>
Rules of the Game is a series of books, and a read on the regulatory landscape: what binds, who it binds, and what it makes an operator actually do. Each edition takes one subject and separates the instrument in force from the rule that is proposed and the intention that has only been announced. The three are not the same thing, and they are routinely reported as though they were.
How AI and agents are governed sovereignly by nations, and what each regime makes you do. A read on the United States, India and the six Gulf states, set against the European instrument they are all drafted beside, and on the one gap every single regime shares. A read on the national rulebooks for artificial intelligence across the United States, India and the six Gulf states, and on the gap every one of them shares.
Every rule carries its instrument, its issuing body and its date. A date that is scheduled rather than in force is said to be scheduled, and a regime announced without an instrument is recorded as an announcement and kept off the deadline board. Where a primary source exists it is used, and where only commentary could be reached the commentator is named.
There is no single AI rulebook, and three patterns account for almost every national position a company now has to comply with.
No single AI rulebook exists, and none is close. Three patterns account for almost every national position as at 6 November 2026. The first is the comprehensive binding statute, and the European Union holds the only working example of it. Regulation (EU) 2024/1689 has been in force since 1 August 2024 (European Parliament and Council, 13 June 2024) and applies in tranches running to 2 August 2028. The second pattern is sectoral and state-led, with no horizontal statute anywhere in sight. That is the American position. The third is strategy before law, and it is the commonest of the three.
China sits outside those three and is better read as a fourth position, binding and registration-led. Four departmental instruments of the Cyberspace Administration of China are in force, with an agent-specific implementation opinion added on 8 May 2026 (CAC, NDRC and MIIT, 8 May 2026). The United Kingdom is regulator-led by deliberate choice, and its only binding AI-specific secondary legislation requires a regulator to write a code rather than regulating any system (SI 2026/425, in force 12 May 2026). India has published voluntary guidelines over a floor of existing law. The Gulf states put strategy first. Statute may follow.
Strategy is now near-universal, which is why it explains so little. UN Trade and Development counted 89 national AI policies worldwide by the end of 2023, of which only six came from least developed countries (UNCTAD, August 2025). About two thirds of developed countries had adopted one by then. The OECD.AI Policy Navigator draws on more than 80 jurisdictions and records over 900 national AI policies and initiatives, of which 63 sit in the active national strategy category (OECD.AI, 6 November 2026). That figure counts initiatives, not jurisdictions. Japan alone holds two of the 63, so the true number of countries is lower.
Enacted law is far rarer. Across the G20, 11 AI-related bills were passed into law in 2025, against none at all in 2016 (Stanford HAI, 2026 AI Index Report). The cumulative count for 2016 to 2025 puts the United States at 25 and South Korea at 17, with France and Japan at 10 apiece and Italy at nine. The Index cautions that the smaller values in that series should be read as indicative rather than exact. On its figures China and India show one enactment each over the decade, which is a useful corrective to the impression that volume of law tracks volume of deployment.
No authoritative count of jurisdictions holding a binding AI-specific statute is published by anyone. The IAPP Global AI Law and Policy Tracker, chapter 8 of the Stanford AI Index and the OECD.AI Policy Navigator were each checked, and not one of them carries such a number. The IAPP tracker mixes comprehensive legislation with use-case legislation and national strategies, and its own landing page concedes that it does not cover every AI initiative in every jurisdiction (IAPP, updated 6 October 2026). Third-party counts therefore vary with the definition of binding. The absence is the finding.
Outside the European Union the statute route has been taken twice, and the penalty shows the ambition. South Korea's Framework Act on the Development of Artificial Intelligence and Establishment of Trust commenced on 22 January 2026 and caps administrative fines at KRW 30,000,000 per violation, attached to three procedural duties only, with no turnover percentage anywhere in the Act (Law No. 20676, promulgated 21 January 2025). Japan's AI Promotion Act, enacted in May 2025, carries no penalty provision at all. The American numbers run the other way on volume and the same way on bite, with 150 state bills enacted in 2025 and 58 federal AI-related regulations recorded in the same year, 28 of them issued by the Executive Office of the President.
The European statute is the reference text for every other drafter, and its hardest obligations have just been pushed into 2027 and 2028.
The European Union's AI Act is the text every other drafter now works against, and its high-risk tier has just moved. Regulation (EU) 2026/1744, the Digital Omnibus on AI, was signed at Strasbourg on 8 July 2026, published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026 (European Parliament and Council, 8 July 2026). Two dates moved. The Annex III high-risk obligations went from 2 August 2026 to 2 December 2027, a delay of 16 months, and the Annex I obligations from 2 August 2027 to 2 August 2028, a delay of 12 months. Recital (40) gives the reason as the delayed availability of standards and the delayed establishment of national competent authorities.
Plenty of the Act binds today. Chapters I and II, including the Article 5 prohibitions as originally adopted, have applied since 2 February 2025. The chapters on general-purpose AI models, on governance and on penalties, along with Article 78, have applied since 2 August 2025, with Article 101 carved out of that tranche. The general date of application was 2 August 2026, which brought in the Article 50 transparency duties and Article 101 itself (Publications Office, consolidated text of 27 July 2026). Articles 102 to 110 have applied since 27 July 2026. What is not yet in force is the substance of the high-risk regime, and that is the part most commentary treats as the Act.
Two obligations fall due on 2 December 2026, and the omnibus created both. The first is a legacy marking duty under a new Article 111(4), which requires providers of systems generating synthetic audio, image, video or text content placed on the market before 2 August 2026 to comply with Article 50(2) by that date. Recital (38) describes the interval as a transitional period of four months. The second is a pair of new Article 5 prohibitions on non-consensual intimate imagery and on child sexual abuse material. A new Article 5(1a) narrows them, so that they bite on a provider only where such generation is the intended purpose or a reasonably foreseeable and reproducible outcome. Neither existed in the 2024 text.
Exactly one harmonised standard supports the Act. EN 18286:2026, on quality management systems for AI Act regulatory purposes, was cited in the Official Journal by Commission Implementing Decision (EU) 2026/2274 of 12 October 2026, published two days later (European Commission, 14 October 2026). The Annex to that decision lists a single entry. CEN-CENELEC calls the standard the first in support of the Act and describes it as one of the first building blocks in a wider standardisation landscape still to come (CEN-CENELEC, 15 October 2026). The citation carries real weight, because under Article 40 conformity with a cited standard raises a presumption of conformity with the requirements that standard covers. One standard, against a 16-month deferral justified by the absence of standards.
The guidelines that would tell a firm whether it is high-risk at all are still in draft. Article 6(5) required the Commission to publish them, with a comprehensive list of practical examples of use cases inside and outside the high-risk category, no later than 2 February 2026. Draft guidelines appeared on 19 May 2026 in three downloadable parts. The Commission's policy page, last updated 19 October 2026, still records that consultation feedback is going to be incorporated in the final version before adoption (European Commission, 19 October 2026). Article 6(5) was expressly carved out of the deferral of Chapter III. The deadline was missed.
The penalties explain why the rest of the world reads the Act so closely. Breach of the Article 5 prohibitions attracts up to EUR 35,000,000 or 7 per cent of total worldwide annual turnover, whichever is higher. Operator duties and the Article 50 transparency obligations sit at EUR 15,000,000 or 3 per cent, and supplying incorrect or misleading information to an authority at
EUR 7,500,000 or 1 per cent. Providers of general-purpose AI models face EUR 15,000,000 or 3 per cent under Article 101, which the omnibus left untouched. Relief exists for smaller firms. Small and medium-sized enterprises pay the lower of the two figures rather than the higher, and the omnibus extended that treatment to small mid-caps.
China regulates AI at the point of entry through filing and labelling, and the money penalty standing behind those duties rose tenfold in 2026.
China regulates AI before deployment rather than after harm. The mechanism is the filing. It predates generative AI. The Internet Information Service Algorithmic Recommendation Management Provisions, Order No. 9 of four issuing bodies, have been in force since 1 March 2022 (CAC, 31 December 2021). A provider with public opinion properties or social mobilisation capacity must file within ten working days of starting the service. What is filed covers the provider's name and service form, the algorithm type and application field, the self-assessment report and the content proposed for publication. The receiving department then has thirty working days to complete the filing and issue a number, or to refuse with reasons.
The generative AI rules sit on top of that machinery rather than replacing it. The Interim Measures for the Administration of Generative AI Services, Order No. 15 of seven issuing bodies, came into force on 15 August 2023 (CAC, 10 July 2023). Article 17 is the compliance trigger. A provider of generative AI services with public opinion properties or social mobilisation capacity must carry out a security assessment under the relevant state provisions and must complete algorithm filing under the 2022 Provisions. Article 2 fixes scope by where the service is offered, applying the Measures to the provision of such services to the public within Chinese territory. There is no incorporation test and no exemption for an offshore provider.
Labelling has been a legal duty since 1 September 2025. The Measures for Labelling AI-Generated Synthetic Content were circulated by four bodies under a notice dated 7 March 2025 and published on 14 March 2025 (CAC, 14 March 2025). They separate explicit labelling, which the user can perceive, from implicit labelling added by technical means in the content file data. Article 5 requires the implicit label in file metadata to carry the content attribute information, the provider's name or code and the content number. The supporting mandatory national standard GB 45438-2025 took effect on the same day (Standardisation Administration of China and SAMR, 28 February 2025). Platforms carry duties too. Article 6 requires a dissemination platform to check the metadata and to add a graded prompt around published content.
A foreign enterprise serving the Chinese public needs four things, and only three of them are written in the AI rules. It needs a security assessment report lodged with the city-level or higher cyberspace department and public security organ of its locality, under the 2018 Security Assessment Provisions. It needs an algorithm filing through the national platform at beian.cac.gov.cn, directed for generative AI to the provincial cyberspace department, with the filing number then displayed prominently on the service. It needs whatever administrative licence the law requires, inside the foreign-investment limits set by the negative list (NDRC and MOFCOM Order No. 23, in force 1 November 2024). The fourth is practice. A domestic entity holding a PRC business licence and an ICP record is reported consistently across practitioner guidance. No CAC document stating that requirement could be verified.
Non-compliance by an offshore provider is handled by blocking rather than by enforcement against the entity. Article 20 of the Interim Measures directs the national cyberspace authority to notify the relevant bodies to take technical measures and other necessary measures. No published enforcement action under the Measures against a named foreign provider could be found. Foreign-brand filings do happen. Han Kun Law Offices reports assisting HP China in completing the deep synthesis service algorithm filing for the assistant built into its AI PC products. It describes its client as among the first foreign brands to complete such a filing (Han Kun Law Offices, 7 August 2024).
The money moved in 2026. The AI instruments mostly carry no fine of their own, referring out instead to the Cybersecurity Law and the Personal Information Protection Law among others. The amending decision on the Cybersecurity Law was adopted on 28 October 2025 and took effect on 1 January 2026, raising the top tier from RMB 1,000,000 to RMB 10,000,000, with up to RMB 1,000,000 on responsible officers (NPC Standing Committee, 28 October 2025). The Personal Information Protection Law remains heavier still at RMB 50,000,000 or 5 per cent of previous-year turnover. Only two of the AI instruments set their own fines. The newer of them, the Anthropomorphic AI Measures in force since 15 July 2026, reaches RMB 200,000 where life and health safety is endangered with harmful consequences.
Voluntary frameworks matter only where a contract or a statute points at them, and the clearest American procurement pointer was deleted in 2025.
The soft layer binds only where something harder points at it, and in 2025 the clearest pointer was cut rather than tightened. OMB Memorandum M-24-18 of 24 September
2024 directed United States federal AI acquisition to the NIST AI Risk Management Framework by name (OMB, 24 September 2024). It cited the framework's appendix on how AI risk differs from traditional software risk, and drew its generative AI risk list from the framework's generative profile. OMB Memorandum M-25-22 of 3 April 2025 rescinded and replaced that memorandum. NIST is not mentioned once. Contracts must instead ensure compliance with the minimum risk management practices required under OMB's own M-25-21 (OMB, 3 April 2025).
The framework that the memorandum stopped citing is itself in suspension. NIST released AI RMF 1.0 on 26 January 2023 as NIST AI 100-1, built around four functions and intended throughout for voluntary use (NIST, 26 January 2023). Its development drew about 400 sets of formal comments from more than 240 organisations, following a direction from Congress. America's AI Action Plan of July 2025 then instructed NIST to revise the framework, specifically to remove references to misinformation, to diversity, equity and inclusion, and to climate change. No draft exists. As at 6 November 2026 there is no version 2.0, no published target date and no version number, and the NIST pages state twice that the framework is being revised.
ISO/IEC 42001:2023 is the one soft instrument that can be audited. Published in December 2023 at 51 pages by ISO/IEC JTC 1/SC 42, it sets requirements for an AI management system rather than guidance about one (ISO, December 2023). What makes it contractable is ISO/IEC 42006:2025, published in July 2025, which sets requirements for the bodies that audit and certify against it. Uptake cannot be stated. The ISO Survey is no longer published by ISO and now sits behind a login on IAF CertSearch, and the International Accreditation Forum itself ceased operations on 1 January 2026, so no authoritative count of certificates could be verified. Secondary estimates circulate and are not printed here.
The OECD Recommendation on Artificial Intelligence is the widest-signed AI instrument and the least binding. OECD/LEGAL/0449 was adopted by the Council at ministerial level on 22 May 2019 and revised twice, on 8 November 2023 and again on 3 May 2024 (OECD, 3 May 2024). The second revision added the handling of misinformation and information integrity in generative AI, together with safe override, repair or decommissioning of systems. It has 47 adherents. A Council Recommendation is not legally binding, and the OECD describes it as a political commitment carrying an expectation that adherents will do their best to implement it. Its definition of an AI system is nonetheless the one the European Union, the Council of Europe and the United Nations have adopted.
The treaty written to be binding is the one that binds nobody. The Council of Europe Framework Convention on Artificial Intelligence was adopted on 17 May 2024 and opened for signature at Vilnius on 5 September 2024, with ten signatories on the day (Council of Europe, 5 September 2024). It now has 21 signatories. It has one Party, the European Union, which ratified on 15 May 2026. Article 30 sets entry into force three months after five signatories have ratified, at least three of them Council of Europe member states. The Convention is therefore not in force, and the United Kingdom, which signed on the opening day, has not ratified.
Three mechanisms give the soft layer real force in 2026, and not one of them is a framework. The first is the AI Act's Article 40 presumption of conformity, which turns a standard cited in the Official Journal into legal cover. The second is ISO/IEC 42006:2025, which makes certification against ISO/IEC 42001 saleable and therefore specifiable in a contract. The third is a contract clause. Proposed GSAR clause 552.239-7001 would define its supply-chain roles by reference to NIST AI RMF 1.0 and would require system documentation consistent with the framework's guidelines (General Services Administration, 17 June 2026). Its final text and effective date could not be verified. British procurement guidance points at none of the three, citing only domestic instruments.
Autonomous agents are almost absent from binding AI law, and what does reach them was written before agentic AI existed.
Autonomous agents sit outside nearly all of the AI rulebooks, and the counts are the proof. The word agent, in any form, appears zero times in Regulation (EU) 2024/1689. For calibration, the same 591,474-character text carries 484 instances of high-risk. The pattern repeats. The published ETSI baseline for AI cyber security contains no instance of the word either, in the technical specification or the European Standard that superseded it. Not one of the 97 titles in the ISO/IEC JTC 1/SC 42 work programme contains agent or agentic. The Texas Responsible Artificial Intelligence Governance Act returns zero for agent and zero for autonomous across its 43,308-character enrolled text.
What the European statute uses instead is autonomy, and it uses it eleven times. Article 3(1) defines an AI system as a machine-based system designed to operate with varying levels of autonomy. The human oversight provisions require measures commensurate with the level of autonomy. The penalty criteria take account of the extent to which a system acts autonomously and whether a human can override a decision. The Digital Omnibus then produced the first appearance of the phrase Agentic AI in binding European law, exactly once, as code AIH 0401 in a new Annex XIV listing the designation codes for notified bodies (European Parliament and Council, 8 July 2026). It imposes nothing.
One instrument anywhere in the world regulates agent autonomy as such, and it is Chinese. The Implementation Opinions on the Standardised Application and Innovative Development of Agents were published on 8 May 2026 (CAC, NDRC and MIIT, 8 May 2026). Item 6 is the operative rule. It requires three decision modes to be delineated, covering decisions reserved to the user personally, decisions requiring the user's authorisation and decisions the agent takes on its own. It then provides that the user must retain the right to know and the final decision, and that an agent's execution of operations must not exceed the scope of the user's authorisation. The caveat matters. This is an administrative implementation opinion giving effect to the State Council's AI Plus policy. It binds state organs and steers industry. It carries no penalties of its own.
The clearest binding rule reaching an autonomous agent is twenty-six years old. Section 14 of the Uniform Electronic Transactions Act, verified here from the Michigan enactment of Act 305 of 2000, immediately effective on 16 October 2000, provides that a contract may be formed by the interaction of electronic agents of the parties. That holds even if no individual was aware of or reviewed the agents' actions or the resulting terms (Michigan Legislature, 16 October 2000). An electronic agent is defined as a computer program or automated means acting without review by an individual. So a contract an agent forms in the United States without human approval already binds the principal, who cannot escape it on the ground that nobody saw it. It was written for electronic data interchange.
Two further hooks exist. Neither was drafted for AI. Commission Delegated Regulation (EU) 2017/589 has been in force since 3 January 2018 and governs systems that transact without per-order human approval, imposing pre-trade controls and a mandatory ability to cancel all unexecuted orders (European Commission, 19 July 2016). It is framed around algorithmic trading by investment firms. The UNCITRAL Model Law on Automated Contracting, adopted on 11 July 2024, is the only international text written deliberately for non-deterministic autonomous systems in contract. Its Article 5(1) denies any objection to a contract on the sole ground that no natural person intervened in its formation (UNCITRAL, 11 July 2024). It is a model law, so it binds nowhere until a state enacts it.
Everything else on agents is voluntary or unfinished. Ten agent-relevant instruments have been published and every one of them is voluntary. They include the OWASP Top 10 for Agentic Applications of December 2025 and the OWASP Agent Control Standard v0.1.0 of 1 September 2026. Singapore's IMDA framework for agentic AI belongs there too, in a version 1.5 of 20 May 2026 that supersedes the launch edition (IMDA, 20 May 2026). Eight more instruments are announced or in draft, including all three ETSI agent work items. Only three bind. Of those, only the Chinese opinion was written for AI agents at all. The sharpest regulatory statement of 2026 came from a competition authority rather than a legislature, with the UK Competition and Markets Authority telling businesses that if an AI agent you use does something illegal, you are responsible (CMA, 9 March 2026).
One free-standing federal AI statute exists, and its subject is intimate imagery. Executive orders and budget guidance carry the rest.
The United States has one free-standing federal statute on artificial intelligence, and its subject is intimate imagery. Everything else is executive action. The TAKE IT DOWN Act, S. 146 of the 119th Congress, became Public Law 119-12 on 19 May 2025. It criminalises the knowing publication of non-consensual intimate visual depictions and of digital forgeries, meaning depictions created through software or machine learning that are indistinguishable from authentic ones. From 19 May 2026 a covered platform must honour a valid removal request within 48 hours, and failure is an unfair or deceptive practice enforceable by the Federal Trade Commission.
What was rescinded matters, because the record is often reported wrongly. Executive Order 14110 of 30 October 2023 was revoked by Executive Order 14148 of 20 January 2025, at section 2(ggg), published at 90 FR 8237. EO 14179 did not revoke it. Its own section 5(a) refers to actions taken under the already revoked order, so commentary crediting EO 14179 with the revocation should be set aside. EO 14179 of 23 January 2025 remains in force. It declares a policy of sustaining American AI dominance and required OMB to revise its two AI memoranda within 60 days.
The orders that fill the gap are about building, exporting and procuring rather than licensing. Three were signed on 23 July 2025, covering federal permitting of data centre infrastructure (EO 14318), procurement of large language models under two Unbiased AI Principles (EO 14319) and export of the American AI technology stack (EO 14320). The same day brought the AI Action Plan, a policy document. EO 14409 of 2 June 2026 went further into security, requiring a classified benchmarking process for covered frontier models. Then it closed the door. Section 3(c) states that nothing in the section authorises a mandatory licensing, preclearance or permitting requirement for new AI models.
One order of 2026 changed the vocabulary rather than the law. Executive Order 14434, "Inaugurating the Era of Super Intelligence", was signed on 29 September 2026 and published on 2 October 2026 at 91 FR 63129. The executive branch is to use "Super Intelligence" and "SI" in place of "Artificial Intelligence" and "AI" in non-statutory documents, while previously issued regulations and contracts need not be altered. Agencies have complied. OCC Bulletin 2026-54 carries a footnote defining the new term by reference to 15 USC 9401(3) and citing the order.
What binds federal agencies is budget-office guidance. OMB Memorandum M-25-21 of 3 April 2025 replaced M-24-10 and reaches all executive branch departments and agencies, including the independent regulatory agencies. It defines high-impact AI as AI whose output is a principal basis for decisions with significant effect on matters such as civil rights, credit or employment, and lists fifteen categories presumed to qualify. Seven minimum practices follow for those uses, among them pre-deployment testing, an impact assessment with independent review and a signed risk acceptance, and human oversight with a fail-safe. M-25-22 of the same date governs acquisition, and requires contracts permanently to prohibit the use of non-public agency data and results to train commercially available AI algorithms absent agency consent.
Pre-emption of state law has been attempted and has not succeeded. The Senate voted 99 to 1 in the early hours of 1 July 2025 to strike a proposed ten-year moratorium from the budget reconciliation bill. Executive Order 14365 of 11 December 2025 took the executive route, requiring an AI Litigation Task Force, a Commerce evaluation of onerous state AI laws by 11 March 2026 and a broadband policy notice on the same date. Neither deliverable was published. The Federal Trade Commission did deliver, issuing a proposed policy statement on the suppression of accuracy in AI systems on 1 July 2026 under docket FTC-2026-0859. No final statement was located, and no statute pre-empting state AI law has been enacted.
Five states carry the operative American rulebook, and the headline Colorado statute never came into force at all.
Colorado's AI Act, the law most often cited as America's first, never took effect. The sequence is worth following. SB 24-205 was due to commence on 1 February 2026, and SB 25B-004 of 28 August 2025 pushed that to 30 June 2026. In April 2026 the Attorney General undertook, in a motion filed jointly with xAI and the Department of Justice, not to enforce it pending the conclusion of rulemaking. It was repealed on 14 May 2026, six weeks before the deferred date arrived. The replacement, SB 26-189 on automated decision-making technology, was enacted the same day and commences on 1 January 2027.
California legislates faster than any other state, and in two directions at once. The first threshold is compute. SB 53, the Transparency in Frontier Artificial Intelligence Act, was signed on 29 September 2025 and took effect on 1 January 2026, binding developers of foundation models trained with more than 10^26 operations and imposing the heavier duties on groups with annual gross revenues above USD 500 million. A critical safety incident must be reported to the Office of Emergency Services within 15 days of discovery, and within 24 hours to law enforcement where there is imminent risk of death. Civil penalties reach USD 1,000,000 per violation.
The provenance regime moved more slowly. SB 942, the California AI Transparency Act, was signed on 19 September 2024, and AB 853 of 13 October 2025 moved its operative date from 1 January 2026 to 2 August 2026. Large online platform duties follow on 1 January 2027, and capture-device duties from 1 January 2028, at USD 5,000 per violation. The privacy agency's automated decision-making regulations took effect on 1 January 2026, with compliance for significant decisions from 1 January 2027.
Texas took the opposite approach to proof. HB 149, the Texas Responsible Artificial Intelligence Governance Act, states in section 10 that it takes effect on 1 January 2026. Section
552.056(b) reaches a developer or deployer acting with the intent to discriminate unlawfully against a protected class, and section 552.056(c) adds that a disparate impact is not sufficient by itself to demonstrate intent. Enforcement is exclusive to the Attorney General and there is no private right of action. Penalties run from USD 10,000 to 12,000 for curable violations left uncured and from USD 80,000 to 200,000 for uncurable ones.
Utah is the one state whose cross-sector AI statute carries an expiry date. The sunset is the point. SB 149, the Artificial Intelligence Policy Act, took effect on 1 May 2024 and is codified at Utah Code Title 13 Chapter 72. Its general duty is disclosure on request rather than disclosure by default, and a heavier duty to disclose prominently applies to regulated occupations. SB 226 narrowed it from 7 May 2025 to a high-risk artificial intelligence interaction. SB 332 extended the Act's sunset from 7 May 2025 to 1 July 2027.
New York's hiring rule is municipal, and the state's own frontier law is not yet live. NYC Local Law 144 became operative on 1 January 2023, with enforcement from 5 July 2023 under rules at 6 RCNY 5-300 and following. Enforcement rests with two attorneys. The State Comptroller's audit, Report 2024-N-6 of 2 December 2025, found two complaints in two years and a single instance of non-compliance among 32 companies surveyed, where auditors reviewing the same companies found at least 17 potential instances. The RAISE Act, signed as Chapter 699 of 2025, was repealed and re-enacted by Chapter 96 of 2026 on 27 March 2026, commencing 1 January 2027.
Any count of state AI law needs its counting rule published alongside it. The Transparency Coalition's 2026 State AI Legislation Report of 27 October 2026 records 115 new AI-related laws passed or enacted across 30 states between 1 January and 15 October 2026, excluding data-centre measures. MultiState's tracker records 1,561 AI-related bills introduced in 45 states as of March 2026. The NYU Center on Technology Policy counted 109 AI laws plus 28 separate data-centre laws to 1 July 2026. The Center for Democracy and Technology, counting public-sector AI bills alone, found three enacted out of 69 introduced. The range is three to 115.
India's data law commences in stages to May 2027, while the labelling and takedown duties on synthetic content already bind.
India's binding floor for AI is its data protection law, and most of it is not yet in force. The Digital Personal Data Protection Act, 2023 (22 of 2023) was enacted by Parliament on 11 August 2023. The Rules followed as G.S.R. 846(E) of 13 November 2025, made under sections 40(1) and 40(2). Commencement is staged. Rules 1, 2 and 17 to 21 took effect on notification, covering definitions and the constitution of the Data Protection Board. Rule 4, on Consent Managers, commences on 13 November 2026, and the substantive duties at Rules 3, 5 to 16, 22 and 23 commence eighteen months after notification, on 13 May 2027.
The May 2027 date carries almost everything a model builder would care about. It brings notice, security safeguards, breach intimation, retention limits, children's data rules, the obligations of a Significant Data Fiduciary, data principal rights and the cross-border transfer power. The penalties are stiff. They reach Rs 250 crore for failure to maintain reasonable security safeguards, Rs 200 crore each for an unnotified breach and for children's-data breaches, and Rs 50 crore otherwise. Appeals from the Board lie to the Telecom Disputes Settlement and Appellate Tribunal.
One clause reaches algorithms directly. Rule 13(3) requires a Significant Data Fiduciary to observe due diligence to verify that technical measures including algorithmic software adopted by it for hosting, display, publishing, transmission, storage or sharing of personal data are not
likely to pose a risk to the rights of Data Principals. Rule 13(1) and (2) add an impact assessment and an audit once every twelve months from designation, with significant observations reported to the Board. There is no AI-specific exemption and no text-and-data-mining exception. The clause was read in a specialist reproduction rather than in the Gazette file, which was truncated on retrieval.
The adjudicator is a body in law and an absence in fact. The Data Protection Board of India is established under section 18 of the Act as a body corporate and an independent adjudicatory authority. A MeitY vacancy circular of 6 May 2026 invited applications for one post of Chairperson and four posts of Member, with consolidated monthly salaries of Rs 4,50,000 and Rs 4,00,000. No appointment could be verified. A dated check by Privacy Law Network on 21 September 2026 found no Gazette notification, ministry document or press release announcing one.
Synthetic content is where India moved first and moved hard. G.S.R. 120(E) of 10 February 2026 amended the Information Technology (Intermediary Guidelines and Digital Media Ethics Code) Rules, 2021, with effect from 20 February 2026. New Rule 2(1)(wa) defines synthetically generated information as audio, visual or audio-visual information artificially or algorithmically created or altered using a computer resource so that it appears real and is likely to be perceived as indistinguishable from a natural person or a real-world event. Text is outside the definition. Carve-outs cover routine good-faith editing and the ordinary creation of documents and training materials where no false record results.
Prevention comes first, labelling second. Rule 3(3)(a)(i) requires reasonable and appropriate technical measures, including automated tools, to prevent users generating synthetic content that breaks the law, naming the IT Act, the Bharatiya Nyaya Sanhita 2023, the POCSO Act 2012 and the Explosive Substances Act 1908. Rule 3(3)(a)(ii) then requires all other synthetic content to be prominently labelled in the visual display, or prefixed with an audio disclosure, and embedded with permanent metadata including a unique identifier for the service used. Rule 3(3)(b) forbids enabling removal of the label or the identifier.
The same instrument cut four clocks. On actual knowledge from a court order or a reasoned written intimation, Rule 3(1)(d) now requires removal within three hours where it allowed thirty-six, and the intimation must come from an officer not below Joint Secretary, or Deputy Inspector General where issued by the police. A complaint about content exposing private areas, nudity or impersonation runs on two hours under Rule 3(2)(b), down from twenty-four.
A removal complaint under the proviso to Rule 3(2)(a)(i) runs on thirty-six hours, down from seventy-two, and general grievance resolution falls from fifteen days to seven. No enforcement action under these provisions could be located.
Six Indian regulators are at work on AI, and only one of them has issued an instrument that binds anybody.
India decided not to legislate horizontally, and said so plainly. MeitY unveiled the India AI Governance Guidelines on 5 November 2025 under the IndiaAI Mission, built on seven
guiding principles, recommendations across six pillars and an action plan on short, medium and long-term timelines. The Secretary, MeitY, Shri S. Krishnan, put the policy in one line: our focus remains on using existing legislation wherever possible. The Guidelines are non-binding and carry no penalties. They propose an AI Governance Group supported by a Technology and Policy Expert Committee, and neither body is reported as constituted.
The Reserve Bank of India has produced a report and a draft, and no circular. Its FREE-AI committee report, on responsible and ethical enablement of AI in the financial sector, was released on 13 August 2025 and runs to seven sutras and twenty-six recommendations across six pillars. It is a committee report. It creates no obligation on a regulated entity. The draft Guidance on Regulatory Principles for Model Risk Management, 2026 was issued for comment on 24 June 2026, with comments closing on 24 July 2026. It remains a draft, and no final circular could be located.
The securities regulator has announcements, not rules. No SEBI artificial intelligence or machine learning instrument could be located. On 10 October 2026 the Chairman, Tuhin Kanta Pandey, said guidelines for capital markets would issue shortly, describing a tiered approach with accountability, data controls, kill switches and human oversight, and stating that every regulated entity remains responsible for any AI tool it uses. The oldest AI-touching Indian instrument found is a 2019 SEBI circular requiring market infrastructure institutions to make quarterly disclosures on AI and machine learning applications.
Medical device software is the one place an Indian regulator has issued a final AI-aware document. The Central Drugs Standard Control Organisation issued its Guidance Document on Medical Device Software, Doc No. CDSCO/MD/GD/MDSW/01/2026, by circular dated 21 July 2026 under the Medical Devices Rules, 2017. It replaces the earlier split between standalone and embedded software with one term, which takes the risk class of the hardware it drives. It recognises AI and machine learning based software with a medical purpose and requires post-market surveillance for performance, bias and drift. An Algorithm Change Protocol may be devised, wherever applicable. So it is discretionary. Whether such guidance binds at all is disputed in the sources, and the CDSCO text could not be opened.
Insurance and telecoms are earlier still. IRDAI constituted a seven-member working group on AI governance in insurance by office order dated 17 June 2026, reporting in three months, and separately directed insurers to file an action taken report on AI cyber readiness by 22 May 2026. No IRDAI AI regulation has issued. TRAI recommended in July 2023 that an independent statutory Artificial Intelligence and Data Authority of India be created, and it was not. On 30 October 2026 TRAI released draft guidelines on the responsible use of AI in telecommunication networks, with comments due on 27 November 2026.
A horizontal law is held in reserve rather than ruled out. On 8 October 2026 the Union Minister for Electronics and Information Technology, Shri Ashwini Vaishnaw, announced that MeitY would release a regulation-focused consultation paper within a month, built on safety, a human-first approach, skilling and leaving no person behind. He called the approach techno-legal. The paper could not be verified as issued. Reports disagree on whether it leads to a standalone AI law, amendments to existing law, or further refinement of the techno-legal framework.
No Indian rule reaches an autonomous agent. No definition of an AI agent appears in the DPDP Act or Rules, in the IT Rules, or in any located instrument of the financial and telecom regulators. The nearest analogue in a binding instrument is the DPDP Consent Manager, a person registered with the Board who lets a Data Principal give, manage and withdraw consent through an interoperable platform, and that duty commences on 13 November 2026. The second nearest is the Account Aggregator, a class created by RBI Master Direction of 2 September 2016, with seventeen companies registered.
Saudi Arabia has issued six AI documents in under three years. None is a law, and what binds is the data regime.
In Saudi Arabia the enforceable instrument is the data law, and the AI instruments sit above it as guidance. The Personal Data Protection Law entered into force on 14 September 2023, with a one-year grace period for controllers ending 14 September 2024, and its implementing regulations were published by SDAIA on 7 September 2023. SDAIA is the competent authority. Commentary records that SDAIA enforcement committees issued 48 formal decisions during 2025 against organisations found in violation, a figure that could not be traced to an SDAIA publication. The position is the inverse of the UAE's, where the data law waits on its regulations.
SDAIA's AI output is advisory. The AI Ethics Principles were issued in final form as Version 1.0 in September 2023, with a Version 2.0 since, and they are not legally binding, although non-compliance may still attract enforcement under other instruments such as the data law. Generative AI Guidelines for government entities followed on 10 January 2024. Deepfakes Guidelines were issued on 8 May 2026, covering protection of personal data, explicit consent, digital watermarks and source verification. The Saudi Press Agency described that document as mandating adherence to strict standards, yet named no enforcement mechanism.
The most recent framework is the national AI risk management framework, launched on 14 July 2026. It is described as a unified national methodology for identifying, assessing, treating and monitoring AI risks, intended to enable public and private entities to adopt AI safely. Its
methodology runs to four phases, ending in continuous monitoring and review. Risk levels are calculated on a matrix linking probability of occurrence to scale of impact. The framework rests on seven core principles and classifies AI risks into seven main types. The announcement nowhere states that the framework is mandatory. It uses the language of enablement.
One policy might have changed the advisory character. It remains unfinished. SDAIA opened a consultation on a draft Responsible AI Policy on 3 April 2026 and closed it on 3 May 2026. No source opened reports a finalised policy or an adoption date, and the risk tiers reported by commentators during the consultation window could not be confirmed against an SDAIA document. The reference number and document date circulating for the risk management framework rest on secondary databases alone, because the SDAIA site rejected direct retrieval.
Sector regulators have moved in narrow places. SAMA's rulebook states, in its fraud risk provisions, that where machine learning or artificial intelligence is used the system should not be a black box and should be capable of being audited. There is no standalone unified Saudi framework governing AI use in financial institutions, so banks fall under existing rules on data governance, cybersecurity, risk management and customer data protection. The Capital Market Authority consulted on regulating robo-advisory in August 2025 and approved the framework in March 2026. Robo-advisory portfolios reached SAR 6.4 billion, up 86 per cent year on year, on the Authority's 2025 annual report.
The medical device picture is sharpest. The Saudi Food and Drug Authority's guidance on review and approval of artificial intelligence and big data based medical devices, as posted on the national consultation platform, carries the document number MDS - G00, the version number 0.0 and a version date of 00/00/2020. It is, on its face, an unnumbered and undated draft. It would set a four-tier version control scheme, under which a change in operating principles, intended use or performance beyond the approved accuracy range requires change approval, while a training data change that stays within that range is left to the manufacturer's own control.
This is not for want of time. Vision 2030 was launched in 2016, SDAIA was created in 2019 and the National Strategy for Data and AI was announced in 2020, so the policy architecture predates most of the world's AI statutes. The International Centre for Artificial Intelligence Research and Ethics is affiliated with SDAIA, and the GCC General Secretariat names it as a
partner in Gulf AI cooperation. Six years on, the binding layer is still the data law. An organisation deploying a model in the Kingdom answers for its personal data, not for its model.
There is no federal UAE AI law. The binding obligations sit in a financial free zone and in two emirate health regulators.
There is no federal AI law. The UAE's newest national body is drawn in terms of government itself. On 14 June 2026 Sheikh Mohammed bin Rashid Al Maktoum approved the establishment of the Artificial Intelligence and Data Authority, described as the single national body for data, artificial intelligence and digital government, reporting directly to the Cabinet. It consolidates the former AI office, the digital government sector of the telecommunications regulator and the UAE Data Office, under Omar Sultan Al Olama as Chairman. Its mandate covers proposing policies and standards, and ensuring compliance across federal entities. The Cabinet statement confers no rule-making power over the private sector.
What sits at the federal level is a charter and an unfinished data law. The UAE Charter for the Development and Use of Artificial Intelligence was published in July 2024 and sets out twelve general principles. It describes itself as a guiding framework. It contains no obligations, no addressees, no thresholds and no enforcement provision, and its twelfth principle refers compliance back to other instruments. Federal Decree-Law No. 45 of 2021, the data protection law, remains in force but its Executive Regulations have still not been issued, which one adviser summarised in June 2026 as in force but unenforceable.
The one instrument in the region drafted for autonomy rather than automation is a free zone rule. Regulation 10 on the processing of personal data through autonomous and
semi-autonomous systems was enacted with amendments to the DIFC Data Protection Regulations announced on 7 September 2023, and guidance dated 1 September 2023. Purely automated systems are deliberately excluded, because systems with no degree of autonomy whose operation is deterministically controlled by humans were not meant to be caught. Autonomy is the trigger, not the technology.
Two new roles replace controller and processor. A Deployer is the person or entity that authorises or benefits from the operation of a System and its output, carrying the general responsibilities of a controller. An Operator is the technical service provider acting on the Deployer's instructions. The guidance states the analogy outright, that a System operating autonomously under the authority and for the benefit of its Deployer is substantially similar to an employee. Purposes the System generates for itself must rest on an exhaustive set of principles hard coded by humans, which the System cannot change.
The high-risk gate is closed and the key has not been cut. Regulation 10.3.3 prohibits making a System available for commercial use in high-risk processing unless audit and certification requirements have been established, the System complies with them, it processes personal data solely for human-defined or human-approved purposes, and an Autonomous Systems Officer is appointed. On DIFC's own account, no System may be used for high-risk processing until the Commissioner promulgates those certification requirements. Two sub-committees were required to recommend the criteria by 30 April 2024. Consultation Paper No. 3 of June 2026 was still proposing the machinery, roughly 33 months after enactment. Whether those amendments have been enacted could not be verified.
ADGM has gone no further. It has no AI-specific instrument, although section 20 of its Data Protection Regulations 2021 covers automated decision-making and profiling much as the DIFC's do. Digital Dubai published AI Ethics Principles and Guidelines with a self-assessment tool in January 2019, and the official document states that the guidelines are non-binding. Entities in either free zone must still comply with federal health obligations under Federal Law No. 2 of 2019, which requires local storage and processing of health data subject to stated exceptions.
Health is where the obligation bites. The Abu Dhabi Department of Health published its policy on the use of artificial intelligence in the emirate's healthcare sector on 30 April 2018, the region's first emirate-level AI governance framework, reaching licensed providers, insurers, researchers and any end user working with Abu Dhabi patient data. It sets six guiding principles, requires regular audits with reporting, and retains the power to impose sanctions. A Responsible Artificial Intelligence Standard followed in 2025. The Dubai Health Authority issued its own policy in August 2021 as circular CIR-2021-00000141, enforced through its Health Regulation Sector. Both require graceful degradation on malfunction.
One Qatari AI instrument plausibly binds, and it reaches only entities licensed by the central bank.
Only one Qatari AI instrument plausibly binds. The Qatar Central Bank issued its Artificial Intelligence Guideline on 4 September 2024, regulating the use of AI by licensed entities, and it is recorded as in force from that date. Qatar News Agency reported the same day that the guideline was issued under the Third Financial Sector Strategy and the FinTech Strategy. It applies to all licensed entities that develop, procure, outsource, operate or provide AI systems, so the trigger is the licence rather than the technology. Secondary sources describe it as legally binding on those entities. That characterisation rests on commentary, because the text itself could not be reached.
The obligations read conventionally. Prior central bank approval is required before launching a new AI system as a provider or materially modifying an existing one, and again before signing a high-risk purchase, licensing or outsourcing agreement. A licensed entity must keep an
updated register of AI systems, run risk assessments that classify high-risk systems, adopt board-approved AI governance policies, maintain human oversight of high-risk systems and give customers a mechanism to query and request review of an AI-driven decision affecting them. The clause numbers and the obligation list come from a regulatory database, not from the central bank, and that database marks its own entry as unconfirmed.
Penalties are not separately provided. The position is described as subject to the central bank's general supervisory enforcement powers rather than a dedicated AI penalty regime. A licensed entity faces supervisory consequences rather than a stated civil penalty per violation. For a bank already inside a prudential relationship, the practical effect may be stronger than a fine. For anybody outside that relationship, there is no effect at all.
The capital markets regulator has an intention. In May 2025, during the 50th Annual Meeting of IOSCO held in Qatar from 12 to 14 May, the Qatar Financial Markets Authority said a draft regulation on AI use in the financial market would be issued soon, framed around transparency and consistency and tied to its Strategic Plan 2023 to 2027. No source opened confirms that the draft has been published for consultation or adopted. The most recent coverage located is from May 2025, and it describes an intention rather than an instrument.
The data law predates the AI debate. Qatar's Personal Data Privacy Protection Law is Law No. 13 of 2016, published in the Official Gazette on 29 December 2016 and effective in 2017. It affords extra protection to data relating to children, health, ethnic origin and religious beliefs, and processing those categories requires permission from the Competent Department. The reported maximum penalty is QAR 5 million. Sources disagree on which body regulates onshore, naming either a division of the national cyber security agency or a national data protection office, and neither account could be traced to a primary source.
As in the UAE, a free zone runs its own regime. The Qatar Financial Centre operates under the QFC Data Protection Regulations 2021, applying to entities incorporated or registered in the Centre, supervised by the QFC Data Protection Office and requiring a data protection impact assessment for high-risk processing. So an organisation in Qatar must establish which
of two data regimes applies before it establishes what applies to its models. The answer turns on registration. Neither regime is AI-specific, and neither sets a labelling or provenance duty of the kind India brought into force in February 2026.
Above that, everything is strategy or draft. Qatar's National Artificial Intelligence Strategy was published by the then Ministry of Transport and Communications and is built on six pillars, providing recommendations rather than binding rules. Its date is disputed, given as 2019 by one source and early 2022 by another, and the discrepancy is unresolved. A draft National AI Policy exists and is expressly a consultation draft, with an adviser writing in November 2025 describing Qatar as moving toward establishing both the policy and a national centre. A draft Cabinet decision approved in September 2025 would create that centre.
All three have published AI strategies. None has a binding AI instrument that reaches a private company.
Bahrain's AI bill was not enacted. The drafting appears to have started again. A thirty-eight article bill was approved unanimously by the Shura Council in April 2024, with a requirement that the government draft it properly and refer it to Parliament within six months. It then met resistance in the Council of Representatives, where the government proposed a regulatory sandbox instead, and it had not been enacted as of July 2026. A report of 19 June 2026 records the Shura Council's legislative committee as still drafting legislation on artificial intelligence, drawing on both the European framework and the Arab model law. No source opened records a promulgated Bahraini AI law.
In Bahrain a government policy binds. The Information and eGovernment Authority issued the General Policy for the Use of Artificial Intelligence, Version 1.0, dated 20 May 2025, approved by the Ministerial Committee for Information and Communication Technology that day, from which date all government entities must comply. Its scope does not reach the private sector. It sets eleven principles, the first requiring that the final decision in important matters remain under human control. Negative effects of an AI solution are attributed to the responsible individuals, not to the machine.
The policy points at other laws. Its first pillar requires compliance with a named set of existing instruments, among them the Personal Data Protection Law of 2018, the Cybercrime Law of 2014, the decree on cloud computing services to foreign parties, the Open Data Policy and the GCC Artificial Intelligence Ethics Guideline. The data law was enacted on 12 July 2018 and came into force on 1 August 2019. The policy records that it includes provisions on processing by automated means and the conditions that make such processing lawful.
Oman's position is settled once the two instruments are read apart. The Ministry of Transport, Communications and Information Technology issued the Public Policy for Safe and Ethical Use of Artificial Intelligence on 1 April 2025, prepared in line with Islamic principles, with oversight shared between the ministry and sector regulators. UNESCO describes a risk-based framework applicable to public entities and relevant private-sector actors, and notes that enforcement procedures remain in early operational stages. Public-sector procurement of AI requires pre-deployment assessments, formal approval by the competent authority and lifecycle monitoring. No AI-specific liability regime has been adopted.
The Omani royal decree that is often read as AI regulation is not. Royal Decree No. 50/2026 of 30 April 2026 established an Artificial Intelligence Special Zone in Muscat, extending special economic zone incentives including tax breaks to projects in the zone. Advisers writing in June 2026 state plainly that the decree does not introduce specific artificial intelligence regulation. A binding instrument about AI therefore exists in Oman, and it is an investment measure. Separately, the Omani data law was promulgated by Royal Decree No. 6/2022 on 9
February 2022 and entered into force a year later, with the compliance window extended to 5 February 2026.
Kuwait has neither an AI law nor an AI regulator. Its data instrument is the Data Privacy Protection Regulation of April 2021, administered by the telecommunications regulator, which carries a 72-hour breach notification duty and fines up to KWD 50,000. The citation is contested. DLA Piper cites a 2024 administrative decision and states that Kuwait has no dedicated personal data protection law applying to all natural and juristic persons. Chambers records that the Kuwaiti regulation contains no provisions specifically addressing AI. A National Artificial Intelligence Strategy for 2025 to 2028 is recorded, and the regulator has issued a non-binding guide on AI ethics.
The bloc has named the gap itself. The GCC General Secretariat approved the formation of an artificial intelligence and emerging technologies team in June 2021, which has produced three named frameworks. A unified Gulf guiding strategy was still in preparation as of 7 October 2025. The Secretary General named the absence of unified legislative standards as the first remaining challenge, and said that member states' AI investments have reached tens of billions of dollars over the past few years, with plans to increase them to hundreds of billions by 2030. No GCC state has enacted an omnibus AI statute.
Every Gulf government is at once the author of its AI rules and the sponsor of the systems those rules would govern.
No Gulf instrument addresses the case where the state owns the model. Research to 6 November 2026 located six national AI strategies across the six states of the Gulf Cooperation Council and no national AI framework law reaching the private sector. The gap is structural. Where a binding rule exists it sits below the national line, in a free zone or inside a sector regulator, and it speaks to licensees rather than to owners.
Saudi Arabia shows the shape of the problem. The Saudi Data and Artificial Intelligence Authority writes the national AI instruments and also administers enforcement of the Personal Data Protection Law, under which, on commentary from Captain Compliance, its enforcement committees issued 48 formal decisions during 2025. Its AI Ethics Principles do not bind. Non-compliance may still attract enforcement under the data law, which is an indirect route and not a rule about models.
The Emirates went further. On 14 June 2026 the Cabinet approved an Artificial Intelligence and Data Authority, described in its own announcement as the single national body responsible for data, artificial intelligence and digital government, reporting directly to the Cabinet. The mandate covers national policy and legislation as well as standards for data and AI management, with a duty of ensuring compliance across federal entities. One body proposes the rule and then checks the entity applying it.
The same statement does not confer rule-making power over the private sector. That matters to a buyer, because a state-backed developer incorporated as a company sits outside the federal-entity perimeter the Authority is drawn around. Nothing located fills the space. The UAE Charter of July 2024 sets twelve principles and contains neither addressees nor thresholds, and no enforcement provision at all.
Where a Gulf instrument does reach the state, it reaches the state as a user. Bahrain's General Policy for the Use of Artificial Intelligence, approved on 20 May 2025, applies to all government entities and not to the private sector. It attributes any negative effect of an AI solution to the responsible individuals rather than to the machine. Oman's Public Policy of 1 April 2025 works the same way, with compliance and oversight sitting in the ministry and in sector regulators. Neither speaks to ownership.
The machinery behind those policies is young. UNESCO's 2026 country profile of Oman records that evaluation and enforcement procedures remain in early operational stages. It also records that no certified registry of AI vendors is maintained and that no AI-specific liability regime has been adopted. A buyer cannot lean on apparatus that is not yet running.
International law offers a reference point the Gulf has not taken up. The Council of Europe Framework Convention, opened for signature on 5 September 2024, covers the use of AI systems by public authorities and by private actors acting on their behalf. Its signatories number 21. No Gulf state is among them, and the treaty has one Party, the European Union, so the nearest thing to a rule about the state as owner has not entered into force.
Saudi Arabia's champion is barely documented in the public record, and the rules it would answer to are advisory by design.
The research behind this book could not establish the basic facts about Humain. No primary source opened for the Gulf fact base records its ownership, its governance or any commitment it has made. The investment totals in circulation, reported as USD 100 billion and 2.2 gigawatts, were traced only to low-quality secondary sources and set aside. That absence is the finding. A buyer cannot contract against a balance sheet it cannot source.
What can be sourced is the bloc's own account of scale. The Secretary General of the Gulf Cooperation Council stated on 7 October 2025 that member states' investments in AI have reached tens of billions of dollars, with plans to raise them to hundreds of billions by 2030. That is the figure to use. It fixes the order of magnitude without pretending to a company-level number that no primary source supports.
Against that scale, the rules are soft. The Saudi Data and Artificial Intelligence Authority has issued five instruments since September 2023, beginning with AI Ethics Principles version 1.0 and ending with a national AI risk management framework launched on 14 July 2026. Latham & Watkins records that the Principles are not legally binding. The Saudi Press Agency announcement of the risk framework uses the language of enablement and nowhere states that the framework is mandatory.
One instrument strains that description. The authority's deepfakes guidelines of 8 May 2026 are reported by the Saudi Press Agency to mandate that developers and creators adhere to
strict ethical and regulatory standards, including watermarking and source verification. No enforcement mechanism is named in the announcement. A duty with no addressee and no penalty remains a statement of expectation, whatever verb the press release reaches for.
The binding layer is the data law. The Personal Data Protection Law entered into force on 14 September 2023 and became fully enforceable a year later, with the AI authority as competent regulator. Commentary records 48 enforcement decisions by its committees during 2025. A Saudi developer training on personal data answers to a live supervisor, though for data handling rather than for model behaviour.
Sector rules reach further. The Saudi Central Bank rulebook provides that where machine learning or artificial intelligence is used in fraud risk, the system should not be a black box and should be capable of being audited. The Capital Market Authority consulted on robo-advisory from August 2025 and approved the framework in March 2026. Robo-advisory portfolios had reached SAR 6.4 billion, up 86 per cent on the year, on the regulator's 2025 annual report.
The medical device picture is weaker still. The Food and Drug Authority's guidance on reviewing AI and big data based medical devices, as posted to the public consultation platform, carries the document number MDS-G00, version number 0.0 and a version date of 00/00/2020. On its face it is an undated draft. A champion selling clinical software in Saudi Arabia would be assessed against a document that has never been finalised.
So the answer on applicable rules is short and unsatisfying. A Saudi state-backed developer answers to the data protection law and to whichever sector rulebook licenses its customer, beneath an advisory national stack written by the same authority that enforces the data law. Ownership changes none of it. No located instrument attaches a single additional obligation to a model the state happens to own.
No Qatari state-owned developer appears in the fact base, so the question becomes which Qatari rule would reach one at all.
Qatar's champion cannot be described from the record assembled for this book. Nothing in the Gulf fact base names QAI, or any other Qatari state-owned model developer, or any commitment made by one. The entity is simply absent. What follows is the rule set that would apply to such a developer, and it is narrow and sits almost wholly in finance.
One Qatari AI instrument plausibly binds. The Qatar Central Bank issued its Artificial Intelligence Guideline on 4 September 2024, and it applies to licensed entities that develop, procure, outsource, operate or provide AI systems. Its binding status rests on commentary. The bank's own text, reported to be published as Annex 160 to the Instructions to Banks 2024, could not be reached during the research pass.
The reported obligations matter more to the buyer than to the builder. Commentary records a requirement of prior approval from the central bank before a licensed entity signs a high-risk AI purchase, licensing or outsourcing agreement, and before it launches or materially modifies an AI system as a provider. A state-owned vendor is not the addressee. Its customer is, and that is the practical constraint on any sale into a Qatari bank.
Penalties follow the general supervisory route. Commentary describes the position as subject to the central bank's ordinary enforcement powers rather than to any dedicated AI penalty regime. There is no AI fine to quote. The onshore data protection law, Law No. 13 of 2016, carries a reported maximum penalty of QAR 5 million, and it is the only Qatari money figure the fact base supports.
Everything above it advises. The National Artificial Intelligence Strategy offers recommendations rather than binding rules, and sources disagree on whether it dates from 2019 or early
2022. Ministry guidance on AI ethics is described as voluntary. A draft National AI Policy exists and is expressly a consultation draft, with Eversheds Sutherland placing its finalisation some time after November 2025.
One reported feature of that draft goes straight to the governance question. The draft is said to exclude public bodies that already require specialised arrangements, naming the National Cyber Security Agency and the Qatar Central Bank. The claim comes from a search extract. It was not traced to a primary source and cannot be relied upon. If it survives drafting, Qatar's horizontal policy will carve parts of the state out of scope at the moment of adoption.
The supervisory map is unsettled as well. Sources disagree on which body administers the onshore data protection law, one naming a division of the National Cyber Security Agency and another a national data protection office under the ministry. Neither account reaches a primary source. A buyer cannot name its counterparty's supervisor from the public record, which is an odd sentence to have to write about a regulated market.
The institutional build-out is still under way. A draft Cabinet decision approved in September 2025 would establish a National Center for Artificial Intelligence under the ministry. Until that body exists, Qatari AI governance runs through a financial regulator and a data law of 2016. Compute policy sits outside both of them, which is where the governance question for a sovereign compute programme actually lives.
Neither G42 nor MGX appears in the fact base, so the only documented constraint on a two-track position is American procurement.
The record assembled for this book contains nothing on G42 or MGX. Neither entity is named in the Gulf, United States, India or global fact bases, and the Stargate UAE capacity figures in circulation were traced to low-quality secondary sources and excluded. The alignment story cannot be evidenced here. What can be evidenced is the mechanism through which American alignment is enforced, and that mechanism is procurement.
The United States conditions federal purchase on properties of the model itself. Executive Order 14319 of 23 July 2025 permits agency heads to procure only large language models developed in accordance with two principles, truth-seeking and ideological neutrality. The implementing guidance followed. OMB Memorandum M-26-04 is dated 11 December 2025, and agencies must write compliance terms into any solicitation or order for a large language model issued after that date.
The disclosure annex is the part that bites a foreign champion. Where an agency needs more, M-26-04 reaches any training or development that occurred outside the United States, including the type of activity and the country in which it happened. It also reaches any modifications made to comply with a regulation from a government other than the American federal government. This is foreign-influence disclosure by contract. It asks a state-backed vendor to describe its own home obligations.
The binding clause is milder. On 28 September 2026 the General Services Administration issued GSAR 552.239-7001 as a regulation deviation, effective 19 October 2026, and
commentary from Crowell and Moring records that it replaced the detailed unbiased-AI obligations with a single reasonable-efforts standard. Decommissioning liability is capped at 25 per cent of the affected order value. The lever loses force at each handover from order to memorandum to clause.
Revocability is the sharper constraint. On 27 February 2026 the General Services Administration announced that it was removing Anthropic from USAi.gov and from the Multiple Award Schedule, in support of a presidential directive to cease all use of that company's technology. No statute and no rulemaking were involved. Access to a vehicle sitting inside more than USD 116 billion of federal contracts was withdrawn by direction, which is the precedent any foreign champion should price.
The domestic track is thinner. There is no federal AI framework law, and Federal Decree-Law No. 45 of 2021 on personal data remains in force without its Executive Regulations, which Trowers & Hamlins described in June 2026 as leaving the law unenforceable. Breach notification, transfer controls and officer thresholds are unfixed. A vendor's home obligations are hard to disclose when they have not yet been written down.
One Emirati instrument is drafted for autonomy rather than for automation. Regulation 10 of the Dubai International Financial Centre assigns controller-style responsibility to the
Deployer that authorises or benefits from a system, and processor-style responsibility to the Operator. The centre's own guidance compares the system's position to that of an employee. Its high-risk gate is shut. No system may be used for high-risk processing until the Commissioner promulgates certification requirements, which were still at consultation in mid-2026, roughly 33 months after enactment.
The two-track position is governed asymmetrically. The American side is written down, conditioned on properties of the model and withdrawable at will. The Emirati side consists of an Authority approved on 14 June 2026 and a free-zone rule whose machinery remains unbuilt. A buyer weighing an Emirati champion should read the American clause first, because it is the only text in this chapter with a working enforcement history.
India funds AI capability generously and attaches no obligation to the funding, which is very nearly the whole answer.
India has a mission rather than a champion, at least in the record assembled here. The India fact base names no state-owned or state-funded model. That is the first finding. The sovereign model question cannot be answered on this evidence, and what the base does document is the money and the absence of any condition attached to it. The IndiaAI Mission was approved by the Union Cabinet on 7 March 2024 with an outlay of Rs 10,371.92 crore over five years.
The Mission funds capability. It is implemented by the IndiaAI Independent Business Division under the Digital India Corporation and imposes no obligation on any private party. Its Safe
and Trusted AI pillar funds work rather than regulating it. No instrument located attaches a safety or audit condition to a model built with that public money.
The obligations that exist are general. The Digital Personal Data Protection Rules 2025 were notified on 13 November 2025, with consent manager registration commencing on 13 November 2026 and the substantive duties on 13 May 2027. Rule 13(3) is the provision that reaches a model directly. A Significant Data Fiduciary must verify that algorithmic software it adopts is not likely to pose a risk to the rights of data principals.
One Indian rule does reach a model builder directly. The amended Information Technology Rules, in force from 20 February 2026, require an intermediary offering a tool that can generate synthetic audio or visual content to label the output prominently and to embed permanent provenance metadata carrying a unique identifier. Text is outside the definition. A state-funded Indian image or speech model would carry that duty on the same terms as a commercial one.
Penalties under the Act are large and generic. Failure to maintain reasonable security safeguards carries up to Rs 250 crore, failures on breach notification and on children's data up to Rs 200 crore each, and any other breach up to Rs 50 crore. None of them is AI-specific. A state-funded developer is exposed on exactly the terms a private one is, which is consistent and still not a governance answer.
The enforcement body is not yet staffed. The ministry advertised one chairperson post and four member posts for the Data Protection Board on 6 May 2026, and a check on 21 September 2026 found no appointment notified. The Board exists in law alone. The substantive duties it would supervise take effect on 13 May 2027, which leaves a short window in which to constitute an adjudicator.
The horizontal choice is on record. The ministry unveiled non-binding India AI Governance Guidelines on 5 November 2025, and its Secretary stated that the focus remains on using existing legislation wherever possible. The governance group and expert committee those guidelines propose are not reported as constituted. A consultation paper on regulation was announced on 8 October 2026 with a one-month horizon, and nothing located shows that it has issued.
So the honest finding for India is a negative one. Public money is specific, year-stamped and traceable to a Cabinet decision of March 2024. The duties that would govern what the money builds are horizontal, and they draw no distinction between a state-funded model and any other. The chapter is short for a reason.
In every Gulf state the body that writes the AI rule also sponsors the programme, which leaves the contract doing the work.
No national regulator in the Gulf answers for a state-owned model, on the evidence assembled here. Research to 6 November 2026 located no binding AI instrument at federal or national level reaching the private sector in any of the six member states. One
free-zone rule was found. Three sector-regulator instruments were found alongside it, two further policies bind government entities only, and at least eight national frameworks are advisory.
The conflict is structural. The Saudi Data and Artificial Intelligence Authority writes the national AI frameworks and administers enforcement of the data protection law. The new Emirati Authority proposes legislation and sets standards, and it carries a duty of ensuring compliance across federal entities. Oman's ministry runs the national AI programme and also holds compliance and oversight for the policy governing it.
The binding rules address the buyer. The Qatar Central Bank guideline is reported to require prior approval before a licensed entity signs a high-risk AI purchase or outsourcing agreement. Regulation 10 of the Dubai International Financial Centre places controller-style liability on the Deployer, with guidance comparing the system's position to that of an employee. Bahrain's policy attributes negative effects to responsible individuals inside the government entity rather than to the machine.
That asymmetry is not an accident of drafting. A regime built on data protection concepts reaches whoever processes personal data, and a state-owned developer selling a model is frequently not that person. The obligation lands downstream. A champion can be commercially dominant and regulatorily peripheral at the same time, and in the Gulf of 2026 several of them are.
So the buyer carries the governance that the state has not written. The first question is which instrument binds the vendor, named and dated, as distinct from the national strategy the vendor will cite. The second is whether the vendor's supervisor is a different body from its sponsor. Both questions are factual. Six strategies were located across the Gulf and not one of them answers either.
Three contract questions follow. Ask who is Deployer and who is Operator, because Regulation 10 allocates liability on that distinction alone. Ask whether high-risk processing is lawful at all, given that the Dubai gate stays shut until certification requirements are promulgated. Ask what happens when the model changes, which the Saudi medical device draft answers in four tiers and the Indian device guidance leaves to discretion.
Disclosure terms are available off the shelf. OMB Memorandum M-26-04 reaches training or development conducted outside the United States and modifications made to satisfy a foreign government's regulation, which is a usable template for anyone buying a state-backed model. The European Commission's model contractual AI clauses were updated on 5 March 2025. The United Kingdom's Procurement Policy Note 02/24 supplies optional questions for identifying AI use in a supply chain.
Two gaps deserve stating plainly. The fact bases name none of the Gulf champions, so nothing in this part describes Humain, QAI, G42 or MGX as entities, and nothing here reports their commitments. No Gulf instrument located uses the word agent or agentic in an operative provision. Both absences are findings a buyer can act on, because they describe precisely what the contract has to supply.
Every regime's first duty is the same, because nothing binds until a system has been found and classified.
Every AI rulebook starts in the same place. Before any duty can attach, a system has to be found and sorted into a class. The European Union makes classification the hinge of its whole statute, fixing obligations by reference to Annex I and Annex III of Regulation (EU) 2024/1689. Article 6(5) required the Commission to publish practical classification guidelines, with worked examples of what is and is not high-risk, no later than 2 February 2026. It did not.
Draft guidelines appeared on 19 May 2026 in three parts, and a targeted consultation closed on 23 July 2026. The Commission's policy page, last updated 19 October 2026, still records that feedback will be incorporated before adoption. So the test by which a European firm decides whether it runs a high-risk system is itself unadopted. The duties have moved too. Regulation (EU) 2026/1744 of 8 July 2026 pushed the Annex III obligations to 2 December 2027 and the Annex I obligations to 2 August 2028, and recital (40) attributes the move to delayed standards and delayed national competent authorities.
Other regimes ask for the register first and argue about classes later. OMB Memorandum M-25-21 of 3 April 2025 requires every United States federal agency to publish an AI use case inventory annually. It defines high-impact AI by the consequence of its output, and lists fifteen categories presumed to fall inside that definition. The Qatar Central Bank's Artificial Intelligence Guideline of 4 September 2024 requires licensed entities to maintain an updated AI system register and to classify high-risk systems by risk assessment. The pattern repeats. DIFC Regulation 10, in force since September 2023, calls its own register a basic map for warp speed processing.
The most useful classification test published anywhere is two days older than this edition's cut-off. OCC Bulletin 2026-54 of 4 November 2026 sets out four illustrative tiers for American banks, and says a system's tier turns on what it could cause and how much agency has been transferred to it, rather than on the technical sophistication of the model. An autonomous payment agent sits in Tier 1 Critical. Decision support with a mandatory human in the loop sits in Tier 3. The test is behavioural, not architectural.
What a usable inventory contains is therefore settled by convergence rather than by any single instrument. Colorado's insurance regulation 10-1-1, amended with effect from 15
October 2025, requires a version-controlled model inventory inside a fourteen-component governance framework. The New York Department of Financial Services, in Insurance Circular Letter No. 7 of 11 July 2024, requires a model inventory and annual output testing for drift. China asks the same, as a filing. Article 24 of the Algorithmic Recommendation Provisions requires the provider's name, service form, application field, algorithm type and self-assessment report within ten working days of launch.
One instrument that used to catch AI models in banks no longer does. SR 11-7 of 4 April 2011 required a maintained model inventory and independent validation on conceptual soundness. It was superseded on 17 April 2026 by SR 26-2, issued alongside OCC Bulletin 2026-13.
The replacement is narrower. That bulletin states that generative and agentic AI models are novel and rapidly evolving and are therefore not within its scope. A firm that built its inventory on model risk management now finds its newest systems outside the thing it built.
A regulator does not ask what you believe about your system. It asks for the file you kept while building it.
Regulators ask for files, not assurances. The European technical documentation duties and the Indian impact assessment and audit duties are written in different legal traditions and arrive at nearly the same list. Both want a record made at the time, by a named person, that would let somebody else reconstruct the decision. Neither accepts a summary written afterwards. The real difference is who holds the file and who may call for it.
In Europe the documentation duty is the thing the deferral did not touch. EN 18286:2026, on quality management systems for EU AI Act regulatory purposes, was cited in the Official Journal by Commission Implementing Decision (EU) 2026/2274 of 12 October 2026, published on 14 October 2026. CEN-CENELEC describes it as the first harmonised standard supporting the Act, and says it turns the quality management requirements of Article 17 into practical, auditable requirements for providers of high-risk systems. It is also the only one. The Annex to that implementing decision lists a single entry.
The legal consequence of the citation is the reason to read the standard now. Recital (1) of the implementing decision records that systems conforming to a harmonised standard whose reference has been published in the Official Journal are presumed to conform with the requirements that standard covers. The presumption runs from publication. So one document, cited in October 2026, is the whole of the auditable evidence route into a statute whose high-risk duties were put back sixteen months on the ground that standards were not ready.
India states the evidence duty as a cadence rather than as a standard. Rule 13 of the Digital Personal Data Protection Rules 2025, notified as G.S.R. 846(E) on 13 November 2025, requires a Significant Data Fiduciary to undertake a Data Protection Impact Assessment and an audit once every twelve months from designation, and to cause the assessor or auditor to furnish a report of significant observations to the Data Protection Board. Rule 13(3) reaches AI directly. It requires due diligence to verify that technical measures including algorithmic software are not likely to pose a risk to the rights of Data Principals.
Two dates govern anyone building that file. Rule 4, on the registration and obligations of Consent Managers, commences one year after publication, on 13 November 2026. The substantive rules, being Rules 3, 5 to 16, 22 and 23, commence eighteen months after publication, on 13 May 2027. Three years and nine months will then have passed since the Act was enacted on 11 August 2023. The Board that receives those audit reports was constituted in law on 13 November 2025, had its Chairperson and four Member posts advertised by MeitY circular of 6 May 2026, and had no appointee recorded as at 21 September 2026.
The penalties make the file worth keeping. Failure to maintain reasonable security safeguards attracts up to Rs 250 crore under the Act. Breach notification failures and children's data failures attract up to Rs 200 crore each, and any other breach by a Data Fiduciary up to Rs 50 crore. Appeals from the Board lie to the Telecom Disputes Settlement and Appellate Tribunal. A firm with no dated impact assessment and no audit trail has nothing to put in front of either body.
The European Union frames oversight by autonomy rather than by action. Regulation (EU) 2024/1689 contains no instance of the word agent, and uses autonomy instead, eleven times. Article 3(1) defines an AI system as one designed to operate with varying levels of
autonomy. The oversight provisions require measures commensurate with the risks, level of autonomy and context of use, and the penalty criteria include the extent to which a system acts autonomously and whether a human could override its decision. The framing is deliberate. Those duties sit in Chapter III, which does not apply until 2 December 2027.
Human oversight appears in every instrument and means something different in each, from a job title to approval of a single action.
Human oversight is the one clause every rulebook contains, and no two mean the same thing by it. At one end sits a named officer. At the other sits approval before a single action may proceed. The cost of compliance runs on that difference, because a firm can appoint an officer in a week and cannot staff a per-action gate at all without redesigning the workflow the system was bought to remove.
The DIFC turned oversight into a post. Regulation 10 on the Processing of Personal Data through Autonomous and Semi-Autonomous Systems, enacted with effect from 1 September 2023, requires a Deployer or Operator engaged in high-risk processing to appoint an Autonomous Systems Officer with competencies substantially similar to those of a Data Protection Officer. The Commissioner's own guidance states the employment analogy plainly. A system operating under the authority and for the benefit of its Deployer, it says, sits in a position substantially similar to that of an employee inside the Deployer's organisation.
The gate is nonetheless closed. Regulation 10.3.3 bars commercial use of a system for high-risk processing until audit and certification requirements have been established, and the guidance records the regulatory intent that no system may be so used until the Commissioner has promulgated them. The two sub-committees charged with recommending the criteria were to report by 30 April 2024. DIFC Consultation Paper No. 3 of June 2026, which closed on 18 July 2026, was still proposing that machinery, roughly thirty-three months after the deadline passed.
The closest thing to a per-action approval expectation is voluntary and Singaporean. The IMDA Model AI Governance Framework for Agentic AI, launched on 22 January 2026 and now served as version 1.5 of 20 May 2026, says organisations should define checkpoints or action boundaries requiring human approval, naming irreversible actions such as permanently deleting data or making payments. It asks for denial by default. The trigger is a failure of the approval infrastructure, including a supervisor who cannot be reached. It also asks firms to audit the overriding itself, tracking the human override rate, because a low rate may signal rubber-stamping and a short response time may signal review fatigue.
Binding versions of the same idea are narrow and late. California SB 947 was chaptered on 30 September 2026 and becomes operative on 1 July 2027, barring an employer from using an automated decision system output for discipline or termination where that output cannot be corroborated. Colorado SB 26-189, signed on 14 May 2026 and in force from 1 January 2027, gives a consumer the right to request human review and reconsideration following an adverse outcome. One is a duty before the decision, confined to employment. The other arrives after it.
Three labelling regimes now bind, and they disagree about what counts as content and about where the mark must go.
Three labelling regimes bind today and they do not agree on the basics. Europe marks at the provider. It labels visibly at the deployer. India labels prominently and embeds provenance metadata, but only for audio and visual material. China does both and then puts
a verification duty on the platform that distributes the file. A firm serving all three cannot satisfy them with one implementation of one standard.
The European route is Article 50, and it has applied since 2 August 2026. Article 50(2) requires providers of systems generating synthetic audio, image, video or text to mark outputs in a machine-readable format detectable as artificially generated, with technical solutions that are effective and interoperable as far as technically feasible. Article 50(4) binds deployers. They must disclose deep fakes, and must disclose AI-generated text published to inform the public on matters of public interest unless a person held editorial responsibility for it.
One European date is close. New Article 111(4), inserted by Regulation (EU) 2026/1744, requires providers of systems generating synthetic content that were placed on the market before 2 August 2026 to take the necessary steps to comply with Article 50(2) by 2 December 2026. Recital (38) calls that a transitional period of four months. The same day brings the two new Article 5 prohibitions. The Commission adopted guidelines on the Article 50 obligations on 20 July 2026, and about 190 companies and organisations had signed the Code of Practice on Transparency of AI-generated Content by the end of that month.
India's route is narrower and much faster. G.S.R. 120(E), dated 10 February 2026 and in force from 20 February 2026, inserted Rule 2(1)(wa) into the Information Technology Rules 2021, defining synthetically generated information as audio, visual or audio-visual material created or altered so as to appear real. Text generated by AI falls outside the definition. Rule 3(3)(a)(ii) requires all other such material to be prominently labelled in the visual display, or prefixed by an audio disclosure, and embedded with permanent metadata or other provenance mechanisms including a unique identifier for the computer resource used.
The same Indian instrument moved the clocks, which is where the compliance cost actually sits. It cut the takedown period on a court or government order from thirty-six hours to three, the period for a complaint about intimate imagery or impersonation from twenty-four hours to two, and the period for a removal complaint under rule 3(1)(b) from seventy-two hours to thirty-six. Grievance resolution fell to seven days. Significant social media intermediaries
carry a further duty under Rule 4(1A) to require a user declaration and then to verify it by technical means.
China's route is the most prescriptive and the longest in operation. The Measures for Labelling AI-Generated Synthetic Content were issued on 7 March 2025 and took effect on 1 September 2025, alongside the mandatory national standard GB 45438-2025. They require explicit labels by modality and an implicit label in the file metadata carrying the content attribute information, the provider's name or code and the content number. Article 6 then moves verification downstream. The dissemination platform must verify that metadata and add a graded prompt, distinguishing content that is generated, that may be generated and that is merely suspected of being generated.
Where legislatures have not acted, buyers have, and the contract clause has become the most reliable instrument of AI regulation.
The sharpest AI rules now in force in the United States are contract terms, not statutes. One free-standing federal AI statute has been enacted, the TAKE IT DOWN Act, approved on 19 May 2025 as Public Law 119-12. Almost everything else that binds a supplier binds it because a buyer wrote it down. The design is deliberate. The hook into the voluntary standard was cut in 2025 and restored in 2026 by a different hand.
OMB Memorandum M-24-18 of 24 September 2024 directed federal AI acquisition to the NIST AI Risk Management Framework by name. M-25-22 of 3 April 2025 rescinded and replaced it, and contains no reference to NIST or to the framework at all, hooking instead into
OMB's own M-25-21 for high-impact use cases. The framework itself is under revision on the instruction of America's AI Action Plan of July 2025, which asked NIST to strip out references to misinformation, diversity and climate change. No draft has been published.
The General Services Administration then put it back, by clause. GSAR 552.239-7001, on basic safeguarding of data within large language model AI systems, was issued as a GSA Regulation Deviation on 28 September 2026 and took effect on 19 October 2026. It defines its supply-chain roles by reference to the framework's Appendix A task categories, and flows down to any subcontractor performing those tasks that handles government data. A second route reaches the same place. Section 552.105 of Texas HB 149, effective 1 January 2026, makes substantial compliance with the framework's Generative AI Profile a defence to a civil penalty.
What the clause actually asks for is the useful part for any buyer. Applicability turns on a two-part test, requiring that language model functionality be a significant feature and that government data be submitted to or produced by it. Incident notification is narrowed to the model used in contract performance, with a safe harbour where a FedRAMP or CISA report carries substantially the same information. The government may suspend use. Decommissioning liability is capped at a quarter of the affected task or delivery order value.
M-25-22 supplies the diligence and documentation list sitting behind that clause. Agency processes must ensure contracts permanently prohibit the use of non-public inputted agency data and outputted results to further train publicly or commercially available algorithms absent explicit consent. Mandatory award terms cover intellectual property and government data, privacy, protection against vendor lock-in, compliance with M-25-21 for high-impact uses and ongoing monitoring. Contracts must also detail the vendor's testing procedures, and must not prohibit agencies from internally disclosing how the testing was done or what it found. The testing term is the sharpest.
Buyers elsewhere do considerably less. The European Commission's model contractual AI clauses, updated on 5 March 2025 with a factsheet of 23 January 2026, are templates and guidance rather than requirements. The United Kingdom's Guidelines for AI Procurement of 8 June 2020 have not been substantively revised in six years, and the Crown Commercial Service buyer's guidance cites only British instruments, mentioning neither the NIST framework nor ISO/IEC 42001. Oman asks for more than either. Its public policy of 1 April 2025 requires pre-deployment assessment and formal approval by the competent authority before a public body procures an AI system.
The standards gap is admitted by the bloc itself. On 7 October 2025 the GCC Secretary General named the absence of unified legislative standards as the first of the remaining challenges, while stating that member states' AI investments have reached tens of billions of dollars with plans to raise them to hundreds of billions by 2030. The consequence for a buyer is that the counterparty's obligations are largely self-declared. Oman is explicit about it. UNESCO's 2026 country profile records that Oman maintains no certified registry of AI
vendors and that no systematic monitoring mechanism has yet assessed how consistently its lifecycle governance requirements are implemented.
When the counterparty is also the regulator's shareholder, diligence has to ask questions a standard vendor assessment never reaches.
Buying from a national champion changes the diligence, not the law. Several Gulf states have built their AI capability through bodies that also write the guidance those bodies' customers follow, and in one case through the body that enforces the data protection statute. The Saudi Data and Artificial Intelligence Authority administers the Personal Data Protection Law, in force since 14 September 2023. It also issued the national AI risk management framework launched on 14 July 2026. The roles sit together.
The Emirates consolidated their machinery in 2026. On 14 June 2026 Sheikh Mohammed bin Rashid Al Maktoum approved the establishment of the Artificial Intelligence and Data Authority. The Cabinet describes it as the single national body responsible for data, artificial intelligence and digital government, reporting directly to the Cabinet. It absorbs the Office of Artificial Intelligence, the Digital Government Sector at the telecommunications regulator and the UAE Data Office. Omar Sultan Al Olama, Minister of State for Artificial Intelligence, was appointed its Chairman.
Read the mandate carefully, because its limits are the point. The Authority proposes national policies and legislation, and sets standards and guidelines for data and AI management, carrying the duty of ensuring compliance across federal entities. The Cabinet statement confers no rule-making power over the private sector. Meanwhile Federal Decree-Law No. 45 of 2021 remains in force with its Executive Regulations still unissued, so breach notification timing, cross-border transfer controls and data protection officer thresholds are not yet fixed in binding detail.
Three questions follow from that structure. The first is which capacity the counterparty acts in when it makes a representation, because a body that issues guidance can revise it. The second is what happens to the contract if state policy turns, which is not hypothetical. On 27 February 2026 the General Services Administration announced the removal of a named AI vendor from USAi.gov and from the Multiple Award Schedule in support of a presidential directive, with no statute and no rulemaking behind it, on a schedule that accounted for over USD 52.5 billion in sales in the 2025 financial year.
The third question is where the gate sits, and whether anybody holds the key. DIFC announced in April 2026 that it intended to become the world's first AI-native financial centre, while the prohibition in its own Regulation 10.3.3 still depends on certification requirements the Commissioner has not promulgated. Oman's Royal Decree No. 50/2026 of 30 April 2026 established an Artificial Intelligence Special Zone in Muscat with tax breaks and customs exemptions, and the decree introduces no AI regulation at all. Incentive and obligation issue from the same hand, at very different speeds.
A dated list of what already applies and what lands next, counted forward from 6 November 2026.
Four things bind today that did not bind a year ago, and the next twenty-two months carry more commencements than the previous four years. The pace has changed. In Europe the general date of application passed on 2 August 2026, bringing the Article 50 transparency duties and the Article 101 fines on general-purpose model providers into force. In India the synthetic content rules took effect on 20 February 2026. In China the amended Cybersecurity Law raised the ceiling tenfold on 1 January 2026. In Texas the Responsible Artificial Intelligence Governance Act commenced the same day.
The nearest date is four weeks away. On 2 December 2026 two European duties fall due together. New Article 111(4) requires providers of systems generating synthetic content placed on the market before 2 August 2026 to comply with the Article 50(2) marking duty, a transitional period recital (38) of Regulation (EU) 2026/1744 describes as four months. The same day the two new Article 5 prohibitions apply, reaching non-consensual intimate imagery and material within Article 2 points (c) and (e) of Directive 2011/93/EU.
India's own nearest date will be a week old by the time this is read. Rule 4 of the Digital Personal Data Protection Rules 2025, on the registration and obligations of Consent Managers, commences on 13 November 2026, one year after the Rules were notified. Substantive duties follow on 13 May 2027. In between, the Telecom Regulatory Authority of India closes comments on 27 November 2026 on its draft guidelines on responsible use of AI in telecommunication networks, released on 30 October 2026.
1 January 2027 is the heaviest single date on the list. Four regimes land at once. Colorado's SB 26-189 takes effect, having been signed on 14 May 2026 to replace an Act that never commenced at all. New York's RAISE Act commences as substituted by Chapter 96 of the Laws of 2026, signed on 27 March 2026. The Illinois Artificial Intelligence Safety Measures Act commences, approved on 6 July 2026. California's automated decision-making technology compliance date arrives under the Privacy Protection Agency regulations, as do the platform duties added by AB 853.
After that the dates thin out. The Monetary Authority of Singapore issued its Guidelines on Artificial Intelligence Risk Management on 7 October 2026, with sections 3 and 4 taking effect on 7 October 2027 and sections 5 and 6 by 7 October 2028. The European high-risk obligations apply to Annex III systems from 2 December 2027 and to Annex I systems from 2 August 2028, with legacy public-authority systems running to 2 August 2030. Utah's Artificial Intelligence Policy Act expires on 1 July 2027 unless extended again.
Two things are worth noticing about the shape of that list. Almost every near date is a transparency or labelling duty, and almost every far date is a conformity duty. Engineering can meet the near ones. The far ones cannot be met without the standards nobody has yet written. One harmonised standard has been cited in the Official Journal, and the deferral that bought sixteen months for the Annex III duties was justified by the absence of the rest.
No statute anywhere is framed around AI agents, and the law that already reaches them was written for electronic data interchange.
The word is missing from the statutes. Regulation (EU) 2024/1689 contains no instance of agent in any sense, against 484 instances of high-risk in the same text. Texas HB 149 contains none, and no instance of autonomous or autonomy either. The published ETSI baseline contains none. The ninety-seven titles in the ISO/IEC JTC 1/SC 42 work programme contain none. The single exception in binding European law is a conformity-assessment code.
That code is worth naming because it is so small. Annex XIV, inserted into the AI Act by Article 1 point 43 of Regulation (EU) 2026/1744, fixes the scope of designation for notified bodies, and code AIH 0401 covers systems based on other emerging AI technologies not covered by other codes, including Agentic AI. It imposes no obligation on anybody. On the evidence gathered it is the first appearance of that phrase in binding European law, and it is a taxonomy entry rather than a duty.
One instrument anywhere regulates agent autonomy as such, and it is Chinese. The Implementation Opinions on the Standardised Application and Innovative Development of Agents, published by the Cyberspace Administration of China with the National Development and Reform Commission and the Ministry of Industry and Information Technology on 8 May 2026, define an agent as an intelligent system possessing autonomous perception, memory, decision-making, interaction and execution capabilities. Item 6 is the authority rule. It requires three decision modes to be delineated, and states that an agent's execution of operations must not exceed the scope of the user's authorisation.
Its legal character matters too. It is an implementation opinion giving effect to the State Council's AI Plus Opinions of 21 August 2025, so it binds state organs and steers industry rather than carrying penalties of its own. Item 11 sets tiered governance, with cyberspace departments and sector regulators fixing which scenarios are open in sensitive fields, and self-testing and industry self-discipline in lower-risk areas. Item 4 proposes an agent registration platform with digital identity management and capability declaration.
What binds an agent today in the West is older than the technology. Section 14 of the Uniform Electronic Transactions Act, enacted in Michigan as Act 305 of 2000 and immediately effective on 16 October 2000, provides that a contract may be formed by the interaction of electronic agents even if no individual was aware of or reviewed their actions. E-SIGN is narrower still. It is a validity rule that presupposes legal attribution rather than supplying it, and the attribution question is the gap.
Two more binding regimes already reach systems that transact without per-action approval, and both are sectoral. Commission Delegated Regulation (EU) 2017/589, in force since 3 January 2018, imposes organisational requirements on investment firms engaged in algorithmic trading, including pre-trade controls and the ability to cancel unexecuted orders. FINRA has solved the accountability question once, by registering a named human responsible for the design of an algorithmic trading strategy, mandatory since 30 January 2017. Neither was written for consumer agents. Both show what a workable answer looks like.
Build once to the strictest version of each duty, because the strictest version is rarely in the same jurisdiction twice.
Building five compliance programmes for five rulebooks costs more than building one to the highest bar, and the arithmetic is not close. The duties overlap heavily in substance and diverge in detail, so the expensive part is never the control itself but the five versions of the evidence. The harder question is where the highest bar sits for each duty, because it moves. No jurisdiction holds every column.
On documentation the highest bar is European and auditable. EN 18286:2026 turns the Article 17 quality management requirements into auditable requirements for providers, and Article 40 of Regulation (EU) 2024/1689 gives conformity with a cited harmonised standard a presumption of conformity running from publication in the Official Journal. ISO/IEC 42001:2023 is certifiable, which ISO's own catalogue establishes through the existence of ISO/IEC 42006:2025 setting requirements for certification bodies. No statute examined demands either. Both are the cheapest way to produce one evidence set rather than five.
On labelling the bar is Chinese. It is the only one with a mandatory technical standard behind it. GB 45438-2025 was issued on 28 February 2025 and took effect on 1 September 2025 as a mandatory national standard, specifying methods for explicit labels and for implicit labels embedded in file metadata. Article 6 of the Labelling Measures then pushes verification downstream to the dissemination platform. A firm that builds to the Chinese metadata requirement has satisfied the European machine-readable duty and the Indian provenance duty on the way.
On human oversight the highest bar is voluntary and Singaporean, which is awkward but true. The IMDA agentic framework asks for defined action boundaries requiring human approval before irreversible actions. It asks for denial by default when the approval infrastructure fails, and for audit of the override rate itself. Nothing binding goes that far. California SB 947 requires corroboration before discipline or termination from 1 July 2027, and Colorado SB 26-189 gives a right to human review after an adverse outcome from 1 January 2027.
The European tiers set the ceiling. They are what should size the programme. Non-compliance with the Article 5 prohibitions attracts up to EUR 35,000,000 or 7 per cent of worldwide annual turnover, whichever is higher. Operator duties and the Article 50 transparency obligations attract up to EUR 15,000,000 or 3 per cent, and the same tier applies to general-purpose model providers under Article 101, in force since 2 August 2026. China's Personal Information Protection Law reaches RMB 50,000,000 or 5 per cent of turnover, and the Indian Act reaches Rs 250 crore.
Two cautions belong with the recommendation. The first is that building to the highest bar does not discharge the lowest, because some duties are procedural rather than substantive,
and a Chinese algorithm filing or an Indian Consent Manager registration cannot be satisfied by a better control somewhere else. The second is that the highest bar can be lowered. The European high-risk obligations moved out by sixteen and twelve months in July 2026, and the Colorado duty of care was repealed on 14 May 2026, six weeks before the deferred date on which it was due to apply.
What this edition covers, how a fact was accepted, and what was deliberately left out of it.
This edition states the position as at 6 November 2026. Nothing later is included. Every date and figure rests on a primary instrument or on a named, dated secondary source, and where only commentary could be reached the commentator is named in the text. Where two sources disagree, the primary one governs and the disagreement is recorded rather than resolved silently. Four fact bases sit behind these chapters, covering the global picture, the United States, India and the Gulf.
Three words carry a fixed sense throughout. In force means a provision applies to somebody today. Scheduled means a commencement date is fixed in an instrument that has already been made, as with the Indian substantive rules on 13 May 2027 or the European Annex III duties on 2 December 2027. Announced means a body has said it will do something without issuing the instrument, as with the Securities and Exchange Board of India's statement of 10 October 2026 that guidelines on AI and machine learning would follow shortly.
Every count carries its rule. The number of American state AI laws enacted during 2026 ranges from three to 115 depending on definition, so every figure used here names its source and its window. No count of jurisdictions holding a binding AI-specific statute appears
anywhere in this edition. None is published by the International Association of Privacy Professionals tracker, by the OECD Policy Navigator or by the Stanford AI Index.
Four things were left out deliberately. Data centre and compute policy sits outside scope, as does export control. Copyright and training data appear only where an instrument imposes a disclosure duty, which is why the British report of 18 March 2026 is noted for dropping a broad exception with opt-out and not discussed further. Litigation is noted where it changed what applies and otherwise left alone. Figures that could not be traced to a page actually opened were excluded, including certification counts for ISO/IEC 42001, for which no public authoritative figure exists.
The work a buyer should have in place, how an engagement of this kind is run, and what it produces.
A buyer facing four rulebooks needs four things in place before it needs advice. The first is an inventory naming every system, its owner and the decision its output feeds. The second is a classification decision for each one, recorded with the test applied and the date. The third is a documentation set that would survive a regulator's request without reconstruction. The fourth is a named human accountable for oversight of every system above the lowest tier. That is the floor. Volge builds those four, in that order, and then maintains them against the dates in the preceding chapter.
The common failures are consistent. Inventories are built once and never maintained, so a classification that was accurate on the day it was signed is wrong within a quarter.
Classification is delegated to procurement, which sees the contract rather than the decision the output feeds. Evidence is written after the fact, which is the one thing no regulator accepts. Oversight is assigned to somebody who has neither the time nor the authority to refuse an action.
Every body cited in this read, with the instrument and the date carried in the text.
Volge is an engineering firm providing enterprise technology transformation services to clients across the Middle East, India, APAC, Europe, and North America.
Delivery stands on two pillars: Forward Deployed Engineering, with senior specialists embedded in the client’s own team, and Intelligent Command Centre Operations, run from our Global Delivery Centres, which keeps what they build running. Senior engineers stay with an engagement from the first diagnosis through to production.
Our services span Strategy, Data, Platforms, AI and Automation, Cyber Trust, Technology Operations, and Applied Functional Engineering. Engagements begin with a diagnosis. The firm is built on working autonomy: engineers decide, innovate, co-create, and own their work. Motivated engineers build the systems that sustain our clients’ growth.
This edition may be shared with organisations evaluating a Volge engagement. It may not be altered or republished without Volge’s permission, and figures carry their stated basis and date. Volge, Intelligent Command Centre, and Intelligent Command Centre Operations are marks of Volge; other names are the trademarks of their respective owners.