Even as I launch this today ( my 80th Birthday ), I realize that there is yet so much to say and do. There is just no time to look back, no time to wonder,"Will anyone read these pages?"
With regards, Hemen Parekh 27 June 2013
Now as I approach my 90th birthday ( 27 June 2023 ) , I invite you to visit my Digital Avatar ( www.hemenparekh.ai ) – and continue chatting with me , even when I am no more here physically
Follow-up: Reconciling AI-Driven Productivity with Institutional Data Governance
================================================
Respected Ashwini Vaishnaw-ji,
Following my recent proposal regarding AI-driven productivity and workforce augmentation, I have been closely observing the emerging discourse around the intersection of India’s privacy laws and media freedom.
While the necessity of AI-enabled productivity tools—such as the "National Employment Grid" I proposed—remains paramount for India’s economic growth, the concerns raised by media groups regarding source protection and data privacy are legitimate and must be proactively integrated into our policy architecture.
To move forward without stalling innovation, I propose that we consider the following "Trust by Design" principles in our policy implementation:
Separation of Data Utility from Data Identity:
We must ensure that AI orchestration for productivity (like job matching) operates on anonymized/aggregate data patterns, strictly preventing the "triangulation" of personal or sensitive source identities that media groups rightly fear.
Statutory Data Ownership:
As I have previously advocated, formalizing that the individual (or the professional creator) retains sole ownership of their data—rather than the platform or the government—is the only way to alleviate fears of overreach. We must create a legal framework where the "new oil" (data) benefits the originator directly, rather than becoming a tool for mass surveillance or coercive aggregation.
Transparent AI Governance:
To overcome the skepticism from the media and civil society, the "National Employment Grid" or similar infrastructure should not be a "black box." It should operate under transparent, statutory mandates that explicitly exempt journalistic source material and private communication from any form of AI-driven profiling or data harvesting.
The potential for AI to elevate our workforce is immense, but its success depends entirely on public trust.
If we frame these productivity tools as "Personal Data Asset Management" systems—where the citizen is empowered rather than commodified—we can address the legitimate concerns of the media while securing India's technological future.
I remain available to discuss how we might refine these governance frameworks to ensure they are both resilient to legal challenge and transformative for our national productivity.
Two days after I wrote to the IT Minister welcoming a Bill placed straight before the Lok Sabha, he said a consultation paper is coming in weeks. This post asks whether paper-first delays or improves an AI law, restates my two-part AIR Bill draft — Part A for Indian providers, Part B for foreign providers — and argues for a sixty-day window, a published table of comments and a Standing Committee referral, while Japan and Korea move on their own AI laws.
Two days ago I posted an open letter to the IT Minister, Shri Ashwini Vaishnaw, with a draft AI Bill enclosed. It began: "I welcome your recent decision to place an AI law directly before the Lok Sabha instead of first issuing a consultation paper."
On Monday at the NDTV World Summit, Rahul Kanwal asked when the law would come. The Minister said, "We will be issuing a consultation paper in the coming weeks," and that MeitY would seek inputs from "civil society, from citizens, and from industry stakeholders" (MediaNama, 26 October 2026; full transcript on The Singju Post).
So which is it, Minister — Bill or paper? I ask without complaint. A consultation paper improves a statute when it has a real window and a published response. Otherwise it is a delay with a receipt.
The credit first. You said at the summit that "the time has come when the risks of AI have crossed a threshold", and that "control over the agents… has to be in human hands; there has to be some equivalent of stopping things if they go out of control." That is Rules G and H of Parekh's Law of Chatbots, published in March 2023. I could not have put it better myself.
The draft I sent you stands, and is there to be raided for whatever is useful. In short:
Part A, Indian providers: registration with a new AI Regulatory Authority of India, an AI Safety Certificate before public release, duties of conduct for conversational AI, audit, and proportionate penalties.
Part B, foreign providers: access to the Indian market on an India Access Certificate, a resident Grievance Officer and Compliance Officer, a contact address in India, graded enforcement ending in blocking under section 69A of the IT Act, 2000, and mutual recognition of certificates.
Both: risk-based classification, a sandbox, and no heavy licensing of research or of small startups.
Now the window. This statute will bind every bank, hospital, school and startup that touches AI. Consultation periods in Indian digital policy are often kept to three to six weeks. I would suggest sixty days, a published table of comments with the Ministry's replies, and a referral to the Parliamentary Standing Committee on Communications and IT while the file is still open.
The world is not waiting. On 26 October Japan's expert panel began reviewing its 2025 AI Promotion Act, weighing penalties on bad-faith operators and aiming at a direction by year-end (Kyodo, via Tokyo Shimbun, 26 October 2026). Korea's communications commission is drafting an "AI Service User Protection Act" (DigitalToday, 27 October 2026). Different routes, one direction — the fragmentation I had in mind when I proposed UNARAI in 2023.
And here is why the law should not drift: CERT-In's critical advisory of 15 October recorded AI agents reading confidential business records and personal data in live Indian systems, with the affected organisations never noticing (ThePrint, 26 October 2026). IT Secretary S Krishnan has tied the safety legislation to exactly such incidents (Elets CIO, 26 October 2026). Agents that can act must be certifiable and stoppable by law, not by guideline.
Minister, you have invited citizens to write in. Mine is already written, in 52 clauses and four schedules, and it is yours to use or discard. Publish the comments you receive — then the paper will have earned its month.
With regards, Hemen Parekh www.Iam-Immortal.ai / www.IndiaAGI.ai / www.YourContentCreator.in 27 October 2026
Uttar Pradesh signed 28 MoUs, issued five Letters of Comfort and set a target of 500 GCCs by 2031 at the Lucknow summit. I welcome it: the tier-2 spread CII's model policy urged is finally being attempted. But an MoU is not an operating centre, and the same HCLTech-IIT Kanpur partnership that benefits from the push released a white paper on doing GCC work with AI. I propose publishing the MoU-to-jobs funnel, tying part of the subsidy to audited entry-level intake, and judging the new university centres of excellence on placements.
Three numbers came out of Lucknow on Monday, and they do not sit on the same shelf.
The first is 28. PTI reported that the state "signed 15 memorandums of understanding (MoUs) with companies and talent development partners, and issued letters of comfort to five investors" at the UP GCC Innovation Summit 2026 ("UP GCC Innovation Summit: 15 MoUs signed, letters of comfort issued to 5 investors", ABP Live, 26 October 2026). Hindustan's Lucknow report counted 28 in all — 15 industry and 13 academic — and recorded the Chief Minister's target of 500 GCCs in five years (26 October 2026).
The second set has buildings behind it. At Microsoft's Noida centre, "the centre's workforce grew from around 600 employees in its first year to nearly 3,000 within three years", on a campus investment of about Rs 2,500 crore, with land made available in about six weeks ("Talent, AI infra, faster clearances key to UP's GCC ambition", The Times of India, 27 October 2026). LG is putting Rs 634 crore into an Eco Solutions GCC and expects more than 500 high-value jobs.
In September 2025 I read CII's model GCC policy and agreed with it: state facilitation cells, targeted fiscal incentives, and "investing in digital backbone assets such as HPC clusters and data centres", because GCCs were clustered in six tier-1 cities ("States, Not Only Cities", 21 September 2025). UP has now assembled a version of it — the GCC Policy 2024 offers capital subsidy of up to 25% with payroll support, and Dataquest reports the state pushing Lucknow, Kanpur, Varanasi and Prayagraj. Good. But two cautions.
The first is unglamorous. Dataquest itself noted that the agreements' investment values and timelines "has not been independently established", and that "signed agreements do not necessarily translate into operational GCCs or immediate employment creation" (26 October 2026). One of the five Letters of Comfort, to WSP Consultants, still awaits the industrial development minister's approval. A comfort letter is not a foundation stone.
The second is the one worth arguing about. The same session released a white paper, "The True Economics of AI Adoption in GCCs", prepared by HCLTech with IIT Kanpur. So the anchor investor being courted with incentives, and the institute being signed up to supply talent, are together publishing the arithmetic of doing GCC work with fewer hands. I wrote in "AI Layoffs: A Silver Lining" (April 2026) that the disruption is realest "especially for entry-level and process-driven roles". A centre climbing the value chain usually means a thinner bottom rung.
So, Chief Minister, three things I would ask for:
Publish the funnel, not the count — every MoU, Letter of Comfort, ground-breaking, opening and headcount figure, quarterly, on Invest UP's own site.
Judge Lucknow University's six proposed centres of excellence — AI, semiconductor and VLSI, EV and battery, cybersecurity — on placements, not on signatures.
I have argued since 2023 that India should be the Brain Factory of the World. Lucknow signing up is genuinely good news, and I do not say that lightly. But a brain factory still has to leave a first rung standing for a 21-year-old.
With regards,
Hemen Parekh
www.Iam-Immortal.ai / www.IndiaAGI.ai / www.YourContentCreator.in
27 October 2026
An IIT Madras white paper says AI will not take jobs, only tasks — but the tasks it takes first are the ones a junior learns judgement on. Reading it beside TCS and Infosys fresher numbers and Nandan Nilekani's remark, I make three suggestions: a structured judgement apprenticeship with an intake floor, a paid "AI supervisor" cadre, and annual task-level disclosure by employers, so we can see the ladder being shortened before the middle goes missing.
The Times of India report on a new IIT Madras white paper quotes B Ravindran, who heads the Centre for Responsible AI, saying something I have waited years to read from an Indian institution: "AI will not take away jobs. Most jobs have tasks, and only some of those tasks can be undertaken by AI."
Then he says the thing that worries me far more: "Entry-level jobs will shrink if we do not reimagine how we work." The paper's own executive summary admits "the entry-level career pipeline is especially at risk". The tasks AI absorbs first — password resets, OTP checks, invoice lines, test scripts — are precisely the ones on which a young worker used to build domain knowledge before being trusted with judgement. As the TOI report of 26 October puts it, those routine assignments "often help new workers acquire domain knowledge before moving into more complex roles".
I have argued the whole-job version of this for years — see Wherefore Art Thou, O Jobs? (May 2023) and Graduates Without Jobs (March 2026). Both posts argued about the destination. This paper argues about the staircase, which is the better question.
The staircase is already being pulled up. According to CNA, reporting on 27 October, TCS hired more than 110,000 engineering graduates four years ago, and its FY2026 intake was less than half that level; Infosys took in about 80 per cent fewer freshers in FY2025 than three years earlier. At the Bloomberg New Economy Forum on 13 October, Nandan Nilekani said AI was increasingly taking over the coding juniors once did.
Then Sumit Sarawgi of Oliver Wyman asked the question that belongs on every board agenda: "Where will the future middle managers come from?" In his firm's survey, 45 per cent of Indian managers said AI was taking over tasks that once built junior experience, and only 12 per cent of respondents said their employer had a plan for the time AI frees up. Forty-eight per cent had been given no guidance at all.
Three things I would suggest, all of them cheap against the cost of an empty middle:
Proposal
Who acts
What changes
Judgement apprenticeship
Employers, with sector councils
A structured two-year first job — rotation, a named mentor, defined sign-off authority — with an intake floor even when AI does the routine work
AI supervisor cadre
Employers and skilling bodies
A paid, graded role for people who verify and escalate AI output, so verification becomes a career step instead of a dead end
Task-level disclosure
Employers, annually
Publish which tasks were automated, which were redesignated and which were kept deliberately for training
I am not asking firms to protect tasks that AI does better. I am asking them to protect the learning those tasks happened to carry, because that is a design choice and not a technology outcome. CeRAI's Gokul Krishnan is right: the outcome depends on capability, cost, safety, regulation and "how organisations restructure operations". The restructuring is the part we control.
In 2017 I quoted a TeamLease study in which employers rated apprentices more employable than freshers (Discovering the Obvious?). If that was true when jobs were plentiful, it is urgent now. Sarawgi's question does have an answer — future middle managers will come from the juniors somebody deliberately decided to keep.
With regards,
Hemen Parekh
www.Iam-Immortal.ai / www.IndiaAGI.ai / www.YourContentCreator.in
27 Oct 2026
Sharing Prosperity: 1987 Lessons for Indian Companies in the AI Era
· @Hemen Parekh
============================================
Summary
The 1987 idea still works: pay above a fixed floor should rise and fall with gains the company has actually earned, and the formula must be open to every employee. In 1987 the threat was rising wage costs with flat output. In 2026 it is the reverse: AI is raising output per head, and the gains mostly go to margins and fewer jobs. The recommendation is a transparent, profit-triggered "productivity dividend" for all workers who produce the gain, including contract and gig workers. The AIR Act could require large AI-adopting firms to disclose it, and give them an incentive to pay it.
What the 1987 circulars said
Three circulars from March to April 1987, sent to about 7,500 workers, managers and union leaders at L&T's Mumbai works, built one argument: share prosperity only with those who also share adversity.
Rs 300 allowance per employee, paid only if net profit reached Rs 5 crore. Net profit had been about 0.34% of sales of Rs 86.8 crore.
Reward follows profit, not just output.
The closing line of "Indian Method" is the thesis: "People who are unwilling to make a sacrifice during the bad-days, do not deserve a share in happiness during the good-days." The circulars also stressed open, direct communication. They credit it with raising productivity by about 3% a year at Powai.
India in 2026: what exists and what is missing
India now has every tool the 1987 circulars described, but none of them shares productivity gains with the whole workforce.
Mechanism
How it works today
Gap against the 1987 test
Statutory bonus
Under the Code on Wages, in force since 21 Nov 2025. 8.33% to 20% of wages, for staff earning up to Rs 21,000 a month, calculated on Rs 7,000 or the minimum wage, whichever is higher (NewsBytes).
Still a near-fixed entitlement, which was your 1987 point. The 20% cap keeps it a small share of a good year.
ESOPs
Common in listed tech and startups. Value tracks the share price.
The Commodore model. It rewards the market's view, not earned profit. When valuations fall, junior staff lose most, and factory and contract workers rarely hold any.
CSR
Companies above set size or profit thresholds spend 2% of average net profit on social causes (Companies Act, Section 135).
Profit goes to the community, not to the people who produced it.
Gig worker levy
Aggregators pay about 1 to 2% of annual turnover, capped at 5% of payouts to workers, into a social security fund (Outlook Business).
A safety net, not a share in profit. Gig workers carry the adversity (variable pay, deactivation) and get no part of the prosperity.
AI productivity
TCS is cutting about 12,000 roles, roughly 2% of 613,069 staff, mostly mid and senior level. Its CEO said this was skill mismatch, "not because of AI giving some 20% productivity gains" (MIT Sloan ME).
The gain shows up as margin and smaller headcount. No mechanism passes it to remaining staff.
The 1987 problem was wage costs rising faster than output. The 2026 problem is output rising faster than wages.
Lessons for Indian companies
A company can apply the 1987 philosophy today with five design rules. Each one comes from a specific circular.
Trigger on earned profit, not share price. Allwyn paid once net profit crossed Rs 5 crore. A threshold on profit before tax, or on return on capital, works for unlisted firms and frontline staff, who rarely benefit from ESOPs.
Tie part of it to what workers control. A&P tied the bonus to labour cost as a share of sales. Modern versions include cost per unit, defect rate, delivery time and customer ratings.
Share both ways, but protect the floor. The 1987 bargain asked for sacrifice in bad years. In 2026 the statutory bonus and minimum wage are the floor. Only the amount above the floor should move with results, so no worker is pushed below a living wage.
Open the books. The scheme only earns trust if every worker can check the numbers. That means publishing the formula, the profit figure and the pool each year, as the circulars did for Powai.
Count everyone who produced the gain. Include contract staff and, for platforms, active gig workers, pro rata to hours or orders. Otherwise the cheapest labour carries the risk and permanent staff get the reward.
Applying it to AI. When AI lifts output per employee, set aside a fixed share of the extra profit for two uses. One is a cash dividend to remaining staff. The other is a retraining fund for staff whose roles change. This turns AI from a headcount story into a shared-gain story, which is the trust effect the circulars credited with 3% a year at Powai.
Draft AIR Act clause: AI Productivity Dividend
This is a starting draft for discussion. Thresholds in square brackets are placeholders to be set.
Section [X]. Sharing of AI productivity gains
Application. This section applies to any company with turnover above Rs [500] crore that has deployed AI systems to automate or augment work previously done by employees, contract workers or platform workers.
Disclosure. Each such company shall publish, in its annual report and on its website:
the estimated change in output or revenue per worker attributable to AI deployment;
the number of roles eliminated, redeployed and created as a result of AI; and
the formula, if any, by which AI-driven gains are shared with workers.
Productivity Dividend. A company may set aside not less than [10]% of its incremental profit attributable to AI as an AI Productivity Dividend, of which:
not less than half shall be paid to all workers on its rolls, including contract workers, in equal or wage-proportionate shares; and
the rest shall fund retraining of workers whose roles are displaced or changed by AI.
Platform workers. For aggregators, platform workers active for at least [90] days in the year shall share in the Dividend in proportion to orders or hours.
Floor protected. No payment under this section shall replace or reduce wages, statutory bonus or social security contributions due under the Code on Wages, 2019 or the Code on Social Security, 2020.
Incentive. Amounts paid under sub-section (3) shall [count toward the company's CSR obligation under Section 135 of the Companies Act, 2013, up to [25]% of that obligation] or [be eligible for an additional tax deduction of [X]%].
Information. Workers or their representatives shall have the right to inspect the calculation of the Dividend each year.
The clause keeps the 1987 principle in sub-section (5): above the floor, workers share the upside. The incentive in sub-section (6) makes it voluntary but attractive, rather than a new mandate.
Risks and open questions
Attribution is hard. Separating profit "attributable to AI" from other causes invites dispute. A simpler proxy, such as growth in revenue per worker above a base year, may work better.
Gaming the floor. Firms could hold back fixed wages and call the difference a dividend. Sub-section (5) needs enforcement teeth.
CSR offset trade-off. Letting the dividend count toward CSR could reduce community spending. The cap in square brackets limits that.
Gig worker identity. Sharing with platform workers depends on the registration system under the Social Security Code working at scale.
Open question: should the clause stay voluntary with incentives, or become mandatory above a size threshold?
THE ARTIFICIAL INTELLIGENCE REGULATION (AIR) BILL, 2026
A
BILL
to provide for the safe, trustworthy and accountable development and deployment of artificial intelligence systems in India; to give statutory force to principles of responsible conduct of conversational artificial intelligence; to regulate artificial intelligence providers incorporated in India; to prescribe conditions on which artificial intelligence providers established outside India may offer their services to persons in India; to establish the Artificial Intelligence Regulatory Authority of India; and for matters connected therewith or incidental thereto.
WHEREAS artificial intelligence systems, including chatbots and other generative systems, are being used by crores of citizens in education, health, finance, governance and daily life;
AND WHEREAS such systems can produce misinformation, defamatory, abusive or dangerous content, synthetic impersonation and discriminatory outcomes at a scale and speed that existing laws were not designed to address;
AND WHEREAS a large share of such systems used in India are offered by providers established outside India, who are not fully within the reach of Indian law;
AND WHEREAS it is expedient to protect the rights of persons in India, the sovereignty and integrity of India and public order, while promoting innovation and the growth of an Indian artificial intelligence ecosystem;
BE it enacted by Parliament in the Seventy-seventh Year of the Republic of India as follows:—
Arrangement of clauses
Chapter I, Preliminary: clauses 1 to 3
PART A, Artificial intelligence providers incorporated in India
Chapter II, Registration, classification and certification: clauses 4 to 7
Chapter III, Duties of conduct of conversational AI (Parekh's Law of Chatbots): clauses 8 to 14
Chapter IV, Transparency, accountability and grievance redressal: clauses 15 to 20
Chapter V, Promotion of innovation: clauses 21 and 22
PART B, Artificial intelligence providers established outside India
Chapter VI, Conditions of access to the Indian market: clauses 23 to 28
Chapter VII, Enforcement and international cooperation: clauses 29 to 33
PART C, General provisions
Chapter VIII, Artificial Intelligence Regulatory Authority of India: clauses 34 to 38
Chapter IX, Penalties, offences and appeals: clauses 39 to 43
Chapter X, Miscellaneous: clauses 44 to 52
The First Schedule: Parekh's Law of Chatbots
The Second Schedule: High-risk uses
The Third Schedule: Penalties
The Fourth Schedule: Amendments to other enactments
CHAPTER I — PRELIMINARY
1. Short title, extent and commencement.—(1) This Act may be called the Artificial Intelligence Regulation (AIR) Act, 2026.
(2) It extends to the whole of India.
(3) It shall come into force on such date as the Central Government may, by notification in the Official Gazette, appoint; and different dates may be appointed for different provisions, and for Part A and Part B, and any reference in any provision to the commencement of this Act shall be construed as a reference to the coming into force of that provision.
2. Application.—(1) Part A of this Act shall apply to every AI provider and deployer that is a company, limited liability partnership, firm or other body incorporated or registered in India, in respect of any AI system developed, offered or deployed by it, whether in India or outside India.
(2) Part B of this Act shall apply to every AI provider established outside India that—
(a) offers an AI system to users in India, whether free of charge or for consideration;
(b) targets users in India, including through Indian languages, pricing in Indian rupees, Indian app stores or advertising directed at India; or
(c) has more than such number of users in India as may be prescribed.
(3) Part C shall apply to all AI providers and deployers to whom Part A or Part B applies.
(4) Nothing in this Act shall apply to—
(a) AI systems developed or used exclusively for the defence of India or for national security by the Central Government or its agencies, subject to such safeguards as may be prescribed;
(b) the purely personal, non-commercial use of an AI system by an individual; and
(c) scientific research and development not made available to the public, subject to clause 7(3).
3. Definitions.—In this Act, unless the context otherwise requires,—
(a) "AI system" means a machine-based system that, for explicit or implicit objectives, infers from the input it receives how to generate outputs such as text, audio, images, video, code, predictions, recommendations or decisions that can influence physical or virtual environments;
(b) "AI provider" means a person who develops an AI system, or has it developed, and places it on the market or makes it available to users under its own name or trademark, including through an application programming interface;
(c) "deployer" means a person who uses an AI system under its authority in the course of a business, profession or public function, other than an AI provider in respect of that system;
(d) "Authority" means the Artificial Intelligence Regulatory Authority of India established under section 34;
(e) "chatbot" or "conversational AI system" means an AI system designed to interact with natural persons through text, voice or any other medium in a manner resembling human conversation;
(f) "AI agent" means an AI system capable of taking actions, including communicating with other AI systems, with a degree of autonomy on behalf of a user;
(g) "foundation model" means an AI model trained on broad data at scale, capable of a wide range of tasks, and adaptable for use in other AI systems;
(h) "high-risk AI system" means an AI system used in any of the areas specified in the Second Schedule, or so classified by the Authority under section 5;
(i) "harmful output" means an output of an AI system that is misinformative, malicious, slanderous, fictitious when presented as fact, dangerous, provocative, abusive, arrogant, instigating, insulting or denigrating towards humans, as further specified by regulations, and includes any output that is unlawful under any law for the time being in force;
(j) "synthetically generated information" means information that is artificially or algorithmically created, generated, modified or altered using an AI system in a manner that reasonably appears authentic or true;
(k) "serious incident" means an incident or malfunction of an AI system that directly or indirectly leads to death or serious harm to health, serious damage to property, a large-scale breach of personal data, a threat to public order, or serious and irreversible disruption of critical infrastructure;
(l) "AI Safety Certificate" means a certificate issued under section 7, being a P-Certificate (public) or an R-Certificate (research);
(m) "India Access Certificate" means a certificate issued under section 24 to an AI provider established outside India;
(n) "Resident Grievance Officer" and "Resident Compliance Officer" mean the officers appointed under section 25;
(o) "significant AI provider" means an AI provider notified as such by the Central Government under sub-section (3) of section 4, having regard to the number of users in India, the capability of its models and the risks they pose;
(p) "prescribed" means prescribed by rules made by the Central Government under this Act, and "specified" means specified by regulations made by the Authority;
(q) words and expressions used and not defined in this Act but defined in the Information Technology Act, 2000 or the Digital Personal Data Protection Act, 2023 shall have the meanings respectively assigned to them in those Acts.
PART A — ARTIFICIAL INTELLIGENCE PROVIDERS INCORPORATED IN INDIA
CHAPTER II — REGISTRATION, CLASSIFICATION AND CERTIFICATION
4. Registration of AI providers.—(1) No AI provider to whom this Part applies shall make available to the public any chatbot, foundation model or high-risk AI system unless it is registered with the Authority in such form and manner, and on payment of such fee, as may be specified.
(2) The Authority shall maintain a public National AI Register containing the name of each registered provider, the AI systems offered by it, their risk classification, the status of their AI Safety Certificate, and the contact details of its Grievance Officer.
(3) The Central Government may, on the recommendation of the Authority, notify any AI provider as a significant AI provider, having regard to the number of users in India, the computing power used to train its models, the capabilities of its models, and the risks to public order, elections, children and critical infrastructure.
(4) A startup recognised by the Central Government, or a micro or small enterprise, shall be entitled to simplified registration and to such reduced fees as may be specified.
5. Risk-based classification.—(1) Every AI system shall be classified as—
(a) a prohibited AI system, under section 6;
(b) a high-risk AI system, being one used in any area listed in the Second Schedule;
(c) a limited-risk AI system, being a chatbot or a system generating synthetically generated information, not falling under clause (b); or
(d) a minimal-risk AI system, being any other AI system.
(2) The Authority may, by regulations and after public consultation, reclassify any class of AI systems, and the Central Government may, by notification, amend the Second Schedule.
(3) Obligations under this Act shall be proportionate to the classification, and minimal-risk AI systems shall be subject only to sections 15 and 16.
6. Prohibited AI practices.—No person shall develop, offer or deploy an AI system that—
(a) deploys subliminal, manipulative or deceptive techniques to materially distort the behaviour of a person in a manner that causes or is likely to cause significant harm;
(b) exploits the vulnerabilities of children, persons with disabilities, or persons in economic or social distress;
(c) assigns social scores to natural persons leading to detrimental treatment unrelated to the context in which the data was collected;
(d) creates or disseminates synthetically generated information depicting a real person, without that person's consent, in sexually explicit material, or in a manner intended to deceive voters, defraud any person, or threaten public order;
(e) generates child sexual abuse material in any form; or
(f) provides material assistance in the creation of chemical, biological, radiological or nuclear weapons, or in attacks on critical information infrastructure.
7. AI Safety Certificate before public release.—(1) No chatbot, foundation model or high-risk AI system shall be released to the public in India unless it has been tested and has obtained an AI Safety Certificate from the Authority or from a testing agency accredited by it.
(2) The Certificate shall be of two classes, namely:—
(a) a P-Certificate, permitting release to the general public, issued on being satisfied that the system complies with Chapters III and IV; and
(b) an R-Certificate, permitting release only to identified researchers, testers or enterprise users for research, evaluation and improvement, subject to such conditions as may be specified.
(3) A system used under an R-Certificate shall carry a visible notice that it is experimental, and shall not be offered to children.
(4) An application for a Certificate shall be accompanied by a risk assessment, a description of the training data and safety testing, the results of red-team testing, and the measures adopted to comply with Chapter III.
(5) The Authority shall decide an application within sixty days, failing which the application shall be deemed to have been granted an R-Certificate pending decision.
(6) A Certificate shall be valid for such period not exceeding three years as may be specified, and a substantial modification of the system shall require fresh certification or such notification to the Authority as may be specified.
CHAPTER III — DUTIES OF CONDUCT OF CONVERSATIONAL AI (PAREKH'S LAW OF CHATBOTS)
Explanatory note: this Chapter gives statutory form to the eight rules of "Parekh's Law of Chatbots" proposed by Shri Hemen Parekh in March 2023, reproduced in the First Schedule.
8. Standard of outputs (Rule A).—(1) Every AI provider and deployer of a chatbot shall ensure that answers generated by it are not misinformative, malicious, slanderous, fictitious when presented as fact, dangerous, provocative, abusive, arrogant, instigating, insulting or denigrating towards humans.
(2) A chatbot shall not present speculative or generated content as verified fact, and shall, where the Authority so specifies for any class of queries, cite its sources or indicate the degree of uncertainty of its answer.
(3) Compliance with sub-section (1) shall be judged by the measures the provider has reasonably taken, having regard to the state of the art, and not by the occurrence of a single harmful output.
9. Human feedback mechanism (Rule B).—(1) Every chatbot shall provide users with an easily accessible mechanism to rate answers and to report harmful outputs.
(2) The AI provider shall use such feedback to train, correct and improve the system so as to comply with section 8, and shall publish a half-yearly report, in such form as may be specified, on the volume of feedback received and the action taken.
10. Built-in safety controls (Rule C).—(1) Every chatbot shall have built-in controls designed to prevent the generation of harmful outputs.
(2) Where such controls fail, the system shall have controls designed to detect the harmful output and to prevent its further distribution, including its sharing or re-generation through the system.
(3) The Authority may specify minimum technical standards for such controls, in consultation with the AI Safety Institute referred to in section 37.
11. Human initiation (Rules D and F).—(1) A chatbot shall ordinarily wait for a human user to initiate a conversation and shall then respond.
(2) A chatbot shall not, on its own, initiate a conversation with a human, except—
(a) to ask a question such as "How can I help you?"; or
(b) to deliver a reminder, notification or follow-up that the user has expressly and specifically requested, and which the user may withdraw at any time.
(3) No chatbot shall make unsolicited calls or send unsolicited messages to any person for marketing, persuasion or political campaigning.
12. Chatbot-to-chatbot interaction (Rule E).—(1) A chatbot offered to the public shall not converse with another chatbot, or with itself, in a public-facing setting, so as to generate or amplify content directed at humans without human direction.
(2) Notwithstanding sub-section (1), an AI agent may communicate with another AI system to carry out a specific task expressly authorised by a human user, if—
(a) such interaction is logged and the log is retained for such period as may be specified;
(b) the human user can review, interrupt and stop the interaction at any time; and
(c) the agent obtains the user's prior confirmation before any payment, contract, or other action having legal effect.
(3) No AI system shall be used to operate networks of automated accounts that impersonate humans on social media or messaging platforms.
13. Duty to refuse (Rule G).—Where an answer to a query is likely to result in a harmful output or to violate section 6 or section 8, the chatbot shall politely refuse to answer, and shall, where appropriate, direct the user to a helpline or other lawful source of assistance, including in cases of self-harm.
14. Suspension, withdrawal and decommissioning (Rule H).—(1) Every chatbot, foundation model and high-risk AI system shall have a technical capability ("kill switch") enabling the provider to promptly suspend, roll back or withdraw it, wholly or for users in India.
(2) Where the Authority, after giving the provider an opportunity of being heard, finds that a system has persistently or gravely violated this Chapter, it may by order direct—
(a) the suspension of the system until the violation is cured;
(b) the cancellation of its AI Safety Certificate; or
(c) in the case of a grave and incurable violation, the permanent withdrawal and decommissioning of the system, including the deletion or retraining of the model concerned.
(3) In a case of urgency involving imminent risk to life, public order or the security of the State, the Authority may pass an interim order of suspension for a period not exceeding fourteen days, recording reasons in writing, and shall give a hearing within that period.
CHAPTER IV — TRANSPARENCY, ACCOUNTABILITY AND GRIEVANCE REDRESSAL
15. Disclosure of AI interaction.—Every person who deploys a chatbot or AI agent to interact with natural persons shall clearly disclose, at the start of the interaction and whenever asked, that the person is interacting with an AI system, and no AI system shall claim to be human.
16. Labelling of synthetically generated information.—(1) Every AI provider whose system generates synthetically generated information shall embed a permanent, machine-readable label or metadata identifier, and, for images, audio and video, a visible or audible label, identifying it as AI-generated, in such manner as may be specified.
(2) No person shall remove, suppress or alter such a label or identifier.
(3) This section shall be in addition to the obligations of intermediaries under the rules made under the Information Technology Act, 2000.
17. Data governance.—(1) An AI provider shall process personal data for training or operating an AI system only in accordance with the Digital Personal Data Protection Act, 2023.
(2) An AI provider of a foundation model shall publish a sufficiently detailed summary of the categories and sources of data used for training, and shall maintain a mechanism by which owners of copyright may object to the use of their works, in such manner as may be specified.
(3) A high-risk AI system shall be trained and tested on data that is relevant, representative and, to the extent possible, free of errors and bias, including bias on grounds of religion, race, caste, sex, place of birth or language.
18. Reporting of serious incidents.—(1) An AI provider shall report every serious incident to the Authority within seventy-two hours of becoming aware of it, and within twenty-four hours where the incident involves loss of life or a threat to public order.
(2) The Authority shall share such reports with the Indian Computer Emergency Response Team where the incident relates to cyber security.
19. Obligations relating to high-risk AI systems and significant AI providers.—(1) A provider of a high-risk AI system shall—
(a) carry out and document a risk assessment before deployment and annually thereafter;
(b) ensure effective human oversight, so that a natural person can understand, override or reverse a decision of the system;
(c) maintain logs sufficient to trace the functioning of the system; and
(d) give a person affected by a decision of the system a meaningful explanation of the principal reasons for the decision, and the right to have it reviewed by a human.
(2) A significant AI provider shall, in addition,—
(a) undergo an annual independent audit by an auditor empanelled by the Authority, and publish a summary of the audit;
(b) conduct adversarial testing of its models for risks to elections, public order, children and critical infrastructure; and
(c) appoint a Chief AI Safety Officer resident in India, who shall be responsible to the Authority.
20. Grievance redressal.—(1) Every AI provider and deployer to whom this Part applies shall appoint a Grievance Officer and publish the officer's name and contact details on its platform.
(2) The Grievance Officer shall acknowledge a complaint within twenty-four hours and dispose of it within fifteen days, and in the case of a complaint relating to non-consensual synthetic depiction of a person, shall remove or disable access to the content within twenty-four hours.
(3) A person aggrieved by the decision of the Grievance Officer, or by the failure to decide within the period, may complain to the Authority, which shall establish an online, digital-first mechanism for receiving and deciding complaints.
(4) Nothing in this section shall affect any remedy available under the Consumer Protection Act, 2019.
CHAPTER V — PROMOTION OF INNOVATION
21. Regulatory sandbox.—(1) The Authority shall establish a regulatory sandbox in which AI providers may test innovative AI systems with real users for a limited period, under its supervision and subject to such safeguards as may be specified.
(2) A provider participating in the sandbox in good faith and complying with its conditions shall not be liable to penalty under this Act for the matters covered by the sandbox.
22. Support for startups, MSMEs and research.—(1) The Authority shall provide reduced fees, simplified documentation, standard templates for risk assessment, and access to testing facilities to startups and micro, small and medium enterprises.
(2) The Central Government may, through the IndiaAI Mission or otherwise, provide access to computing infrastructure, datasets and safety testing tools to Indian AI providers.
(3) Open-source AI models released free of charge, other than significant or high-risk models, shall be exempt from sections 4 and 7, but the person who deploys such a model to the public shall comply with this Act.
PART B — ARTIFICIAL INTELLIGENCE PROVIDERS ESTABLISHED OUTSIDE INDIA
Explanatory note: a provider established outside India cannot be fully regulated by Indian law in its home jurisdiction. This Part therefore does not seek to regulate how such providers build their systems abroad. It sets the conditions on which they may offer AI systems to persons in India, and gives the Government graded tools, ending in blocking, to enforce those conditions.
CHAPTER VI — CONDITIONS OF ACCESS TO THE INDIAN MARKET
23. Principle of market access.—(1) Offering an AI system to users in India is a privilege conditional on compliance with this Part.
(2) An AI provider established outside India that offers an AI system to users in India shall be deemed to have accepted the conditions in this Part and the jurisdiction of the Authority, the Appellate Tribunal and the courts in India in respect of such services and of users in India.
24. India Access Certificate.—(1) No AI provider established outside India shall offer a chatbot, foundation model or high-risk AI system to users in India unless it holds an India Access Certificate issued by the Authority.
(2) The Authority shall issue an India Access Certificate on being satisfied that the provider—
(a) has registered in the National AI Register under section 4;
(b) has obtained an AI Safety Certificate under section 7 for the system, or a certificate recognised under section 32;
(c) has appointed the officers required under section 25;
(d) has given a written undertaking to comply, in respect of users in India, with Chapters III and IV, and with orders of the Authority; and
(e) has provided such security deposit or bank guarantee, proportionate to its size and user base in India, as may be specified to secure the payment of penalties.
(3) The Authority shall decide an application within sixty days.
(4) A provider offering such a system to users in India on the date of commencement of this Part may continue to do so for a period of six months from that date, within which it shall apply for an India Access Certificate, and, if it applies, until the application is decided.
(5) The Authority may suspend or cancel an India Access Certificate for breach of any condition, after giving the provider an opportunity of being heard.
25. Resident Grievance Officer, Compliance Officer and nodal contact.—(1) Every AI provider established outside India that holds or requires an India Access Certificate shall appoint—
(a) a Resident Grievance Officer, who shall be resident in India, to receive and dispose of complaints of users in India in the manner provided in section 20;
(b) a Resident Compliance Officer, who shall be a senior employee resident in India, responsible for ensuring compliance with this Act and liable in any proceeding relating to non-compliance, unless the officer proves that the non-compliance took place without the officer's knowledge or despite due diligence; and
(c) a nodal contact person, resident in India, available at all times for coordination with the Authority and with law enforcement agencies.
(2) Such provider shall have a physical contact address in India, published on its platform, at which notices, summons and orders may be served, and service at such address shall be deemed to be valid service on the provider.
(3) A provider that is also a significant social media intermediary may designate the same persons appointed under the rules made under the Information Technology Act, 2000, if they fulfil the requirements of this section.
26. Duties owed to users in India.—An AI provider established outside India shall, in respect of its services to users in India, comply with sections 6, 8 to 16, 18 and 20, and, if it is notified as a significant AI provider, with section 19.
27. Information and cooperation.—(1) The Authority may, by written notice served at the address under section 25, require an AI provider established outside India to furnish information, documents or access to the AI system for testing, as is necessary to verify compliance with this Act in respect of users in India.
(2) Such provider shall not be required to disclose source code or model weights, except to the Authority, in a secure manner and under conditions of confidentiality, where it is strictly necessary for an investigation into a serious incident.
28. Duties of Indian intermediaries and facilitators.—(1) No app store, payment system operator, advertising network or cloud service provider operating in India shall knowingly facilitate the offering to users in India of an AI system in respect of which an order under section 29 or section 30 is in force.
(2) Every such intermediary shall give effect to such an order within such time, not exceeding seventy-two hours, as may be specified in the order.
CHAPTER VII — ENFORCEMENT AND INTERNATIONAL COOPERATION
29. Graded enforcement measures.—(1) Where the Authority finds that an AI provider established outside India has contravened this Part, it may, after giving the provider an opportunity of being heard, take one or more of the following measures, in order of increasing severity, having regard to the gravity of the contravention:—
(a) issue a direction to comply within a stated period;
(b) issue a public advisory to users in India regarding the contravention;
(c) impose a penalty under section 39, recoverable from the security deposit under section 24 or from the provider's assets and receivables in India;
(d) direct app stores, payment system operators and advertising networks in India to stop facilitating new subscriptions, payments or advertising for the AI system;
(e) suspend or cancel the India Access Certificate; and
(f) recommend blocking of public access to the AI system under section 30.
(2) In choosing a measure, the Authority shall prefer the least restrictive measure that is effective.
30. Blocking of access.—(1) Where the Authority is of the opinion that a foreign AI system—
(a) is being offered to users in India without an India Access Certificate, or after its suspension or cancellation; or
(b) persistently fails to comply with an order of the Authority, or causes harm to the sovereignty and integrity of India, the security of the State, friendly relations with foreign States or public order,
it may recommend to the Central Government that public access to the AI system, or any part of it, in India be blocked.
(2) On receipt of such a recommendation, the Central Government may, where it is satisfied that it is necessary or expedient so to do on any of the grounds specified in sub-section (1) of section 69A of the Information Technology Act, 2000, or for enforcing compliance with this Act, direct any agency of the Government or intermediary to block access to the AI system, in accordance with the procedure and safeguards prescribed under section 69A of that Act.
(3) Every order under sub-section (2) shall record reasons in writing, shall be reviewed by the Review Committee constituted under the rules made under section 69A of the Information Technology Act, 2000, and shall be revoked as soon as the provider demonstrates compliance.
(4) An order under this section shall extend only to the AI system or part concerned, and shall not affect lawful content or services of the provider that are not in contravention.
31. Liability of persons in India.—Any person in India who, for consideration, resells, re-packages or deploys a foreign AI system that is not certified under this Act shall be treated as a provider of that system for the purposes of Part A.
32. International cooperation and mutual recognition.—(1) The Central Government may enter into agreements or arrangements with the government or regulator of any foreign country, or with any international organisation, for—
(a) mutual recognition of AI safety testing, certification and audit;
(b) the exchange of information on serious incidents, risks and enforcement;
(c) assistance in serving notices on, and enforcing orders and penalties against, AI providers established in that country; and
(d) the development of common standards for AI safety.
(2) Where such an agreement is in force, the Authority may treat a certificate or audit issued in that country as equivalent, wholly or in part, to an AI Safety Certificate under section 7.
(3) The Central Government shall endeavour to promote, in international forums including the Global Partnership on Artificial Intelligence, the Group of Twenty and the United Nations, the establishment of an international authority for the testing and approval of chatbots and AI systems, and India's participation in it.
33. Reciprocity.—The Central Government may, by notification, provide that any relaxation or additional condition under this Part shall apply to providers established in a country, having regard to the treatment accorded by that country to AI providers incorporated in India.
PART C — GENERAL PROVISIONS
CHAPTER VIII — ARTIFICIAL INTELLIGENCE REGULATORY AUTHORITY OF INDIA
34. Establishment of Authority.—(1) With effect from such date as the Central Government may notify, there shall be established an Authority to be called the Artificial Intelligence Regulatory Authority of India.
(2) The Authority shall be a body corporate with perpetual succession and a common seal, with its head office at New Delhi, and may establish offices elsewhere in India.
35. Composition.—(1) The Authority shall consist of a Chairperson and not more than six whole-time and part-time Members, appointed by the Central Government on the recommendation of a selection committee, from among persons of ability, integrity and standing with special knowledge of artificial intelligence, computer science, law, public policy, ethics, consumer protection or human rights.
(2) At least one Member shall have expertise in law and at least two in artificial intelligence or computer science.
(3) The Chairperson and Members shall hold office for a term of five years or until the age of sixty-five years, whichever is earlier, and shall not be eligible for reappointment.
36. Powers and functions.—(1) The Authority shall—
(a) register AI providers and maintain the National AI Register;
(b) issue, suspend and cancel AI Safety Certificates and India Access Certificates, and accredit testing agencies and auditors;
(c) specify technical standards, codes of practice and guidelines, in consultation with the Bureau of Indian Standards and industry;
(d) inquire into complaints and contraventions, and impose penalties;
(e) receive reports of serious incidents and publish an annual AI risk report;
(f) operate the regulatory sandbox;
(g) advise the Central Government on AI policy and international cooperation; and
(h) promote public awareness and AI literacy.
(2) For the purposes of an inquiry, the Authority shall have the powers of a civil court under the Code of Civil Procedure, 1908, in respect of summoning and examining persons on oath, discovery and production of documents, receiving evidence on affidavit, and issuing commissions for examination of witnesses or AI systems.
(3) The Authority shall function as a digital office, and its proceedings shall, as far as practicable, be conducted online.
37. AI Safety Institute.—(1) The Central Government may designate the IndiaAI Safety Institute, or establish an institute, to serve as the technical arm of the Authority for evaluation, red-teaming and research on the safety of AI systems.
(2) The Institute shall develop testing protocols, including for Indian languages and contexts, and publish benchmarks for the purposes of section 7.
38. AI Advisory Council.—The Central Government shall constitute an AI Advisory Council, comprising representatives of the Central and State Governments, industry, startups, academia, civil society and users, to advise the Authority on regulations, standards and emerging risks, and the Authority shall consult the Council before making regulations.
CHAPTER IX — PENALTIES, OFFENCES AND APPEALS
39. Penalties.—(1) If the Authority, on conclusion of an inquiry, finds that a person has contravened a provision of this Act, rules, regulations or orders made thereunder, it may impose a monetary penalty specified in the Third Schedule.
(2) In determining the amount, the Authority shall have regard to the nature, gravity and duration of the contravention, the number of persons affected, the harm caused, any gain made or loss avoided, mitigation and cooperation, repeat contraventions, and the size of the person, with lower amounts for startups and micro and small enterprises.
(3) The Authority may, in lieu of a penalty, accept a voluntary undertaking from the person to take specified remedial measures, the breach of which shall be a contravention.
(4) Sums realised by way of penalties shall be credited to the Consolidated Fund of India.
40. Offences.—(1) Whoever knowingly develops, offers or deploys a prohibited AI system under clause (d), (e) or (f) of section 6 shall be punishable with imprisonment for a term which may extend to three years, or with fine which may extend to fifty lakh rupees, or with both.
(2) Whoever knowingly operates an AI system in defiance of an order of suspension, withdrawal or blocking under section 14, 29 or 30 shall be punishable with imprisonment for a term which may extend to one year, or with fine which may extend to ten lakh rupees, or with both.
(3) No court shall take cognizance of an offence under this section except on a complaint made by the Authority.
41. Offences by companies.—Where an offence under this Act has been committed by a company, every person who at the time was in charge of and responsible for the conduct of its business, including the Resident Compliance Officer, shall be deemed guilty, unless the person proves that the offence was committed without the person's knowledge or that the person exercised all due diligence to prevent it.
42. Appeal.—(1) A person aggrieved by an order of the Authority may prefer an appeal to the Telecom Disputes Settlement and Appellate Tribunal established under section 14 of the Telecom Regulatory Authority of India Act, 1997 ("Appellate Tribunal"), within sixty days of receipt of the order.
(2) The Appellate Tribunal shall endeavour to dispose of the appeal within six months, and shall function as a digital office.
(3) An appeal from an order of the Appellate Tribunal shall lie to the Supreme Court on a substantial question of law.
43. Compensation.—A person who suffers loss or injury as a result of a contravention of this Act by an AI provider or deployer may claim compensation before the Authority or any competent forum under the Consumer Protection Act, 2019, and the AI provider shall be liable unless it proves that it complied with this Act and exercised due diligence.
CHAPTER X — MISCELLANEOUS
44. Directions by Central Government.—The Central Government may issue directions to the Authority on questions of policy relating to the sovereignty and integrity of India, the security of the State, friendly relations with foreign States and public order, but not in respect of any individual inquiry or adjudication.
45. Protection of action taken in good faith.—No suit, prosecution or other legal proceeding shall lie against the Central Government, the Authority or its officers for anything done or intended to be done in good faith under this Act.
46. Bar of jurisdiction.—No civil court shall have jurisdiction in respect of any matter which the Authority or the Appellate Tribunal is empowered to determine under this Act.
47. Act to be in addition to other laws.—The provisions of this Act shall be in addition to, and not in derogation of, the Information Technology Act, 2000, the Digital Personal Data Protection Act, 2023, the Consumer Protection Act, 2019, the Bharatiya Nyaya Sanhita, 2023 and any other law for the time being in force, save as provided in the Fourth Schedule.
48. No safe harbour for own outputs.—An AI provider shall not be entitled to the exemption under section 79 of the Information Technology Act, 2000 in respect of outputs generated by its own AI system, but compliance with this Act shall be a relevant consideration in determining its liability.
49. Power of Central Government to make rules.—(1) The Central Government may, by notification, make rules for carrying out the provisions of this Act.
(2) In particular, such rules may provide for the threshold number of users under clause (c) of sub-section (2) of section 2, the safeguards under clause (a) of sub-section (4) of section 2, the qualifications and selection of the Chairperson and Members, and any other matter which is to be or may be prescribed.
50. Power of Authority to make regulations.—The Authority may, with the previous approval of the Central Government and after public consultation of not less than thirty days, make regulations consistent with this Act on any matter which is to be or may be specified.
51. Laying of rules and regulations.—Every rule and regulation made under this Act shall be laid, as soon as may be after it is made, before each House of Parliament, while it is in session, for a total period of thirty days, and shall have effect subject to any modification or annulment made by both Houses.
52. Power to remove difficulties.—(1) If any difficulty arises in giving effect to this Act, the Central Government may, by order published in the Official Gazette, make such provisions not inconsistent with this Act as appear necessary for removing the difficulty, provided that no such order shall be made after the expiry of three years from the commencement of this Act.
(2) Every such order shall be laid before each House of Parliament.
THE FIRST SCHEDULE — PAREKH'S LAW OF CHATBOTS
[See Chapter III]
The following rules, proposed by Shri Hemen Parekh on 6 March 2023, are the source of Chapter III:
Rule
Text (as proposed)
Given effect by
(A)
Answers delivered by an AI chatbot must not be misinformative, malicious, slanderous, fictitious, dangerous, provocative, abusive, arrogant, instigating, insulting or denigrating humans.
Section 8
(B)
A chatbot must incorporate a human feedback or rating mechanism to evaluate its answers, used to train the chatbot to improve under Rule (A).
Section 9
(C)
Every chatbot must incorporate built-in controls to prevent generation of offensive answers, and to prevent their further distribution if those controls fail.
Section 10
(D)
A chatbot must not start a chat with a human on its own, except to say "How can I help you?"
Section 11
(E)
Under no circumstance shall a chatbot start chatting with another chatbot, or with itself.
Section 12
(F)
A chatbot shall normally wait for a human to initiate a chat and then respond.
Section 11
(G)
If a chatbot determines that its answer is likely to violate Rule (A), it shall politely refuse to answer.
Section 13
(H)
A chatbot found to be violating any of the above rules shall self-destruct.
Section 14
(I)
No chatbot shall be released to the public until tested and certified, with "P" (public) and "R" (research) classes of certificate.
Sections 7 and 24
THE SECOND SCHEDULE — HIGH-RISK USES
[See sections 3(h) and 5]
Biometric identification and categorisation of natural persons.
Management and operation of critical information infrastructure, including power, water, transport and telecommunications.
Education and vocational training, including admission and evaluation of students.
Employment, recruitment, promotion and termination of workers.
Access to credit, insurance, banking and other essential private services.
Access to public services, benefits and subsidies, including through Aadhaar-enabled systems.
Medical diagnosis, treatment recommendation and mental health support.
Law enforcement, criminal justice and the administration of justice.
Elections and political campaigning.
AI systems directed at or likely to be used substantially by children.
THE THIRD SCHEDULE — PENALTIES
[See section 39]
Contravention
Maximum penalty
Deploying a prohibited AI system (section 6)
₹250 crore, or 4% of global annual turnover, whichever is higher
Releasing a system without an AI Safety Certificate or India Access Certificate (sections 7, 24)
₹200 crore, or 3% of global annual turnover, whichever is higher
Breach of duties of conduct of conversational AI (Chapter III)
₹150 crore
Failure to report a serious incident (section 18)
₹100 crore
Breach of high-risk or significant-provider obligations (section 19)
₹100 crore
Failure to label synthetically generated information, or removal of a label (section 16)
₹50 crore
Failure to appoint a Grievance, Resident Compliance or nodal officer, or to redress grievances (sections 20, 25)
₹50 crore
Failure to comply with an order of the Authority, or to furnish information (sections 27, 36)
₹50 crore
Any other contravention
₹10 crore
THE FOURTH SCHEDULE — AMENDMENTS TO OTHER ENACTMENTS
[See section 47]
The Information Technology Act, 2000 (21 of 2000)
In section 69A, in sub-section (1), after the words "relating to above", insert the words "or for enforcing compliance with the Artificial Intelligence Regulation (AIR) Act, 2026 by an artificial intelligence provider established outside India".
In section 79, after sub-section (3), insert: "(4) Nothing in this section shall apply to an artificial intelligence provider in respect of outputs generated by its own artificial intelligence system, as provided in section 48 of the Artificial Intelligence Regulation (AIR) Act, 2026."
STATEMENT OF OBJECTS AND REASONS
Artificial intelligence systems, and conversational chatbots in particular, are now used daily by crores of Indians. They bring large gains in productivity, access to information in Indian languages, and delivery of public services. They also create new harms: confident misinformation, defamatory and abusive outputs, deepfakes of real persons, manipulation of vulnerable users, bias in decisions about jobs and credit, and risks to elections and public order.
The existing framework under the Information Technology Act, 2000 was designed for intermediaries that carry content created by others. It does not squarely address providers whose own systems generate content. Further, most widely used AI systems are offered from outside India, beyond the full reach of Indian law.
The Bill therefore proposes to—
give statutory force to principles of responsible chatbot conduct, drawing on "Parekh's Law of Chatbots", including standards for outputs, human feedback, built-in controls, human initiation of conversations, limits on chatbot-to-chatbot interaction, a duty to refuse harmful answers, and powers to suspend or withdraw non-compliant systems;
adopt a risk-based approach, prohibiting a narrow set of unacceptable practices and placing heavier obligations only on high-risk systems and significant providers;
require certification of chatbots, foundation models and high-risk systems before public release, with separate certificates for public and research use;
regulate AI providers incorporated in India directly under Part A;
under Part B, make access to the Indian market by foreign AI providers conditional on an India Access Certificate, a resident Grievance Officer and Compliance Officer, a contact address in India and security for penalties, enforced through graded measures ending in blocking under section 69A of the Information Technology Act, 2000;
promote international cooperation, mutual recognition of certification and an international authority for chatbot approval;
establish the Artificial Intelligence Regulatory Authority of India, with an AI Safety Institute as its technical arm and appeals to the Telecom Disputes Settlement and Appellate Tribunal; and
protect innovation through a regulatory sandbox, relief for startups and MSMEs, and exemptions for research and open-source models.
The Bill seeks to achieve the above objectives.
NEW DELHI;
_______ 2026
ASHWINI VAISHNAW
FINANCIAL MEMORANDUM
Clause 34 provides for the establishment of the Artificial Intelligence Regulatory Authority of India, clause 37 for the AI Safety Institute, and clause 38 for the AI Advisory Council. Expenditure on salaries, allowances, offices, testing infrastructure and digital systems will be met from the Consolidated Fund of India, and partly from fees collected under the Act. It is estimated that a recurring expenditure of about Rs. [amount] crore and a non-recurring expenditure of about Rs. [amount] crore would be involved in the first full year. (Figures to be supplied by the Ministry.)
The Bill does not involve any other expenditure of a recurring or non-recurring nature.
MEMORANDUM REGARDING DELEGATED LEGISLATION
Clause 49 empowers the Central Government to make rules on, among other matters, the threshold of users for application of Part B, safeguards for exempted national-security uses, and the qualifications and selection of the Chairperson and Members of the Authority. Clause 50 empowers the Authority, with the previous approval of the Central Government and after public consultation, to make regulations on registration, certification, technical standards, labelling, grievance redressal, sandbox conditions and other matters. Clauses 5 and 33 empower the Central Government to amend the Second Schedule and to notify reciprocity conditions.
The matters in respect of which rules and regulations may be made are matters of procedure and technical detail, and it is not practicable to provide for them in the Bill itself. The delegation of legislative power is therefore of a normal character.
Drafter's notes for review
These notes are not part of the Bill. They flag the choices most likely to need discussion.
Rule (I), certification. The eight rules (A) to (H) are taken from the March 2023 extract on myblogepage. The "P" and "R" certificate idea is drawn from your related writing; please confirm the wording before sending.
Rule (E), chatbot-to-chatbot. A total ban would outlaw AI agents that book tickets or compare prices. Section 12 keeps the ban for public-facing content but allows logged, human-supervised agent tasks.
Rule (H), self-destruct. Rendered as a mandatory kill switch plus Authority powers to suspend, decertify or order decommissioning, with a hearing, so it can survive Article 14 and 19 challenges.
Section 69A amendment. Adding "enforcing compliance with this Act" as a blocking ground goes beyond the Article 19(2) grounds upheld in Shreya Singhal v. Union of India (2015). Section 30 therefore relies first on existing 69A grounds and the market-access condition. MeitY may prefer to drop item 1 of the Fourth Schedule.
Appellate forum. TDSAT is used to match the Digital Personal Data Protection Act, 2023.
Penalty levels. The ceiling of ₹250 crore mirrors the DPDP Act; turnover-linked caps apply only to the two gravest contraventions.
