AI regulation is becoming sector-specific and operational. IOSCO, Hong Kong, Singapore, the UAE and the EU have advanced measures covering capital markets, licensed firms, agentic AI, healthcare and AI Act deadlines. Legal force and effective dates remain the key compliance questions.
Summary
- [Global capital markets] IOSCO’s toolkit is official supervisory material, not domestic law. It gives member regulators methods for examining AI use. A firm faces direct duties only through applicable law, licence conditions, rules, or adopted guidance.
- [Hong Kong] SFC Circular 26EC32 sets current expectations for licensed firms facing AI-enabled cyberattacks. Senior management and the MIC-IT retain responsibility. Existing licensing, internal-control, notification, and disciplinary rules give the circular practical force.
- [Hong Kong] SFC Circular 24EC55 already sets supervisory expectations for AI language-model use linked to regulated activities. Investment recommendations, advice, and research are generally high risk. High-risk adoption may require notification, validation, human review, repeated disclosure, and continuing monitoring.
- [Singapore] IMDA’s Model AI Governance Framework for Agentic AI, version 1.5, is voluntary guidance. It does not displace contract, negligence, electronic-transactions, or sector law. Its controls may inform reasonable-care evidence and contract drafting.
- [United Arab Emirates] The Cabinet approved a health AI policy and directed officials to draft federal legislation. The announcement did not enact the announced licensing, accreditation, liability, patient-rights, or coordination rules.
- [European Union] Regulation (EU) 2026/1744 was published on 24 July 2026. It enters into force on 27 July 2026. It changes specified dates, while the AI Act’s general 2 August 2026 application date remains.
- [Cross-border] One enterprise control set can support all five measures. Local legal force, notification triggers, data rules, operator roles, and commencement dates still require separate mapping.
Analysis by Issue
Global capital-markets supervision through IOSCO
The IOSCO toolkit does not itself bind regulated firms. IOSCO is an international association of securities regulators, not a legislature. Its 25 May 2026 final report gives regulators practical methods for supervising AI systems used by regulated entities. The stated objectives concern investor protection, market integrity, and financial stability. IOSCO, Supervisory Toolkit for AI Use in Capital Markets, Final Report FR/02/2026 (25 May 2026).
Domestic law remains the source of enforceable duties. A member regulator may use the toolkit during inspections, thematic reviews, supervisory dialogue, or later rulemaking. Those uses can affect the evidence a regulator expects before formal local adoption. They cannot create a penalty, private claim, or licence condition without a valid local basis.
The toolkit therefore has persuasive supervisory weight rather than direct legal force. A capital-markets firm should prepare an evidence record that local supervisors can test. That record should identify AI systems, owners, use cases, third parties, testing, monitoring, incidents, disclosures, and material changes. The exact obligation depends on the firm’s home and host jurisdictions.
IOSCO cannot convert recommendations into binding national law. Each regulator retains its statutory powers, procedures, and territorial reach. Firms should therefore track local adoption instead of treating the toolkit as a universal rulebook.
Hong Kong licensed firms and AI controls
SFC Circular 26EC32 has practical supervisory force through existing Hong Kong duties. It does not create a separate AI statute. It applies to licensed corporations, SFC-licensed virtual asset service providers, and their associated entities. The circular states that senior management, including the MIC-IT, remains ultimately responsible. SFC Circular 26EC32, paras. 1–3 (2 June 2026).
The circular directs firms to examine five control areas. They are patching and vulnerability management, access and privilege controls, detection and monitoring, third-party supply chains, and incident response. Firms should keep current technology inventories and support same-day prioritisation when new threats emerge. High-impact actions should use maker-checker controls. Material cyber incidents require prompt SFC notification. Id. paras. 5–19.
These directions connect to binding and disciplinary sources. Licensed persons must remain fit and proper under the Securities and Futures Ordinance. They must comply with applicable subsidiary legislation, codes, and SFC guidance. The SFC may discipline misconduct or a loss of fitness and properness. Available sanctions include licence action, prohibitions, reprimands, and financial penalties. Securities and Futures Ordinance (Cap. 571), ss. 129, 193–194.
Circular 24EC55 adds direct AI-use expectations for licensed corporations. It applies when AI language models provide services or functions linked to regulated activities. Source, ownership, and open-source status do not remove the circular’s reach. The SFC generally treats investment recommendations, advice, and research as high-risk uses. SFC Circular 24EC55, paras. 6–8 (12 November 2024).
For those uses, the firm should validate the model before approval and after material changes. It should test the full input-to-output process and document results. It should keep a human reviewer in the output path. Client interfaces should identify AI interaction and output limits on each relevant interaction. The firm remains accountable when it uses an external or open-source model. Id. paras. 9–29.
High-risk adoption can also trigger a filing duty. Circular 24EC55 points to significant changes in business nature or service types. The Information Rules require notice for specified changes under section 135 of the Ordinance. Securities and Futures (Licensing and Registration) (Information) Rules (Cap. 571S), s. 4 and Sch. 3; SFC Circular 24EC55, paras. 30–31.
The strongest limitation concerns legal form and proportionality. The circulars use risk-based language and contain several “should” statements. They do not make every listed control an identical statutory duty for every firm. Any disciplinary case must still rest on the Ordinance, subsidiary rules, applicable codes, or fitness-and-properness findings. The SFC must also follow its disciplinary process. The MIC regime creates no additional liability by itself.
A Hong Kong firm should treat both circulars as current examination criteria. It should map each use to the regulated activity, risk tier, responsible manager, controls, notification analysis, and supporting records. A board-level approval alone will not establish operational compliance.
Singapore agentic AI and existing private law
IMDA’s version 1.5 document is voluntary official guidance. It does not create a licence, civil claim, offence, or statutory safe harbour. The Model AI Governance Framework for Agentic AI addresses organisations that build or deploy AI agents. It states that humans remain ultimately accountable. IMDA, Model AI Governance Framework for Agentic AI, version 1.5 (2026).
The document organises deployment controls around four dimensions. Organisations should bound use cases, autonomy, tools, and data access. They should assign meaningful human approval points. They should test through the agent lifecycle and restrict external services. They should disclose agent identity and train affected users. Version 1.5 adds multi-agent systems, third-party agents, automation bias, and deployment case studies.
Existing law still determines legal responsibility. The Electronic Transactions Act 2010 recognises electronic contract formation. Section 15 prevents invalidity solely because automated message systems acted without review for each action. Parties may vary several statutory effects by agreement under section 5. Electronic Transactions Act 2010, ss. 5, 11, 15.
Section 16 supplies only a narrow human-input-error remedy. It applies when a natural person makes an input error and lacks a correction opportunity. Prompt notice and the absence of received benefit are required. It does not grant a general right to reverse an agent’s unwanted or mistaken transaction. Id. s. 16.
Quoine Pte Ltd v B2C2 Ltd confirms that deterministic algorithms can form contracts. The Court of Appeal examined programmer knowledge for unilateral mistake. The court stressed the programmed parameters and the relevant knowledge period. The reasoning does not resolve every issue raised by learning or adaptive agents. Quoine Pte Ltd v B2C2 Ltd [2020] SGCA(I) 2 at [96]–[105].
Negligence remains fact-sensitive. Singapore applies factual foreseeability, legal proximity, and policy controls when deciding whether a duty exists. Contracts can affect that analysis by defining rights, liabilities, and reliance. Spandeck Engineering (S) Pte Ltd v Defence Science & Technology Agency [2007] SGCA 37 at [73]–[85].
No located Singapore enactment gives an AI agent separate legal personality. Responsibility therefore stays with natural or legal persons under contract, tort, statute, agency, employment, and sector rules. The result depends on authority, control, representations, fault, causation, loss, and agreed risk allocation.
IMDA’s guidance may inform evidence of reasonable precautions. It does not prove due care by itself. Non-adoption does not establish negligence by itself. Organisations should translate its measures into permissions, approval gates, logs, monitoring, vendor terms, incident procedures, and transaction limits.
United Arab Emirates health AI policy and current law
The 18 May 2026 Cabinet announcement did not enact a unified health AI law. The Cabinet approved the National Policy for Advancing Digital Healthcare Services and Artificial Intelligence in the Healthcare Sector. It also approved drafting a future federal law. The announced bill would address development, licensing, accreditation, operation, health data, safety, liability, patient rights, and federal-local coordination.
A direction to draft legislation is not enacted legislation. No located official text supplies a law number, publication date, commencement date, transitional rule, offence, or penalty for the proposed regime. The policy may guide public administration, procurement, and later approvals. It does not, on the available record, create the announced unified private-sector licence or liability code.
Current federal health laws still apply. Federal Law No. 2 of 2019 covers health information technology across the UAE, including free zones. It regulates confidentiality, system access, health-data storage, overseas processing, and retention. Covered health data generally requires at least 25 years of retention from the last health procedure. Federal Law No. 2 of 2019, arts. 2, 4, 12–16, 20.
The general federal data law does not displace sector health-data legislation. Federal Decree-Law No. 45 of 2021 excludes personal health data governed by separate legislation. A health AI operator must therefore begin with the health-sector rules, then test any remaining data law. Federal Decree-Law No. 45 of 2021, art. 2(2)(e).
Professional and product rules also apply before the future bill. Physicians need licences and must act within their licensed scope. A physician may not diagnose or treat through an unexamined method without Health Authority approval. Federal Law No. 5 of 2019, arts. 2, 4, 23–24.
Federal Decree-Law No. 38 of 2024 expressly includes qualifying AI-based products within “Medical Equipment.” Product classification turns on intended purpose and statutory criteria. A diagnostic or therapeutic AI product may therefore face current product approval and circulation duties. Federal Decree-Law No. 38 of 2024, arts. 1–2.
Federal medical-liability law also remains relevant to clinical use. It governs professional fault and the medical-liability process. The announced policy does not replace that law. Federal Decree-Law No. 4 of 2016; Cabinet Resolution No. 40 of 2019.
The main implementation risk is premature reclassification of policy as law. A healthcare business should not wait for the future bill before addressing present duties. It should map health-data location, professional approvals, facility licensing, product classification, clinical validation, consent, records, and incident handling now. It should separately monitor the bill’s official publication and commencement.
European Union AI Act dates
Regulation (EU) 2026/1744 is enacted but not yet in force on the as-of date. The Official Journal published it on 24 July 2026. Article 4 sets entry into force on the third day after publication. The entry date is therefore 27 July 2026. Regulation (EU) 2026/1744, art. 4, OJ L 2026/1744, 24 July 2026.
Until 26 July 2026 ends, the original Article 4 of Regulation 2024/1689 remains operative. It requires providers and deployers to pursue a sufficient AI-literacy level “to their best extent.” From 27 July, the replacement Article 4 requires measures supporting AI-literacy development. It states that no provider or deployer must guarantee a specific individual level. Regulation (EU) 2024/1689, art. 4; Regulation (EU) 2026/1744, art. 1(5).
The amendment does not postpone the entire AI Act. The general application date remains 2 August 2026. The amendment defers specified high-risk provisions in Chapter III, Sections 1 to 3. Annex III systems under Article 6(2) move to 2 December 2027. Annex I product systems under Article 6(1) move to 2 August 2028. Article 6(5) is excluded from those deferrals. Regulation (EU) 2026/1744, art. 1(40).
Other new dates also matter. Newly added Article 5 prohibitions apply from 2 December 2026. Providers of legacy synthetic-content systems must meet Article 50(2) by that date. The legacy rule covers systems placed on the market before 2 August 2026. Articles 102 to 110 apply from 27 July 2026. Id. art. 1(39)–(40).
Territorial application still depends on Article 2. A non-EU provider can fall within scope through EU market placement or EU-used output. Operator classification also matters. Providers, deployers, importers, distributors, product manufacturers, and authorised representatives carry different duties. Regulation (EU) 2024/1689, art. 2.
The main error would be saying that “the EU AI Act was delayed.” That statement is too broad. The enacted amendment changes named provisions and dates. All other applicable dates require a provision-by-provision check.
Organisations should preserve evidence for both legal states. They should record the Article 4 change on 27 July and the general 2 August date. They should separately classify Annex I, Annex III, synthetic-content, prohibited-practice, and operator issues. A central calendar should link each date to an owner and proof of completion.
Cross-border implementation
Legal-force register
Classify every item as binding law, binding licence condition, supervisory guidance, voluntary guidance, policy, proposal, or enacted future change. Record the issuer, scope, effective date, territory, and enforcement path. This prevents an IOSCO report or UAE policy from being treated as a statute.
Named accountability
Assign a business owner, technical owner, risk owner, and final approver for each system. Hong Kong requires specific senior-management and MIC-IT attention. Singapore’s guidance also places accountability on people, not agents.
Action and access boundaries
Define permitted tools, data, systems, transaction values, and reversible actions. Require human approval for material, irreversible, client-facing, clinical, or privileged actions. Use least privilege and maker-checker controls where the use justifies them.
Testing and operational evidence
Maintain pre-deployment tests, change tests, prompt-variation tests, red-team results, monitoring records, incident logs, and decommissioning records. Keep inventories for models, components, data sources, connectors, and vendors. Evidence should match the legal duty and risk tier.
Incident and notification matrix
Link each incident type to internal escalation, regulator notice, client notice, data-breach duties, and preservation steps. Hong Kong material cyber incidents and high-risk business changes need specific analysis. UAE health data and EU operator duties require separate branches.
Contracts and data
Allocate agent authority, approval limits, audit rights, model changes, data use, security, continuity, indemnities, and termination rights. Contract terms do not erase statutory duties. They can determine recourse and shape negligence analysis.
Deadline control
Use a dated obligations register rather than a single global launch date. The EU amendment alone requires separate entries for 27 July, 2 August, 2 December 2026, 2 December 2027, and 2 August 2028. Each entry should identify the affected systems and evidence owner.
