published in the Journal of Empirical Legal Studies found that leading commercial legal research tools, purpose-built, retrieval-equipped, and sold to careful lawyers precisely to avoid fabrication, still produced false citations in seventeen to thirty-three per cent of responses. And a Princeton benchmark published in June 2026 found that citation hallucination rates across successive model generations are not monotonically improving; one frontier model line regressed from just over one per cent to over six. The tool the careful lawyer bought to avoid being fooled will fool him. The countermeasure must therefore operate on the advocate's process, not the platform's features.

Second, and this is the correction: the machine can verify; it cannot be responsible. The same Princeton work showed that current systems, run as multi-step verification agents against legal databases, detect more than eighty per cent of fabricated citations, and in the course of doing so surfaced citation errors made by human appellate lawyers in briefs filed years before these tools existed. A Stanford study of judgment review found that human reviewers assisted by a grounded tool were more accurate than unaided humans and faster. I take those findings seriously, and I state their consequence plainly: any version of my framework that reads as a claim about machine incapacity is wrong, and I no longer defend it. What survives, and what the evidence strengthens rather than weakens, is the claim I should have led with all along. Responsibility for verification cannot be delegated, because a machine cannot be a respondent before a disciplinary committee, cannot owe a duty of candour to the court, and cannot be sanctioned. Verification may be assisted by the machine; it must be owned by the human. Stage 4 of the Supervised Intelligence Method, correctly stated, vests sole human responsibility for the verification outcome, with grounded tools available as instruments. A framework whose ground is capability expires with the next model release. A framework whose ground is responsibility does not expire at all.

Third, the unit of governance is the individual advocate. Rules are coming; the Supreme Court has now directed that they be framed. But rules operate on conduct after the fact. The only governance that operates at the moment of drafting, at eleven at night, in a chamber, with a filing due in the morning, is the standard the advocate has personally adopted. Voluntary standards travel faster than mandatory ones, and they build the professional culture that makes the eventual rules enforceable rather than aspirational.

Fourth, words matter more now, not less. A fabricated judgment is not a precedent, and I have stopped letting the loose phrase pass. Precedent is a judicial decision that binds a later court; authority is a decision an advocate cites in support of a proposition; a hallucinated citation is neither. And within the pathology itself, three defects must be kept distinct, because they carry different detection burdens and different culpability: fabrication, where the cited object does not exist; misattribution, where the object exists but the citation points elsewhere; and misrepresentation, where the citation is correct and the proposition is not supported. The July record before the Supreme Court contained all three in a single order. Counterfeit currency, passed in the one marketplace whose entire value is trust, which is why the Court's formulation, that a decision built on it is no decision in the eyes of the law, is not rhetoric but strict doctrinal accuracy.

Fifth, standards built inside practice hold; standards built about practice do not. Everything in my framework that has survived contact with real matters survived because it was extracted from real matters, and the one part the evidence corrected was, tellingly, the part that reached beyond what practice can observe into what only measurement can settle. The credibility of any professional standard is proportional to the author's exposure to its consequences, and, I would now add, to the author's willingness to be corrected by the data.

The work now moves along four lines.

The adversarial line. Volume II of the series, The Adversarial Standard, enters drafting after Bangkok. Volume I taught the advocate to govern their own use of the machine. Volume II teaches the advocate to confront everyone else's: detecting and challenging AI-generated submissions, cross-examining machine-produced evidence, and contesting algorithmic state action, together with model professional guidelines, practice directions, and curriculum in a form that the Bar Council, courts, and law schools can adopt as drafted. The Supreme Court's July direction to the Bar Council makes this the most immediately consequential part of the project: the profession's regulator has been ordered to build, and the materials exist.

The constitutional line. The deeper question beneath every case in Part II is not professional but constitutional: what is the individual's position before decisions produced by systems that give no reasons, name no officer, and hold no hearing? Here Indian law is not starting from nothing. For seventy years our administrative jurisprudence has required the speaking order, has voided decisions taken under dictation, and has treated reasons as the evidence that a mind was applied. The machine did not create the unexaminable decision; it industrialised it, and the cure was already on the books. My submission under review frames the modern form of that guarantee as Cognitive Sovereignty: the entitlement of every person, not only every advocate, to determinations affecting their rights that are the product of accountable human judgment, with reasons as its audit trail.

The empirical line. The correction recorded above imposed an obligation. A standard that claims the authority of practice must generate evidence from practice, not merely argument. The next stage of this work is therefore empirical: a structured record of how AI output actually behaves in a working Indian chamber, and a systematic dataset of the documented Indian incidents classified by defect type, opened as the Register of AI-Fabricated Authority in Indian Proceedings. The field's evidence base is currently American. The jurisdiction whose Supreme Court has gone furthest deserves its own.

The international line. The cognitive mechanisms are universal, the judicial responses are converging, and the Indian record, which now includes the strongest apex-court intervention on AI misuse in any common law jurisdiction, deserves to be argued into the global conversation rather than footnoted in it. Bangkok in November is the next step on that road.

Eighteen months ago, the sentence at the centre of my book read as a thesis. After this year, it reads as a case note. The machine processes. The advocate authors. And the discipline this year has added to it: the scholar verifies. Everything I have built, and everything still to come, exists to keep those sentences true.