Law by BogdanovaThe Encyclopedia of Law and Legal Studies

RESEARCH DOSSIER

Comparative Technology Governance

Abstract. Comparative technology governance examines the legal architecture through which public authorities and private institutions authorize, constrain, monitor, and remedy the design and use of consequential technologies. This monograph compares artificial intelligence, online platforms, data-intensive infrastructures, automated public administration, cybersecurity, and product safety across the European Union, the United States, the United Kingdom, China, South Korea, Canada, and selected international organizations. Its method is functional and doctrinal: it identifies the regulated activity, the institution controlling it, the applicable legal source and date, the duty-bearer, the required evidence, the supervisory mechanism, and the route to an enforceable remedy. The study analyzes the European Union's Artificial Intelligence Act in its consolidated 2026 form, the Digital Services and Digital Markets Acts, data-protection and cybersecurity instruments, American agency and state-based governance, British sectoral regulation and automated-decision safeguards, Chinese generative-AI measures, and South Korea's operative AI framework. International treaty commitments and nonbinding recommendations are assessed at their proper legal levels. Across jurisdictions, the analysis establishes that technological risk categories acquire legal consequences only through institutions, defined obligations, applicable procedural rules, and systems of enforcement. The central comparative finding is that effective governance depends on the continuity of evidence and responsibility across technology development, deployment, public decision-making, market intermediation, and redress. The conclusions are original doctrinal synthesis of established authorities, without a claim to a newly invented legal theory. The research cutoff is November 21, 2026.

Keywords: Comparative law; technology governance; artificial intelligence regulation; digital platforms; data protection; automated decision-making; administrative accountability; algorithmic transparency; digital markets; cybersecurity; product safety; human rights; international regulatory coordination; legal remedies

Introduction and research problem

Technology governance is the legal ordering of decisions made through technical systems and of decisions made about their creation and use. Its field includes substantive restrictions on harmful conduct, procedural entitlements of affected persons, the institutional powers of supervisory authorities, duties of firms and public bodies, and mechanisms for proving and correcting violations. Artificial intelligence constitutes a central contemporary case because one computational model may enter distinct legal relationships: consumer advertising, medical treatment, employment, policing, education, financial services, and public-benefit administration. The same model capability can be subject to materially different legal obligations as its role, user, purpose, and consequences change.

The research question is therefore precise: how do legal systems distribute authority, obligations, proof, and remedies over the design, deployment, and effects of emerging technologies? A complete answer requires comparison across legal fields as well as territories. The European Union's Artificial Intelligence Act establishes obligations around classified systems and institutional roles, while its Digital Services Act addresses intermediary services and platform processes. The American AI Risk Management Framework supplies a technically influential but ordinarily voluntary assessment vocabulary. China's Interim Measures for Generative AI Services regulate designated provision of services through a national administrative architecture. These texts cannot be made equivalents by calling all of them AI policies.

The problem has four dimensions. First, legal systems classify activities differently: provider obligations, deployer duties, contractual disclosures, unfair practices, public-authority procedures, product-safety requirements, and fundamental-rights restrictions. Second, each system builds a distinctive chain of supervision involving public regulators, designated firms, courts, independent auditors, or ministerial bodies. Third, technical artifacts often cross both borders and institutional roles without preserving the evidence an injured person needs. Fourth, legal categories change over time. An enacted but deferred obligation is legally real and prospectively consequential, yet it cannot be described as presently applicable in the same manner as a rule already commenced.

This monograph approaches those dimensions through comparative legal institutions rather than a ranking of supposedly permissive and restrictive countries. A jurisdiction with elaborate ex ante documentation may supply weak practical redress; an enforcement-based regime may impose substantial liability even without a single comprehensive AI code. Actual consequences depend on applicable law, regulators' capacity, litigation access, and evidentiary conditions. The method follows Hildebrandt's account of the rule of law under algorithmic regulation, while treating its normative insight as a proposition to test against positive legal structures.

Comparative method and the status of sources

Comparative analysis must fix its units before drawing similarities. The primary unit adopted here is a legally consequential decision involving a technological system: release of a product, operation of a platform, collection of personal data, generation of a credit score, administrative deprivation, or processing of a complaint. For each decision the inquiry identifies its jurisdiction, legal basis, addressee, triggered duty, evidence obligation, oversight institution, and legal remedy. This makes it possible to compare, for example, an EU conformity assessment with an American consent-order compliance obligation without treating their different legal sources as interchangeable.

The source hierarchy begins with enacted legislative instruments, treaties whose binding operation is established for the jurisdiction considered, directly applicable regulations, domestic implementing law, and judicial holdings. It then distinguishes delegated regulations, agency rules, published enforcement decisions, administrative policies, technical standards, recommendations, proposals, and scholarly commentary. A regulator's press release is evidence of an agency's announced action and rationale, but the legally operative duties must be traced to the underlying statute, rule, or order. A discussion paper signals deliberation, and an introduced bill establishes a possible trajectory rather than current law.

This distinction has immediate consequences for historical claims. Veale and Zuiderveen Borgesius's analysis of the draft EU AI Act addressed the 2021 legislative proposal. Its analysis remains useful to reconstruct drafting choices but is not a substitute for the 2024 enacted regulation or its 2026 amendments. Likewise, Brazil's legislative record for PL 2338/2023 documents a continuing legislative process; provisions debated there cannot be attributed to a binding Brazilian AI act. The method retains the historical significance of such proposals while withholding claims of their legal operation.

Functional comparability requires attention to background doctrine. A common-law claim for negligent design demands elements such as duty, breach, causation, and loss; a European administrative-law challenge addresses competence, procedure, reviewability, and proportionality. They may correct the same technical injury, but their tests and evidentiary burdens differ. Similarly, an instrument's geographical reach must be separated from its political influence. A multinational enterprise's internal adoption of a nonbinding standard does not transform that standard into a legislative duty; its relevance may instead arise through contract, professional practice, procurement, or evidence of reasonable care.

The evidence used here is a targeted corpus of officially hosted primary instruments, public judicial decisions, institutional guidance, and selected scholarly studies. The approach performs doctrinal source comparison, not a statistical survey of regulator behavior or a claim to have examined confidential compliance records. Official consolidated legislation is preferred where amendment timing matters. This manuscript uses dates as of November 21, 2026 and explicitly separates already effective obligations from scheduled future application. English-language translations of non-English rules are described at the access level available; their domestic originals remain authoritative.

Historical development and regulatory functions

Legal control of technological activity precedes contemporary machine learning. Product liability, communications regulation, public procurement, professional negligence, constitutional process, and competition law each developed institutional devices for allocating the risks of powerful tools. The novelty of contemporary governance lies in the interaction of these inherited techniques with systems that are rapidly updated, statistically trained, interconnected, and often difficult to explain through ordinary records. The legal history relevant here is an institutional history of verification: what must be established before entry into a market, what must be documented during operation, and what must be demonstrated when a person contests an outcome.

Data-protection law expanded this history by treating information processing as a legally governed activity. The European General Data Protection Regulation ties processing to specified legal bases, principles of fairness, rights of access and challenge, and regulatory remedies. Its rights cannot be reduced to a general demand that a model explain itself: different obligations attach to different kinds of information processing and decisions. Digital intermediary regulation then addressed institutional arrangements through which platforms structure visibility, content circulation, commerce, and access to other markets. The Digital Markets Act exemplifies ex ante obligations for designated gatekeepers, a competition-related legal response to durable intermediary power.

Artificial-intelligence governance adds system-specific lifecycle vocabulary: training and data governance, testing, risk management, human oversight, incident reporting, and model evaluation. Several of these functions are familiar from safety engineering; they acquire specific legal force only where an applicable enactment, contract, authorization, or duty of care adopts them. The distinction matters for evaluating the NIST AI RMF: its Govern, Map, Measure, and Manage functions can organize a firm's internal controls and regulatory dialogue while remaining distinct from compulsory statutory commands.

Three recurring legal functions emerge from this chronology. Authorization governs whether a system may enter a market or public workflow. Supervision governs its continuing conduct, monitoring, reporting, and correction. Redress governs how affected persons can obtain an explanation, reversal, damages, or other legally appropriate relief. Each function requires institutional competence and records capable of supporting a decision. Contemporary systems create particular challenges where authorization falls on one firm, operational knowledge remains with another, and the affected person must litigate against a third.

The European Union's layered technology acquis

The European Union supplies the most developed example here of simultaneous sectoral, horizontal, and cross-cutting technology instruments. Its legal structure is not reducible to the AI Act. The GDPR protects personal-data processing; the Digital Services Act structures intermediary obligations and systemic platform risk; the Digital Markets Act regulates designated gatekeeper conduct; the NIS2 Directive addresses cybersecurity duties through national transposition; and the Cyber Resilience Act introduces product-oriented cybersecurity obligations subject to staged application. Their different legal techniques must be read together when a service performs overlapping functions.

The distinction between a regulation and a directive is central. EU regulations ordinarily apply directly according to their terms and commencement provisions, while directives establish obligations for Member States to achieve prescribed results through national measures, subject to the applicable doctrines of EU law. It would be inaccurate to treat a directive's transposition deadline as evidence that every national implementation has identical operative language. Equally, the presence of an EU regulation does not eliminate the need to identify national competent authorities, complaint procedures, and rules governing investigation or judicial review.

The AI Act works mainly through categorized uses and actors; the GDPR follows personal-data processing and its legal bases; the DSA follows intermediary service status and relevant scale; the DMA follows formal gatekeeper designation; and cybersecurity instruments follow products, entities, or covered network activities. The intersection is a legal fact rather than a drafting inconvenience. An AI-assisted recommendation system deployed by a large platform may implicate platform transparency and systemic risk, personal-data rules, and—depending on its specific functionality—AI requirements. The correct analysis asks which obligations apply concurrently, which authority can enforce them, and which evidence obligations are genuinely complementary.

The EU Data Act also illustrates the separation of data access and data protection: contractual and statutory rights to access or share certain data operate with continuing constraints on personal-data processing. A contractual demand for connected-product data is therefore not a universal justification to disclose personal information. This interaction is a general lesson for technology governance: legal entitlements concerning data portability, competition, interoperability, and safety must be interpreted within the wider legal order.

The Artificial Intelligence Act as amended in 2026

The legal chronology of the EU AI Act is an essential case study in comparative precision. Regulation (EU) 2024/1689 entered into force in 2024 with multiple staged dates. The EU legislature then enacted Regulation (EU) 2026/1744, amending the implementation timetable and related provisions. The consolidated AI Act records the operative amended framework. Assertions that the original 2024 timetable remains unmodified would misstate the law at this research cutoff.

The amended scheme specifies December 2, 2027 for the high-risk AI requirements concerning systems classified under Article 6(2) and Annex III, and August 2, 2028 for relevant Article 6(1) and Annex I product-related systems. These are future applicability dates relative to November 21, 2026. Other portions of the Act, including important prohibitions and certain general-purpose AI obligations, have already reached their prescribed stages, subject to particular transitional rules. The Commission's legislative summary of the revised schedule and the official consolidated text should be read together before asserting the date of an obligation for a specified system.

This staging alters how compliance claims should be assessed. A firm could be subject to one presently applicable prohibition, an existing model-related rule, and an upcoming high-risk conformity duty for another application. The fact that an obligation will apply later does not render a planned control legally irrelevant: prospective design, procurement, and documentation may need to anticipate it. But a legal report must distinguish preparation for a future duty from a currently enforceable contravention. The dated analysis of particular use cases should state their classification, relevant role, market status, transitional treatment, and trigger date.

The regulatory architecture also differentiates prohibited practices, high-risk systems, transparency obligations, general-purpose model responsibilities, and institutional oversight. Each category has its own elements. A financial credit-scoring application requires examination of the statutory classification and applicable exceptions; a general-purpose model requires attention to the model provider's specific obligations; synthetic-content disclosures fall under distinct transparency provisions. In legal analysis, a broad assertion that all AI is high-risk is equally defective as the assertion that a model's general-purpose character exempts every downstream use.

A comparative point follows. European ex ante controls may organize evidence before deployment, but accountability ultimately also requires a route to challenge an adverse decision and obtain material records. Documentation becomes legally productive when its contents, custodians, retention arrangements, and accessibility correspond to duties that the appropriate court or regulator can enforce. Compliance labels detached from a particular version, use, and date supply a weaker basis for legal accountability than a documented chain connecting system design to a contested outcome.

Data protection and automated individual decisions

Data governance supplies a distinct legal vocabulary for technologically mediated conduct. Information about identifiable persons is subject to processing rules whose application turns on controller and processor roles, legal bases, purposes, minimization, accuracy, security, and rights. In the European GDPR's substantive provisions, automated individual decision-making must be evaluated through Article 22 and the law's broader provisions, including transparency, access, lawful processing, special-category data, and remedies. Reading Article 22 as a free-standing universal right to inspect all algorithmic source code would ignore its defined conditions and the surrounding legal safeguards.

The Court of Justice addressed one important boundary in SCHUFA Holding, C-634/21. The case concerned automated probability scoring used in a credit context and the circumstances in which production of a score could constitute automated individual decision-making within Article 22. Its significance lies in treating commercially distributed scoring as part of an actual downstream decision chain. A formal partition between a score supplier and the institution making the final decision cannot settle whether an individual is subjected in substance to automated decision-making. The judgment still requires careful application to each factual setting; it does not declare every statistical prediction unlawful.

This holding illuminates a recurring governance problem. A company may assign a person to an eligibility class through a combination of third-party score, internally chosen threshold, and automatic rejection. If each participant asserts that the decisive act occurred elsewhere, legal responsibility risks disappearing into the interfaces between organizations. Doctrinal reconstruction must therefore trace information flows and identify which actor determines the relevant purpose and means, who adopts the threshold, whether human intervention is meaningful, and which person can obtain review or challenge.

The United Kingdom illustrates a related but distinct approach. Section 80 of the Data (Use and Access) Act 2025 amended the UK framework for significant decisions based solely on automated processing, with relevant commencement addressed in the 2026 commencement regulations. A comparative report must use the UK provisions actually in force, rather than importing the EU's current Article 22 text into post-Brexit law. The substance of safeguards, available challenges, and protected processing categories needs to be read from the operative UK framework.

Data-protection impact assessments are likewise specific legal instruments rather than generic statements of ethical intention. Their content and triggering conditions follow applicable law, including the risks to rights and freedoms of natural persons. An internal AI impact assessment may overlap with a GDPR assessment without automatically satisfying it. Conversely, a privacy-compliant operation may still raise workplace discrimination, consumer deception, or safety concerns. The correct legal conclusion is cumulative: data governance constitutes one necessary layer of the analysis wherever personal information is processed, while other causes of action and supervisory powers remain independently relevant.

Digital platforms, intermediary power, and market contestability

Platform regulation treats digital infrastructures as institutions that distribute speech, access, visibility, commerce, and dependency. The European Digital Services Act differentiates intermediary services and imposes further duties according to relevant service characteristics and scale. It contains procedural duties associated with content moderation, transparency, complaint handling, advertising, and certain systemic-risk assessments. Its focus is the governance of intermediary activity, including automated systems used in that activity; the fact that a platform relies on AI does not erase its position under intermediary law.

A provider's content-ranking algorithm can influence what users encounter without expressly prohibiting any item of speech. This makes information about recommender parameters, systemic risk, and the possibility of independent scrutiny significant. The European Commission's DSA enforcement explanation describes investigatory authority, access to relevant platform data and algorithms, and sanctions under applicable conditions. These powers demonstrate the legal importance of regulatory access to technical evidence. They do not constitute a general public entitlement to obtain every platform dataset or inspect proprietary code without statutory limits.

The Digital Markets Act pursues a different legal objective: contestability and fairness of designated core platform services. Duties concerning conduct by designated gatekeepers address structural commercial relationships, including access, data practices, self-preferencing, and other specified behaviors. A single recommendation model may be relevant to both DSA and DMA inquiries, yet the legally decisive questions differ. Under the DSA, the regulator may examine illegal-content responses and systemic risks; under the DMA, the inquiry may concern the competitive effects of a gatekeeper's conduct under identified obligations.

Outside the EU, statutory and constitutional approaches to moderation diverge. United States regulation proceeds through a different configuration of federal statutes, constitutional rights, state initiatives, agency authority, and judicial decisions. It is therefore misleading to label the DSA equivalent to American intermediary immunity doctrines. Such doctrines serve different purposes, and constitutional limits on government intervention are not automatically equivalent to a platform's private contractual choices. Institutional function offers a more reliable comparison: who can compel disclosure, what rights can users assert, and which specific wrong can a regulator or court remedy?

Platform governance also presents a transnational conflict of norms. A service may host content lawful in one country and unlawful in another, contract with users under a third country's terms, and operate recommender infrastructure across several territories. The legal answer involves jurisdiction, applicable law, public policy, regulatory cooperation, and the precise scope of any remedial order. It cannot be derived solely from the technical ability to remove content globally. Governance requires territorial reasoning alongside technical implementation.

Cybersecurity, resilience, and connected products

Cybersecurity law is an essential component of technology governance because failures of software and infrastructure can produce physical, economic, privacy, and public-service harms. The European NIS2 Directive develops cybersecurity risk-management and reporting requirements for specified categories of entities, with domestic transposition and supervisory arrangements determining much of their practical operation. A legal claim about one Member State's NIS2 obligations should identify the country's implementing law and the regulated entity's category. The directive's existence alone does not establish the final form of each national procedure.

The Cyber Resilience Act, Regulation (EU) 2024/2847, adds cybersecurity requirements for products with digital elements, including lifecycle considerations and obligations connected to manufacturers and related economic actors. Its timetable is staged: some vulnerability and incident-reporting provisions operate ahead of the regulation's broad December 2027 applicability. Treating the entire Act as fully applicable at the 2026 research cutoff would erase this distinction. Effective compliance assessment demands a provision-specific date and a determination of the product's scope.

Cybersecurity duties often have a temporal dimension. A product can be secure when released and later become vulnerable through newly discovered attack methods, third-party library changes, or altered operating conditions. Statutory reporting and vulnerability-handling rules reflect the need for continuing maintenance rather than a single point-in-time certification. Procurement documents, update policies, software-component inventories, incident-response records, and vulnerability disclosures may become central evidence when regulators investigate whether a vendor discharged applicable obligations.

Connected products also create overlapping data and security questions. The European Data Act addresses defined rights and conditions concerning data generated through connected products and related services. A right to access or port data does not eliminate duties to secure systems, protect trade secrets under relevant conditions, or comply with personal-data rules. Nor does cybersecurity automatically justify refusing every legitimate data request. Governance consists of interpreting the distinct applicable tests and implementing technical controls that can satisfy them concurrently.

The relationship between general AI governance and cybersecurity is similarly specific. A model may assist in detecting security threats, generate exploit-like material, or itself be vulnerable to adversarial manipulation. Existing cybersecurity laws may apply because of the entity, product, or network service involved, independently of the model's classification under AI-specific legislation. Conversely, the presence of an AI risk framework does not automatically establish conformity with a statutory cybersecurity regime. Documentary equivalence must be proven against the legal elements of each obligation.

United States governance across agencies and states

The United States presents a distributed governance system shaped by constitutional division of powers, sector-specific statutes, federal agency mandates, state legislation, common-law claims, and private enforcement. A central comparison with the EU is institutional rather than rhetorical: the U.S. regime lacks a single comprehensive federal AI statute that simply reproduces the structure of the EU AI Act. Yet federal and state law can impose materially significant obligations on particular conduct. Consumer deception, discrimination, unsafe products, unauthorized data practices, and unlawful public decisions remain legally cognizable when their governing elements are established.

The Federal Trade Commission illustrates enforcement through existing authority. In the FTC's Rite Aid matter, the agency challenged alleged deficient uses of facial recognition and obtained relief through a consent-order process. An FTC complaint expresses allegations, and a settlement establishes the order's terms; neither should be described as a contested judicial determination of every alleged fact. Nonetheless, the case demonstrates how agency supervision can address an AI-related practice through existing unfair-or-deceptive-practices authority and specific remedial commitments. Its significance is the concrete power of a regulator to demand corrections, not the creation of a general AI statute through one case.

The National Institute of Standards and Technology's AI Risk Management Framework 1.0 exemplifies another governance channel. The framework's functions can guide risk identification, monitoring, and institutional management across sectors. It has technical influence but is generally voluntary as a NIST publication; legally enforceable duties arise when statutes, regulations, contracts, procurement obligations, or other binding arrangements incorporate relevant provisions. A certificate that a firm has aligned with NIST does not by itself answer whether a particular bank, hospital, or employer has fulfilled mandatory legal obligations.

State regulation supplies a further layer. California enacted SB 53's frontier-AI safeguards in 2025, with obligations regarding designated frontier developers, safety policies, reporting, and related protections under its applicable terms. California's 2026 AI safeguard legislation demonstrates the continuing accumulation of state-specific controls. The precise enactment, commencement, coverage thresholds, and enforcement provisions must be examined for each statute. State enactment is neither a federal mandate nor proof that every AI provider operates within its scope.

Colorado offers a caution about legislative chronology. Its AI legislation has been subject to subsequent measures on implementation timing, recorded in the official SB25B-004 legislative record. A previously proposed or amended date must not be repeated without checking the final signed enactment and later law. This distinction is not administrative trivia: firms planning safeguards and persons seeking remedies need to know whether a duty is presently operative or still prospective.

Federalism makes legal fragmentation a source of both institutional experimentation and compliance complexity. A provider operating nationally may need to identify which duties apply because of the customer's location, protected activity, regulated sector, sale or deployment context, and the reach of a particular state provision. Where competing duties exist, the question is one of statutory interpretation, constitutional competence, preemption, and applicable law. It cannot be resolved by assuming that the strictest state rule governs every transaction throughout the country.

The United Kingdom's sectoral and public-law arrangements

The United Kingdom governs artificial intelligence through a combination of existing regulators, data-protection legislation, safety and consumer law, public-law duties, and sectoral supervision. This institutional configuration differs from the EU's horizontal AI Act, but the absence of an equivalent omnibus statute does not establish the absence of law. The state can impose obligations through the statutory competence of data, competition, communications, financial, equality, and sector-specific regulators.

The Online Safety Act 2023 structures specified duties concerning online services, safety processes, and regulatory supervision, with provisions subject to staged commencement and detailed codes or guidance. Its legal object is online-service conduct under the Act rather than artificial intelligence as a standalone category. A platform using automated content identification must therefore examine its online-safety obligations, relevant speech and privacy law, and any applicable data-processing safeguards as separate but interacting questions.

Public-sector transparency illustrates the mixture of policy and law. The government's Algorithmic Transparency Recording Standard guidance describes methods for reporting information about algorithmic tools in applicable public-body settings. It is important to distinguish government policies directing particular institutions from general statutory rights binding all actors. Publication of an algorithmic record may help establish a system's purpose, responsible agency, and risk controls, but legality also depends on statutory authority, rationality, proportionality where applicable, and procedural fairness.

Judicial review gives such obligations their practical significance. In R (Bridges) v South Wales Police, the Court of Appeal examined the lawful use of police facial recognition under the relevant rights, data-protection, and public-sector-equality frameworks. The holding demonstrates that even where automated surveillance technology is technically capable of identifying individuals, deployment requires a sufficiently lawful and constrained institutional basis. The case must be understood through its specific reasoning rather than converted into a universal prohibition of facial recognition.

UK data protection has also changed. The Data (Use and Access) Act 2025, with the relevant 2026 commencement instrument, modified automated-decision rules within the UK legal framework. Comparative commentary written before commencement must be checked against the operative amended law. A general description of the EU GDPR is not a sufficient account of current UK rules, even though historical language and underlying regulatory concepts overlap.

The legal advantage of this model is that established regulators and courts can apply domain-specific expertise to AI-enabled conduct. Its challenge is that responsibilities may become dispersed across agencies, procurement teams, vendors, and complaints bodies. Effective legal supervision demands a coordinated record of who authorized deployment, what problem the system was intended to solve, how accuracy and equality were assessed, and which person may correct a mistake. The comparative issue is thus one of demonstrable responsibility within a sectoral architecture.

China's administrative governance of generative systems

China's technology governance is organized through a distinct combination of nationally administered information-system controls, sector-specific responsibilities, data rules, service-provider obligations, and institutional supervisory powers. The Cyberspace Administration of China's Interim Measures for Generative AI Services, adopted in 2023, establish a legally important framework for provision of designated generative-AI services to the public. The legal category is service provision within the measures' specified scope, and application depends on the measure's definitions, exclusions, and relationship to other rules.

The Measures address matters including the conduct of service providers, training and data considerations, user rights and safety, and supervisory responsibilities. China's official English-language policy announcement conveys the government's explanation of the initiative. For doctrinal analysis, the authoritative Chinese text governs; an English summary supports the description of regulatory objectives but is not a substitute for examination of each operative clause when a specific legal consequence is asserted.

One distinctive feature is the interaction between content-related administrative duties and the governance of technology development and distribution. In a platform setting, the application of generative AI to public communication may create obligations under the service framework and other information-governance rules. This differs institutionally from a European analysis that first separates the AI Act's risk classification from DSA intermediary duties and GDPR processing obligations. The functional questions remain comparable: who is identified as the responsible service provider, what conduct triggers supervision, which records establish compliance, and how affected persons obtain a lawful response.

The jurisdictional comparison should avoid treating regulatory ambitions as proven outcomes. The existence of supervision does not establish the quality of every enforcement intervention, and official statements about innovation objectives do not independently demonstrate market effects. A complete empirical evaluation would require enforcement decisions, institutional data, and controlled evidence regarding service behavior. This monograph draws a narrower doctrinal conclusion: China's current administrative model specifies regulatory obligations through national service governance while retaining an important role for authorities that supervise networked information services.

The comparison also cautions against superficial analogies. An EU provider–deployer distinction cannot simply be imposed on all Chinese legal categories, just as a Chinese service-provider responsibility cannot be imported into United States consumer law without identifying an American cause of action. Comparative legal scholarship must translate the legal functions and preserve their institutional origins. The object of comparison is the consequence assigned to an actor under a legal instrument, not the availability of a shared English phrase.

South Korea and the emerging Asian regulatory comparison

South Korea's AI framework provides a useful comparison because it establishes broad national legislation with an express combination of innovation policy, trust, and institutional governance. According to the Ministry of Science and ICT's implementation announcement, the AI Basic Act and its Enforcement Decree took effect on January 22, 2026. The framework creates public institutional responsibilities alongside obligations attached to particular AI services and applications. The date establishes an operative legal regime as of this study, rather than a future proposal.

The Ministry's AI transparency guidance further illustrates the relationship between statutory rules and implementing explanation. Guidance can supply practical information about interpreting the law, while the underlying statutory provisions and subordinate regulations determine formal legal obligations. The comparison requires particular attention to what the official ministry materials state and what the legislative text actually requires. It would be inaccurate to assume that every transparency practice recommended in guidance carries the same sanction or scope.

The Korean arrangement is significant for its explicit conjunction of industrial development and trust-based protection. This creates possible institutional tension: the public body responsible for promoting AI capacity may also administer regulatory measures intended to protect users or the public. The existence of both responsibilities is not proof of compromised independence; it makes transparent criteria, reasoned decisions, and appropriate review mechanisms especially important for assessing the institution's operation. Comparative analysis can examine whether supporting legislation and administrative structures supply those mechanisms.

The EU model likewise combines innovation support with detailed compliance rules, while the United States often separates innovation promotion and enforcement across institutions. South Korea's consolidated national framework shows that governance instruments cannot be grouped adequately into a simple pair of restrictive and permissive systems. The relevant distinctions concern enforceable duties, reporting procedures, administrative powers, and legal remedies, as well as the stage in a technology's lifecycle that prompts government intervention.

Singapore offers a contrasting governance channel through voluntary guidance and cross-sector practice. The Infocomm Media Development Authority's generative AI governance materials demonstrate institutional efforts to organize evaluation and accountability without making each recommended practice a universal statutory duty. Their persuasive and operational importance can be substantial in procurement and organizational risk management, but the basis of legal enforceability must still be independently identified.

Taken together, the examples establish diversity within Asian governance. China, South Korea, and Singapore differ in their choice of source, institutional authority, and the function assigned to technical standards and guidance. Geographical grouping does not create a unified regulatory family. Lawful comparison requires country-specific primary text, accurate translation, and separation of statutory coverage from voluntary implementation materials.

Canada and automated public administration

Automated public administration raises a particular legal question: what authorizes an agency to make a decision affecting a person's rights or interests through a computational process? Canada's federal Directive on Automated Decision-Making is a prominent instrument directing covered administrative practices. It operates within the federal administrative policy framework and must be distinguished from an act of parliament that independently confers or removes an agency's substantive legal powers.

The directive's algorithmic impact-assessment architecture helps organize risks by considering system effects and corresponding procedural safeguards. Such a framework can be valuable because agencies often procure systems from third parties and lack sufficient in-house access to design documentation. It creates an administrative basis for examining what the agency knows about the system, how the system affects a decision, whether review processes are available, and how the institution communicates with affected individuals.

Yet a policy-driven assessment does not confer legal authority to impose a burden that a statute does not permit. Administrative legality rests first on the agency's statutory mandate and applicable procedural requirements. If a procurement contract promises explainability but the agency cannot obtain the evidence required in review proceedings, the practical value of that promise is doubtful. Contractual governance should therefore be structured to preserve documentation and access rights necessary for the agency's public-law duties.

The comparison with the United Kingdom is instructive. Both systems can rely on government-wide administrative instruments while retaining familiar judicial review principles and sectoral legislation. The UK algorithmic-transparency guidance emphasizes public disclosure of information about tools; Canada's directive places weight on administrative impact assessment and procedural protections for covered departments. The two can be compared as complementary administrative-governance techniques without asserting that they confer identical legal rights.

A broad lesson emerges: automated administration demands records that correspond to the reasons an agency must lawfully provide. It is insufficient to publish a generic description of a technology if the legal dispute concerns whether a particular claimant's evidence was ignored, whether a decisive statutory criterion was met, or whether a discriminatory proxy influenced the result. The record must connect the general design of the system to the particular official action at issue.

Constitutional rights and technologically mediated power

Constitutional and fundamental-rights protections structure the permissible uses of technology by public authorities and, through relevant doctrines, private actors. Their application varies by jurisdiction and legal mechanism. Privacy, data protection, expression, equality, procedural fairness, property, and access to an effective remedy do not operate as an undifferentiated right to responsible AI. Each right has identifiable bearers, duty-bearers, scope, standards of justification, and enforcement institutions.

The Dutch District Court's 2020 SyRI judgment concerning a state fraud-detection system is instructive. The court examined the system's statutory and operational arrangement under the right to private life protected by Article 8 of the European Convention on Human Rights, with attention to transparency and proportionality. Its doctrinal significance is that a public-interest objective such as combating fraud does not by itself settle whether a data-intensive mechanism is justified. Legality, necessity, safeguards, and the balance of interests require case-specific assessment. The holding belongs to its legal and factual context rather than a universal judicial rule against predictive risk scoring.

In the United Kingdom, Bridges likewise connects automated facial recognition to public-law constraints. The legal issues implicated discretionary criteria, data-protection impact assessment, and equality duties. Both cases show that algorithmic opacity can become relevant to established constitutional or administrative requirements, but neither demands a newly invented constitutional right in order to adjudicate the contested governmental practice.

Rights-based comparison must distinguish the legal protection of decisions from the legal protection of data. A public authority may lawfully possess data yet employ them in an unauthorized decision, or may reach an otherwise permissible conclusion through a processing operation that infringes privacy rules. Similarly, a mathematically accurate predictor can produce a legally impermissible discriminatory effect if the relevant equality law prohibits the practice. Courts investigate the elements and burdens of the specific claim; predictive performance is evidence only where legally relevant.

The ability to challenge a decision requires practical access to information. A right to reasons offers little protection if reasons consist solely of a score that omits the facts, criteria, and human decision responsible for the outcome. At the same time, an explanation need not disclose every parameter of a complex model to answer the legally material question. A person contesting benefits denial may require the statutory ground, the factual inputs attributed to the applicant, and the process for correcting errors. The required level of disclosure is determined by the governing legal scheme, competing interests, and the particular dispute.

The relationship between technology and rights is therefore procedural as well as substantive. A legal system's ability to detect an infringement depends on independent adjudication, standing, discovery or access rights, burden allocation, admissible evidence, preservation of records, and timely relief. Governance systems that regulate design but lack credible routes to contest harmful uses may underperform their stated rights-protection purposes.

Administrative legality, reason-giving, and judicial review

Public authorities are bound by the legal rules defining their jurisdiction and methods of action. Deploying an algorithm does not enlarge an agency's statutory discretion, change eligibility criteria, or replace the requirement to consider legally relevant evidence. The correct starting point for the lawfulness of automated administration is the authority's enabling statute, procedural obligations, relevant rights, and the standards governing judicial or administrative review.

Australia's Robodebt Royal Commission report documents an important failure of public administration involving automated debt assessment and unlawful government action. The report reveals the institutional significance of legal advice, procedural oversight, records, internal challenge, and the relationship between technology and bureaucratic decision-making. It is not a judicial precedent establishing that all automation violates administrative law; it is an authoritative official investigation of a specific program and the institutional conditions in which unlawful conduct occurred.

This record establishes a broader analytical point. A technical process may generate confident numerical results while implementing an invalid legal assumption. Errors of this type cannot be corrected solely by retraining a model or improving statistical accuracy. The agency must verify that the computational step corresponds to the statute's actual criteria, that the available data can legitimately establish the required fact, and that the individual has a procedure for contesting adverse inferences.

Judicial review may address questions of legality, reasonableness, procedural fairness, equal protection, statutory interpretation, and proportionality, depending on the legal system. An assessment of algorithmic use must specify which of these grounds is invoked and how the challenged decision meets its elements. A generalized objection that a model is opaque will carry different legal implications where a law specifically requires reasons, where transparency is conditioned on processing categories, and where the court's concern is the decision-maker's improper reliance on irrelevant factors.

Algorithmic procurement is another point of public responsibility. If a vendor retains decisive information as proprietary, the agency remains obligated to comply with its public-law requirements. A contract designed without audit access, retention of historical versions, or capacity to reconstruct individual decisions can frustrate that obligation. Procurement therefore becomes a legal-governance activity: contracting authorities need enforceable provisions that preserve their capacity to meet statutory duties, defend lawful decisions, and correct unlawful ones.

A practical review file for an adverse automated decision should establish the governing authority and eligibility criterion; the identity and version of the system used; the relevant inputs and their provenance; the rules and human intervention applied; the resulting communication to the person; the record of objections and corrections; and the available appeal channel. These elements form an analytical checklist drawn from existing legal functions, not a universal statutory checklist. Their exact necessity depends on the dispute and the applicable legal regime.

Liability, causation, evidence, and civil remedies

Civil liability supplies a distinct response to technology-related injury. It may involve product liability, negligence, contract, statutory consumer protection, discrimination, privacy, professional responsibility, or other recognized causes of action. The diversity matters because a technological harm does not automatically establish a defendant's legal duty, breach, causation, or available remedy. A careful analysis identifies the claimant, the defendant's role, the alleged wrongful conduct, the damage or other protected interest, the causal link, and the burden and standard of proof.

The EU's Product Liability Directive 2024/2853 modernizes the European product-liability framework to address products including relevant software and associated technical complexity. It is a directive requiring implementation through Member State law, and its prospective applicability must be distinguished from the legal regime governing products before relevant dates. As of November 21, 2026, the December 2026 transposition and application milestone is still future. It is legally inaccurate to treat every substantive provision of the new regime as already governing past injuries in all Member States.

Technology complicates proof in identifiable ways. A system may combine models, training data, integrations, human adjustments, and continuing updates. The claimant may lack access to the precise version used when the injury occurred. A defect might concern design, inadequate instructions, failure to update, unauthorized changes by a downstream operator, or negligent reliance on a system in a regulated profession. The relevant legal category determines which facts must be established and whose custody of the records affects procedural fairness.

Consider a clinical decision-support tool that contributes to an incorrect diagnosis. Possible inquiries include product status and defect, whether the supplier's warnings were adequate, whether the clinician fulfilled professional obligations, and whether the hospital operated appropriate monitoring. Several legally relevant acts may coexist. An assertion that the software made the decision would obscure the human and organizational relationships necessary to establish liability under the applicable law.

Comparative liability law also reveals the limits of technical transparency. Detailed logs may make a factual causal pathway easier to reconstruct, but liability cannot be inferred automatically from log completeness. Conversely, incomplete logs may impede proof and sometimes trigger applicable disclosure or evidentiary consequences where law provides for them. Courts must determine the relevance, admissibility, reliability, and burden allocation under the governing procedure.

Remedies vary. An administrative body may order cessation of a practice, a consumer regulator may require corrective disclosures, a court may grant damages, a privacy authority may order processing changes, and an affected person may obtain reconsideration of a public decision. The remedies are not interchangeable: an administrative fine paid to the state does not automatically compensate an injured individual, and an internal audit does not itself reverse an unlawful denial of benefits. Comprehensive governance requires both preventive controls and a legally accessible path to the remedy suitable for the right infringed.

Auditing, standards, and the limits of verification

Auditing makes technical conduct legible to organizations, supervisors, courts, and affected persons, but its legal significance depends on scope and authority. An internal evaluation, independent conformity assessment, statutory inspection, external assurance report, and litigation expert opinion answer different questions. Their evidence can overlap while the legal consequences diverge. The presence of a sophisticated audit process does not prove that every affected individual has a remedy, nor that a regulator has accepted the system.

The NIST AI RMF supplies useful risk-management functions, while Ada Lovelace Institute's analysis of algorithmic assessments examines the prospects and limits of assessment instruments. Their value for comparative law is to identify the informational predicates that governance needs: clear system boundaries, affected populations, intended use, anticipated harms, accountability assignment, and follow-up procedures. The legal analyst must still determine the statute, contract, or policy that requires a given assessment and the consequences of deficient performance.

One fundamental audit issue is version dependence. The characteristics of a trained system can change through model updates, altered prompts, new data, threshold adjustment, or integration into another workflow. A report describing one frozen configuration may cease to support claims about a changed deployment. Valid evidence should identify which version and operational setting were assessed and what alterations trigger further review. This conclusion follows from ordinary evidentiary relevance and the technical reality of evolving systems, rather than from a new doctrine of algorithmic evidence.

Another issue is the difference between aggregate performance and individual legality. A model may achieve high average accuracy while producing systematic adverse outcomes for a protected group, failing at a rare but critical task, or applying an unauthorized decision criterion. Statistical performance is therefore only one dimension of scrutiny. Legal evaluation may require proof concerning discrimination, procedural safeguards, permissible purposes, specific product defects, or the fairness of a particular decision. A generalized safety score cannot resolve all these distinct legal questions.

Independent verification can improve credibility only when independence, access, competence, and consequences are adequately defined. An auditor whose access excludes the data, configurations, or decisions most relevant to risk may provide useful but limited assurance. Publication of an executive summary without methods can inform the public while leaving crucial questions unresolved. Regulators and courts need proportional means of obtaining the underlying evidence, subject to legitimate confidentiality and privilege constraints.

The analytical implication is institutional: technical standards and audits should be connected to defined legal triggers and follow-up responsibilities. A company must know who receives a report, who is required to act on findings, how corrective actions are recorded, and what occurs when deployment changes. The effectiveness of audit governance rests on this chain of custody and responsibility, together with the possibility of meaningful independent challenge.

Cross-border jurisdiction and institutional cooperation

Emerging technologies routinely connect actors located in different territories. A model may be developed in one jurisdiction, incorporated into a service in another, and applied to users or institutions in a third. A platform may process content and personal information across regions while engaging separate regulatory authorities. Cross-border activity therefore requires a disciplined sequence of questions: which conduct is regulated, what connection brings it within a law's territorial scope, which authority has jurisdiction, which substantive provisions apply, and where any remedial order can be enforced.

Some instruments define jurisdiction by factors such as offering services or placing products on a market, rather than the developer's headquarters alone. The European AI Act and GDPR both contain territorial-scope provisions, but with their own wording and conditions. An assertion that any internet-accessible service is automatically subject to every European technology rule would be false. A legal opinion must establish the specified territorial nexus for each instrument.

Regulatory jurisdiction and private international law are related but separate. A regulator may exercise a statutorily authorized supervisory power, while a private litigant confronts choice of forum, governing law, service of process, evidence collection, and recognition of a judgment. Territorial reach of a regulatory standard does not automatically guarantee a usable private remedy in every court. Similarly, a defendant's contractual choice-of-law clause may not displace overriding mandatory protections in the relevant legal system.

Cross-border governance also requires attention to conflicting duties. A platform may receive legally issued demands for data access, preservation, deletion, or disclosure from different jurisdictions. The governing analysis must consider statutory prohibitions, lawful bases, international-transfer mechanisms, confidentiality restrictions, conflicts rules, and relevant public policy. An abstract command to maximize transparency supplies no resolution where one legal order requires disclosure and another restricts it. Institutional cooperation can reduce friction, but it must itself operate under lawful powers.

The Council of Europe AI Convention illustrates the potential of treaty-based convergence at the level of public-law commitments. The OECD AI Recommendation offers a nonbinding vocabulary that can assist coordination. The latter cannot settle a conflict of national laws merely because regulators share its principles; the former must be evaluated state by state under treaty law and domestic law. The legal effects of international commitments turn on their status, scope, and implementation.

An effective cross-border compliance program should therefore record, for each significant use, the relevant legal nexuses, the actors with operational control, the locations of critical evidence, applicable transfer restrictions, and the institutions capable of ordering relief. This is a legal inventory rather than a technological map. It can identify when a service's architecture deprives a domestic regulator or court of evidence even though the service remains directed toward users in that jurisdiction.

Comparative case studies

Consumer credit scoring across organizational boundaries

Imagine a consumer who is denied a loan after a financial institution relies heavily on a model-generated score obtained from an external provider. A complete legal inquiry begins by identifying whether the score is personal data, what determines the decision, the applicable consumer and financial rules, and whether a protected legal interest is affected. Under the European framework, SCHUFA demonstrates the need to examine the material role of automated scoring rather than the formal statement that a human institution issues the final loan denial.

If the case arises in the United States, the analysis instead turns to the applicable fair-credit, consumer-protection, data, and state law, the precise conduct involved, and available administrative or civil remedies. It would be a category mistake to apply the EU Article 22 test as the direct rule of decision. If a related operation falls within the scope of the EU AI Act's relevant classifications, its staged high-risk timetable must additionally be assessed according to the system and deployment dates. This case shows the importance of distinguishing the legal consequences of a technology's functional role from the regulatory techniques chosen in each jurisdiction.

The evidentiary problem is similar across regimes: when a claimant asks why credit was denied, relevant data may be dispersed among the model vendor, score supplier, lender, and data brokers. Establishing the legally decisive reason may require the inputs, decision threshold, contractual allocation of control, and records of human intervention. But access to each component depends on the particular legal power or right invoked. A judicially useful technical explanation is one that answers the elements of the claim, rather than one that simply provides more mathematical description.

Public benefits and statistical inference

Consider an agency estimating a person's eligibility or alleged overpayment by applying a statistical assumption to incomplete earnings data. The first legal question is whether the relevant statute authorizes that assumption as a basis for a debt or adverse determination. Australia's Robodebt Commission illustrates the institutional costs when automated calculations are used within a legally defective program. Improving the precision of software cannot establish statutory power that the legislature did not confer.

A corresponding European privacy issue may arise if a public agency also combines large datasets and predictive profiling. The Dutch SyRI judgment shows that an asserted public-interest objective must be evaluated against applicable rights and safeguards. A Canadian agency may additionally have obligations under the federal Automated Decision-Making Directive, depending on the agency and system's coverage. These are legally different controls directed to a shared governance problem: public decisions must remain connected to lawful authority, intelligible evidence, and accessible correction.

The decisive record should contain the source of power, the criteria governing entitlement, the factual basis and uncertainty of the calculation, the method for resolving inconsistent data, the identity of the responsible official, and a procedure for challenge. Such a record cannot guarantee that all decisions will be correct. Its significance is that errors become attributable and susceptible to legal review. Where the agency cannot reconstruct why a particular person received an adverse decision, the problem is institutional evidence governance as well as technical quality.

Facial recognition in public and commercial settings

Facial recognition exposes differences in legal competence and context. The UK Bridges judgment concerned police use of a facial-recognition system under public-law and rights constraints. The FTC Rite Aid proceeding concerned the agency's allegations about a private retail operator and consequent settlement relief. Both address risks around biometric identification, but they rest on different legal sources and produce different forms of review.

The legally relevant inquiry in a police deployment includes the public authority's statutory competence, the scope of discretion, equality obligations, privacy safeguards, and available judicial oversight. In a retail deployment, unfair or deceptive commercial conduct, representations to consumers, training and monitoring procedures, and the terms of an enforcement order may be more directly relevant. The presence of the same facial-recognition technique does not create a uniform cause of action.

Technical validation must also be specific. False positives can produce different harms depending on whether a match leads to an employee checking an alert, denial of access, a police stop, or arrest. A lawfully obtained biometric image may be used in an unlawful way if the separate requirements for the resulting intervention are absent. Model accuracy needs to be assessed in the context of legally material consequences and procedural opportunities to correct errors.

Platform recommendation and market power

A large online platform ranks content, recommends products, selects advertisements, and provides access to dependent businesses. The European DSA's intermediary framework and the DMA's gatekeeper framework may both be relevant, but their tests are distinct. DSA obligations can address content governance, disclosures, and systemic platform risks, while the DMA addresses specified unfair or contestability-limiting conduct by designated gatekeepers.

If the platform uses personalized recommendations based on personal data, the GDPR may also apply. A regulator investigating manipulation of user attention needs different evidence from one investigating discrimination in market access or unlawful data processing. One repository of recommendation logs can be relevant to multiple legal issues, but procedural protections, retention, and disclosure permissions must be considered for each inquiry.

A comparative analysis of American platforms must separately identify federal statutory protections, agency authority, constitutional limits, and state law. It is neither accurate nor informative to state that the DSA's duties automatically govern every American platform interaction. What can be compared is the practical capacity of institutions to investigate ranking effects, prevent prohibited conduct, hear complaints, and impose remedies authorized by law.

General-purpose model distribution and downstream deployment

Suppose a developer releases a general-purpose model to commercial integrators, who use it in employment screening, medical triage, and public-facing communication. A supply-chain assessment should identify the model provider, each integrator, the operating deployer, the affected population, and the legally consequential use. Under the EU amended AI Act, general-purpose-model obligations and high-risk application obligations occupy distinct legal categories with different compliance and transition dates.

In the United States, the same technology may trigger sectoral duties or state measures depending on its use. California's frontier-AI enactment illustrates regulation directed toward specified frontier development and safety practices, whereas federal agency authority may address misleading consumer representations or harmful regulated conduct. The NIST risk framework can inform testing across the supply chain without creating a universal legal authorization for each downstream use.

If an integrator modifies a model or combines it with a proprietary scoring layer, responsibilities require reevaluation. A document describing the original model may supply historical evidence but may fail to prove the behavior of the integrated system. Procurement arrangements should preserve the ability to identify versions, limitations, changes, and operational incidents, subject to applicable law. This is a practical conclusion about evidence and attribution, not a declaration that every legal system imposes the same supply-chain audit requirement.

Incident response across product and network rules

Consider a connected device whose software vulnerability causes service disruption and exposes personal information. The legal analysis may implicate product design and continuing security duties, covered-entity reporting, data-protection notification rules, and contractual rights. In the EU, the Cyber Resilience Act, NIS2 Directive, and GDPR may each be relevant according to their separate scope and applicability.

An incident report prepared for one regulator need not satisfy another instrument's elements, deadlines, addressee, or content requirements. A sound compliance process distinguishes the occurrence of an incident, the factual inquiry into its cause, the determination of whether a legal reporting trigger is met, the lawful sharing of evidence, and the implementation of remedies. The legal importance of cybersecurity documentation is therefore both prospective—demonstrating reasonable system controls where required—and retrospective—allowing authorities to reconstruct an event.

The case demonstrates why sectoral overlap is often necessary. Product safety may address the maker's design and support, cybersecurity regulation the resilience of critical services, privacy law the unauthorized exposure of personal data, and civil liability the consequences suffered by affected persons. Their separate legal objectives require a coordinated institutional response without merging distinct tests into a fictional general duty of technological responsibility.

Research findings and institutional implications

The comparative analysis supports several connected findings. The first is that legal systems govern technologies primarily through relationships and conduct. A classifier or model becomes legally significant because an actor deploys it within a field where specified laws apply. This explains why the same general-purpose tool may carry provider documentation obligations, downstream administrative restrictions, sectoral discrimination duties, and private-law exposure without a single rule supplying the entire answer.

The second finding is that legal diversity lies in allocation of governance functions. European Union law has increasingly articulated horizontal and cross-cutting duties, while retaining data, platform, competition, product, and cybersecurity layers. American regulation distributes many controls across federal agencies, states, sectors, and courts. The United Kingdom and Canada illustrate important administrative and regulator-led techniques; China and South Korea demonstrate distinctive nationally administered statutory approaches. These descriptions are comparisons of institutional design, not unverified judgments about relative enforcement success.

The third finding is that enactment, commencement, and enforcement must be treated as separate temporal properties. The 2026 amendment to the EU AI Act's timetable is a vivid example: a statute can establish future obligations whose practical preparation is urgent while their full application lies more than a year ahead. Similarly, legislative proposals, administrative guidelines, and international recommendations have significance without necessarily imposing immediately enforceable duties. Reliable research records these statuses on the same date as the analysis.

The fourth finding concerns evidentiary continuity. Technical testing at development, deployment records at operational use, administrative reasons for particular decisions, regulatory inspection materials, and litigation evidence often reside with different parties. Rights and remedies depend on legally authorized access to the documents needed to establish a violation. Across regimes, a persistent accountability risk appears whenever the actor with legal responsibility cannot obtain necessary technical evidence or the injured person cannot reconstruct a legally decisive decision.

The fifth finding is that institutional independence and competence affect the meaningfulness of oversight. A rule requiring an audit can be weak if the auditor lacks the relevant access, if the authority cannot order a remedy, or if the audit excludes the actual system version used. Conversely, familiar legal institutions can provide significant protection through enforceable duties of fair dealing, equal treatment, public-law legality, and product safety. Evaluating governance therefore requires examination of powers, procedures, evidence, and outcomes rather than the length or stated ambition of a national AI strategy.

These findings suggest a disciplined institutional practice. Public authorities should record statutory competence before procurement, preserve legally relevant testing and version histories, ensure contractual access necessary for review, and distinguish policy standards from binding obligations. Firms should map duties to provider, deployer, intermediary, controller, manufacturer, and other applicable roles, maintain date-sensitive legal inventories, and ensure that complaints and incident reports reach actors empowered to correct the relevant conduct. Courts and regulators should ask for evidence targeted to the legally material elements of the case.

Each recommendation follows from comparative doctrinal reasoning and remains subject to sectoral legal requirements and institutional resources. It is not presented as a newly invented governance theory or as empirical proof of regulatory effectiveness. Its purpose is to connect established legal controls across the lifecycle of technical activity in a manner that can be examined, contested, and corrected.

Conclusion

Comparative technology governance is the study of how legal orders convert technological conduct into enforceable relationships among developers, deploying organizations, markets, public authorities, and affected persons. Its legal content lies in the specific combination of permission, restriction, documentation, supervision, adjudication, and remedy established by the applicable legal system. General claims about whether a jurisdiction favors innovation or control explain little unless grounded in the binding sources and the institutions that implement them.

As of November 21, 2026, the European Union maintains a complex set of overlapping digital and AI instruments with materially important staged obligations under the amended AI Act. United States governance relies substantially on agencies, states, sectors, and established causes of action. The United Kingdom combines statutory and administrative techniques with judicial review, China regulates designated generative service provision through national administrative measures, South Korea operates a comprehensive AI framework, and Canada demonstrates policy-driven safeguards for covered federal automated decisions. International instruments contribute convergence and shared vocabulary according to their distinct legal force.

The decisive substantive conclusion is that responsibility must remain traceable from the legally regulated activity to its institutional decision-maker and available remedy. Effective governance requires evidence capable of supporting this traceability over system versions, contractual divisions, national borders, and transitions between development and deployment. Comparative law makes visible the points at which established duties meet and the points at which accountability is weakened by missing information or fragmented institutional competence.

The research contributes a source-specific synthesis of governance functions across legal regimes, with explicit attention to amendments, applicability dates, and the legal status of recommendations and proposals. The study establishes an analytical basis for future case-law and empirical investigation, including scrutiny of implementation by specific national regulators and the practical accessibility of remedies. Its conclusions remain grounded in the examined instruments and cases. Any expansion to additional jurisdictions, later amendments, or measured regulatory outcomes calls for renewed legal verification rather than extrapolation from the present record.

References