Policy & Standards Library
82 regulations · 443 obligations · 29 standards · 23 jurisdictions · 431 official sources
Content as of 2026-09-12 · corpus 2026.09|Compiled from public official sources with links and review dates for compliance decision support. Rules change; have licensed professionals review before external disclosure, regulatory filing or contractual commitments.
EU Artificial Intelligence Act
The first comprehensive horizontal AI law, setting obligations by risk tier. Unacceptable-risk practices are banned. High-risk systems must meet requirements on risk management, data governance, technical documentation, logging, transparency, human oversight, accuracy and cybersecurity, undergo conformity assessment, bear the CE marking and be registered in the EU database. Providers of general-purpose AI models owe technical documentation, downstream documentation, a copyright policy and a public training-content summary; models meeting the systemic-risk threshold additionally owe model evaluation, risk mitigation, serious-incident reporting and cybersecurity protection. Limited-risk uses trigger transparency notices and synthetic-content marking. The AI Omnibus (Regulation (EU) 2026/1744, in force July 2026) postponed the high-risk application dates, simplified documentation for SMEs and small mid-caps, added two prohibited practices, and permits processing of special categories of personal data for bias detection under strict conditions.
AI Omnibus (Digital Omnibus on AI)
A targeted simplification of the AI Act that also amends the EU Aviation Safety Regulation (EU) 2018/1139 and the Machinery Regulation (EU) 2023/1230. Key content: postponement of the high-risk application dates (Annex III to 2 Dec 2027, Annex I to 2 Aug 2028); extension of SME-only reliefs to small mid-cap companies (SMCs); simplification of the AI literacy duty with a stronger promotional role for the Commission and Member States; clarification of the "safety component" boundary, so that AI used solely for non-safety-related user assistance, performance optimisation, service efficiency, automation, convenience or quality control does not qualify; new prohibitions on generating non-consensual intimate imagery and child sexual abuse material; a new Art. 4a permitting processing of special categories of personal data for bias detection and correction under strict conditions; expanded regulatory sandboxes including an EU-level sandbox; a single application and unified assessment procedure for notified bodies; and a presumption that systems complying with the Cyber Resilience Act meet the AI Act Art. 15 cybersecurity requirement.
EU General Data Protection Regulation (GDPR)
The horizontal foundation of EU personal data protection. It is built on seven principles — lawfulness, fairness and transparency, purpose limitation, data minimisation, accuracy, storage limitation, integrity and confidentiality, and accountability — and sets out controller and processor duties (records of processing, data protection by design and by default, security measures, processor contracts, data protection impact assessments, data protection officers, EU representatives), data subject rights (information, access, rectification, erasure, restriction, portability, objection and protection against solely automated decisions), personal data breach notification, and the Chapter V transfer mechanisms (adequacy decisions, standard contractual clauses, binding corporate rules, certification and codes of conduct, and derogations). For AI companies the three highest-frequency risk areas are the lawful basis for training data, the erasability of personal data inside training sets, and explanation and human intervention for automated decisions.
Personal Information Protection Law
China’s foundational personal-information statute. It sets the lawful bases for processing (Art. 13), the notice obligation (Art. 17), separate consent and necessity requirements for sensitive personal information (Arts. 28–30), transparency and opt-out for automated decision-making (Art. 24), personal information protection impact assessments (Arts. 55–56), the three outbound-transfer routes (Art. 38), and penalties of up to “RMB 50 million or 5% of the prior year’s turnover” (Art. 66).
Data Security Law
The foundational data statute. It establishes classified and graded national data protection with stricter treatment of national core data (Art. 21), requires processors to build whole-lifecycle data security management systems (Art. 27), obliges important-data processors to run periodic risk assessments and file the reports (Art. 30), and in Art. 36 imposes a blocking rule: data stored in China may not be provided to foreign judicial or law-enforcement bodies without approval from the competent PRC authority. Art. 25 confirms that data qualifying as a controlled item is subject to export control.
Regulations on Network Data Security Management
The first comprehensive administrative regulation sitting under the Cybersecurity, Data Security and Personal Information Protection Laws, converting their principles into executable actions. Six points matter most to outbound-facing AI companies: (1) Art. 10 requires a report within 24 hours where a product or service risk endangers national security or the public interest; (2) Art. 12 requires records of providing or entrusting personal information and important data to be kept for at least three years; (3) Art. 19 requires generative AI providers to strengthen the security management of training data and training-data processing; (4) Art. 26 requires an offshore processor’s PRC entity or representative under PIPL Art. 53 to be reported to the municipal cyberspace authority; (5) Art. 28 pulls processors of more than 10 million individuals’ data into part of the important-data regime; and (6) Art. 35 enumerates every lawful basis for transferring personal information abroad.
Measures for Security Assessment of Outbound Data Transfers
The highest-threshold and longest-running of the three outbound routes: the processor applies through its provincial cyberspace authority to the CAC, submitting the application form, a self-assessment report on outbound transfer risk, and the draft legal instrument with the offshore recipient. The provincial authority completes a completeness check within five working days of receipt (Art. 7); once the CAC accepts the filing it issues a written acceptance notice, and the assessment must be completed within 45 working days of that notice, extendable where the case is complex or materials need correction (Art. 12). Art. 9 lists six matters the legal instrument must cover at minimum. The validity of an approval was extended to three years by CAC Order No. 16.
Measures on the Standard Contract for Outbound Transfer of Personal Information
The most widely used of the three outbound routes: execute the standard contract with the offshore recipient strictly in the form annexed to the Measures (Art. 6 requires the annexed template to be followed and its core clauses not to be altered), run a personal information protection impact assessment before the transfer, and file with the provincial cyberspace authority within ten working days of the contract taking effect. The filing is a post-execution notification rather than a prior approval, but provincial or higher authorities may summon the processor for an interview and require remediation where they identify significant risk or an incident.
Provisions on Promoting and Regulating Cross-border Data Flows
The master switch of the current outbound-data regime. Arts. 7 and 8 set the volume thresholds across the assessment, standard contract and certification routes; Arts. 3–6 create four families of exemption (data from international trade and similar activities containing no personal information or important data; personal information collected abroad, processed in China and sent back out; the four specific cases listed in Art. 5; and data outside a free-trade-zone negative list); Art. 9 fixes the validity of an assessment approval at three years, extendable by three more; Art. 2 confirms that data not notified or publicly designated as important data need not be declared as such; and Art. 13 provides that these Provisions prevail over CAC Orders No. 11 and No. 13 where inconsistent. Any outbound scenario should be analysed from these Provisions, not from the 2022 assessment measures.
Interim Measures for the Management of Generative AI Services
China’s first departmental rule dedicated to generative AI services. Art. 2 scopes it to “services that use generative AI technology to provide generated text, images, audio, video and similar content to the public within the PRC”, and Art. 22 makes clear that a provider includes an organisation or individual that supplies the service through a programmable interface. The core duties sit in training data (Art. 7), annotation (Art. 8), service conduct (Arts. 9–15), and filing and assessment (Art. 17). Art. 20 empowers the CAC to have technical measures taken against non-compliant services supplied into China from abroad.
Provisions on the Administration of Deep Synthesis in Internet Information Services
The departmental rule governing internet information services that apply deep synthesis technology inside China, and the technical bedrock of Chinese AI content governance. Art. 23 defines deep synthesis technology as the use of deep learning, virtual reality and other generative or synthetic algorithms to produce text, images, audio, video and virtual scenes, listing six families: passage generation and question-answer dialogue; text-to-speech and voice conversion; music and ambient-sound generation; face generation, face swapping, attribute editing and pose manipulation; image generation, enhancement and restoration; and 3D reconstruction and digital simulation. Duties fall on both the “deep synthesis service provider” and the “deep synthesis technical supporter” (Arts. 14, 15 and 22), so teams that supply only the underlying model without facing end users are also in scope.
Provisions on the Administration of Algorithmic Recommendation in Internet Information Services
The institutional hub of Chinese algorithm governance and the origin of the term “algorithm filing”. It applies to the use of algorithmic recommendation technology to provide internet information services inside China (Art. 2). Three layers matter: accountability (the eight management systems in Art. 7, periodic review and validation in Art. 8, and user-tag management in Art. 10); user rights (Art. 16 requires prominent notice and publication of the basic principles, purpose and main operating mechanism; Art. 17 requires an option not targeted at personal characteristics or a convenient off switch; Art. 21 prohibits discriminatory pricing based on user profiling); and the regulatory interface (graded and classified supervision in Art. 23, filing and publication in Arts. 24–26, and the security assessment in Art. 27).
Measures for Labelling AI-Generated and Synthetic Content
This turns the labelling duties in Art. 12 of the Generative AI Interim Measures and Arts. 16–17 of the Deep Synthesis Provisions into an executable technical regime. Art. 2 scopes it to network information service providers falling within the Algorithmic Recommendation Provisions, the Deep Synthesis Provisions or the Generative AI Interim Measures; Art. 3 defines AI-generated and synthetic content as text, images, audio, video, virtual scenes and similar information produced or synthesised using AI. The regime has two legs: a human-visible explicit label (Art. 4, with placement specified separately for text, audio, images, video and virtual scenes) and a machine-readable implicit label (Art. 5, written into the file metadata and carrying the generated-content attribute, the provider’s name or code and a content identifier). Art. 11 adds that the labelling must also meet the applicable mandatory national standard, namely GB 45438-2025.
Export Control Law
China’s basic export-control statute. Art. 2 defines controlled items as dual-use items, military products, nuclear items and other goods, technologies and services relating to national security and interests and to non-proliferation obligations, and states that “controlled items as referred to above include technical data and other data relating to the items”. Its third paragraph defines export control as prohibitions or restrictions on transferring controlled items from within China to outside it, and on PRC citizens, legal persons and unincorporated organisations providing controlled items to foreign organisations and individuals — the “providing to foreign organisations and individuals” limb brings deemed exports into scope. Art. 12 establishes the licensing regime and the catch-all. Art. 18 creates the control list. Art. 44 provides for liability where offshore organisations and individuals breach the export-control rules and harm PRC national security and interests.
Regulations on Export Control of Dual-Use Items
The principal administrative regulation implementing the Export Control Law for dual-use items, replacing the four earlier item-specific regulations. Key mechanics: Art. 14 imposes a licensing regime on dual-use exports; Art. 15 requires an individual licence, a general licence or a registration-based export voucher; Art. 17 gives MOFCOM 45 working days from acceptance to grant or refuse; Art. 12 caps each round of temporary controls at two years; Art. 24 requires an end-user and end-use certificate from the end user, and MOFCOM may additionally require a certificate issued or authenticated by a government body in the end user’s jurisdiction; Art. 27 requires related records to be kept for at least five years; Art. 38 requires immediate reporting of any foreign-government request for export-control-related visits or on-site inspections, which may not be accepted without MOFCOM’s consent; and Art. 49 allows MOFCOM to require operators to follow the Regulations for dual-use items manufactured abroad that contain, integrate or incorporate specified Chinese-origin dual-use items, or that are made using specified Chinese-origin technology.
U.S. DOJ Bulk Sensitive Personal Data Rule (Data Security Program)
Issued under the International Emergency Economic Powers Act, the rule imposes two tiers of control on four transaction types (data brokerage, vendor agreements, employment agreements and investment agreements) between U.S. persons and countries of concern (China including Hong Kong and Macau, Cuba, Iran, North Korea, Russia and Venezuela) or covered persons subject to their jurisdiction, where the transaction involves government-related data or bulk U.S. sensitive personal data. Data brokerage and transactions involving bulk human omic data or biospecimens are prohibited outright. Vendor, employment and investment agreements are restricted transactions, permitted only where the CISA security requirements are fully implemented and the compliance program, audit, recordkeeping and reporting duties are met. DOJ itself describes the program as effectively an export control.
U.S. Export Administration Regulations (advanced computing, AI model weights, FDP rules and the Entity List)
The EAR determines licensing by item classification (ECCN), destination country group and end user. Four control points matter most to AI companies: license requirements and tiered review policy for advanced computing hardware and associated software and technology (3A090, 4A090 and the .z entries) and semiconductor manufacturing equipment; the worldwide license requirement for AI model weights (ECCN 4E091) and the strict conditions of License Exception AIA; the Foreign Direct Product rules, which can bring chips, equipment and model weights produced entirely outside the United States under EAR jurisdiction; and the Entity List and related restricted end-user regimes, where the 50 percent affiliates rule is currently stayed.
Illinois Biometric Information Privacy Act (BIPA)
BIPA is the most litigation-heavy biometric privacy statute in the United States. A private entity in possession of biometric identifiers or biometric information must develop and publish a retention and destruction policy, destroying the data when the initial purpose is satisfied or within three years of the individual last interaction, whichever comes first. Before collecting, capturing, purchasing or otherwise obtaining such data it must give written notice that data is being collected or stored, give written notice of the specific purpose and term, and obtain a signed written release. It may not sell, lease, trade or otherwise profit from the data, and may not disclose, redisclose or disseminate it except as the statute allows. Critically, BIPA provides a private right of action with liquidated damages for each violation of $1,000 for negligent and $5,000 for intentional or reckless violations, or actual damages if greater, plus attorney fees and costs, which has made it a long-running source of class actions. P.A. 103-769 (eff. 2024-08-02) made two decisive changes: section 14/20(b) and (c) collapse repeated collection from the same person using the same method, and repeated disclosure to the same recipient, into a single violation with at most one recovery, ending the per-scan damages theory; and section 14/10 brings an electronic signature within the definition of written release.
Singapore Personal Data Protection Act
The PDPA is Singapore’s general personal data protection statute, built around consent, notification, purpose limitation, accuracy, protection, retention limitation, transfer limitation, access and correction, and accountability, with a mandatory breach-notification regime layered on from 2021. It distinguishes organisations from data intermediaries: under s.4(2) an intermediary processing personal data on behalf of and for the purposes of another organisation **pursuant to a contract evidenced or made in writing** carries only the s.24 protection obligation, the s.25 retention-limitation obligation and the duties in ss.26C(3)(a) and 26E — without a written contract the carve-out does not apply and the full set of obligations bites. Section 4(3) confirms that outsourcing does not reduce the principal organisation’s own liability. PDPC enforcement runs through directions plus financial penalties, with turnover-linked caps since October 2022.
PDPC Advisory Guidelines on Use of Personal Data in Generative AI
The guidelines clarify how the PDPA applies across the generative AI lifecycle. Development: web-scraped data that is genuinely publicly available may rely on the Publicly Available Exception in paragraph 1 of Part 2 of the First Schedule instead of consent, but data behind digital barriers (paywalls, registration, authentication, geo/age gates, anti-scraping controls) requires a documented case-by-case assessment; where an organisation trains on its own User Data, a general “product improvement” statement is not sufficient consent and an AI-Specific Notification is required. Deployment: stakeholders are classified as Model Providers, System Providers and System Deployers, each mapped to organisation or data intermediary status. Post-deployment: access and correction obligations apply to personal data used for model or system development. The guidelines are advisory and the PDPA prevails in the event of inconsistency.
Japan Act on the Protection of Personal Information (APPI)
The APPI is Japan’s general personal information statute. Core duties include specifying and publicising the purpose of use, not exceeding that purpose, obtaining prior consent for sensitive (“special care-required”) personal information, security control measures, supervision of contractors, consent and record-keeping for third-party provision, and special restrictions on provision to third parties in foreign countries. Article 171 establishes extraterritorial scope: a business operator abroad is covered where it handles the personal information of persons in Japan in connection with supplying goods or services to persons in Japan. A major amendment Act was promulgated in July 2026 but is not yet in force.
Japan 2026 Act Partially Amending the Act on the Protection of Personal Information
This is the second major overhaul under the APPI’s three-yearly review cycle. It was promulgated on 17 July 2026 and is not yet in force; commencement is in principle within two years of promulgation, on a date to be fixed by Cabinet Order. There are four strands. First, promoting appropriate data use: a new exception for the creation of statistical information removes the consent requirement for third-party provision of personal data and for acquisition of publicly available sensitive information where the data will be used solely to create statistical information — which the PPC expressly states includes AI development that can be characterised as statistics creation. Second, risk-calibrated rules: children under 16, face-feature data, contractor duties, and a relaxation of individual breach notification. Third, preventing improper use: a new prohibition on improper use and improper acquisition of personally referable information that enables contact with a specific individual, plus verification duties for opt-out list businesses. Fourth, enforcement effectiveness: broader recommendation and order powers, higher statutory criminal penalties, and a new administrative surcharge.
Korea Personal Information Protection Act (PIPA)
PIPA is Korea’s general personal information statute. The 2023 overhaul established five routes for cross-border transfer, the right to refuse and to an explanation of automated decisions, and surcharges of up to 3% of total revenue. The 2026 amendment (Act No. 21445, in force 11 September 2026) adds three things that matter to foreign entrants: a higher surcharge tier of up to 10% of total revenue (or KRW 5 billion where revenue is absent or not ascertainable) for serious or repeated violations; express designation of the business owner or representative as the ultimate person responsible for personal information processing and protection, with reinforced CPO staffing, budget and designation-filing requirements; and expansion of breach notification to forgery, alteration and damage, plus a new duty to notify where there is a possibility of a leak. Mandatory certification under Art. 32-2 applies from 1 July 2027.
Korea AI Framework Act
Korea’s AI Framework Act is the first comprehensive AI statute in the Asia-Pacific, in force from 22 January 2026. It is both a promotion and a regulatory statute: Chapters 1–3 cover promotion and infrastructure, Chapters 4–5 impose business duties. The core chain is: transparency (Art. 31) — prior notice to users that a product or service using high-impact or generative AI runs on that AI, labelling of generative outputs as AI-generated, and clear notice or labelling of synthetic audio, images or video hard to distinguish from reality; safety (Art. 32) — for AI systems meeting all three conditions in Decree Art. 24 (cumulative training compute of 10^26 FLOPs or more, use of the most advanced technology, and a risk level capable of broad and serious effect on life, bodily safety and fundamental rights), lifecycle risk identification, assessment and mitigation plus a risk-management system that monitors and responds to AI safety incidents, with results submitted to the Minister of Science and ICT; high-impact AI (Arts. 33–35) — a mandatory prior self-check with an optional request for ministerial confirmation, the six items of Art. 34(1) (risk management plan, explanation plan, user protection plan, human oversight, creation and retention of documentation, and any other matter resolved by the Committee), and a duty to endeavour to conduct a fundamental-rights impact assessment; and the domestic representative (Art. 36) — required in writing, with notification to the Minister, for operators with no Korean address or place of business that meet any Decree Art. 29 threshold (prior-year revenue of KRW 1 trillion or more, AI-services revenue of KRW 10 billion or more, a daily average of 1 million domestic users over the three months to the end of the prior year, or a prior administrative fine). Article 4(1) is explicitly extraterritorial: the Act applies to acts done abroad that affect the domestic market or users.
UK data protection regime (UK GDPR and DPA 2018 as amended by the Data (Use and Access) Act 2025)
The UK kept the UK GDPR and Data Protection Act 2018 after Brexit; the Data (Use and Access) Act 2025 is the largest rewrite since 2021. It replaces Article 22 with new Articles 22A–22D, defining a solely automated decision by the absence of “meaningful human involvement”, keeping an outright restriction only for significant decisions based on special category data, and otherwise requiring a mandatory safeguards package. It adds Schedule 4 “recognised legitimate interests” and Schedule 5 compatible purposes; converts the adequacy standard into a “data protection test” (protection not materially lower than the UK’s) and lets controllers and processors, “acting reasonably and proportionately”, judge whether an Article 46 safeguard meets it; and creates a statutory complaints route to controllers. The ICO refreshed its international transfer guidance (January 2026) and automated decision-making guidance (March 2026) accordingly.
Saudi Regulation on Personal Data Transfer Outside the Kingdom
PDPL Article 29 sets a two-layer structure: the transfer must first fall within a permitted purpose (performing an obligation under an agreement to which the Kingdom is a party, serving the Kingdom’s interests, performing an obligation to which the data subject is a party, or another purpose set out in the Regulations), and must then satisfy three conditions (no prejudice to national security or the Kingdom’s vital interests; an adequate level of protection outside the Kingdom, at least equivalent to that guaranteed by the Law and Regulations as assessed by the Competent Authority; and limitation to the minimum personal data needed). This Regulation specifies the “other purposes” as three: necessary central processing that enables the controller to conduct its activities; providing a service or benefit to the data subject; and scientific research and studies. Article 4 lists five cases in which a controller is exempt from the adequacy and minimisation conditions (or either of them), provided it implements one of three appropriate safeguards — standard contractual clauses, binding common rules, or a certificate of accreditation. Article 5 confirms that onward transfers remain subject to the Law and Regulations. Article 6 revokes the exemption where the controller fails to implement the safeguards or the Competent Authority finds them inadequate, requiring the controller to halt the transfer and notify the recipients. Article 7 mandates a prior risk assessment in two cases: transfers made under the Article 4 exemptions, and continuous or widespread transfer or disclosure of sensitive data outside the Kingdom.
EU Machinery Regulation
The Machinery Regulation replaces the Machinery Directive and writes digital risks — self-evolving behaviour, autonomy, data dependency — into the essential health and safety requirements. For the first time it lists safety components with fully or partially self-evolving behaviour using machine learning that ensure safety functions (and machinery embedding such systems) in Annex I Part A, making third-party conformity assessment mandatory. Technical documentation must be kept for 10 years and source code or programming logic must be produced on a reasoned request; anyone carrying out a substantial modification is treated as the manufacturer.
UN Regulation No 155: cybersecurity and cybersecurity management system
UN R155 turns vehicle cybersecurity into a two-layer approval of organisational capability plus vehicle-type evidence. The manufacturer must first obtain a Cybersecurity Management System (CSMS) certificate of compliance, valid for a maximum of three years, and then demonstrate for each vehicle type that an exhaustive risk assessment has been performed, that the Annex 5 mitigations have been implemented, and that continual monitoring, detection and response capability covers vehicles after first registration. Approval authorities must refuse cybersecurity type approval where the manufacturer has not performed the exhaustive risk assessment or has not considered all Annex 5 Part A threats.
UN Regulation No 156: software updates and software update management system
UN R156 brings OTA updates into the type-approval order. The manufacturer must hold a Software Update Management System (SUMS) certificate of compliance, valid for a maximum of three years, and must operate an RXSWIN (type-approval relevant software identification number) mechanism per vehicle type so that the mapping between software version and approval stays traceable, verifiable and protected against unauthorised modification. An auditable register must describe all software relevant to each RXSWIN before and after an update, including versions and integrity validation data. Before an update the manufacturer must assess whether it affects any other system needed for safe and continued operation, or adds or alters functionality compared with registration, and decide accordingly whether an extension or a new approval is required.
UN Regulation No 157: Automated Lane Keeping Systems (ALKS)
ALKS is the first automated driving system regulation inside the international type-approval system: once activated the system performs the driving task, manages all situations including failures, and may be overridden by the driver at any time. It sets requirements for system safety and failsafe response, for safe handover of the driving task to the driver including a minimum risk manoeuvre where the driver does not respond appropriately, and for a human-machine interface that prevents misunderstanding and misuse — for example on-board displays used for non-driving activities must be automatically suspended as soon as a transition demand is issued. Every ALKS-equipped vehicle must be fitted with a Data Storage System for Automated Driving (DSSAD) enabling determination of the interactions between the ALKS and the human driver, and Annex 4 requires a safety management system process audit of the manufacturer processes for managing safety and continued compliance across the product lifecycle.
Several Provisions on Automotive Data Security Management (Trial)
These Provisions are the foundation of Chinese automotive data governance. They set four advocated principles — in-vehicle processing, no collection by default, appropriate precision range, and de-identification — and define sensitive personal information to include vehicle trajectory, audio, video, images and biometric characteristics. Important data is enumerated in six categories: geographic information and people or vehicle flow data for militarily or politically sensitive areas; vehicle flow and logistics data reflecting economic conditions; operating data of the vehicle charging network; off-vehicle video and image data containing facial or licence plate information; personal information involving more than 100,000 personal information subjects; and other data designated by the authorities. Around important data the Provisions impose four gates: risk assessment reporting, domestic storage with a security assessment for outbound transfers, a prohibition on exceeding the purpose, scope, means, data types and volumes confirmed in that assessment, and annual reporting.
AI medical software under the EU Medical Devices Regulation
AI software under the MDR is classified chiefly by Rule 11 in Annex VIII section 6.3: software intended to provide information used to take decisions with diagnosis or therapeutic purposes is class IIa, rising to class III where those decisions may cause death or an irreversible deterioration of health, or class IIb where they may cause a serious deterioration of health or a surgical intervention; software intended to monitor physiological processes is class IIa, or class IIb where it monitors vital physiological parameters whose variations could result in immediate danger; all other software is class I. Because everything above class I requires a notified body, most clinical decision support AI entering the EU must go through notified body conformity assessment, on top of quality management system, UDI, EUDAMED registration, post-market surveillance, periodic safety update reporting and vigilance obligations.
U.S. export controls on advanced computing and AI model weights — sector application points for AI hardware
For AI hardware going abroad, the sector-relevant U.S. export control points fall into three lines. First, the item line: advanced computing integrated circuits and the related software and technology, including high bandwidth memory under 3A090.c with 3D001 and 3E001, require a licence for Macau and Country Group D:5 destinations. Second, the model line: AI model weights under ECCN 4E091 require a licence for all destinations worldwide, and applications are reviewed under a presumption of denial for end users headquartered, or whose ultimate parent is headquartered, outside the destinations in supplement no. 5 to part 740. Third, the end-use line: §744.23 requires a licence, even for items not otherwise listed, where the exporter has knowledge that the item is destined for the development, production, operation, installation, maintenance, repair, overhaul or refurbishing of a supercomputer located in or destined to Macau or a D:5 destination, or for an advanced-node IC production facility, or, for Category 3 items, for a facility where the technology node is unknown. This entry records sector application points only and does not restate the control text in full.
EU Data Act
A horizontal rulebook for the EU data economy covering six areas: access for users to data generated by connected products and related services (Chapter II); making data available to data recipients under statutory obligations (Chapter III); control of unfair contractual terms in business-to-business data contracts (Chapter IV); public sector access to private sector data on exceptional need (Chapter V); switching and interoperability for data processing services such as cloud and edge (Chapter VI); and safeguards against unlawful third-country governmental access to non-personal data held in the Union (Chapter VII). For Chinese companies Chapters II and VI bite hardest: the first requires smart hardware to let data out, the second requires cloud and PaaS services to let customers leave.
EU NIS 2 Directive
The horizontal EU cybersecurity baseline. Scope is set by Annex I (sectors of high criticality, including digital infrastructure and business-to-business ICT service management) and Annex II (other critical sectors, including online marketplaces, online search engines, social networking platforms and research organisations), with entities split by size into "essential" and "important" and supervised with different intensity. The core duties are the Art. 21 all-hazards cybersecurity risk-management measures (at least ten categories, including supply chain security and multi-factor authentication), the Art. 20 management body approval and training duty, the Art. 23 staged incident reporting (24-hour early warning, 72-hour notification, one-month final report), and the Art. 27 registry duty for certain entities. Being a directive, it requires national transposition and thresholds and enforcement details vary by country.
EU Cyber Resilience Act
Horizontal cybersecurity legislation for products with digital elements, applying to products made available on the EU market whose intended purpose or reasonably foreseeable use includes a direct or indirect logical or physical data connection to a device or network, including software and remote data processing solutions. The core requirements are: secure design, development and production against the Annex I Part I essential cybersecurity requirements; a documented cybersecurity risk assessment; vulnerability handling under Annex I Part II including a software bill of materials; determining and publishing a support period, which must be at least five years unless the product is expected to be in use for less; selecting a conformity assessment procedure under Art. 32, with stricter routes for important products in Annex III classes I and II and for critical products; drawing up Annex VII technical documentation, issuing an EU declaration of conformity and affixing the CE marking; and the Art. 14 vulnerability and incident reporting.
EU revised Product Liability Directive (covering software and AI systems)
Extends the EU no-fault product liability regime into the digital age. Art. 4(1) makes clear that "product" includes electricity, digital manufacturing files, raw materials and software, so standalone AI software and AI components embedded in products are both covered. Arts. 4(3) and (4) bring in "related services" and "components", tangible or intangible, raw materials or related services. Art. 4(5) defines "manufacturer’s control" to cover integration, inter-connection or supply of a component including software updates or upgrades that the manufacturer performs, authorises or consents to, modification of the product including substantial modification, and the manufacturer’s ability to supply software updates or upgrades itself or via a third party. In assessing defectiveness, Art. 7(2) expressly requires account to be taken of the effect on the product of any ability to continue to learn or acquire new features after it is placed on the market or put into service, of the reasonably foreseeable effect of other products expected to be used together with it including by inter-connection, and of relevant product safety requirements including safety-relevant cybersecurity requirements. Art. 11(2) removes certain defences: an economic operator cannot rely on the defect not existing when the product was placed on the market where the defectiveness is due to a related service, software including updates or upgrades, a lack of updates or upgrades necessary to maintain safety, or a substantial modification, provided these are within the manufacturer’s control.
EU ePrivacy Directive
The sector-specific privacy law for electronic communications, applying alongside the GDPR and taking precedence on the matters it covers. Three provisions matter most for AI and digital products. Art. 5(1) protects confidentiality of communications, prohibiting listening, tapping, storage or other kinds of interception or surveillance of communications and related traffic data by persons other than users without the consent of the users concerned, except when legally authorised, and without preventing technical storage necessary for conveyance. Art. 5(3) makes storing information, or gaining access to information already stored, in a subscriber’s or user’s terminal equipment conditional on that person’s consent after being provided with clear and comprehensive information, the only exceptions being technical storage or access for the sole purpose of carrying out the transmission of a communication, or as strictly necessary to provide an information society service explicitly requested by the subscriber or user. Art. 13 restricts unsolicited communications: automated calling and communication systems without human intervention, fax or electronic mail may be used for direct marketing only in respect of subscribers or users who have given prior consent, subject to the soft opt-in for existing customers in Art. 13(2).
Cybersecurity Law
China’s foundational statute for cyberspace, amended by the NPC Standing Committee in October 2025. The core corporate duties sit in Chapter III: the five security obligations under the multi-level protection scheme (Art. 23(1)–(5), including retention of network logs for no less than six months); network products and services must meet mandatory national standards, with immediate remediation and reporting of vulnerabilities (Art. 24); certification or testing of critical network equipment and dedicated cybersecurity products (Art. 25); real-name authentication of users (Art. 26); and incident response plans and reporting (Art. 27). The amendment adds Art. 20, setting out state support for AI alongside ethics norms and security supervision.
Measures on Certification for Outbound Transfer of Personal Information
This operationalises the “personal information protection certification” route in PIPL Art. 38 §1(2) into a complete regime: an accredited certification body assesses whether a processor’s outbound transfers and related processing conform to the applicable laws, administrative regulations, departmental rules, standards and technical specifications, and issues a certificate where they do. The certification standards and technical specifications are set by the CAC together with the national data administration and other departments, while the certification rules and the unified certificate and mark are set by SAMR together with the CAC. Certification bodies must file with the CAC within ten working days of obtaining accreditation.
Measures for Science and Technology Ethics Review (for Trial Implementation)
China’s unified procedural law for science and technology ethics governance. Art. 4 places enterprises alongside universities, research institutes and medical institutions as responsible entities and requires that “entities engaged in life sciences, medicine, artificial intelligence and other science and technology activities whose research touches ethically sensitive areas shall establish a science and technology ethics (review) committee”. Art. 2 identifies three categories requiring review: activities involving human research participants, including the use of human biological samples and personal information data; activities involving laboratory animals; and activities that involve neither directly but may raise ethical risks for life and health, the ecological environment, public order or sustainable development. Art. 25 establishes a list of activities requiring expert re-review; the annexed list has seven items, three of which bear directly on AI.
Technology export administration: the prohibited/restricted technology Catalogue and the Technology Import and Export Regulations
The second control line over technology leaving China, running in parallel with the Export Control Law. It operates under the Foreign Trade Law and the Technology Import and Export Administration Regulations, with MOFCOM and the science-and-technology authority classifying technologies, in catalogue form, as prohibited, restricted or free for export. It differs from the dual-use regime in that it governs the outward transfer of the technology itself: Art. 2 of the Regulations defines technology export as transferring technology abroad by way of trade, investment or economic and technical cooperation, expressly including assignment of patent rights, assignment of the right to apply for a patent, patent licensing, transfer of know-how, technical services and other means — no physical carrier required. So even with no hardware shipped at all, moving a listed technology through licensing, a contribution of technology as capital, an offshore R&D entity or outbound technical services still requires the permit process. The three classes are handled differently (Arts. 32, 33 and 39): prohibited technologies may not be exported; restricted technologies are subject to licensing and may not be exported without a permit; free technologies are subject to contract registration, and such a contract takes effect when validly formed, registration not being a condition of effectiveness. The restricted-class permit runs in two stages (Arts. 34–38): the applicant first files with MOFCOM, which reviews together with the science-and-technology authority and decides within 30 working days of receipt, issuing a “letter of intent for a technology export licence” if approved; only after obtaining that letter may the applicant enter substantive negotiations and sign the contract; after signature the applicant files the letter of intent, a copy of the contract, an export list of the technical materials and proof of both parties’ legal status, and MOFCOM decides within 15 working days, with the contract taking effect on the date the licence is issued. Where a confidentiality review is required, national rules apply. A change to the main terms requires the permit or registration to be redone, and termination must be filed (Art. 43). The Catalogue adjustment announcement of 28 August 2020 adds that military-civil dual-use technologies fall under the export-control regime. VERSION STATUS: the Catalogue version verified here is the annex to the 2008 Order No. 12 as adjusted by Announcement No. 38 of 2020; that Order was repealed by MOFCOM/MOST Order No. 2 of 2023 with effect from 21 Dec 2023 and no official full text of a replacement could be retrieved in this round, so no specific technology entries or control points are stated below.
Regulations on the Security Protection of Critical Information Infrastructure
The top tier of China’s cybersecurity regime, establishing how critical information infrastructure (CII) under the Cybersecurity Law is identified and what its operators must do. Identification is not self-assessed: the competent and supervisory departments for each important sector (the “protection departments”) draw up identification rules, file them with the Ministry of Public Security, identify the CII in their sector against those rules, notify the operator of the result and report it to the MPS (Arts. 8–10); where a material change to the CII may affect that result, the operator must report promptly and the protection department completes a re-identification within three months of receiving the report (Art. 11). The operator duties sit in Chapter 3: security measures must be planned, built and brought into use simultaneously with the CII itself (Art. 12); a cybersecurity protection system and accountability regime must be established, with the principal bearing overall responsibility (Art. 13); a dedicated security management body must be set up and its head and key-post personnel put through security background vetting, assisted by the public security and state security organs (Art. 14); that body carries eight statutory duties including risk assessment, emergency plans and drills, appraisal of key posts, education and training, responsibility for personal information and data security, security management of design, construction, operation and maintenance services, and reporting of incidents and significant matters (Art. 15); it must be funded and staffed, and its personnel must take part in cybersecurity and informatisation decisions (Art. 16); the operator must, itself or through a cybersecurity service provider, conduct at least one security test and risk assessment per year, remediate promptly and report as the protection department requires (Art. 17); major incidents and major threats must be reported to the protection department and the public security organs (Art. 18); secure and trustworthy network products and services must be given priority, and procurement that may affect national security must pass cybersecurity review (Art. 19); a confidentiality agreement must be signed with the product or service provider specifying its technical support and security-confidentiality obligations and liabilities, and its performance must be supervised (Art. 20); and mergers, divisions and dissolutions must be reported promptly with the CII handled as the protection department requires (Art. 21). Art. 31 adds a prohibition binding on everyone: without approval from the CAC or the MPS, or authorisation from the protection department or the operator, no individual or organisation may carry out vulnerability probing, penetration testing or similar activities against CII; such activities against basic telecommunications networks must be reported to the telecommunications authority in advance.
Measures for the Administration of Personal Information Protection Compliance Audits
Turns the PIPL requirement to “regularly audit compliance of personal information processing with laws and administrative regulations” into an executable regime with thresholds, cycles, working papers and a filing standard. It runs on two tracks. SELF-INITIATED AUDITS (Art. 3): conducted regularly by the processor’s internal function or by an engaged professional institution, with the mandatory cycle in Art. 4 — a processor handling more than 10 million individuals’ personal information must run an audit at least once every two years. AUDITS MANDATED BY THE PROTECTION DEPARTMENT (Art. 5): the CAC and other departments performing personal information protection duties may require a processor to engage a professional institution in three situations — processing activities carrying significant risk such as serious impact on individual rights or a serious lack of security measures; processing that may infringe the rights of numerous individuals; or a personal information security incident causing the leakage, alteration, loss or destruction of more than 1 million individuals’ personal information or more than 100,000 individuals’ sensitive personal information. The same incident or risk may not be used to require an audit twice. A mandated audit adds: selecting the professional institution as the department requires and completing the audit within the time limit, extendable with approval where the matter is complex (Art. 9); providing necessary support and bearing the cost (Art. 8); filing the institution’s audit report on completion, signed by the institution’s principal and by the audit lead and bearing the institution’s seal (Art. 10); and remediating as required, filing a remediation report within 15 working days of completing the fixes (Art. 11). Both tracks follow the annexed Personal Information Protection Compliance Audit Guidelines (Art. 6), whose 27 sections set out review points covering lawful basis, processing rules, notice, joint processing, entrusted processing, transfer, provision to others, automated decision-making, public disclosure, image collection in public places, already-public personal information, sensitive personal information, minors under 14, outbound transfers, the right to deletion, safeguarding individual rights and explaining the rules, internal management systems and operating procedures, security technical measures, education and training, the personal information protection officer’s performance, impact assessments, emergency plans and incident response, and — for major platforms — platform rules and the social responsibility report. On the organisational side: a processor handling more than 1 million individuals’ personal information must appoint a personal information protection officer responsible for the compliance audit work, and a processor providing important internet platform services with a very large user base and complex business types must set up an independent body composed mainly of external members to supervise the audit (Art. 12). On the institution side: it must have the audit personnel, premises, facilities and funding for the work and is encouraged to obtain certification (Art. 7); it owes confidentiality over the personal information, trade secrets and confidential business information obtained and must delete them promptly once the work ends (Art. 13); it may not sub-delegate the audit (Art. 14); and the same institution and its affiliates, and the same audit lead, may not audit the same subject more than three times consecutively (Art. 15). The Measures do not apply to audits of state organs or of organisations authorised by law to administer public affairs (Art. 19).
U.S. Outbound Investment Security Program
The program requires U.S. persons to notify Treasury of certain investments involving covered foreign persons and prohibits another class outright. Covered activities span semiconductors and microelectronics, quantum information technologies and artificial intelligence. On the AI side the lines are drawn by compute and end use: developing an AI system trained with more than 10 to the power of 25 operations, or more than 10 to the power of 24 operations using primarily biological sequence data, is prohibited; developing an AI system trained with more than 10 to the power of 23 operations, or intended for cybersecurity applications, digital forensics, penetration testing or the control of robotic systems, is notifiable; and developing AI designed exclusively for, or intended for, military, government intelligence or mass-surveillance end uses is prohibited. U.S. persons must also take all reasonable steps regarding controlled foreign entities and must not knowingly direct such transactions.
Colorado Artificial Intelligence Act
The act centres on algorithmic discrimination and imposes a duty of reasonable care on developers and deployers of high-risk AI systems, offering a rebuttable presumption in exchange for compliance: a developer that supplies the required documentation, publishes the required statement and gives the 90-day algorithmic discrimination notice is presumed to have used reasonable care, and so is a deployer that implements the risk management policy and program, completes impact assessments and provides consumer notice with correction and appeal rights. Compliance with a recognized framework such as the NIST AI Risk Management Framework or ISO/IEC 42001, combined with discovering and curing violations through feedback, adversarial testing or internal review, is an affirmative defence. Enforcement rests exclusively with the Attorney General, a violation constitutes an unfair trade practice under the Colorado Consumer Protection Act, and there is no private right of action.
California Transparency in Frontier Artificial Intelligence Act (TFAIA)
The act targets frontier models (foundation models trained with more than 10 to the power of 26 integer or floating-point operations, counting subsequent fine-tuning, reinforcement learning and other material modifications) and large frontier developers (annual gross revenues with affiliates above $500,000,000 in the preceding calendar year). Large frontier developers must publish and comply with a frontier AI framework. All frontier developers must publish a transparency report before or with deployment, report critical safety incidents to the Office of Emergency Services within 15 days and disclose imminent risks of death or serious physical injury to an appropriate authority within 24 hours. Large frontier developers must additionally transmit internal-use catastrophic risk summaries every three months. The act prohibits materially false or misleading statements about catastrophic risk and creates whistleblower protections plus an anonymous internal disclosure channel.
California Generative AI Training Data Transparency Act (AB 2013)
The act requires developers of generative AI systems or services to post training-data documentation on their websites covering dataset sources or owners, how the datasets further the intended purpose, the number and types of data points, whether the data is protected by copyright, trademark or patent or is entirely public domain, whether it was purchased or licensed, whether it includes personal information and aggregate consumer information, whether it was cleaned or processed and why, the collection period and whether collection is ongoing, the first date of use in development, and whether synthetic data generation is used. The trigger is making the system publicly available to Californians, regardless of compensation, and the documentation must be refreshed before each public release or substantially modified release. Exemptions cover systems whose sole purpose is security and integrity, systems whose sole purpose is the operation of aircraft in the national airspace, and national security, military or defence systems made available only to a federal entity.
California AI Transparency Act (SB 942, as amended by AB 853)
The act is built around content provenance. Covered providers, meaning producers of generative AI systems with more than 1,000,000 monthly visitors or users that are publicly accessible in California, must offer a free AI detection tool including an API, embed in generated image, video and audio content a latent disclosure that is consistent with widely accepted industry standards, detectable by their own detection tool and permanent or extraordinarily difficult to remove, and offer users the option of a manifest disclosure. Where the system is licensed to a third party the provider must contractually preserve the disclosure capability and must revoke the licence within 96 hours of learning it was disabled. As amended by AB 853 the regime extends downstream: from 2027-01-01 large online platforms with more than 2,000,000 unique monthly users must detect, display and allow inspection of provenance data and must not strip it, and GenAI hosting platforms must not make available systems that omit disclosures; from 2028-01-01 capture device manufacturers must embed latent disclosures by default.
California Consumer Privacy Act and the ADMT, risk assessment and cybersecurity audit regulations
On top of the existing CCPA and CPRA consumer rights and business duties, the regulation package completed by the California Privacy Protection Agency in September 2025 adds three strands. Risk assessments: a business must complete one before initiating processing that sells or shares personal information, processes sensitive personal information, uses ADMT for a significant decision, or uses automated processing to infer capability, performance, health, behaviour or movements from systematic observation of job applicants, employees, students and similar groups, and must submit summary information to the Agency. Automated decisionmaking technology: a business using ADMT for significant decisions must provide a Pre-use Notice, an opt-out mechanism and access request channels, and must be compliant by 2027-01-01. Cybersecurity audits: qualifying businesses must complete a first audit report on a revenue-phased schedule of 1 April 2028, 2029 or 2030, and annually thereafter.
Texas Responsible Artificial Intelligence Governance Act (TRAIGA)
TRAIGA uses an intent-based prohibition structure with generous cure provisions, in clear contrast with the Colorado reasonable-care and impact-assessment model. The core prohibitions are: not developing or deploying AI in a manner that intentionally aims to incite or encourage self-harm including suicide, harm to another person or criminal activity; not developing or deploying AI with the sole intent to infringe, restrict or impair constitutional rights; not developing or deploying AI with the intent to unlawfully discriminate against a protected class, with the statute stating that disparate impact alone is insufficient to show intent; and not developing or distributing AI with the sole intent of producing or distributing unlawful sexual material or deep fakes, or intentionally developing or distributing a system that engages in text-based sexual conversation while imitating a child under 18. Social scoring and non-consensual biometric identification prohibitions apply to governmental entities. Consumer disclosure duties fall mainly on government agencies and healthcare providers. Enforcement rests exclusively with the Attorney General, with a 60-day cure period, a presumption of reasonable care and several defences, and no civil penalty action may be brought for a system that has not been deployed.
Illinois Human Rights Act provisions on AI in employment (HB 3773)
The provision brings employment-related use of AI into the civil rights violation framework of the Illinois Human Rights Act. Two violation types exist: using AI that has the effect of subjecting employees to discrimination on the basis of protected classes, or using ZIP codes as a proxy for protected classes, with respect to recruitment, hiring, promotion, renewal of employment, selection for training or apprenticeship, discharge, discipline, tenure, or the terms, privileges or conditions of employment; and failing to provide notice to an employee that AI is being used for those purposes. The Department of Human Rights must adopt rules on implementation and enforcement, including the circumstances and conditions requiring notice, the time period and the means. Compared with the intent standard in the Texas TRAIGA, this is an effects standard with a materially lower evidentiary bar.
FTC Act Section 5 enforcement posture on artificial intelligence
Section 5 of the FTC Act prohibits unfair or deceptive acts or practices in commerce and is the only federal enforcement tool that applies to all AI commercial conduct without thresholds or sector limits. In Operation AI Comply the FTC stated that using AI tools to trick, mislead or defraud people is illegal and that there is no AI exemption from the laws on the books. Enforcement runs along two lines. Claims and substantiation: assertions about accuracy, autonomy, substituting for professional credentials and earnings must be supported by evidence held in advance. Deployment and safeguards: the Rite Aid matter established a baseline that deploying automated systems using biometric information to track or flag consumers requires comprehensive safeguards, that the technology must be discontinued where potential risks to consumers cannot be controlled, and that the related security program must be overseen by senior company executives.
PDPC Advisory Guidelines on Use of Personal Data in AI Recommendation and Decision Systems
The guidelines explain how organisations may lawfully use personal data to build and deploy recommendation and decision systems embedding machine learning models. They cover four areas: the two consent exceptions available for development, testing and monitoring — the Business Improvement Exception and the Research Exception, each with conditions; the notification duty at deployment, requiring disclosure of the product function, the types of personal data processed, how that processing relates to the feature, and the features most likely to influence the output; the accountability duty to document AI practices and safeguards in written policies and to publish them proactively; and procurement, listing service-provider best practices such as data mapping and labelling and provenance records tracing training-data lineage.
Singapore Model AI Governance Framework for Generative AI
The framework is a voluntary governance baseline with no penalties, but it is widely used by Singapore regulators, corporate buyers and investors as the skeleton of generative AI due-diligence questionnaires. It sets out nine dimensions: Accountability (allocating responsibility along the development chain), Data (training-data governance and provenance), Trusted Development and Deployment (baseline evaluation and disclosure practice), Incident Reporting (detection and response), Testing and Assurance (a third-party testing ecosystem), Security (new attack surfaces specific to generative AI), Content Provenance (labelling and watermarking), Safety and Alignment R&D, and AI for Public Good. It stresses that no single intervention is a silver bullet and the dimensions must be viewed in totality.
India Digital Personal Data Protection Act, 2023
The DPDP Act is India’s first comprehensive personal data statute, structured around Data Fiduciaries, Data Processors and Data Principals. Core duties include giving notice and obtaining consent in clear and plain language; ensuring accuracy and completeness; taking reasonable security safeguards to prevent breaches; notifying the Board and affected Data Principals of a personal data breach; additional duties for children’s data; heightened duties where the Central Government notifies an entity as a Significant Data Fiduciary — appointing a Data Protection Officer based in India, appointing an independent data auditor and conducting periodic Data Protection Impact Assessments; and, under s.16, a power for the Central Government to restrict transfers to notified countries or territories. Section 3 is extraterritorial: the Act applies to processing of digital personal data outside India where that processing is in connection with any activity related to offering goods or services to Data Principals within India.
Thailand Personal Data Protection Act
Thailand’s PDPA is structurally close to the GDPR: lawful bases and consent, data-subject rights, controller and processor duties, cross-border transfer restrictions, a DPO regime and administrative fines. Four points matter most to foreign entrants: the extraterritorial reach of s.5 (a controller or processor outside Thailand is covered where it offers goods or services to data subjects in Thailand or monitors their behaviour in Thailand); the 72-hour breach notification in s.37(4); the s.37(5) requirement that a controller caught by the second paragraph of s.5 designate in writing a representative in Thailand, authorised to act on its behalf without limitation of liability in respect of the controller’s collection, use or disclosure purposes (with s.38 paragraph one exempting public authorities and controllers that neither process s.26 sensitive data nor hold personal data above the Committee’s volume threshold, and paragraph two applying the same rule to processors); and the transfer rules in ss.28 and 29. Enforcement is chiefly by administrative fines, up to THB 5 million.
Australia Privacy Act 1988 and the 2024 reforms
Australia’s regime centres on the Privacy Act 1988 and the Australian Privacy Principles in Schedule 1. The 2024 reforms did three structural things. First, they rebuilt the civil penalty architecture: alongside serious interference with privacy, a new s.13H civil penalty provision covers interference with privacy generally and s.13K identifies provisions for which infringement notices can issue, giving the Commissioner tools proportionate to smaller breaches. Second, from 10 June 2025 a statutory tort for serious invasions of privacy lets individuals sue directly. Third, from 10 December 2026 new APP 1.7–1.9 require disclosure of automated decisions in the privacy policy. For a body corporate the maximum penalty under s.13G is the greatest of A$50 million, three times the benefit obtained from the contravening conduct, or, where that benefit cannot be determined, 30% of the entity’s adjusted turnover during the breach turnover period.
OAIC Australia AI privacy guidance
OAIC has issued two AI-facing privacy guidance documents: one on developing and training generative AI models, one on using commercially available AI products. The development guidance takes five core positions. First, developers must take reasonable steps to ensure accuracy under APP 10, commensurate with the likely increased level of risk in an AI context, including using high-quality datasets, undertaking appropriate testing and using disclaimers appropriately for high privacy risk uses. Second, just because data is publicly available or otherwise accessible does not mean it can legally be used to train or fine-tune generative AI models or systems; developers must consider whether the data they intend to use or collect, including publicly available data, contains personal information and comply with their privacy obligations, and may need additional steps such as deletion. Third, particular care is required with sensitive information, which generally requires consent to collect — many photographs or recordings of individuals, including artificially generated ones, contain sensitive information and therefore may not be scraped from the web or collected from a third-party dataset without establishing consent. Fourth, where a developer seeks to use personal information it already holds for AI training and that was not a primary purpose of collection, it must carefully consider its obligations: absent consent for that secondary AI-related purpose, it must be able to establish that the secondary use would be reasonably expected by the individual, taking particular account of their expectations at the time of collection, and that it is related — directly related for sensitive information — to the primary purpose, or that another exception applies. Fifth, where that cannot be clearly established, the developer should, to avoid regulatory risk, seek consent for the use and/or offer individuals a meaningful and informed ability to opt out. The guidance also states that developers should update their privacy policies and notifications with clear and transparent information about their use of AI generally, and that personal information includes inferred, incorrect or artificially generated information produced by AI models, such as hallucinations and deepfakes, where it is about an identified or reasonably identifiable individual.
UAE Federal Personal Data Protection Law (Federal Decree-Law No. 45 of 2021)
The UAE’s first horizontal federal data protection law, structurally close to the GDPR: consent-based with eleven enumerated consent exemptions; general controller and processor duties; a processing record; a data protection officer; data subject rights (information, portability, correction and erasure, restriction, cessation, objection to automated processing); security measures; data protection impact assessments; breach notification; and two cross-border transfer routes depending on whether an adequate level of protection exists. Its scope expressly reaches controllers and processors outside the UAE that process the data of subjects inside it. Note the carve-outs: government data and government entities, personal data held by security and judicial authorities, health and banking/credit data governed by dedicated legislation, and companies and establishments located in free zones that have their own personal data protection legislation — the last being the legal basis for the separate DIFC and ADGM regimes.
DIFC Data Protection Law 2020 and the DIFC Data Protection Regulations (including Regulation 10 on AI systems)
The DIFC was the first GCC jurisdiction to legislate on data protection. The Data Protection Law 2020 closely tracks the GDPR in structure and vocabulary: lawful bases, special categories, consent conditions, accountability and notification to the Commissioner, records of processing, DPOs, DPIAs and prior consultation, joint controller and processor chains, an adequacy list and safeguards for exports, information provision, data subject rights (including Article 38 on automated decisions and profiling and Article 39 on non-discrimination), breach notification, plus the Schedule 2 fines table and the uncapped general fine in Article 62(3). Regulation 10, in force since 1 September 2023, extends governance from data to systems: wherever personal data is processed for use in, or to train, an autonomous or semi-autonomous system, the Deployer and Operator must give clear notice on first use covering processing that is not human-initiated, the human-defined purposes and the principles and limits within which the system may define further purposes, the output and how it is used, safeguards built in by design, and the codes or certifications the system was built to. The system itself must be designed to be ethical, fair, transparent, secure and accountable, may process personal data only for human-defined or human-approved purposes, and — where used for High Risk Processing Activities — may not be commercially deployed until the Commissioner has established audit and certification requirements, with an Autonomous Systems Officer appointed.
ADGM Data Protection Regulations 2021 (Abu Dhabi Global Market)
The ADGM Regulations are the closest UAE analogue to the GDPR in both numbering and substance: processing principles and lawful bases, consent conditions, special categories, data subject rights (sections 10–20, including section 20 on automated individual decision-making and profiling), controller responsibility and data protection by design and by default, the Data Protection Fee and notification to the Commissioner (section 24), joint controllers and processors, records of processing, security, cessation of processing, breach notification (section 32, 72 hours), communication to data subjects (section 33), DPIAs (section 34), DPOs (sections 35–37), codes of conduct and certification, and the export regime in sections 40–44 (adequacy decisions, appropriate safeguards, binding corporate rules, derogations). Section 55 allows the Commissioner to impose a fine in an amount considered appropriate, capped at USD 28 million for a contravention and at USD 28 million in total for several contraventions arising from the same or linked processing operations.
Saudi Personal Data Protection Law (PDPL) and its Implementing Regulation
The PDPL is Saudi Arabia’s first comprehensive personal data protection statute. Article 2 applies it to any processing of personal data taking place in the Kingdom by any means, “including the processing of personal data related to individuals residing in the Kingdom by any means from any party outside the Kingdom”. Its spine is controller obligations: privacy policy and notice; a record of processing activities (Article 31, which must state whether data has been or will be transferred or disclosed outside the Kingdom); security measures (Article 19); notification of breaches to the Competent Authority and to data subjects (Article 20, with a 72-hour deadline in the Implementing Regulation); a personal data processing impact assessment (Article 22); deadlines for responding to data subject requests (Article 21); enhanced controls on health and credit data (Articles 23–24); a ban on copying official documents identifying data subjects (Article 28); cross-border transfer (Article 29); supervision and the national register of controllers (Article 30); and Article 33(4), which directs the Competent Authority to set tools and mechanisms for monitoring controllers and processors outside the Kingdom and to define extraterritorial enforcement procedures. Article 32 of the Implementing Regulation and the DPO Appointment Rules define the three cases requiring a personal data protection officer: a public entity providing services involving large-scale processing; core activities requiring regular and systematic monitoring of data subjects; and core activities based on processing sensitive personal data.
SDAIA AI Ethics Principles and Generative AI Guidelines
The AI Ethics Principles apply a four-tier risk scheme: systems of little or no risk (such as spam filters) face no restrictions though ethical compliance is recommended; limited-risk systems are subject to the ethics principles in the document; high-risk systems must undergo pre- and post-conformity assessments and, beyond ethics, meet the relevant statutory requirements; and systems posing an unacceptable risk to people’s safety, livelihood and rights — such as social profiling, exploitation of children, or likely distortion of behaviour — are not allowed. The document sets seven principles and breaks the controls down across the AI system lifecycle (plan and design, prepare input data, build and validate, deploy and monitor): Fairness; Privacy & Security; Humanity; Social & Environmental Benefits; Reliability & Safety; Transparency & Explainability; and Accountability & Responsibility. The Privacy & Security section states directly that the decision criteria of automated technology must not be based on personally identifying characteristics or information; that designing AI systems that profile individuals or communities may occur only if approved by the Chief Compliance and Ethics Officer or Compliance Officer, or in compliance with a code of ethics and conduct developed by a national regulatory authority for that sector; that data classification and profiling should be planned at the data architecture stage to define protection and usage levels for personal data; and that de-identification mechanisms should be planned for sensitive or personal data with read/write/update authorised by group. The Generative AI Guidelines for the Public address all stakeholders designing, developing, deploying, using or affected by GenAI systems in the Kingdom, covering GenAI principles, risks and mitigations.
Brazilian General Personal Data Protection Law (LGPD)
The LGPD is Brazil’s horizontal data protection statute — GDPR-descended but locally distinct: ten legal bases (including credit protection, a Brazil-specific basis), heightened rules for sensitive personal data and children’s data, a required data processing officer (encarregado), records of processing, a personal data protection impact report (RIPD), security incident notification, and nine data subject rights including the right to request review of decisions taken solely on the basis of automated processing. Article 33 sets nine routes (I–IX) for international transfers: (I) to countries or international organisations the ANPD has determined provide an adequate degree of protection; (II) where the controller offers and demonstrates guarantees of compliance through specific contractual clauses for a given transfer, standard contractual clauses, global corporate rules, or regularly issued seals, certificates and codes of conduct; (III) where necessary for international legal cooperation between intelligence, investigation and prosecution bodies under instruments of international law; (IV) where necessary to protect the life or physical safety of the data subject or a third party; (V) where the national authority authorises the transfer; (VI) where the transfer results from a commitment made in an international cooperation agreement; (VII) where necessary to execute a public policy or a legal attribution of public service, published under Article 23(I); (VIII) where the data subject has given specific and highlighted consent to the transfer, having been informed beforehand of its international character and with that purpose clearly distinguished from others; and (IX) where necessary to meet the situations in Article 7(II), (V) and (VI). Recent ANPD enforcement has centred on platforms and children’s data.
Brazilian AI Bill PL 2338/2023 (not yet law)
This is Brazil’s horizontal AI bill; the Senate-approved text is built on risk tiering. Article 12 lets an AI agent carry out a preliminary assessment before placing a system on the market or putting it into use, treating that assessment as good practice that can earn benefits. Article 13 lists “excessive risk” cases where development, implementation and use are prohibited. Chapter IV sets governance measures and internal processes for high-risk systems (from Article 18). Section V of Chapter IV governs general-purpose and generative AI: the developer must, in addition to the relevant development documentation, carry out a preliminary assessment identifying expected risk levels including potential systemic risk; and for general-purpose and generative systems with systemic risk, before commercial release, must meet the seven requirements in Article 30: describe the model, document the tests and analyses performed to identify and manage reasonably foreseeable risks, document residual non-mitigable risks after development, process and incorporate only datasets collected and handled in accordance with legal requirements and subject to adequate data governance (and, for personal data, the LGPD and Chapter II of this Law), publish a summary of the training dataset as regulated, design the systems using applicable standards to reduce energy and resource use and waste, and produce intelligible technical documentation and instructions for use — with Article 30 §1 applying all of that whether the system ships as a standalone model, embedded, under free and open-source licences, as a service, or through other distribution channels. Article 80 provides that the Law enters into force 730 days after official publication, with Article 13, Section V of Chapter IV and the rules in Section IV of Chapter X taking effect after 180 days (Article 62 immediately), and Chapter IX (except Article 50) and Sections III and V of Chapter X taking effect on publication. Article 77 also provides that regulation of online content circulation that may affect freedom of expression, including AI used for content moderation and recommendation, may only be done by specific legislation.
Personal Information Protection and Electronic Documents Act (PIPEDA)
PIPEDA applies to organisations that collect, use or disclose personal information in the course of commercial activities, and its substantive requirements sit in the ten fair information principles of Schedule 1: accountability, identifying purposes, consent, limiting collection, limiting use, disclosure and retention, accuracy, safeguards, openness, individual access, and challenging compliance. Two matter most for exporters. Schedule 1 clause 4.1.3 makes an organisation responsible for personal information in its possession or custody, including information transferred to a third party for processing, and requires contractual or other means to provide a comparable level of protection while the third party processes it — the hook that governs handing data to a Chinese parent or an offshore vendor. And section 10.1 creates the breach regime: where it is reasonable in the circumstances to believe that a breach of security safeguards creates a real risk of significant harm to an individual, the organisation must report to the Commissioner in the prescribed form and manner as soon as feasible after determining the breach occurred, and must notify affected individuals as soon as feasible with enough information to let them understand the significance of the breach and take steps to reduce or mitigate the risk of harm; section 10.2 requires notifying other organisations or government institutions that could reduce or mitigate the harm; section 10.3 requires keeping records of breaches. Section 28 makes it an offence to knowingly contravene section 10.1 (among others) or to obstruct the Commissioner, punishable on summary conviction by a fine of up to CAD 10,000 or on indictment by a fine of up to CAD 100,000.
Québec Law 25 (Act respecting the protection of personal information in the private sector, CQLR c. P-39.1)
Law 25 turned Québec’s private-sector privacy statute into one of North America’s strictest regimes, with three features that stand out against the other jurisdictions in this file. First, section 12.1 requires an enterprise that uses personal information to render a decision based exclusively on automated processing to inform the person of that fact no later than when it informs them of the decision; on request, to inform them of the personal information used, the reasons and the principal factors and parameters that led to the decision, and their right to have that information corrected; and to give them the opportunity to submit observations to a member of personnel in a position to review the decision. Second, section 17 front-loads a mandatory privacy impact assessment before any communication of personal information outside Québec, taking into account in particular the sensitivity of the information, the purposes for which it is to be used, the protection measures including contractual ones, and the legal framework of the destination State including the personal information protection principles applicable there; the information may be communicated only if the assessment establishes it would receive adequate protection, and the communication must be the subject of a written agreement that takes into account the assessment results and any agreed risk-mitigation terms — the same applies when entrusting a person or body outside Québec with collecting, using, communicating or keeping the information. Third, the penalties are severe: section 90.12 caps administrative monetary penalties at CAD 50,000 for a natural person and, in all other cases, the greater of CAD 10 million or 2% of worldwide turnover for the preceding fiscal year; section 91 sets penal fines of CAD 5,000–100,000 for a natural person and, in all other cases, CAD 15,000–25,000,000 or, if greater, 4% of worldwide turnover. Section 3.3 also requires a privacy impact assessment for any project to acquire, develop or overhaul an information system or electronic service delivery system involving the collection, use, communication, keeping or destruction of personal information, which an AI deployment will almost always trigger.
EU General Product Safety Regulation
The GPSR is the EU safety backstop for consumer products: any consumer product not fully covered by specific EU harmonisation legislation must meet the general safety requirement. It hard-codes an EU-established responsible person, accident notification through the Safety Business Gateway, recall notices with minimum remedies, and an internal complaints register, plus dedicated duties for online marketplaces. For consumer robots, home and companion robots, drone accessories, smart wearables and AI hardware it applies alongside product-specific law and directly drives cross-border e-commerce listing and recall costs.
ISO 10218 series on industrial robot safety (2025 editions)
The ISO 10218 series is the core type-B/C standard set for industrial robot safety. Part 1 covers inherently safe design, risk reduction measures and information for use of the robot itself and treats the robot as incomplete machinery; Part 2 covers integration of industrial robot applications and robot cells. The 2025 editions of both parts took effect in CEN member bodies on 1 April 2025, superseding the 2011 editions. Part 1 expressly does not apply to underwater, law-enforcement, military (defence), airborne and space robots, nor to medical robots, healthcare robots, prosthetics and other aids for the physically impaired, service robots, consumer products, or equipment lifting or transporting people.
EU implementing regulation on type-approval of the automated driving system (ADS) of fully automated vehicles
This is currently the only EU legislation dedicated to type-approving the automated driving system of fully automated, unsupervised vehicles. It applies to categories M and N for three use cases: carriage of passengers or goods on a predefined area, hub-to-hub carriage on a predefined route with fixed start and end points, and automated valet parking (AVP). It covers vehicles operating without driver supervision and dual mode vehicles that have a driver seat and can switch between manual and fully automated mode, with transitions permitted only at standstill; as amended by Implementing Regulation (EU) 2026/481 the AVP use case is no longer confined to dual mode vehicles and instead covers vehicles with an ADS feature for parking applications within an ODD. Approval rests on three annexes: the Annex I information document entries, the Annex II technical specifications and the Annex III assessment method, with AVP features additionally subject to Annex V. The Annex III compliance assessment has five parts: the traffic scenarios to consider; assessment of the ADS safety concept and audit of the manufacturer safety management system; tests for the most relevant traffic scenarios; credibility assessment principles for using a virtual toolchain in ADS validation; and the system established by the manufacturer to ensure in-service reporting. The Regulation also defines the operational design domain (ODD) and unreasonable risk as an overall risk level increased compared with a manually driven vehicle in comparable transportation services and situations within the ODD.
China pilot for market access and road operation of intelligent connected vehicles
This pilot is China institutional route taking L3/L4 automated driving from road testing and demonstration to product access plus road operation, in two stages. In the product access stage the pilot manufacturer refines the access testing and safety assessment scheme, which MIIT and the Ministry of Public Security confirm before it is executed under the supervision of provincial authorities and the city government where the vehicles operate, with MIIT commissioning technical service institutions to assess the scheme, its execution and its results; only after passing may the manufacturer file the product access application, which MIIT decides after acceptance, examination and public notice, announcing the validity period and the restrictions such as the area of operation. In the road operation stage the approved product operates within a defined area, and the pilot user entity must insure the vehicles, register them and monitor their operating status, and must hold the corresponding transport business qualification where the vehicles are used commercially. The notice also sets out suspension and exit mechanisms.
FDA Predetermined Change Control Plans (PCCP) for AI-enabled devices
The PCCP addresses the central commercial question for AI devices: whether every model iteration needs a new submission. Under section 515C and this final guidance, a sponsor can include a predetermined change control plan in a marketing submission (510(k), De Novo or PMA) describing the planned modifications to the AI-enabled device software functions, the associated methodology to develop, validate and implement those modifications, and an assessment of their impact. FDA reviews the PCCP as part of the marketing submission so that each modification later implemented within the plan needs no additional marketing submission, while the device remains safe and effective. The final guidance also recommends what PCCP information to include in labelling and publicly available decision summaries, how to implement modifications consistent with an authorised PCCP, and postmarket surveillance.
DORA obligations bearing on AI and ICT third-party service providers
DORA binds financial entities directly, but it reaches ICT third-party service providers — including providers of cloud computing services, software, data analytics services and data centre services — through contractual terms and the oversight framework. Financial entities must maintain, at entity, sub-consolidated and consolidated level, a register of information on all contractual arrangements on the use of ICT services and make the full register or specified sections available to the competent authority on request. Contracts with providers must set out the rights and obligations and the service level agreements in one written document and must contain at least the elements in Article 30(2), with additional elements under Article 30(3) for services supporting critical or important functions. For critical ICT third-party service providers established in a third country, Article 31(12) provides that a financial entity may only use their services if the provider has established a subsidiary in the Union within the 12 months following designation.
EU Digital Services Act
A horizontal framework for intermediary services (mere conduit, caching, hosting) layering obligations from intermediary services to hosting services, online platforms and very large online platforms and search engines. The AI-relevant core duties are: terms and conditions must describe the policies, procedures, measures and tools used for content moderation, including algorithmic decision-making and human review; the main parameters of recommender systems must be disclosed with options to modify them; advertising must be identifiable in real time with disclosure of targeting parameters; minors must be protected and profiling-based advertising to minors is prohibited; and very large platforms and search engines must assess, annually and before deploying functionalities likely to have a critical impact on identified risks, the systemic risks stemming from the design or functioning of their service and related systems including algorithmic systems, and put in place mitigation measures which expressly include testing and adapting their algorithmic systems. Providers not established in the Union must designate an EU legal representative.
EU Data Governance Act
Establishes a trusted intermediation and governance framework for data sharing in the EU. It has three pillars: conditions for the re-use of protected data held by public sector bodies; a notification regime and conditions for providing data intermediation services (Arts. 10 to 12); and a voluntary registration regime for data altruism organisations. Most directly relevant to AI companies are the three categories of data intermediation service defined in Art. 10: intermediation between data holders and potential data users, including providing the technical or other means to enable it, which may include bilateral or multilateral data exchanges, creation of platforms or databases enabling exchange or joint use of data, and establishment of other specific infrastructure interconnecting data holders with data users; intermediation between data subjects, or natural persons seeking to make non-personal data available, and potential data users, in particular enabling the exercise of GDPR rights; and data cooperative services. Such services must be notified and must meet the structural conditions in Art. 12, the most binding of which are that the data must not be used for purposes other than putting it at the disposal of data users, and that the service must be provided through a separate legal person.
New York City Automated Employment Decision Tools law (Local Law 144 of 2021)
The local law prohibits employers and employment agencies from using an automated employment decision tool unless three conditions are met together: the tool has been subject to a bias audit within one year of its use; information about that bias audit is publicly available; and the required notices have been provided to employees or job candidates. DCWP clarified in its June 2023 revised presentation materials that the notice must be provided 10 business days before use of the tool. DCWP accepts public complaints, including for failure to obtain the required bias audit, failure to post a summary of the audit results and failure to give the required notices.
Japan AI Promotion Act
This is a principles-and-promotion statute, not a regulatory one. It has four chapters and 28 articles plus supplementary provisions: Chapter 1 general provisions (purpose, definitions, basic principles, responsibilities of the State, local governments, research institutions, utilisation businesses and the public, strengthened coordination, legislative measures); Chapter 2 basic measures (promotion of R&D, shared facilities, ensuring appropriateness, workforce, education, research, international cooperation); Chapter 3 the AI Basic Plan; Chapter 4 the AI Strategy Headquarters. A full-text search for “penal provisions” returns zero hits — the Act imposes no penalties and no licensing, filing, assessment or labelling duty on companies. Article 7, addressed to “utilisation businesses”, is a responsibility clause: they should actively use AI in line with the basic principles to make their operations more efficient and create new industries, and must cooperate with measures implemented by the State and local governments. Article 13 provides that the State shall develop guidelines consistent with international norms and take other necessary measures to ensure appropriate R&D and use — so the concrete conduct rules sit in those guidelines and in existing law (the APPI, copyright, product liability, sectoral regulation), not in this Act.
Malaysia Personal Data Protection Act and current regulatory requirements
Malaysia’s regime is built on Act 709 and enforced by the Department of Personal Data Protection (JPDP). On the requirements verifiable from official channels today, three operational strands face companies: (1) entities within the prescribed classes of data controllers must apply to JPDP for registration and hold a valid registration certificate, with JPDP publishing a 14-working-day review period for complete applications and separate guidance for changing, replacing and cancelling certificates; (2) a Data Protection Officer must be designated and registered in line with JPDP’s DPO Competency Guideline, alongside which JPDP has issued a guideline on managing DPO training service providers; and (3) data breaches must be reported through JPDP’s Data Breach Notification (DBN) system, for which JPDP publishes a user manual. At statute level, duties continue to run on the seven principles of notice and choice, consent, disclosure, security, retention, data integrity and access.
Canada’s Artificial Intelligence and Data Act (AIDA, in Bill C-27) — died with the session and never came into force
The sole purpose of this entry is to correct an error that recurs in market materials: AIDA is not in force. It was Part 3 of Bill C-27, bundled with the Consumer Privacy Protection Act that would have replaced PIPEDA and with the Personal Information and Data Protection Tribunal Act. The bill received first reading on 16 June 2022 and second reading and referral to committee on 24 April 2023, then remained at committee stage until the 44th Parliament’s 1st session ended on 6 January 2025 and it died without Royal Assent. Three consequences matter commercially. First, Canada has no horizontal federal AI statute in force and no statutory “high-impact system” duties derived from AIDA. Second, PIPEDA and the ten fair information principles in its Schedule 1 remain the operative federal private-sector rules, so AI compliance hooks come from privacy, consumer protection, human rights and sectoral regulation rather than AIDA. Third, provincial law already imposes concrete enforceable duties on automated decisions — notably section 12.1 of Québec’s Law 25 — so design to provincial law rather than to a lapsed federal bill when entering the Canadian market.
Utah Artificial Intelligence Policy Act
Utah takes a regulatory-learning rather than direct-regulation approach: an Office of Artificial Intelligence Policy sits within the Department of Commerce, creates and administers an AI learning laboratory program, consults businesses and other stakeholders on potential regulatory proposals, and makes rules on required participant disclosures to consumers, reporting to the Office, criteria for limited extension of the participation period and other administration. Participants may apply for a regulatory mitigation agreement or a joint interpretation agreement, which specifies limits on the scope of use, safeguards, any regulatory mitigation granted, required consumer disclosures and reporting to support Office audits. The Office performs regular audits while an agreement is in effect, and participants remain subject to all requirements not expressly waived or modified. The chapter is subject to a sunset and lapses on 2027-07-01 unless extended.
SEC climate-related disclosure rules (stalled and proposed for rescission)
The climate-related disclosure rules the SEC adopted in 2024 would have required registrants to provide climate-related information in registration statements and annual reports. They never took effect: on 2024-04-12 the SEC announced that the effective date is delayed indefinitely pending the consolidated Eighth Circuit review, and on 2026-06-03 the SEC published a proposal to rescind them, with comments closing 2026-08-03. In client-facing materials and deliverables the rules must therefore be described as adopted but not in force and proposed for rescission, never as a live mandatory disclosure obligation.