reportAI in CourtsAI GovernanceAI Hallucinations
Legal AI in India: the second-largest search market, nine court decisions and a rulebook still in draft
India searches 'legal ai' 8,100 times a month against 9,900 in the US; its courts have acted on eight matters built on fake authority; its rulebook is in draft.
Edited and verified by Cognesio LLP
Researched with AI assistance · sources verified by Cognesio LLP · How this was made ↓
In India, the phrase “legal ai” is searched about 8,100 times a month on Semrush’s India database against 9,900 in the United States, and the country’s courts have, between December 2024 and September 2026, set aside, quashed or recalled seven decisions in which an adjudicator relied on authority that does not exist, and let an eighth matter, a petition built on the same kind of citations, be withdrawn. The rules that would govern the tools behind those searches and those decisions are court policies from three High Courts, two Supreme Court judgments, one Bar Council circular and a set of draft regulations that closed for comment on 15 July 2026 and had not been notified on 5 September 2026. This report puts the three records together: what Indians search for, what the courts have decided, and what the guidance says, with the vendor documentation of the seven Indian legal-AI products in the SafeLegalAI tools directory read on the same schema used for the global audit.
It is descriptive. It does not rank tools or advise on compliance. Every count comes from the site’s own collections or from a document read on the date stated, and translation limits are recorded where a source is in an Indian language.
Key findings
- On Semrush’s India database read on 5 September 2026, the “legal ai” seed returns 202 keywords with a combined volume of 50,900 searches a month (Semrush’s total for the set; the published first page of 100 rows sums to 50,470) at an average difficulty of 37, against 15,713 keywords and 569,840 a month for the same seed in the United States. The head term alone is 8,100 a month in India, up from 6,600 recorded on 15 July 2026, against 9,900 in the United States; the long tail is a tenth of the American one.
- Indian demand is dominated by product names, most of them Indian: “libra login” 6,600, “lexis” 3,600, “niyam” 2,900, “legal desk” 2,400, “legora” 2,400, “legis” 1,900, “nyayassist” 1,300 and “lexlegis ai” 390. The one need-shaped query at scale is “ai for legal research” (1,000 a month, difficulty 59); the consumer query “free legal advice online in india” carries 880.
- The SafeLegalAI incident tracker holds eight Indian rows out of 150 as of 5 September 2026, and in seven of them the fabricated authority was in the adjudicator’s own order, not a lawyer’s brief: an income-tax tribunal (Buckeye Trust, 30 December 2024), a City Civil Court judge (Sammaan Capital, 24 March 2025), a faceless assessing officer (KMG Wires, 6 October 2025), the NCLT and NCLAT (Pooja Ramesh Singh, 2 July 2026), a Commissioner of Income Tax (Appeals) (EPFO v PCIT, 17 July 2026), a state tax officer (Faiz Enterprise, 20 August 2026) and an Additional Commissioner of Customs (Gadiya, 2 September 2026). Worldwide, the tracker’s 79 actor-coded rows are 52 lawyers, 21 litigants in person, 4 other and 1 judge; the Indian record is the inverse.
- A ninth decision, read for this report and queued for the tracker, is the first Indian cost order for AI-generated submissions: on 7 January 2026 the Bombay High Court, in Bahry v Heart and Soul Entertainment, found the respondent company’s written submissions, filed by its director appearing in person, “prepared using an AI tool such as Chat GPT or alike”, could not trace the case cited, said the practice “is going to result in costs” and that an advocate doing the same might be referred to the Bar Council, and ordered the party to pay costs of Rs 50,000 to the High Court Employees Medical Fund within two weeks (paragraph 35).
- No Indian decision in the record imposed a monetary sanction on a lawyer for fabricated citations, and none was referred to a bar disciplinary body by a court: the Karnataka High Court placed a trial judge’s order before the Chief Justice, the Supreme Court left a customs officer’s discipline to the appointing authority, and the Delhi High Court let a petition be withdrawn. The Supreme Court’s 2 July 2026 direction to the Bar Council of India to frame AI rules had, by 5 September, produced a circular on advocates’ digital and social-media conduct (17 July 2026) that requires disclosure of generative-AI use in online legal content and prohibits fabricated citations, but no rule on AI in court filings.
- The Regulation Tracker holds eight Indian instruments, all from July 2025 onward: the Kerala High Court’s district-judiciary policy (19 July 2025), MeitY’s national AI governance guidelines (5 November 2025), the Gujarat High Court’s policy (April 2026), the Punjab and Haryana High Court’s directive (reported 6 April 2026), the Supreme Court AI Committee’s draft regulations (3 June 2026), the Pooja Ramesh Singh judgment (2 July 2026), the Bar Council circular (17 July 2026) and the Gadiya judgment (2 September 2026). Three bind judicial officers by High Court policy; none yet binds lawyers’ use of AI in filings.
- The Supreme Court’s own AI programme is administrative, not adjudicative, on the government’s own account: the Press Information Bureau’s explainer of 11 February 2026 records AI transcription of Constitution Bench hearings, SUVAS translation of judgments into 18 languages (36,344 judgments in Hindi and 47,439 in other languages on the e-SCR portal as of 28 March 2025), e-filing defect detection, and says SUPACE “remains in an experimental stage and is not yet deployed for regular judicial use”. eCourts Phase III, approved on 13 September 2023 with an outlay of Rs 7,210 crore, earmarks funds for “Future Technological Advancement” including AI.
- Of the seven Indian legal-AI products in the tools directory, two publish a SOC 2 attestation, one an ISO 27001 certificate, one a public price list (Lexlegis.ai, Rs 9,000 to Rs 30,000 per user per month), five a statement that customer data is not used to train models, and three any accuracy claim, all vendor-generated; two have been verified by the site against primary documents.
Why India
Two numbers explain the choice. The first is demand: on the head term India is within a fifth of the United States, and no English-language publication outside the vendors covers Indian legal AI as a record. The second is the shape of the incident record. In every other jurisdiction in the tracker the fabricated citation arrives in a brief and the court catches it; in India it has arrived, seven times in eight, in the order itself, from tribunals and tax officers who are quasi-judicial adjudicators with no law clerk and no opposing counsel to check them. That pattern produced, in fourteen months, the two most quoted apex-court statements on AI-fabricated authority anywhere: that a decision resting even partly on such material “is no decision in the eyes of the law” (2 July 2026) and that “AI may well serve as training wheels but entrusting it with the pilot’s seat would be both imprudent and dangerous” (2 September 2026).
The site’s earlier India pages are the Pooja Ramesh Singh analysis, the Gadiya news brief and the Kerala policy entry; the judges’ record treats seven of the Indian rows as adjudicator rows. This report is the first to add the demand and vendor layers.
Method and data
Demand. Semrush Keyword Magic Tool, India database, seed “legal ai”, broad match, sorted by volume, read in the browser on 5 September 2026: 202 keywords, of which the first page of 100 was captured and published as legal_ai_in.csv alongside the US and UK sets of the search-demand report. The head-term country split is from Semrush Keyword Overview on the same day. Volumes are Semrush estimates, not Search Console data, and the India database is smaller than the American one, which depresses long-tail counts.
Incidents. The SafeLegalAI incident collection, 150 rows as of 5 September 2026 (/tracker/incidents.json), filtered to jurisdiction: india (eight rows, all status: verified), plus one decision read for this report and not yet a row (Bahry v Heart and Soul Entertainment Ltd, Bombay High Court, 7 January 2026, Indian Kanoon). Primary documents were read for every row: the ITAT order (archived), the Karnataka, Delhi, Bombay and Gujarat High Court orders on Indian Kanoon or the Charlotin archive, the Supreme Court judgments on Indian Kanoon and the Court’s own PDF server. Actor coding follows the tracker’s actor field where set and the order’s own description where not.
Guidance. The Regulation Tracker, 324 records (/regulation/documents.json), filtered to country: IN (eight records; four verified: true, four recorded from reports because the issuing court’s PDF was not retrievable or the instrument is an administrative communication). The Delhi High Court’s 2023 statement on ChatGPT (Christian Louboutin v The Shoe Boutique, 22 August 2023) was read on Indian Kanoon; the Supreme Court’s 4 March 2026 observations in Gummadi Usha Rani are taken from a Verdictum report and marked as such. Government programme facts are from two Press Information Bureau releases (3 April 2025 and 11 February 2026) and the Cabinet approval of 13 September 2023.
Vendors. The tools directory, 129 records (/tools/tools.json), filtered to hq: IN (seven records), read on the schema of the vendor documentation audit: SOC 2, ISO 27001, training-on-customer-data statement, zero retention, private deployment, data residency, public pricing, accuracy evidence and its author. Five of the seven are verified: false, meaning the site has recorded the vendor’s public pages but not confirmed them against a primary certificate or contract. The vendor pages were read for the directory on 3 and 4 September 2026 and were not re-opened for this report; the table repeats the records.
Limits. Semrush’s India database understates regional-language demand and the site did not pull Hindi-language seeds. The incident record is a selection of decisions found and read, not a census of Indian courts; district-court orders are rarely published. Two guidance records (Punjab and Haryana; Bar Council of India) rest on press reports because the issuing body’s text was not located. The report covers no arbitration record. Nothing here is a recommendation.
The search demand
The India “legal ai” set is small and brand-heavy. Of the 100 keywords on the first page, the 20 largest by volume are:
| Keyword | Volume/month | KD | Shape |
|---|---|---|---|
| legal ai | 8,100 | 55 | head term |
| libra login | 6,600 | 8 | product (Libra, an Indian practice tool) |
| lexis | 3,600 | 70 | brand |
| niyam | 2,900 | 35 | product |
| legal desk | 2,400 | 31 | product |
| legora | 2,400 | 55 | brand (Swedish vendor) |
| legis | 1,900 | 27 | product |
| nyayassist | 1,300 | 27 | product |
| ai for legal research | 1,000 | 59 | need |
| free legal advice online in india | 880 | 45 | consumer need |
| lex software | 880 | 44 | product |
| legal ai tools | 720 | 60 | need |
| legora ai | 720 | 39 | brand |
| nyayam | 720 | 33 | product |
| rukmini vasanth takes legal action against ai generated deepfakes | 720 | 26 | news |
| draft check | 590 | 39 | product |
| legalxgen | 590 | 27 | product |
| ai legal research | 480 | 62 | need |
| law bot | 480 | 39 | consumer |
| legal bot | 480 | 47 | consumer |
Three things distinguish it from the American set. First, the brands are Indian: Libra, Niyam, Legal Desk, Legis, NyayAssist, Nyayam, LegalXGen, Lexlegis and Draft Check between them account for more searches than the head term, and none of them appears in the US set. Second, the need-shaped queries are research queries (“ai for legal research”, “ai legal research”, “legal ai tools”, “legal research tools” 210), not the incident- or regulation-shaped queries that carry the American long tail; no Indian keyword on the first page asks about sanctions, hallucination rates or court rules. Third, a celebrity deepfake story (“rukmini vasanth takes legal action against ai generated deepfakes”, 720 and a variant at 320) is the only news-shaped query, and it is about personality rights, not legal practice.
The seed-level totals put the two markets in proportion. India’s 202 keywords carry 50,900 searches a month; the US seed’s 15,713 keywords carry 569,840. The head terms are close (8,100 to 9,900) because “legal ai” is a category search that every market makes; the long tail is where American demand for incident, benchmark and regulation content lives, and India’s database does not yet show it. Whether that reflects the database or the searchers cannot be told from Semrush alone.
The incident record
The nine decisions, in date order. “Actor” is who put the fabricated authority into the record.
| Date | Matter | Court | Actor | What was fabricated | What the court did | Tool named |
|---|---|---|---|---|---|---|
| 22 Aug 2023 | Christian Louboutin v The Shoe Boutique | Delhi High Court | Counsel (ChatGPT answer offered as evidence of reputation) | Not a fabrication; an AI answer offered as proof | Declined to rely on it: ChatGPT “cannot be the basis of adjudication of legal or factual issues in a court of law”; “AI cannot substitute either the human intelligence or the humane element in the adjudicatory process” | ChatGPT (named by counsel) |
| 30 Dec 2024 | Buckeye Trust | ITAT Bengaluru | Tribunal | Three of four citations in the order, two attributed to the Supreme Court | Order later recalled under section 254(2) (tracker; recall order not read) | None in the order; press attributed to ChatGPT |
| 24 Mar 2025 | Sammaan Capital v Mantri | Karnataka High Court | City Civil Court judge | Two Supreme Court judgments “never decided by the Apex Court or any other Court” | Revision allowed; order placed before the Chief Justice “for further action against the learned judge” | None |
| 25 Sep 2025 | Greenopolis v Narender Singh | Delhi High Court | Petitioner | Non-existent precedents and quoted passages absent from real judgments | Petition dismissed as withdrawn after respondents exposed the citations | None in the order; press attributed to ChatGPT |
| 6 Oct 2025 | KMG Wires v NFAC | Bombay High Court | Faceless assessing officer | Precedents “completely non-existent” | Assessment quashed and remanded; AI output “not to be blindly relied upon” | Unspecified |
| 7 Jan 2026 | Bahry v Heart and Soul Entertainment (not yet a tracker row) | Bombay High Court | Respondent company, by its director in person | ”Jyoti w/o Dinesh Tulsiani v Elegant Associates”, which the court “and its law clerks were at pains to find out … but could not find” | Practice deprecated; the court said it “is going to result in costs” and that an advocate doing the same might face “referring to Bar Council”; costs of Rs 50,000 ordered, payable to the High Court Employees Medical Fund within two weeks (para 35) | “an AI tool such as Chat GPT or alike” (court’s inference) |
| 2 Jul 2026 | Pooja Ramesh Singh v J&K Bank | Supreme Court | NCLT and NCLAT | Fake and hallucinated judgments in the tribunals’ reasoning | Both orders set aside; “such a decision is no decision in the eyes of the law”; “zero tolerance for the Bar as well as the Bench”; Bar Council directed to frame rules | None |
| 17 Jul 2026 | EPFO v PCIT | Delhi High Court | Commissioner of Income Tax (Appeals) | Six judgments “none of which actually existed” | Order set aside as a nullity (the order’s word is “nullify”) and “fraud upon the justice delivery system”; remanded; no disciplinary direction because the officer had retired | None (officer said an Inspector supplied the citations) |
| 20 Aug 2026 | Faiz Enterprise v State Tax Officer | Gujarat High Court | State tax officer | Non-existent and misattributed case law | Registration-cancellation order and show-cause notice quashed; verification instructions for all tax officers; contempt warning | AI, admitted by the officer |
| 2 Sep 2026 | Gadiya v Union of India | Supreme Court | Additional Commissioner of Customs | Non-existent judgments and misrepresented authorities behind a Rs 425.28 crore penalty | Penalty order and High Court judgment set aside; remanded to a different officer; discipline left to the appointing authority | None |
Three features stand out. First, the actor. Of the eight tracker rows, seven place the fabrication in the adjudicator’s own order; only Greenopolis is a party’s filing, and the Bombay cost order concerns a company director who appeared in person for the respondent. Across the tracker’s 79 actor-coded rows worldwide, lawyers account for 52, litigants in person 21, a judge 1. India’s record is not a record of careless advocates; it is a record of tribunals and revenue officers deciding with tools no rule yet governs.
Second, the consequence. In every Indian row the reviewing court acted on the decision, not the decision-maker: set aside, quashed, remanded or recalled. The Karnataka High Court’s referral of a trial judge to the Chief Justice and the Supreme Court’s remittal of a customs officer to the appointing authority are the only two instances of a personal consequence being set in motion, and neither outcome is on the record. No Indian court has fined a lawyer for a fabricated citation; the one cost order, Rs 50,000, concerns a company director appearing in person. The sanctions ledger records 45 monetary orders worldwide; none is Indian.
Third, the tool. Only one Indian order records an admission (the Gujarat tax officer); one records the court’s own inference (Bombay, from “green-box tick-marks, bullet-point-marks, repetitive submissions”); the rest do not say. The Supreme Court in Gadiya checked the citations itself and wrote that the case laws relied on “are either non-existent or have fake citations” and that some real ones “do not lay down the ratio deduced therefrom”, without naming a product.
The Supreme Court had signalled the direction four months before Pooja Ramesh Singh. On 4 March 2026, in Gummadi Usha Rani v Sure Mallikarjuna Rao (SLP (C) No. 7575 of 2026), a bench of Justices Narasimha and Aradhe is reported by Verdictum to have said that “a decision based on such non-existent and fake alleged judgments is not an error in the decision making. It would be a misconduct and legal consequence shall follow”. The order itself was not located for this report and the case is not a tracker row.
The guidance record
The eight Indian instruments in the Regulation Tracker, with the earlier Delhi statement and the Supreme Court’s institutional steps, form a timeline that runs from a single judge’s paragraph in 2023 to a draft national rulebook in 2026.
| Date | Instrument | Issuer | Binds | Core rule | Read |
|---|---|---|---|---|---|
| 22 Aug 2023 | Christian Louboutin v The Shoe Boutique, CS(COMM) 583/2023 | Delhi High Court (Prathiba M. Singh J) | Nobody (judgment) | ChatGPT output “cannot be the basis of adjudication”; “at best the tool could be utilised for a preliminary understanding or for preliminary research and nothing more” | Primary (Indian Kanoon) |
| 13 Sep 2023 | eCourts Phase III approved, Rs 7,210 crore over four years | Union Cabinet | Programme | Funds “for Future Technological Advancement including AI, ML and blockchain” (PIB, 11 Feb 2026) | Primary (PIB) |
| 19 Jul 2025 | Policy on AI tools in the district judiciary | Kerala High Court | District judges and staff, Kerala | AI tools “shall not be used to arrive at any findings, reliefs, order or judgment under any circumstances”; outputs verified; no confidential data in public tools | Mirror PDF (The Leaflet); record unverified |
| 5 Nov 2025 | India AI Governance Guidelines | MeitY, IndiaAI Mission | Nobody (non-binding) | National framework with no justice-specific rule (recorded from the government news announcement; the guidelines PDF was not retrieved) | Official news service; PDF not retrieved |
| 11 Dec 2025 | AI Committee reconstituted | Chief Justice of India | Supreme Court administration | Four-member committee chaired by Justice P. S. Narasimha “to oversee initiatives relating to the adoption, development and deployment of AI tools in the judiciary” | Official news service (News on AIR) |
| Apr 2026 | Policy on AI in judicial and court administration | Gujarat High Court | District judiciary, Gujarat | AI “shall never be employed for any form of decision-making, judicial reasoning, substantive order drafting or judgment preparation, bail/sentencing considerations”; permitted for research, translation, administration; independent verification of citations | Primary (court PDF) |
| 6 Apr 2026 | Directive on AI use by judicial officers | Punjab and Haryana High Court (Registrar General) | Judicial officers in Punjab, Haryana and Chandigarh | No AI platforms for writing judgments or orders, or for legal research | Press report (Law Trend); text not located |
| 3 Jun 2026 | Draft Regulations for Use of AI in Courts, 2026 | Supreme Court AI Committee | All courts and tribunals, once notified | AI may assist with research, drafting, transcription, translation and citation checks; may not decide cases, bail or credibility; lawyers to disclose AI use in pleadings; institutional AI committees. Comments closed 15 July 2026 (extended from 20 June by a notice of 16 June 2026); not notified on 5 September | Primary (court PDF) |
| 2 Jul 2026 | Pooja Ramesh Singh v J&K Bank, 2026 INSC 668 | Supreme Court | Bar and Bench (judgment) | Orders founded even partly on fabricated precedent to be set aside; citing unverified AI judgments is misconduct; Bar Council of India to frame rules and discipline | Primary (Indian Kanoon) |
| 17 Jul 2026 (date per the circular number reported, BCI/D/4659/2026) | Circular on digital and social-media conduct | Bar Council of India | Advocates, law students, interns | Disclose generative-AI use in online legal content; no fabricated judgments, fake citations or deepfakes; referral to State Bar Councils or contempt | Press report (Verdictum); official PDF not located |
| 2 Sep 2026 | Gadiya v Union of India, 2026 INSC 947 | Supreme Court | Quasi-judicial authorities (judgment) | “Assistance can never be substituted for adjudication”; orders relying on fabricated or misrepresented AI citations unsustainable | Primary (Supreme Court PDF) |
What the timeline shows is a rulebook built from the bench down. The first binding instruments are High Court policies addressed to the judges below them, issued under the High Courts’ powers over the courts below them (Gujarat’s cites Articles 225 and 227); they bind judicial officers, not lawyers. The two Supreme Court judgments state duties for both Bar and Bench but are judgments, enforceable case by case. The only instrument that would bind lawyers’ use of AI in filings across the country, the draft regulations, has not been notified. The Bar Council’s July circular, issued fifteen days after the Supreme Court directed it to frame rules, governs advocates’ online content and does not reach a brief filed in court; whether it is the Council’s answer to the direction or a first step is not stated in the circular as reported.
The Supreme Court’s own use of AI is administrative and the government says so. The 11 February 2026 explainer from the Press Information Bureau, titled “From Digitisation to Intelligence”, lists transcription of oral arguments in Constitution Bench matters (“gradually extended to regular hearing days”), SUVAS translation into 18 languages, e-filing defect identification, and two research tools, LegRAA and SUPACE, which “operate strictly as research aids, with no role in recommending outcomes or drafting judgments autonomously”; SUPACE “remains in an experimental stage and is not yet deployed for regular judicial use”. The translation numbers are the one hard output: as of 28 March 2025, 36,344 Supreme Court judgments in Hindi and 47,439 in other languages on the e-SCR portal, with High Court AI Translation Committees overseeing quality (PIB, 3 April 2025). None of the eight incident rows involves a court’s own AI system; all involve a general-purpose tool used by an individual.
The academic measurement on record is a single paper: “Evaluating the Role of Large Language Models in Legal Practice in India” (Hemrajani, National Law School of India University, arXiv 2508.09713, 13 August 2025), which tested GPT, Claude and Llama across issue spotting, drafting, advice, research and reasoning against junior lawyers’ work, rated by law students, and found drafting and issue identification often equal to or better than the human work while the models “struggle with specialised legal research, frequently generating hallucinations”. It is the only India-specific benchmark the site has found, and it is a survey experiment, not a citation-accuracy test of the products below.
The vendor record
The tools directory holds seven products headquartered in India. The table applies the audit schema; “unknown” means the vendor’s public pages, as read on 3 or 4 September 2026, do not say.
| Product | Vendor | Category | SOC 2 | ISO 27001 | No training on customer data | Data residency stated | Public pricing | Accuracy evidence | Verified by site |
|---|---|---|---|---|---|---|---|---|---|
| SCC Online AI Pro (preview, Feb 2026) | EBC Publishing | Legal research | unknown | unknown | yes | none stated | no (per-seat) | none | no |
| CaseMine AMICUS AI | Gauge Data Solutions | Legal research | unknown | unknown | unknown | none stated | no (per-seat) | vendor: 88% on All India Bar Exam questions, NLU Delhi observer | no |
| SpotDraft (VerifAI) | SpotDraft | Contract lifecycle | yes | unknown | yes | US, EU, IN, Middle East | no (enterprise) | vendor: “class-leading accuracy”, no metrics | no |
| Jhana | Jhana AI | Legal research | unknown | unknown | unknown (free tier data sharing is opt-out) | none stated | freemium | none | no |
| Lexlegis.ai (MIRA, On-Desk) | Lexlegis.ai | Legal research | yes | yes | yes | none stated (On-Desk is air-gapped) | yes: Rs 9,000 to Rs 30,000 per user per month | none | yes |
| Adalat AI | Adalat AI | Court technology | no (“coming soon”) | no (“coming soon”) | yes | IN | no (enterprise) | vendor: 90%+ transcription accuracy, no method | yes |
| Manupatra AI Search, Manuworks.ai | Manupatra | Legal research | unknown | unknown | yes | IN | no (per-seat) | none | no |
Read against the global audit of 47 verified tools, where 42 publish a SOC 2 attestation and seven a public per-seat price, the Indian seven publish less: two SOC 2, one price list, three accuracy claims all written by the vendor. Two features are specific to India. Data residency inside the country is stated by Adalat AI and Manupatra and offered as a region by SpotDraft; Lexlegis.ai’s On-Desk product (February 2026) puts the model on the customer’s own hardware. And the one court-facing vendor, Adalat AI, which describes deployments in 11 partner states and with more than 6,000 judges on its own pages, publishes neither certification and states an accuracy figure without a method; if the Supreme Court’s draft regulations are notified as drafted, a court-facing transcription system would fall inside the institutional-approval regime they create.
None of the nine Indian decisions names an Indian product. Where a tool is named or inferred it is ChatGPT or “an AI tool such as Chat GPT or alike”.
What to watch
The draft regulations are the pending instrument. Notified as drafted, they would put a disclosure duty on lawyers in every court and tribunal, prohibit AI in the decision itself nationwide, and create the committees through which court-facing tools would be approved; the Regulation Tracker will record the notified text and date. The Bar Council’s rules on AI in filings, directed by the Supreme Court on 2 July 2026, are the second; the July circular is the only Council instrument on record and it does not reach court filings.
On the incident side, the pattern to watch is whether the Gadiya direction to the appointing authority and the Karnataka referral produce a published disciplinary outcome, which would be the first personal consequence for an Indian adjudicator; and whether a High Court imposes costs on an advocate, as the Bombay High Court said it would. The tracker’s Indian rows will be re-swept quarterly; the Heart and Soul decision and the Gummadi order are queued.
On the demand side, the next Semrush read will show whether the India long tail acquires incident- and rule-shaped queries as the American one has, and whether the Supreme Court’s draft regulations register as a search term. On 5 September 2026 no Indian keyword on the first page referred to them.
Three sentences journalists can quote
India searches “legal ai” 8,100 times a month against 9,900 in the United States, but its long tail is brand names and research needs, with no first-page query about sanctions, hallucination rates or court rules.
In seven of the eight Indian decisions in the SafeLegalAI tracker the fabricated authority was in the adjudicator’s own order, and every reviewing court acted on the decision rather than the decision-maker; no Indian lawyer has been fined for an AI citation on the record.
India’s binding AI rules for the courts are three High Court policies for judicial officers and two Supreme Court judgments; the draft national regulations that would bind lawyers closed for comment on 15 July 2026 and had not been notified on 5 September.
Appendix A: data tables
A1. Indian incident rows by actor and outcome (tracker, 5 September 2026)
| Actor | Rows | Matters | Court acted on the decision | Personal consequence set in motion |
|---|---|---|---|---|
| Tribunal or quasi-judicial adjudicator | 6 | Buckeye, KMG Wires, Pooja Ramesh Singh (NCLT/NCLAT), EPFO (CIT(A)), Faiz (tax officer), Gadiya (customs) | 6 | 1 (Gadiya: appointing authority) |
| Judge | 1 | Sammaan Capital | 1 | 1 (order before the Chief Justice) |
| Party | 1 | Greenopolis | withdrawal | 0 |
| Party in person (not yet a row) | 1 | Heart and Soul | costs of Rs 50,000 ordered | 0 |
A2. Indian rows in the global actor coding
Actor (tracker actor field, 79 coded rows worldwide) | Worldwide | India |
|---|---|---|
| lawyer | 52 | 0 |
| litigant in person | 21 | 0 |
| other (adjudicators, officers) | 4 | 3 (EPFO, Faiz, Gadiya) |
| judge | 1 | 0 |
| firm | 1 | 0 |
| uncoded | 71 | 5 |
A3. India “legal ai” set, summary (Semrush IN database, 5 Sep 2026)
| Measure | India | United States (same seed, 5 Sep 2026) |
|---|---|---|
| Keywords returned | 202 | 15,713 |
| Total monthly volume | 50,900 | 569,840 |
| Average KD | 37 | 41 |
| Head term “legal ai” | 8,100 | 9,900 |
| Largest non-head keyword | libra login, 6,600 | lexis, 60,500 |
| Largest need-shaped keyword | ai for legal research, 1,000 | (see the search-demand report) |
A4. Regulation Tracker records with country: IN
| Record | Type | Version date | Verified |
|---|---|---|---|
| kerala-hc-ai-policy-district-judiciary | policy | 19 Jul 2025 | no |
| in-meity-india-ai-governance-guidelines-2025 | guidance | 5 Nov 2025 | no |
| in-gujarat-hc-ai-policy-judicial-court-administration | policy | Apr 2026 | yes |
| in-punjab-haryana-hc-ai-ban-judicial-officers | policy | 6 Apr 2026 | no |
| india-supreme-court-draft-ai-regulations-2026 | regulation (consultation) | 3 Jun 2026 | yes |
| in-sc-pooja-ramesh-singh-v-jk-bank-2026 | judgment-guidance | 2 Jul 2026 | yes |
| in-bci-digital-conduct-circular-2026 | regulation | 17 Jul 2026 | no |
| in-sc-vijay-gadiya-v-union-of-india-2026 | judgment-guidance | 2 Sep 2026 | yes |
Appendix B: sources
- Christian Louboutin SAS v The Shoe Boutique (Shutiq), CS(COMM) 583/2023 (Delhi High Court, 22 Aug 2023): Indian Kanoon
- Deepak Bahry v Heart and Soul Entertainment Ltd, WP 8390/2009, 2026:BHC-AS:828 (Bombay High Court, 7 Jan 2026): Indian Kanoon; report: LiveLaw, 16 Jan 2026; Bar and Bench
- Gummadi Usha Rani v Sure Mallikarjuna Rao, SLP (C) 7575/2026 (Supreme Court, 4 Mar 2026), as reported: Verdictum
- Pooja Ramesh Singh v Jammu and Kashmir Bank Ltd, 2026 INSC 668 (2 Jul 2026): Indian Kanoon
- Vijay Ghanshyam Gadiya v Union of India, 2026 INSC 947 (2 Sep 2026): Supreme Court PDF
- Sammaan Capital v Mantri Infrastructure, CRP 49/2025 (Karnataka High Court, 24 Mar 2025): Indian Kanoon; KMG Wires v NFAC (Bombay High Court, 6 Oct 2025): Indian Kanoon; Greenopolis Welfare Association v Narender Singh (Delhi High Court, 25 Sep 2025): Indian Kanoon; EPFO v PCIT, W.P.(C) 6049/2026 (Delhi High Court, 17 Jul 2026): Charlotin archive PDF; Faiz Enterprise v State Tax Officer, R/SCA 10123/2026 (Gujarat High Court, 20 Aug 2026): Charlotin archive PDF; Buckeye Trust v PCIT, ITA 1051/Bang/2024 (ITAT Bengaluru, 30 Dec 2024): archived order
- Supreme Court of India, Draft Regulations for Use of Artificial Intelligence in Courts, 2026 (3 Jun 2026): PDF; notice of 16 June 2026 extending the comment deadline to 15 July 2026, mirrored with the draft: Allahabad High Court PDF
- Kerala High Court policy (19 Jul 2025), mirror: The Leaflet PDF; Gujarat High Court policy (Apr 2026): court PDF; Punjab and Haryana directive, as reported: Law Trend; Bar Council of India circular BCI/D/4659/2026 of 17 July 2026, as reported: Verdictum; circular number and date as republished: TaxGuru; MeitY guidelines, as announced: News on AIR
- Press Information Bureau, “The Supreme Court is collaborating with the High Courts in translation of e-SCR Judgements in 18 vernacular languages” (3 Apr 2025): pib.gov.in (archive); “From Digitisation to Intelligence: How AI is Enhancing Access to Justice in India” (11 Feb 2026): pib.gov.in (archive); Cabinet approves eCourts Phase III (13 Sep 2023): pmindia.gov.in (archive); Supreme Court reconstitutes AI Committee (11 Dec 2025): News on AIR
- Note: pib.gov.in and pmindia.gov.in return 403 to the site’s link checker; both PIB releases were read with a browser user agent on 5 Sep 2026 and archive links are given.
- Hemrajani, “Evaluating the Role of Large Language Models in Legal Practice in India”, arXiv 2508.09713 (13 Aug 2025): arxiv.org
- Vendor pages as recorded in the tools directory (read 3 to 4 Sep 2026): SCC Online AI Pro; CaseMine AMICUS; SpotDraft security; Jhana FAQ; Lexlegis.ai trust; Adalat AI security; Manupatra legal
- SafeLegalAI datasets: /tracker/incidents.json, /regulation/documents.json, /tools/tools.json, as of 5 September 2026; Semrush India pull: /data/search-demand-2026-09/legal_ai_in.csv
Appendix C: changes to this report
None since first publication.