Reference · The Library
The Library.
A structured library of definitions, frameworks, instruments and case notes on artificial intelligence in legal practice: supervised AI, AI governance, cognitive sovereignty, and the record of fabricated authority in the courts. Read in order or by the question you arrive with.
Twenty-four entries · Growing · Maintained by Shivam Shukla, Advocate, Allahabad High Court
How to read itIn order · Or by question
The Library sets down, one unit at a time, the concepts and instruments that govern artificial intelligence in legal practice. Each entry states a term, why it matters to the reader in front of it, how it works, one decided case, and the rule. Read it from Part I as a single argument, or go to the entry that answers the question you arrived with. The Standard and the Register keep their own pages.
Part IWhat the machines do
Part I. What the machines do
Ground
The four forms a fabricated authority takes A machine-produced authority fails in one of four ways, and each is caught by a different check: a citation that identifies no decision; a wrong citation attached to a decision that exists; a correct citation with a passage or paragraph that does not exist in it; and a decision that exists and is correctly cited but does not hold what it is cited for. Only the last two survive a check that the case exists.
Part IIWhat the instruments require, across jurisdictions
Part II. What the instruments require, across jurisdictions
Ground
Supreme Court of India, draft Regulations for Use of Artificial Intelligence in Courts, 2026 A draft instrument published for public consultation on 3 June 2026 by the Supreme Court's Artificial Intelligence Committee, grounded in human primacy, transparency, accountability, data protection and judicial independence, that would make an officer exclusively accountable for AI-assisted decisions, treat AI output as advisory subject to verification, prohibit algorithmic adjudication and risk scoring, and require parties to disclose AI-assisted filings at submission. A draft as of 25 September 2026, not law.
Ground
Bar Council of India, Circular BCI:D:4657/2026 of 17 July 2026 A circular of the Bar Council of India on the use of social media and digital conduct by advocates, law students and interns, issued 17 July 2026 with Annexures A to F. It governs what an advocate publishes and depicts, not what an advocate files; it reaches fabricated authority only as a content rule; and it places on Indian advocates their first regulator-imposed duty to disclose material AI use, on the publication side.
Ground
EU Artificial Intelligence Act: human oversight, justice as high-risk, and the worker information right Regulation (EU) 2024/1689 classifies AI systems intended to assist a judicial authority in researching and interpreting facts and law as high-risk, requires that such systems be designed so natural persons can effectively oversee them (Article 14), and gives workers' representatives a right to be informed before a high-risk system is used at the workplace (Article 26(7)). Its duties fall on providers and deployers; it confers no right on the person judged.
Ground
GDPR Article 22: the right not to be subject to a solely automated decision Article 22 of Regulation (EU) 2016/679 gives a data subject the right not to be subject to a decision based solely on automated processing which produces legal or similarly significant effects, and, where such processing is permitted, safeguards including the right to obtain human intervention, to express a view and to contest the decision. It is satisfied by a human's intervention; it does not require that the human verified anything.
Ground
UNESCO Recommendation on the Ethics of Neurotechnology, 2025, and the neurorights line The line of instruments that protects the brain and mental states against neurotechnology: the four neurorights proposed by Ienca and Andorno in 2017, Chile's constitutional amendment of 2021 and its Supreme Court's first writ on neural data in 2023, and UNESCO's Recommendation SHS/BIO/REC-NEURO/2025 adopted on 11 November 2025. These are the nearest neighbours of the term cognitive sovereignty, and they concern a different object.
Ground
Bar Standards Board guidance on AI, May 2026: what soft activation looks like Guidance issued by the regulator of barristers in England and Wales, in force from 18 May 2026, mapping the use of artificial intelligence onto the existing Core Duties of the BSB Handbook. It annotates the outcome standard for a new technology and enacts no method standard; it is the clearest example of a professional body responding to AI by guidance rather than by rule.
Ground
The screen guild agreements: the construction pathway in contract The 2023 memoranda of agreement won by the Writers Guild of America and SAG-AFTRA after strikes, which made AI-generated material not literary or source material, barred requiring writers to use AI, required disclosure of AI-generated inputs, and required consent and compensation for digital replicas of performers; and their 2026 successors, which preserved and extended those terms. Bargained instruments binding on signatory employers for their term, not statute; the demonstration that a workforce without a statutory body can build the accountability layer itself.
Ground
The refusal-capacity map: who can say no, by profession and jurisdiction A comparative map of which body, in a given profession and country, can refuse an AI deployment and make the refusal bind, arranged from the strongest enforceable power (a statutory works council veto) through bargained floors, binding professional codes, information rights and guidance, down to workforces with no body at all. Read as a repertoire, not a ranking: every profession stands on an identifiable rung and a documented mechanism exists for climbing.
Part IIIThe method
Part III. The method
Canon
Cognitive Sovereignty The entitlement of a person whose case is decided by a court or tribunal that the reasoning, findings and authorities on which the decision rests were verified and are owned by a human adjudicator, so that no part of what is decided against them rests on machine output that no one checked.
Canon
Supervised Intelligence Model (SIM) A five-stage working method for professional work with generative AI in which three stages are reserved to the human alone (Legal Framing, Verification, Strategic Judgment) and two admit the machine (Pattern Expansion, Doctrinal Reconstruction), so that what leaves the professional's hands is human judgment with machine assistance inside it, never the reverse.
Canon
AI Responsibility Test (ART) A four-step test applied after the fact to a piece of AI-assisted professional work to decide whether the professional supervised or delegated, asking in turn whether the framing was diligent, the supervision adequate, the verification complete and the judgment independent.
Canon
Responsibility conditioned on office, not authorship Responsibility for a professional or judicial act attaches to the office of the person who performs it, not to the authorship of the words. The judge who signs an order owns every citation in it whoever typed them; the advocate who files a pleading owns every authority in it whatever produced them. Supervision, not authorship, is the unit of responsibility.
Canon
Authority, precedent, statutory provision Three terms kept apart. An authority is a judicial decision cited for a proposition. A precedent is a judicial decision that binds a later court. A statutory provision is enacted text. A statute is never called an authority, and a decision is not a precedent merely because it is cited.
Canon
The locator-not-source rule A law report, digest, database entry, news item or machine output only locates an authority; the court's own order or judgment is the source. Nothing is cited, quoted or relied on from the locator, and a locator's own reference number is never cited in place of the neutral citation or the law report.
Canon
Verification ledger A record kept beside any document that cites sources, with one row for each claim that depends on a source: the proposition it carries, the citation as held, the primary source to check, the specific points to confirm, the falsifier that would defeat the claim, and a status of unverified, verified against the primary record on a stated date, or failed.
Canon
The promulgation deficit The condition of a profession in which the outcome standards for AI-assisted work are old and settled (do not file what does not exist), the method standards (what supervision, verification and record constitute adequate professional control of a generative system) have been promulgated by no governing body, and the missing method standard is therefore being written retrospectively, one sanction order at a time, by the tribunals that punish its breach.
Canon
Refusal capacity: the activation and construction pathways The capacity of an institution to refuse a harmful AI deployment on behalf of a profession or workforce and make the refusal bind. It arrives by two pathways that are one process at different stages: activation, where a statutory professional body already holds the power and has not used it; and construction, where no such body exists and a workforce organises, wins recognition and assumes the power itself.
Canon
Answerability The criterion on which an AI deployment in professional or adjudicatory work is accepted or refused: whether, after the machine has acted, an accountable person remains who can be held to answer for the work. A machine may verify, but a machine cannot be a respondent; responsibility cannot be delegated to what cannot be held to account.
Canon
Outcome standard and method standard Two kinds of professional standard kept apart. An outcome standard states what finished work must not contain (a pleading must not cite an authority that does not exist). A method standard states how the work must be done for the outcome to be reliable (what checking, what supervision, what record constitutes adequate control). Generative AI changed neither outcome standard; it created a method that no method standard yet governs.
Part IVWhat has happened
Part IV. What has happened
Record
Gummadi Usha Rani v. Sure Mallikarjuna Rao (Supreme Court of India, 27 February 2026) Case note. SLP(C) No.7575 of 2026. A trial court's order relied on four Supreme Court decisions that do not exist; the High Court found them AI-generated and affirmed on the merits; the Supreme Court issued notice, stayed reliance on the report, declared that a decision based on non-existent judgments "is not an error in the decision making. It would be a misconduct", issued notice to the Attorney General, the Solicitor General and the Bar Council of India, and appointed a senior counsel to assist.
Record
Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd (Supreme Court of India, 2 July 2026) Case note. 2026 INSC 668, Civil Appeal No.11950 of 2025. The National Company Law Tribunal admitted an insolvency application relying on six citations that were non-existent, wrongly attributed or carried non-existent paragraphs, none cited by counsel; the appellate tribunal did not detect it. The Supreme Court set both decisions aside, held that a decision resting on fabricated material "is no decision at all", declared zero tolerance for the Bar and the Bench, and directed the Bar Council of India to prescribe a guiding principle with disciplinary consequences.
Record
Vijay Ghanshyam Gadiya v. Union of India (Supreme Court of India, 2 September 2026) Case note. 2026 INSC 947. A customs adjudicating officer imposed a penalty of Rs.425,27,99,100 by an order relying on judgments and articles that the Supreme Court, verifying them itself, found non-existent, falsely cited, or existing but not holding what was attributed to them. The Court applied the zero-tolerance standard of Pooja Ramesh Singh to executive adjudication, set aside the order and the High Court's confirmation, remanded to a different officer, and left disciplinary action against the author of the order to the appointing authority.
Record
The Charlotin database of AI hallucination cases, and what it counts A public database maintained by Damien Charlotin since April 2025 that tracks judicial decisions in which a court or tribunal addressed reliance on hallucinated AI content in more than passing reference. It recorded 1,809 cases as of late July 2026 and is updated daily. It counts incidents globally by court decision; it does not claim to capture every fabricated citation, and it does not verify against primary Indian records or annotate the standard that was breached, which is what the Register on this site does.
The Library, extended.
New entries, Field Notes, and essays on AI, law, and professional responsibility, by email. Occasional, signal only.
Recorded. You are on the dispatch list.