AI-Proof: What Machines Can Never Be — Volume II: Application, Part V: Conscious Strategy, Chapter 25: Reading the Regulatory Map
Part V looks at strategy. After speed in Chapter 24, this chapter turns to the rules: what the laws for machines ask, why they differ by country and year, and what no law can do.
A report scored a drug company’s systems against duties they did not have, by a deadline that moved. The laws for machines differ by country and year. Each can make someone answer. None can answer in the room.
In Brief For readers who build, buy or run machines that decide about people, and wonder which rules apply. Chapter 25 maps the laws for machines as they stood in October 2026: the European Union, the United States, China, India and its neighbours, Africa and Latin America. An operations head in Hyderabad finds that a report has scored her systems against duties they never had. In the same city, a state welfare algorithm cancelled ration cards without notice. An edict of the emperor Aśoka and the legal philosopher Lon Fuller disagree about whether law cares for people like a nurse or treats them as people who must know the rules. Together they help answer what regulation is for, and what no rule can do.
A law for machines keeps the same commitments in one place and one year, so every map of it must carry a date; a rule can make an institution name who answers and give the people it decides about a way to ask, but it cannot do the answering in the room.
Opening: The Two Scores in Genome Valley
“Ninety-one,” the consultant said. “And twenty-nine.”
It was the third week of April 2026, in a meeting room on the seventh floor of NovaChem’s site in Genome Valley, Hyderabad. NovaChem is a drug company with its head office in Basel. Meera Reddy had been its vice-president of operations in Hyderabad for seven years, and with the company for fifteen. The report on the table had come from the company’s global legal team in Basel. A consultancy had written it, and one of its partners had flown in to present it. It scored seventeen of NovaChem’s machine systems against the European Union’s law on artificial intelligence, the AI Act. Five of the systems ran in Meera’s operations, one of them for the company’s drug-safety function in Basel. On human oversight her five scored ninety-one out of a hundred. On technical documentation they scored twenty-nine.
The five were familiar to her. A model sorted reports of side effects from NovaChem’s clinical trials around the world, for the drug-safety physicians in Basel. A system read batch records in quality assurance. In customer service, a chatbot answered customers’ written questions. A second model listened to their calls and routed them by how upset the callers sounded, and the report scored the two as one system. A forecast predicted demand for the supply chain, and an agent compared suppliers for procurement. Meera read the list twice. Then she asked one question.
“Which of these five does the Act cover?”
“We scored all of them as high-risk,” the partner said. “It is the cautious reading.”
The cautious reading had a cost, and the twenty-nine was part of it. The scores were the consultancy’s own; the Act gives none. They measured her systems against duties that, as Meera would learn within a month, none of them had. An adviser is seldom blamed for finding too much risk, and a low score was also an argument for the consultancy’s next phase of work. The report was dated, too. Its deadline was 2 August 2026, the date on which the Act’s heaviest duties were then due. Before the report was four months old, that date had moved.
So this chapter asks: When the rules for machines differ at every border and change every year, what is regulation for, and what must still be done by the people it governs?
The print edition treated regulation as terrain. A builder was to read it and position early, because early compliance “creates compounding advantage”, and reading it was a skill that “requires consciousness”. This edition withdraws both claims, for reasons this chapter will show. It holds something plainer. Laws for machines are one way a society keeps two of this book’s commitments. The first is that a machine’s makers and users will never be free of what it does. The second is that being intelligent will never give a machine the right to decide for people. Those commitments do not change. The ways of keeping them do. Indian ethics has a phrase for what must be weighed when a rule is applied, and I borrow it for what changes: deśa, kāla, pātra, place, time and person.1 A law is made in one country, in one year, for some people and not others.
That is why a map of these laws must carry a date, the way an honest watch says “water-resistant to fifty metres” instead of “waterproof”. Every claim about a law in this chapter is dated, and the map as a whole is as of late October 2026. When the dates change, the map must be corrected where readers can see it. The map is also selective. It follows places that keep the commitments in different ways, not every country. And it leaves out war and national security, as Europe’s Act itself does for systems used only for military, defence or national-security purposes.2
Two questions run through every section. Each of these laws names someone who must answer for a machine, often more than one: the company that sells it, the one that uses it, a public office. The first question is to whom they must answer: a regulator, a court, a customer, or the person the machine decided about. The second is whether that person can ask why, and be put right. Beside them sits a limit, which the last section argues for. No law can decide in the room, where a person must apply it to the case in front of her.
The chapter’s title needs one correction. As of 2026, machines read the regulatory map very well, and this chapter’s first story shows one doing it. What a machine does not hold is the choice of what to do with the reading. Nor does it answer for the choice.
25.1 The EU AI Act and What It Means for Builders
Europe’s law is the AI Act.2 Formally Regulation (EU) 2024/1689, it entered into force on 1 August 2024 and applies in stages. Its bans and a duty on AI literacy have applied since 2 February 2025. Duties for the makers of general-purpose models have applied since 2 August 2025. So have its rules on enforcement. Each member state must name authorities to enforce the Act. A company that breaks a ban can be fined up to €35 million or 7 per cent of its worldwide yearly turnover, whichever is higher. Small and medium-sized firms face whichever of the two is lower.2
The bans come first in the Act, in Article 5. These are not technical specifications. They are moral declarations. Some uses are not risky but wrong. In plain words, the Act forbids machines that manipulate people in ways that cause them significant harm, or exploit their age, disability or social or economic situation. It forbids social scoring, by companies as well as governments. It forbids predicting that a person will commit a crime from profiling alone. It forbids building face-recognition databases by “the untargeted scraping of facial images from the internet or CCTV footage”. It forbids systems that “infer emotions of a natural person in the areas of workplace and education institutions”, except for medical or safety reasons. It forbids sorting people by their bodies to infer their race, politics or beliefs. And it forbids the police to identify people in public places by face or other bodily features in real time, with narrow exceptions. In 2026 the lines grew. An amending regulation added two bans, from 2 December 2026. One covers machines that make realistic sexual images, video or audio of an identifiable person without consent. The other covers machines that make child sexual abuse material.3 Lines can be added. That is part of what makes them lines and not specifications.
Below the bans sit the high-risk systems, listed in the Act’s Annex III. These are allowed, with heavy duties: risk management, records, technical documentation, human oversight, and an assessment before sale. Annex III names eight areas: biometrics, critical infrastructure, education, employment, essential private and public services and benefits, law enforcement, migration, and justice and democratic processes.
The machine-side call came before the lawyers’ reply. In late April, Meera gave a model the text of Annex III and a page describing each of her five systems, and asked which points applied. It said that none of the five was listed. Reports of side effects from clinical trials are not in the annex. Neither is the reading of batch records, which is governed by the rules of good manufacturing practice that drug regulators write. The forecast and the procurement agent are not listed either. The chatbot would owe customers a plain statement that they were talking to a machine, under Article 50, if its customers were in the EU. They were in India. The model raised one hard case on its own. The Act lists “AI systems intended to be used for emotion recognition” as high-risk, and routing calls by how upset a caller sounds might count. Is a guess at anger from a voice emotion recognition in the Act’s sense? Lawyers would have to decide, if the system were ever used on callers in the EU. If it did count, the same article would also oblige NovaChem to tell those callers it was in use. NovaChem’s lawyers in Basel replied in May. They agreed on all five.
So the machine was right, and the consultancy’s cautious reading was wrong. Being outside Annex III did not make the five harmless, or free of rules. The side-effect model sorted reports that physicians would act on, and the rules on drug safety already asked a great deal of anyone handling such reports. The batch records answered to good manufacturing practice. “High-risk” in the Act is a legal category, made for that law’s purposes, and a system outside it can still hurt someone. The model’s call also has a limit, and it is dated. A model whose training ended before July 2026 would still give the old August date, unless it searched for what had changed. A model’s knowledge is a photograph taken when its training ended, and laws do not stand still for it. Reading well does not mean reading without error. By June 2026, as an earlier chapter reported, one researcher had listed more than a thousand cases in American courts in which filings contained material invented by AI.41 A model’s reading of a law still has to be checked against the law’s own text, as Basel’s lawyers checked this one.
What, then, had the twenty-nine measured? It had measured the documentation the Act asks of high-risk systems: how a system was designed, what data trained it, how it was tested, and what it gets wrong. Meera’s teams could explain all of this in a meeting. Most of it was not written where a stranger could read it. The print edition called documentation “the adjacent thing”, the proof that follows the real work. It is not adjacent. A record is how people who were not in the room can check what was done there. A practice that no outsider can check does not yet answer to anyone outside. A record must also describe the work as it was done. One written only to fill a template lets an outsider check nothing. Some of the paperwork was waste. Not all of it was.
The date in the report moved because the law moved. The Commission proposed a delay on 19 November 2025, as part of a package it called a “Digital Omnibus”. Parliament and Council agreed on 7 May 2026. Regulation (EU) 2026/1744 entered into force on 27 July 2026.3 It moved the duties for the high-risk systems in Annex III to 2 December 2027, and for AI built into regulated products to 2 August 2028. The proposal tied the new dates to the arrival of the technical standards that would show how to comply, and those standards had not been published. The final law fixed the dates instead.4 The same regulation softened the duty on AI literacy: providers and users must now “take measures to support” it, rather than ensure it. It kept Article 50’s rule that machine-made images, audio, video and text be marked in a form machines can read. Systems already on the market have until 2 December 2026 to comply.
The Act also gives something to the people a machine decides about. Article 86 has applied since 2 August 2026. It covers decisions based on most of the high-risk systems listed in Annex III. The decision must have legal or similarly significant effects on a person, who considers it harmful to her health, safety or rights. She can ask whoever used the system for “clear and meaningful explanations of the role of the AI system in the decision-making procedure and the main elements of the decision taken”. But the rules that class a system as high-risk now apply only from 2 December 2027, so until then the right has little to hold on to. It is also a fallback, applying only where EU law gives no such right already. Since 2018 the EU’s data-protection law, the GDPR, has given one for significant decisions made by a machine alone. Article 86 reaches further, to decisions a person makes on a machine’s output.5 People must also be told. Under the GDPR, a company that will decide by machine alone must say so when it collects the person’s data. Its preamble gives an example of such a decision: “e-recruiting practices without any human intervention”. From December 2027 the AI Act adds a duty to tell people when a high-risk system is used in deciding about them, even when a person makes the decision. A tool that filters job applications is one such system.5 Knowing the machine’s part is a start. It does not show that the decision was right, or put it right. A separate law covers one kind of putting right. Under Directive (EU) 2024/2853, software, including AI systems, counts as a product, and its makers are liable for defects without anyone having to prove fault. It applies to products placed on the market after 8 December 2026. A maker cannot limit or exclude this liability towards the injured person by a term in a contract, such as a disclaimer in the terms of sale. The directive gives compensation for death, injury, damage to private property and the loss or corruption of private data, not for a benefit wrongly refused.5 For other harms, contract and consumer law decide, country by country, what a disclaimer is worth.
The Act reaches beyond Europe. It applies to providers and users outside the EU when a system’s output is used inside it.2 NovaChem’s head office is not in the EU at all. Switzerland chose another route. On 12 February 2025 its government decided to bring the Council of Europe’s convention on AI into Swiss law. It would change its other laws sector by sector where needed, with a draft due by the end of 2026.6 That convention is the first binding international treaty on AI. It opened for signature on 5 September 2024, and the European Union, the United States, the United Kingdom, Japan, Canada and Switzerland are among those who signed it. As of October 2026 only the European Union had ratified it, on 15 May 2026, and it was not yet in force. India has not signed it.7
25.2 The US Approach: From Federal Order to State Patchwork
The United States has no general federal AI law. In January 2025 the new administration revoked its predecessor’s order on AI safety and issued its own, titled “Removing Barriers to American Leadership in Artificial Intelligence”.8 In December 2025 a further order set up an AI Litigation Task Force in the Justice Department to challenge state AI laws. It also asked for a federal framework that would override them.8 An executive order directs federal agencies. It does not repeal a state’s law. A state’s law gives way only to the Constitution or to federal law made by Congress or under its authority, and the courts decide when it does. As of early October 2026 Congress had passed no such law.9 Older federal laws still reach machines. In December 2023 the pharmacy chain Rite Aid agreed with the Federal Trade Commission not to use facial recognition for surveillance for five years. The commission had charged that it used the technology without reasonable safeguards.9
So the rules written for AI are made in the states. A colleague of Meera’s in NovaChem’s office in New Jersey once summed them up for her in one line: “There is no map. There is a weather system.”
Colorado shows what that means. In May 2024 it passed the first broad state law on high-risk AI, due to take effect on 1 February 2026. In August 2025 a new law delayed it to 30 June 2026. In April 2026 the AI company xAI sued to stop it, the federal Justice Department joined the suit, and a federal court paused its enforcement with the agreement of the state’s attorney general. On 14 May 2026 the governor signed a third law that repealed the first and replaced it, mostly from 1 January 2027.10 The new law drops the risk programmes and the impact assessments. It keeps something smaller. A consumer must be told before a covered system is used on them. After a decision that goes against them, they can seek an explanation, correct their data, and perhaps obtain a review by a person. The state’s attorney general is to write the rules for this.
The print edition used Colorado to make a point about judgement. A software platform, it said, “can tell you what Colorado requires today. It cannot tell you whether Colorado’s framework will become a national template or an outlier. That judgement is human.” In the same paragraph, human judgement said that Colorado’s law “takes effect in mid-2026”. It never took effect. Forecasting a legislature is a task, it carries a date, and nobody does it well. The call goes to no one, and the claim about human judgement falls with it.
What survived in Colorado is worth noticing. It is answering in its smallest legal form: a notice, a reason, a chance to correct the record and ask for a person. Other American rules are narrower still. New York City has required bias audits of automated hiring tools, and notice to candidates, since July 2023.11 California’s law of September 2025 requires the largest developers of frontier models to publish how they manage catastrophic risks and to report serious incidents.12 One older federal right gives a reason, in one field. Since 1977 a lender that refuses credit has owed the applicant its specific reasons. The rule that carries this out says that a failed score on the lender’s own scoring system is not reason enough.9
A law that changes three times in three years has a cost of its own, and an American legal philosopher named it. In The Morality of Law (1964), Lon Fuller told the story of a king, Rex, who wanted to be a good lawgiver and failed in eight different ways. Fuller drew the lesson that the attempt to make and keep a body of legal rules “may miscarry in at least eight ways”.13 Among them are “rules that require conduct beyond the powers of the affected party”, and “introducing such frequent changes in the rules that the subject cannot orient his action by them”. A company in Denver could not settle its plans by Colorado’s law, because the law kept changing before it began. Companies across Europe were asked to meet high-risk duties before the standards that would show how to meet them existed. As I apply Fuller, putting those duties off was a repair, since the agreed ways of showing compliance did not yet exist. The heads of more than forty European companies had asked for a pause of two years in July 2025, so the repair also suited them.4 Fuller’s test looks only at whether a rule can be followed. It does not ask who wanted the change. The cost was in the timing. From November 2025 until late July 2026 nobody could be certain that the August date would move, and the law that moved it took effect six days before it.
Fuller does not say that law must never change. A law for a technology that moves this fast cannot stand still. The British scholar David Collingridge described the trap in 1980.14 Early on, a technology is easy to change, but we cannot yet see what needs changing. By the time we can see it, change has become slow and costly. The EU wrote early. Most American states waited. Both paid a price. What fails people is not change itself. It is change that leaves them unable to know what is asked of them.
25.3 China’s AI Governance Model
China has made rules for particular uses, one at a time. Rules for recommendation algorithms took effect on 1 March 2022, and rules for synthetic images and voices followed in January 2023.15 Interim measures for generative AI took effect on 15 August 2023. They require providers to “Uphold the Core Socialist Values”, and some must pass security assessments and file their algorithms with the regulator. Measures on labelling took effect on 1 September 2025. They require visible labels on machine-made content and hidden marks in the files.16 In October 2025 China’s legislature added an article on AI to its Cybersecurity Law, in force from 1 January 2026.17 The new article commits the state to support AI research and infrastructure, refine AI ethics and strengthen the monitoring of AI risks.
There is no comprehensive AI law, but one is now planned. The State Council’s legislative plan for 2026, published in May 2026, calls for work to “speed up comprehensive legislation for the healthy development of artificial intelligence”. The legislature’s own plan lists it among bills still being prepared, for research and drafting, with no date set for a first reading.18 As of October 2026 there is no such law.
Read China’s rules with the chapter’s first question. The rules on algorithms name who must answer. A provider files its algorithm, passes an assessment and labels its output. But the provider answers chiefly to the regulator, and the regulator answers to the state. One reading of the labelling rule is that people have a right to know when what they consume was made by a machine. That may be one of its purposes. The same rules also require content to uphold the core socialist values. A reader of the map should not choose the more flattering motive, or assume that the person decided about can ask nothing. Since November 2021 China’s law on personal information has given a person the right to ask for an explanation of an automated decision that significantly affects her rights and interests, and to refuse one made by a machine alone. If the company or office handling her data refuses, she can sue in a people’s court.15 On paper, this is close to what the GDPR has given Europeans since 2018. How far she can use it, against a large platform or an office of the state, the text alone does not show.
25.4 India, Southeast Asia, and the Laws Still Being Written
India has chosen not to write a general AI law, for now. On 5 November 2025 the Ministry of Electronics and Information Technology published guidelines for AI governance.19 They set out seven principles, which the guidelines call sūtras. Two of them are “People First” and “Innovation over Restraint”. The government chose, as one law firm put it, to steer away from “prescriptive regulation at this stage”. In the government’s own words, the guidelines state “that a new horizontal AI law is not required at this stage”: no single law covering every sector. India regulates AI largely by sector, through the regulators of finance, health and other fields. The guidelines “rely on existing laws”, the government said, and “sectoral regulators will remain responsible for enforcement and oversight”.19 As of October 2026 I found no Indian rule like Europe’s ban on machines that read emotions at work and in schools. The “for now” may be short. On 8 October 2026 the minister for electronics and IT, Ashwini Vaishnaw, said his ministry would publish a consultation paper on regulating AI within a month, with safety and deepfakes among its concerns. As of late October it had not appeared.19
The existing laws, dated, are these. The Digital Personal Data Protection Act was passed in 2023. Its rules were notified on 13 November 2025, and most of its duties apply only from 13 May 2027.20 From May 2027 it will give a person the right to a summary of the data held about her and to have it corrected. It names no right to an explanation of a decision. The consent it asks for must be “specific” and given “for the specified purpose”. As I read it, a person who agrees to one use of her data has not agreed to every other use. And section 7(b) will let the state process personal data to give “such subsidy, benefit” and other services without fresh consent, where the person has agreed to such use before or the data sits in records the government has notified. The Act protects only data about an identifiable individual. As I read it, a firm’s own records, such as what a supplier delivered and when, are left to contract unless they also identify an individual.20 The rules for online platforms were amended on 10 February 2026, in force from 20 February. Machine-made content must now be labelled and carry permanent marks in the file, and platforms must act on government and court orders within three hours.21
The guidelines speak to companies. The hardest case in India is the state itself, and it is in Meera’s city.
In 2016 the government of Telangana, the state whose capital is Hyderabad, began using a system called Samagra Vedika. Al Jazeera reported on it in January 2024, with the Pulitzer Center.22 The system had first been built for the Hyderabad police. It linked data from as many as thirty government databases into a profile of each person. Then it flagged claimants who seemed to break a rule of eligibility, such as owning a four-wheeler. By 2018 it was used in most of the state’s welfare schemes. Between 2014 and 2019, the reporters found, Telangana cancelled more than 1.86 million cards for subsidised food and rejected 142,086 new applications, without notice. At first the government called them fraudulent claims. The reporters found several thousand wrongful exclusions among them.
One of them was a widow in Hyderabad. The family’s card had been cancelled in 2016, and their application for a new one sat for four years. In 2021, while her husband, a rickshaw puller, was still alive, the system tagged him as the owner of a car. It had confused him with another man who did own one, and the application was rejected. Her husband later died of cancer. When she brought evidence, officials admitted the error, then said the family’s income was too high. The reporters found that it was not. When the reporters met her, she had been without rations for more than seven years.
The reporters asked, under the Right to Information Act, for the system’s source code. The state refused, because the company that built it held “rights over” the code. A family cannot go to another ration office, yet the code that judged it was held as a company’s property. A state that buys a system it may not show to the people it judges has given away part of its power to answer to them. People who were excluded had to prove their own eligibility, and officials often sided with the algorithm. The state held the databases; the families held whatever papers they had. What the widow needed was smaller than the code: to be told which record said her husband owned a car.
The state’s side should be heard. The principal secretary of its IT department said that the old system, “based on the discretion of officials”, had been “opaque and misused and led to corruption”, and that the new one had “very high levels of accuracy”. The first half is plausible: discretion over rations is where favours can live. In Jharkhand, economists studied reforms that added fingerprint checks to rations. They found that corruption fell, at a real cost to legitimate beneficiaries, some of whom lost their rations for a time.23 A machine can be fairer than a corrupt clerk on average and still be wrong about a particular widow. In Australia a royal commission called that country’s automated welfare debt scheme “a crude and cruel mechanism, neither fair nor legal”.24
What did India’s law give her? It gave her a court, eventually. In April 2022, hearing an activist’s petition, the Supreme Court ordered Telangana to verify every cancelled card in the field. “You cannot cancel before verifying,” Justice L. Nageswara Rao told the state, as LiveLaw reported.25 The order put the work of checking back on the side that held the records. By July 2022, the reporters found, the state had processed 205,734 of 491,899 applications for verification and approved 15,471. Their reporting mentions no repayment of the rations lost in the years between. Stopping a harm is not the same as putting it right.
India’s highest court had already set the frame. In August 2017 nine judges held unanimously that privacy is a fundamental right.26 In September 2018 the court upheld the national identity number, Aadhaar, for welfare, struck down its use by private companies, and required that any intrusion be proportionate. One judge dissented. Justice D. Y. Chandrachud wrote: “Dignity and the rights of individuals cannot be made to depend on algorithms or probabilities.”27 It was a dissent, but it named the danger the widow met.
The Indian lamp for this chapter is older than any of these laws. Aśoka ruled a vast empire in the third century BCE. His edicts, carved on rocks and pillars, survive in more than thirty places in today’s India, Nepal, Pakistan and Afghanistan. After his armies conquered Kalinga, around 262 BCE, the bloodshed changed him, and he spent the rest of his reign trying to govern by Buddhist principles. In the fourth of his pillar edicts, carved twenty-six years after his coronation, he wrote about the officers he called rajjukas, who heard petitions and gave judgement in the countryside. In Ven. S. Dhammika’s translation:28
“Just as a person feels confident having entrusted his child to an expert nurse thinking: ‘The nurse will keep my child well,’ even so, the Rajjukas have been appointed by me for the welfare and happiness of the people in the country. The hearing of petitions and the administration of justice have been left to the Rajjukas so that they can do their duties unperturbed, fearlessly and confidently. It is my desire that there should be uniformity in law and uniformity in sentencing. I even go this far, to grant a three-day stay for those in prison who have been tried and sentenced to death. During this time their relatives can make appeals to have the prisoners’ lives spared.”
The edict is about judges and prisoners in an ancient empire, not about machines, and Aśoka kept the death penalty: the stay was mercy inside it, not its end. I borrow only its arrangement. One law for everyone. Officers trusted to decide in the room. And a pause between a sentence and what cannot be undone, so that someone may ask. Samagra Vedika kept the first, in its way: one set of rules for every claimant. Aśoka had left the hearing to his officers so that they could act “fearlessly”. Telangana’s officials, the reporters found, often deferred to the system, so no one was free to decide in the room. And it had no stay. Cards were cancelled first, and people proved their eligibility afterwards. In effect, in 2022 the Supreme Court put back the stay.
Set the same widow in Europe, and the map changes, though not mainly because of the AI Act. Under the GDPR she could see the data held about her, including the record that called her husband a car owner, and have it corrected. If a machine alone had decided, she could also ask for meaningful information about its logic. In February 2025 the EU’s Court of Justice said what that information must be: the procedure and principles actually applied to her data, in a form she can understand. A trade secret does not simply end the matter. The office that must answer her has to show the withheld material to a regulator or a court, which weighs it against her right.5 From December 2027 the AI Act will add a right for decisions that an official takes on the output of a benefits system, where EU law gives none already. In Telangana, the vendor’s claim ended the reporters’ request for the code, though India’s own Act lets an official release such material when the public interest is greater.20 In Hyderabad, as of October 2026, no law gives her a right to an explanation of the machine’s part. An older duty applies. The Right to Information Act requires public offices to give affected people the reasons for their administrative decisions, yet cards like the family’s were cancelled without notice.22 From May 2027 India’s data law adds a plainer duty. Anyone who uses personal data to make a decision that affects a person must ensure the data’s “completeness, accuracy and consistency”. The duty binds the state too, unless the government exempts an office, as the Act allows.20 The Act’s rights to see and correct one’s data are tied to consent. As I read the text, it does not say plainly whether they reach records the state takes from its own registers. The Act’s fines are paid into the government’s own fund, not to the person harmed. Where a person stands decides what she may ask, and of whom. That is deśa, place, and pātra, the person, in one case.
Elsewhere in Asia, countries have chosen differently again. Vietnam passed an AI law on 10 December 2025, in force from 1 March 2026. It sorts systems into three tiers of risk, requires foreign providers of high-risk systems to keep a contact in Vietnam, and requires machine-made images, audio and video to be marked.29 South Korea’s AI Basic Act took effect on 22 January 2026. It asks for explanation and human oversight for “high-impact” systems and labels on generated content, with a year’s grace before fines.30 Japan passed an act in May 2025 to promote AI, with no penalties.31 Singapore hosted the meeting at which ASEAN, the association of Southeast Asian states, endorsed a regional guide on AI in February 2024.32 Countries can share a sea and not a map.
25.5 Africa and Latin America: Regulation as Leapfrog
Africa and Latin America are not simply behind Europe. In July 2024 the African Union’s Executive Council endorsed a continental strategy on AI in Accra. “AI systems must be adapted to local realities,” the AU commissioner Amani Abou-Zeid said, in a statement the AU published in August 2024.33 Nigeria published a national AI strategy in September 2025. It binds no one, and as of July 2026 several AI bills were before its National Assembly and none had passed.34
A late regulator can skip an early regulator’s mistakes. Kenya shows how, and what it costs. Kenya passed a general law on data protection in 2019, two years before the company that would test it began work there. A project called Worldcoin scanned irises in Nairobi during a pilot from 2021. At its global launch in July 2023 it began scanning Kenyans’ irises at scale, in return for cryptocurrency tokens, and by early August more than 350,000 Kenyans had signed up.35 On 2 August 2023 the interior ministry suspended it. The law was already there to apply. On 5 May 2025 Kenya’s High Court ruled against the company. It could not collect biometric data in Kenya without a lawful assessment of the risks to privacy, as the 2019 Act requires. It also found fault with consent “obtained through inducement by cryptocurrency”. It ordered the company “to permanently delete, within seven days and under the supervision of the Data Protection Commissioner”, the data it had collected unlawfully.36
The leap worked. Kenya did not need an AI law, because it had a data law written for no particular technology. But look at the dates. The irises were scanned from 2021 to 2023. The order to delete them came in 2025. The ministry’s suspension in August 2023 was a stay, but it came after more than 350,000 people had signed up. A law on paper needs an office to enforce it, and here it also needed a rights group to bring the case. Late regulators can skip early mistakes. They can also find that a law written in good time still arrives late in court.
Chile wrote a new right before the harm was common. In October 2021 its Congress amended the constitution. In one English rendering, the amendment says that “scientific and technological development must be at the service of people”, and that the law must protect “cerebral activity” and the information that comes from it.37 On 9 August 2023 the Supreme Court ordered a company that sold a headset reading brain signals to erase the data it held on a former senator. Brazil’s AI bill passed the Senate in December 2024. It has been with the Chamber of Deputies since March 2025, and as of late October 2026 it was still awaiting its rapporteur’s report.38 It is a bill, not a law.
The print edition ended this section with advice. Builders should join regulators’ consultations early, because the relationships would “become a powerful competitive moat”. That is withdrawn. The American economist George Stigler warned in 1971 that regulation tends to be “acquired by the industry and is designed and operated primarily for its benefit”.39 A company that is safe because it is close to the regulators is protected by its access, whatever its quality. Sharing what a builder knows in a public consultation is useful. Turning the relationship into an advantage is capture. With machines there is a quieter kind. A regulator that must ask a company what its system does, and how to test it, depends on that company for the facts.
25.6 How Regulation Reshapes Where to Build
Divergence is not a mistake. Different laws for the same machine follow from different judgements about what a machine may do to a person, and those judgements belong to the people who live under them. Not every state asks them. Other countries’ rules are also not smaller versions of the EU’s. India’s three-hour orders, China’s filing and labels, Vietnam’s local representative and Colorado’s notices are not parts of the EU’s rules. They are different rules. Regulating by sector has its own strength. A law about AI in general must list its uses in advance, while a regulator who knows a trade knows where in the work a person signs. But rules made sector by sector leave a gap where the state is its own user. Samagra Vedika had no regulator but the government that ran it, until a court stepped in.
Building where the rules are light, to avoid the cost of heavier ones, works only while a product stays at home. The EU’s law reaches any system whose output is used in the EU, wherever its maker sits. A product that crosses a border brings that border’s rules with it, and software crosses borders easily.
The map of three models that this chapter follows is the legal scholar Anu Bradford’s. She calls them the American market-driven model, the Chinese state-driven model and the European rights-driven model.40 She also named the Brussels effect: companies that sell everywhere often follow the EU’s rule everywhere, because one standard is cheaper than many. She argues that this happens. It does not make other countries’ laws smaller copies of Europe’s.
For a company, divergence becomes a set of decisions about where to build, and someone makes each one. In June 2026 NovaChem’s legal team in Basel made two such decisions. Systems used across borders, such as the human resources department’s tools for hiring and promotion, would be documented once, to the strictest rule they met. Systems used in one country would be documented for that country’s rules. The first choice costs more for systems that may never reach Europe. The second costs more if they do. Four of Meera’s five fell under the second. The fifth, the adverse-event model, served Basel’s drug-safety physicians with reports from trials around the world, so it fell under the first, and Basel’s drug-safety function would document it. Her teams were told to stop writing to the AI Act’s template, and the consultancy’s next phase was cancelled.
The decisions cost least for the companies that can afford a legal team in Basel. A small exporter fills in each buyer’s questionnaire himself, and each new rule raises the price of entering a trade. The answer to that is simpler ways for small firms to show what they do, not an exemption from showing it.
25.7 The Moral Dimension: What Regulation Is For
The print edition said that the best regulation “draws moral lines first and builds procedure outward from there”. Fuller disagreed with that order, and the disagreement is worth having. For him, procedure is not the servant of the lines. It has a morality of its own. A law that nobody can know, that changes too often to plan by, or that asks the impossible, wrongs the people it governs, even if its aims are kind. Fuller argued that to govern people by rules at all is to treat each of them as a responsible agent.13 Such a person can understand and follow rules, and can be held answerable for failing to. His critics answered that the eight ways describe how to rule well by rules, and say nothing about whether the rules are good. H. L. A. Hart, reviewing the book, said that the poisoner’s art has its principles too, and that to call them a morality would confuse working well for a purpose with being good.13 So both are needed: lines like the bans, and procedure that lets a person know a rule and contest its use.
Aśoka’s edict sees the governed differently. In his image they are a child, and the state is a nurse the king can trust. His officers decide for them, kindly and without fear, and the king keeps a stay in reserve for mercy. Fuller’s subject is not a child. He must be able to know the rules in advance, and to ask why when they are used against him.
The two partly answer different questions. Aśoka asks who should decide in the room, and with what care. Fuller asks what makes a rule a rule. But they pull against each other in the cases this chapter has met. The welfare system in Telangana was the nurse without the stay: it decided for people who could not see its reasons, and it acted before anyone could ask. Colorado’s three laws in three years, and a European date left in doubt until the months before it, failed Fuller’s test from the other side. They asked the governed to plan by rules that would not stay still. Europe’s rights to an explanation treat a person as Fuller’s agent, and so does Colorado’s notice. The Supreme Court’s order to verify before cancelling was Aśoka’s stay, restored by a court.
Here is what regulation is for, on this account. It governs institutions. It can require a company or a public office to name who must answer for a machine. It can require records that a stranger can check, and a notice to people when a machine has decided about them. It can give them a way to ask why, and a remedy when the answer is bad. Courts and tribunals had begun this without any AI law. In 2024 a Canadian tribunal held an airline responsible for what its chatbot had told a grieving customer about fares.41
What regulation cannot do is decide in the room. Aśoka’s edict governed the rajjukas; the rajjuka decided. A law can require a person to oversee a machine and give them the power to override it. It can require the seat. It cannot fill it with someone able to see what the machine missed, or give that person the time to look and the standing to disagree without fear. Those are the institution’s work, and it pays for them in hours or does not.
This is why the book’s commitments come before any of these laws, and outlast them. A society keeps them through law in its own place and year. It can also abandon them. Where no law yet asks, the makers and users of a machine are still not free of what it does. They simply have no regulator to answer to, only the people the machine touched.
The other side
The strongest objection to a seat in the room comes from Ben Green, of the University of Michigan. In 2022 he surveyed 41 policies that require people to oversee the algorithms governments use.41 He found two flaws. First, “evidence suggests that people are unable to perform the desired oversight functions”. Second, such policies “legitimize government uses of faulty and controversial algorithms”. A person named beside the machine makes the system look checked. When it fails, the blame lands on the person, and the system stays.
His remedy is “a shift from human oversight to institutional oversight”. Before a system is adopted, the agency must justify its use, and any oversight it proposes, with evidence. That justification then goes to public review and approval.
Part of this wins. A law that asks only for a person beside the machine can do exactly the harm he describes. Green studied public bodies, and applying his point to companies is my step. My reply is that his remedy keeps the seat. It asks the institution to prove that the seat can be filled, by a person with the evidence, the time and the skill the work needs. That is the institution’s work this section has named, with a duty added to show it. Whether such proof should go to the public, and not only to a regulator or a head office, this chapter leaves open.
Ending: A Name in the Intended Use
In early October 2026 the consultancy’s report was still on a shelf in Meera’s office on the seventh floor. Its cover gave 2 August 2026 as the deadline. That day had passed with nothing due for her five systems.
The paper on her desk was newer, and it was still a draft. The European rules of good manufacturing practice, which NovaChem followed for the medicines it made, were to gain a new annex on AI. Inspectors from Europe’s medicines agencies had drafted it with an international group of inspectorates, and the public consultation had closed in October 2025.42 Where a model gives “an input to a decision made by a human operator”, and where “the effort to test such model has been diminished”, the draft said, “the description of the intended use should include the responsibility of the operator”. It added: “In this case, the training and consistent performance of the operator should be monitored like any other manual process.”
The batch-record review system gave such an input. Whether its testing had been diminished enough for the draft to apply was arguable, and the draft would not have required what Meera did next in every case. She chose not to wait for the argument. The system’s operator was a quality manager in her operations, Madhavi Krishnamurthy. Madhavi had read batch records for about a decade, and she still signed for every batch that left her area. Meera wrote her name in the description of the system’s intended use.
Then she wrote the second part, which was her own choice of means. To monitor an operator’s consistent performance, there had to be a performance to monitor. Madhavi and her analysts would read one batch record in twenty by hand, in full, before opening the system’s reading. Each month Madhavi would set their readings beside the system’s and write down where they differed. The purpose was not to catch what the system missed, and not to train beginners, though it might do both. It was to keep the person named in the seat able to sit in it. The hours came from the time the system had saved. The monthly report to Basel would show them, and show the slower month.
No law yet required the hours, and nobody had asked for them. Meera signed the request and sent it to Basel with the month’s figures.
Practice
The map of one system. Choose one system that makes or shapes decisions about people where you work, or one that decides something about you. On one page, write today’s date. Then write four things. First, which law, in which country, asks anything of this system, and from which date. Second, to whom someone must answer under that law: a regulator, a court, a customer, or the person decided about. Third, whether a person the system decides about is told, what she can ask, of whom, and whether anyone can put a bad decision right. Fourth, what no law asks of the system that you think those people are owed, and whose time would pay for it. Mark each line you took from a machine, and check it against the law’s own text. Then put the page somewhere you will find it, and read it again in a year, when some of it will be wrong. This takes an hour; if the hour is your employer’s, ask for it. It will not make anyone compliant, and it promises nothing about a job. It shows how much of the answering has been written down, and where.
Words to keep
- Annex III: the list in the EU’s AI Act of the areas where AI systems count as high-risk, such as employment, education and public benefits.
- Brussels effect: Anu Bradford’s name for the way companies that sell everywhere often follow the EU’s rules everywhere.
- deśa, kāla, pātra: place, time and person; in this chapter, what differs in how a society keeps the same commitments through law.
- emotion recognition: in the AI Act, a system that infers people’s emotions or intentions from their biometric data; banned at work and in schools except for medical or safety reasons, high-risk elsewhere.
- GDPR: the European Union’s data-protection law (Regulation (EU) 2016/679), which has applied since May 2018; among other things, it gives a person rights about decisions made about her by a machine alone.
- high-risk system: in the AI Act, a legal category for systems that are allowed but bound by heavy duties of risk management, documentation and human oversight. A system outside it can still do harm.
- human in the loop: a person who makes the decision a model gives input to; in the EU’s draft rules for medicine manufacturing, where such a model has been tested less, someone whose responsibility is written down and whose performance is monitored.
- rajjuka: an officer of the emperor Aśoka who heard petitions and gave judgement in the countryside.
- regulatory capture: rules written and run mainly for the benefit of the firms they are meant to govern (Stigler); with machines, also a regulator’s dependence on those firms for the facts about their systems.
- sūtra: a short rule or principle; India’s AI guidelines of 2025 call their seven principles sūtras.
Sources
- Bhagavad Gītā 17.20, where a gift is rightly given deśe kāle ca pātre ca, at the right place, time and person; later ethical writing uses deśa-kāla-pātra for what must be weighed when a rule is applied. See the Prologue, note 16.
- Regulation (EU) 2024/1689 (the Artificial Intelligence Act), Article 2(1)(c) (third-country providers and deployers whose output is used in the Union), Article 2(3) (systems used exclusively for military, defence or national-security purposes excluded), Article 5(1) (prohibited practices; quotations from points (e) and (f)), Article 50 (transparency), Article 86(1) (right to explanation of individual decision-making) and Annex III (high-risk areas; point 1(c), emotion recognition; point 5(a), public assistance benefits); entry into force 1 August 2024; Article 5 and Article 4 from 2 February 2025, general-purpose model duties from 2 August 2025; Article 70(1) (each member state to establish or designate national competent authorities) and Article 99(1), (3) and (6) (penalties: for breach of Article 5, fines “up to EUR 35 000 000” or “up to 7 % of its total worldwide annual turnover for the preceding financial year, whichever is higher”; for SMEs, whichever is lower), from 2 August 2025 under Article 113(b) (artificialintelligenceact.eu, consolidated text with the 2026 amendments, checked 23 October 2026; Articles 70 and 99 checked 30 October 2026).
- Regulation (EU) 2026/1744 of 8 July 2026 (the Digital Omnibus on AI), amending Regulation (EU) 2024/1689: proposed 19 November 2025; provisional agreement 7 May 2026; published 24 July 2026; in force 27 July 2026; Annex III high-risk duties from 2 December 2027 and product-embedded systems from 2 August 2028 (recital 40); new Article 4 (“take measures to support”); new Article 5(1)(ba) and (bb), from 2 December 2026; Article 50(2) for systems already on the market from 2 December 2026 (EUR-Lex; White & Case, “EU agrees Digital Omnibus deal” and “EU AI Omnibus enters into force,” 2026).
- Kommission Arbeitsschutz und Normung (KAN), “Proposed amendments to the AI Act,” 2026 (“the harmonised standards supporting the requirements of the Act have still not been published”); European Parliamentary Research Service, “Parliament’s emerging position on the Digital Omnibus on AI,” At a glance, PE 785.685, March 2026 (new dates tied to the availability of harmonised standards); Gibson Dunn, client alert on the AI Act Omnibus agreement and its postponed high-risk deadlines, 2026 (the final law fixed the dates). Euronews, 3 July 2025 (headline as in its address, “Europe’s top CEOs call for Commission to slow down on AI Act”): the heads of “more than 40 European companies including ASML, Philips, Siemens and Mistral” asked for a “two-year clock-stop”, including on the obligations for high-risk systems.
- Directive (EU) 2024/2853 on liability for defective products (23 October 2024): software, including AI systems, counts as a product; to be transposed by 9 December 2026; Article 2(1) as corrected by the corrigendum of 7 May 2026: it applies to products “placed on the market or put into service after 8 December 2026”; Article 6(1): compensation only for death or personal injury, damage to property, and destruction or corruption of data not used for professional purposes; Article 15: the liability of an economic operator “is not, in relation to the injured person, limited or excluded by a contractual provision or by national law” (EUR-Lex, checked 30 October 2026). See Chapter 15. On the rights EU law already gave: Regulation (EU) 2024/1689, Article 86(3) (the right applies only where Union law does not otherwise provide it), Article 26(11) (deployers of Annex III systems that “make decisions or assist in making decisions related to natural persons” must inform them “that they are subject to the use of the high-risk AI system”; from 2 December 2027) and Annex III, point 4(a) (systems used “to analyse and filter job applications, and to evaluate candidates”); Regulation (EU) 2016/679 (the GDPR), Articles 15(1) (access), 15(1)(h) and 22 (decisions made solely by automated means), 16 (rectification) and 13(2)(f) (information, when the data are obtained, on “the existence of automated decision-making, including profiling, referred to in Article 22(1) and (4)”), with Recital 71 (“e-recruiting practices without any human intervention”), applying since 25 May 2018; Court of Justice of the European Union, Case C-203/22, Dun & Bradstreet Austria, judgment of 27 February 2025 (the controller must explain “the procedure and principles actually applied”; information claimed as a trade secret goes to the supervisory authority or court, which balances the rights at stake), as summarised by A&O Shearman, “CJEU issues judgment on balancing the right of access and protecting trade secrets in automated decision-making processes,” 12 March 2025.
- Federal Office of Communications (OFCOM, Switzerland), “Artificial intelligence”: the Federal Council’s decision of 12 February 2025 to incorporate the Council of Europe convention into Swiss law, to make sector-specific amendments where needed, and to prepare a consultation draft by the end of 2026 (bakom.admin.ch).
- Council of Europe Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law (CETS No. 225), opened for signature 5 September 2024; Council of Europe, “European Union ratifies the Council of Europe Framework Convention on Artificial Intelligence” (15 May 2026); signatures and ratifications as listed in the Netherlands Treaty Database (checked 23 October 2026: the EU the only ratification; no entry into force; India not a signatory).
- Executive Order 14179, “Removing Barriers to American Leadership in Artificial Intelligence,” 23 January 2025 (govinfo.gov), issued after Executive Order 14148 of 20 January 2025 had revoked Executive Order 14110 (govinfo.gov); Executive Order 14365, “Ensuring a National Policy Framework for Artificial Intelligence,” 11 December 2025, sections 3 and 8 (whitehouse.gov).
- Cloud Security Alliance, “State AI Laws Take Hold as Federal Preemption Stalls,” 4 April 2026 (no federal legislation enacted; the Task Force had filed no complaint as of late March 2026); Buchanan Ingersoll & Rooney, “AI Legislative Status,” 31 July 2026 (“Congress has not passed any legislation to establish a federal framework on AI”; the only AI-related law of the 119th Congress concerns AI-generated intimate deepfakes); The Regulatory Review, “The Hidden Costs of AI Preemption,” 7 October 2026 (preemption still a pending proposal). Federal Trade Commission, “Rite Aid Banned from Using AI Facial Recognition After FTC Says Retailer Deployed Technology without Reasonable Safeguards,” press release, 19 December 2023. Equal Credit Opportunity Act, 15 U.S.C. § 1691(d)(2)–(3) (“Each applicant against whom adverse action is taken shall be entitled to a statement of reasons for such action from the creditor”; it must contain “the specific reasons for the adverse action taken”), added by Pub. L. 94-239 of 23 March 1976 and in force for this section from 23 March 1977, twelve months after enactment (law.cornell.edu, effective-date note); Regulation B, 12 CFR § 1002.9(b)(2): statements “that the applicant, joint applicant, or similar party failed to achieve a qualifying score on the creditor’s credit scoring system are insufficient” (consumerfinance.gov, checked 30 October 2026).
- SB 24-205, signed May 2024, effective 1 February 2026; SB 25B-004, signed 28 August 2025, delaying it to 30 June 2026 (Greenberg Traurig, September 2025); SB 26-189 (the Colorado ADMT Act), signed 14 May 2026, mostly effective 1 January 2027, repealing and replacing SB 24-205 (Morrison Foerster, 15 May 2026; Goodwin, June 2026). xAI’s suit against Colorado’s attorney general, filed 9 April 2026 in the US District Court for the District of Colorado, and the Justice Department’s complaint in intervention of 24 April 2026 (National Law Review, 1 May 2026; Jenner & Block, 29 April 2026); the order of Magistrate Judge Cyrus Y. Chung of 27 April 2026 staying enforcement of SB 24-205, on a motion both parties made (Colorado Politics, 28 April 2026).
- New York City Department of Consumer and Worker Protection, “Automated Employment Decision Tools” (Local Law 144 of 2021; enforcement from 5 July 2023).
- California SB 53, the Transparency in Frontier Artificial Intelligence Act, signed 29 September 2025 (Clifford Chance, “SB 53: California sets standards for AI transparency,” October 2025).
- Lon L. Fuller, The Morality of Law, revised edition (New Haven: Yale University Press, 1969; first edition 1964), ch. 2, pp. 33–39 (Rex and the eight ways; the list at pp. 38–39), quoted as block-quoted in James Boyle, “Legal Realism and the Social Contract” (law.duke.edu/boylesite/fuller.htm); p. 162, on the governed as a “responsible agent”, paraphrased. H. L. A. Hart, review of The Morality of Law, Harvard Law Review 78 (1965): 1281–1296, on the poisoner’s art and the difference between efficiency for a purpose and morality, paraphrased.
- David Collingridge, The Social Control of Technology (London: Frances Pinter, 1980), paraphrased; see Chapter 4.
- Provisions on the Management of Algorithmic Recommendations in Internet Information Services, effective 1 March 2022 (Covington, Inside Privacy, January 2022); Provisions on the Administration of Deep Synthesis of Internet Information Services, effective 10 January 2023 (China Briefing, December 2022); Interim Measures for the Management of Generative Artificial Intelligence Services, Articles 4(1), 17 and 24, effective 15 August 2023 (trans. China Law Translate); Personal Information Protection Law of the People’s Republic of China, effective 1 November 2021, Article 24 (an explanation of automated decisions with a significant impact on a person’s rights and interests, and refusal of decisions made solely by automated means) and Article 50 (a lawsuit in a people’s court if a request is refused) (English text as published by Prighter).
- Measures for Labeling Artificial Intelligence-Generated Synthetic Content, issued March 2025, effective 1 September 2025 (Loeb & Loeb, March 2025).
- Cybersecurity Law of the People’s Republic of China, as amended by the NPC Standing Committee on 28 October 2025, new Article 20 on artificial intelligence, effective 1 January 2026 (China Briefing; Caixin Global, 29 October 2025).
- State Council, 2026 Annual Legislative Work Plan, published 11 May 2026: “完善人工智能治理,加快推进人工智能健康发展综合性立法” (my translation), as reported by Chinanews, 3 July 2026; NPC Standing Committee, 2026 Legislative Work Plan, released 11 May 2026, section (三), preparatory review items: legislation on “the healthy development of artificial intelligence” to be researched and drafted, “with review arranged as circumstances permit” (my translation; Chinese text as posted by NPC Observer).
- Ministry of Electronics and Information Technology, India AI Governance Guidelines, 5 November 2025: the seven sūtras, including “People first” and “Innovation over restraint” (S.S. Rana & Co.; AZB & Partners, 11 November 2025, for the sūtras’ capitals and “prescriptive regulation at this stage”); Press Information Bureau, “India AI Governance Guidelines do not allow unrestricted deployment of high-risk AI systems,” 19 December 2025: “The Guidelines state that a new horizontal AI law is not required at this stage”; “the Guidelines rely on existing laws”, which include “the Information Technology Act, the Digital Personal Data Protection Act, and sector-specific regulations”; “sectoral regulators will remain responsible for enforcement and oversight”. The release names no particular use, such as emotion recognition, and I found no Indian law, rule or guideline restricting emotion recognition at work or in schools as of 30 October 2026. On the consultation paper: ThePrint, “India steps up AI regulation push with consultation paper likely next month,” 8 October 2026 (the minister’s remarks that day); MediaNama, “India’s draft AI regulations could be out in November, says IT Minister Ashwini Vaishnaw,” 26 October 2026 (not yet released).
- Digital Personal Data Protection Act, 2023, sections 2(s) (a “person” includes an individual, a company, a firm and “the State”), 2(t) (“personal data” means “any data about an individual who is identifiable by or in relation to such data”; indiacode.nic.in, a2023-22.pdf, checked 30 October 2026), 6(1) (consent “shall be free, specific, informed, unconditional and unambiguous” and shall signify an agreement to processing “for the specified purpose”), 7(b), 8(3) (where personal data is likely to be “used to make a decision that affects the Data Principal”, the Data Fiduciary “shall ensure its completeness, accuracy and consistency”), 11(1)(a) (the right to a summary, against the Data Fiduciary “to whom she has previously given consent, including consent as referred to in clause (a) of section 7”) and 12(1) (correction and erasure of data “for the processing of which she has previously given consent”), 17(4) and 17(5) (exemptions for the State; the power to exempt any Data Fiduciary for a period, by notification, within five years of commencement) and 34 (penalties “shall be credited to the Consolidated Fund of India”) (dpdpa.com, checked 30 October 2026); Digital Personal Data Protection Rules, 2025, notified 13 November 2025, with most duties from 13 May 2027 (Bar & Bench, November 2025), sections 8, 11 to 17 and 28 to 34 among them (G.S.R. 843(E), 13 November 2025, as reproduced by dpdp.myndsolution.com); Right to Information Act, 2005, section 4(1)(d): every public authority shall “provide reasons for its administrative or quasi-judicial decisions to affected persons”; section 8(1)(d): commercial confidence, trade secrets and intellectual property are exempt “unless the competent authority is satisfied that larger public interest warrants the disclosure” (indiacode.nic.in).
- Information Technology (Intermediary Guidelines and Digital Media Ethics Code) Amendment Rules, 2026, notified 10 February 2026, in force 20 February 2026 (Hogan Lovells, February 2026).
- Tapasya, Kumar Sambhav and Divij Joshi, “How an algorithm denied food to thousands of poor in India’s Telangana,” Al Jazeera, 24 January 2024, produced with the support of the Pulitzer Center. The widow is named in the report; she is left unnamed here.
- Karthik Muralidharan, Paul Niehaus and Sandip Sukhtankar, “Identity Verification Standards in Welfare Programs: Experimental Evidence from India,” Review of Economics and Statistics 107, no. 2 (2025): 372–392; NBER Working Paper 26744 (February 2020, revised September 2021), abstract: reforms adding biometric identity requirements, studied with randomised and natural experiments; “Corruption fell but with substantial costs to legitimate beneficiaries”. Chapter 28 tells this study.
- Report of the Royal Commission into the Robodebt Scheme (7 July 2023), as quoted in “Crude, cruel and unlawful: Robodebt Royal Commission findings,” Law Society Journal (NSW), July 2023. See Chapter 6.
- “’You Cannot Cancel Before Verification’: Supreme Court On Telangana Govt Cancelling 19 Lakh Ration Cards; Directs Field Verification,” LiveLaw, 27 April 2022 (Justices L. Nageswara Rao and B. R. Gavai; petition by S. Q. Masood; the remark is Justice Rao’s).
- Justice K. S. Puttaswamy (Retd.) v. Union of India (2017) 10 SCC 1, nine judges, 24 August 2017; Justice K. S. Puttaswamy (Retd.) v. Union of India (Aadhaar), five judges, 26 September 2018 (Columbia Global Freedom of Expression case summary).
- Justice D. Y. Chandrachud, dissenting in Justice K. S. Puttaswamy (Retd.) v. Union of India (Aadhaar) (2019) 1 SCC 1, 26 September 2018, as quoted in Sruthisagar Yamunan and Rohan Venkataramakrishnan, “Majority vs Chandrachud,” Scroll.in, 26 September 2018; see also the Hindustan Times, 27 September 2018.
- Pillar Edict IV (the fourth of the Seven Pillar Edicts), trans. Ven. S. Dhammika, The Edicts of King Asoka (Kandy: Buddhist Publication Society, The Wheel Publication No. 386/387, 1993; Access to Insight edition).
- Baker McKenzie, “Vietnam: Artificial Intelligence Law, foundation and outlook,” February 2026 (passed 10 December 2025; effective 1 March 2026).
- Cooley, “South Korea’s AI Basic Act: Overview and Key Takeaways,” 27 January 2026 (in force 22 January 2026; a one-year grace period before fines).
- White & Case, “Japan’s first AI legislation becomes law: focus is on promoting research and development, no monetary penalties,” 2025 (passed by the Diet on 28 May 2025).
- Personal Data Protection Commission Singapore, “Two Guides Unveiled at ASEAN Digital Ministers’ Meeting (ADGMIN) 2024,” 2 February 2024 (the ASEAN Guide on AI Governance and Ethics); Crowell & Moring, “4th ASEAN Digital Ministers’ Meeting,” 2024 (held in Singapore, 30 January to 2 February 2024; the guide endorsed there).
- African Union, “African Union committed to developing AI capabilities in Africa,” press release, 28 August 2024 (the Continental AI Strategy endorsed by the Executive Council at its 45th Ordinary Session, Accra, 18–19 July 2024; the quotation is Commissioner Amani Abou-Zeid’s).
- Digital Policy Alert, “Nigeria’s AI governance approach,” 5 August 2026 (National AI Strategy 2025–2029, September 2025, non-binding; AI bills under deliberation; no standalone AI law as of July 2026).
- “Kenya becomes first country to suspend Worldcoin,” Fortune, 2 August 2023 (over 350,000 Kenyans signed up, citing Reuters; suspension by the interior ministry that day); for the pilot since 2021, “Worldcoin has been scanning Kenyans’ eyeballs since 2021, officials tell Parliament,” Citizen Digital, 6 September 2023 (Tools for Humanity’s evidence to the parliamentary ad hoc committee: “we have been providing proof of humanness verification since 2021”, at more than 20 locations in Nairobi).
- Katiba Institute, statement on its judicial review application against the Worldcoin Foundation and others, High Court of Kenya (Aburili J), judgment of 5 May 2025, under section 31 of the Data Protection Act, 2019.
- Arent Fox Schiff, “Neural Rights: Landmark Ruling”: Law No. 21.383 of 14 October 2021, amending Article 19 of the Chilean Constitution; the Supreme Court’s judgment of 9 August 2023 against Emotiv (plaintiff Guido Girardi).
- Câmara dos Deputados, PL 2338/2023, ficha de tramitação (received 17 March 2025; “Aguardando Parecer do(a) Relator(a)”; last action 27 October 2026; checked 8 November 2026); passed by the Senate on 10 December 2024.
- George J. Stigler, “The Theory of Economic Regulation,” Bell Journal of Economics and Management Science 2, no. 1 (1971): 3–21.
- Anu Bradford, The Brussels Effect: How the European Union Rules the World (New York: Oxford University Press, 2020); Digital Empires: The Global Battle to Regulate Technology (New York: Oxford University Press, 2023), on “the American market-driven model, the Chinese state-driven model, and the European rights-driven regulatory model” (publisher’s description).
- Moffatt v. Air Canada, 2024 BCCRT 149 (Civil Resolution Tribunal of British Columbia, 14 February 2024), paragraph 27; see Chapter 8. Damien Charlotin, AI Hallucination Cases database (damiencharlotin.com/hallucinations): 1,058 cases in US courts by the start of June 2026, as summarised by the UNC Law Library and told in Chapter 8; 1,514 listed for the USA when checked on 30 October 2026. For The other side in 25.7 (added in this edition): Ben Green, “The Flaws of Policies Requiring Human Oversight of Government Algorithms,” Computer Law & Security Review 45 (2022): 105681 (preprint arXiv:2109.05067), abstract: a survey of “41 policies that prescribe human oversight of government algorithms”; “evidence suggests that people are unable to perform the desired oversight functions”; such policies “legitimize government uses of faulty and controversial algorithms”; the proposal of “a shift from human oversight to institutional oversight”. The agency is to justify, with evidence, both the algorithm and any human oversight it proposes, before public review and approval.
- European Commission, “Stakeholders’ Consultation on EudraLex Volume 4, Good Manufacturing Practice Guidelines: Chapter 4, Annex 11 and New Annex 22” (7 July to 7 October 2025), drafted by the EMA GMDP Inspectors Working Group in cooperation with PIC/S; draft Annex 22, “Artificial Intelligence,” section 3.3, lines 52–56 (PIC/S copy, picscheme.org/docview/9715): “Where a model is used to give an input to a decision made by a human operator (human-in-the-loop), and where the effort to test such model has been diminished, the description of the intended use should include the responsibility of the operator.” Still a draft as of October 2026 (PIC/S news page, checked 23 October 2026; pharmanow.live, 12 August 2026).
Named people in this chapter are composites, and their conversations and circumstances are constructed; the studies, laws and figures cited are real and sourced.
Revised from the print edition (2026)
- What the print edition claimed: that reading regulation “requires consciousness. It is not automatable … reading human systems requires being one”; that “Consciousness is the deeper structure: the disposition of care, attention, and accountability” of which compliance is the expression; that early compliance “creates compounding advantage” and that other jurisdictions’ requirements are “subsets” of the EU’s; that relationships with regulators “become a powerful competitive moat”, that “The opportunity is real. It belongs to the builders who arrive early”, and that regulatory intelligence is “growing more valuable”; that a platform “cannot tell you whether Colorado’s framework will become a national template … That judgement is human”, while dating Colorado’s Act to “mid-2026”; that the EU’s high-risk duties apply “from August 2026” and product duties “by August 2027”; that Article 5 bans “governmental social scoring” and “untargeted facial recognition in public spaces”; that NovaChem’s systems “touched four” Annex III categories; that the EU required “forty-three specific document formats”; that China’s comprehensive AI law was “removed from the national legislative agenda”; that India’s data rules “impose” duties; that Vietnam was “developing” an approach; that Nigeria had published an “AI governance framework” in 2024; that Brazil’s framework was “the most advanced”; that each law expresses its noblest motive (EU dignity, India’s “trust … built through engagement”, China’s “right to know”); that “In Beijing, you listen”; and that “The best regulation draws moral lines first and builds procedure outward from there”. It said Meera had caught the adverse-event bias that “nobody else had thought to look for”, that she “built” the governance team, that she “fought to weight” the safety signals equally, that the trial sites were in “low-income countries”, that she worked on the “eighth floor”, that she had been VP “for six years”, that her Indian skill had made her “complacent”, and that she read Indian rules by feel, the way a textile exporter reads cloth. Its coda had a meeting “one of the most productive of her career” and a sense of what she did not know as “the most valuable thing she had gained all week”. It used the map of Anu Bradford’s Digital Empires without credit, and ended 25.7 on Scanlon’s title as its own question (“What do we owe each other?”).
- What I now hold: laws for machines are one way a society keeps two commitments that do not change (a machine’s makers and users will never be free of what it does; being intelligent will never give a machine the right to decide for people), kept differently by place, time and person; so every map of them is dated and selective, and this one is as of late October 2026. Each law names someone who must answer, often more than one; the questions are to whom, and whether the person decided about is told, can ask why and can be put right. As of 2026 machines read the map well: a model correctly found that none of NovaChem’s five systems in Meera’s operations was high-risk under Annex III, where a consultancy had scored all five as if they were; but “high-risk” is a legal category, and the five still answered to the rules on drug safety and manufacture. Forecasting a legislature is a task that nobody does well (Colorado’s first law never took effect). Documentation is not “adjacent”: it is how people outside can check, if it is written from the work. The EU’s high-risk duties moved to December 2027 and August 2028 under Regulation (EU) 2026/1744, which took effect six days before the date it moved; putting off duties that no one could yet show they met was a repair, and the months of not knowing were the cost. Article 86’s right to an explanation backs up the GDPR’s, which has applied since 2018, and extends it to decisions a person makes on a machine’s output; in February 2025 the Court of Justice said what that explanation must contain, and that a trade secret is weighed by a regulator or court, not simply accepted. An explanation is not a remedy, and the product liability directive pays for death, injury, property and private data, not for a benefit wrongly refused. China’s rules on algorithms name who answers, chiefly to the regulator, and its personal-information law has given a right to an explanation of significant automated decisions since November 2021. India relies on existing laws, applied largely by sector; its hardest case is the state as user (Telangana’s Samagra Vedika), where the Right to Information Act already required reasons, and a vendor’s claim closed the code although the Act allows release in the public interest; the Supreme Court restored a stay before cancellation and put the checking back on the side that held the records; Puttaswamy set the frame. Late regulators can skip early mistakes, and a law written in good time can still arrive late in court (Kenya and Worldcoin, scanning from 2021 to 2023, deletion ordered in 2025). Divergence is a set of decisions, each someone’s; Bradford’s Brussels effect is a tendency, not a law of subsets; regulating by sector knows where a person signs and leaves a gap where the state is its own user. Capture can be a matter of knowledge. Aśoka’s edict (law as a nurse’s care, officers trusted in the room to judge fearlessly, a stay for appeal) and Fuller (law’s own morality; the governed as responsible agents) partly answer different questions and pull against each other in the chapter’s cases, and Fuller’s eight ways do not make a rule good (Hart). Regulation governs institutions; it cannot decide in the room. A machine’s reading of a law is checked against the law’s own text, and the Omnibus’s repair also suited the companies that asked for it.
- What changed my mind: the Prologue’s rule against defining people by what machines cannot yet do, and its three nevers, two of them commitments kept by deśa, kāla, pātra; Chapter 15’s withdrawal of regulation as a moat (Stigler) and its point about documentation; the law itself in 2025 and 2026: the Omnibus, Colorado’s repeal, China’s 2026 plan, India’s guidelines and rules, Vietnam’s and Korea’s laws, Kenya’s High Court; Al Jazeera’s reporting on Samagra Vedika and the Supreme Court’s order of 2022; the audit’s corrections (the dates, Article 5, Annex III, the invented formats, Colorado, China, India, Vietnam, Brazil); Fuller’s eight ways and Aśoka’s fourth pillar edict; Hart’s review of Fuller; the GDPR and the Court of Justice’s judgment of February 2025 on explaining automated decisions; China’s Personal Information Protection Law; India’s Right to Information Act, section 4(1)(d); three outside readings of an earlier version of this chapter, in October 2026, which pointed out that a legal category is not a measure of risk, that an explanation is not a remedy, that “only Congress” and “the only binding treaty” were too wide, that legality is not justice, that the opening scores needed an owner and the cautious reading an incentive, that the state had bought code it would not show, and that Kenya’s order came years after the scanning.
- What survives: the two numbers that open the chapter; “These are not technical specifications. They are moral declarations.”; the regulatory map, dated (EU, US, China, India and Southeast Asia, Africa and Latin America), now with East Asia and the Council of Europe treaty; “There is no map. There is a weather system.”; “the rules that matter tomorrow are not always the rules on the books today”, now the method of the whole chapter; India as the chapter’s hardest case, now with the state as user; that Africa and Latin America are not simply behind Europe; regulation as leapfrog, narrowed; that building where the rules are light works only while a product stays at home; that regulation should not depend on whether one person happens to ask; what regulation is for.
- Section titles in the print edition: 25.4 was “India, Southeast Asia, and the Emerging Landscape”, and 25.5 was “Africa and Latin America: Regulation as Leapfrog Opportunity”. The other five keep their print titles.