Introduction: One Technology, Many Legal Answers
Artificial intelligence (AI) is increasingly deployed across borders, but the legal answers to AI-related harm are far from uniform. The European Union, the United States, and key Asian jurisdictions (such as China, Japan, South Korea and Singapore) are developing distinct liability models for AI systems. These models differ not only in their regulatory techniques, but also in their underlying legal culture and policy priorities: fundamental rights vs. innovation, ex ante regulation vs. ex post liability, and public law vs. private ordering.
This article offers a comparative overview of AI liability models in the EU, US and Asia, with a focus on how each region approaches civil and (to a lesser extent) criminal responsibility for AI-induced harm. The aim is not to provide an exhaustive survey, but to highlight structural differences that will shape the future of AI governance and cross-border compliance.
1. Conceptual Starting Points: Who (or What) Can Be Liable?
Across all three regions, one common thread remains: AI systems themselves are not treated as legal or criminal persons. Responsibility is attached to human and corporate actors who design, deploy, operate or benefit from AI systems.
However, the starting questions differ:
- The EU asks: Which actors along the AI value chain should bear ex ante duties and ex post liability to protect fundamental rights and safety?
- The US asks: How can existing tort, contract and sectoral laws handle AI, and when is new law really necessary?
- Asian jurisdictions often ask: How can we harness AI for economic growth while maintaining social stability and state control?
These different starting points lead to different liability architectures, even if many doctrinal tools (negligence, product liability, corporate liability) look superficially similar.
2. The EU Model: Risk-Based Regulation + Structured Liability
2.1. Ex Ante: The AI Act and Risk-Based Duties
The EU’s approach is anchored in the EU Artificial Intelligence Act (AI Act), a horizontal regulation that introduces a risk-based regulatory framework for AI systems. High-risk systems (e.g. in healthcare, critical infrastructure, employment, credit, policing, justice) are subject to extensive obligations regarding:
- Risk management,
- Data quality and governance,
- Technical documentation,
- Logging and traceability,
- Transparency,
- Human oversight.
These ex ante duties are not themselves “liability rules” in the traditional sense, but they strongly influence how negligence or fault will be interpreted when harm occurs.
2.2. Ex Post: Civil and Corporate Liability
In parallel, the EU is moving towards modernized civil liability for AI, including:
- Adjustments to product liability rules to cover AI-enabled products and software,
- Provisions easing the burden of proof for victims where AI opacity makes causation hard to establish,
- More robust corporate liability regimes for systematically unsafe AI deployment.
In many Member States, criminal liability can also attach to human or corporate actors where gross negligence or deliberate misconduct in AI development or deployment leads to serious harm (e.g. in product safety, workplace safety or data protection contexts). The EU model is thus characterized by a tight coupling between:
- A public law layer of detailed pre-market and post-market obligations, and
- A private law layer of fault-based and strict liability mechanisms.
3. The US Model: Incremental, Sectoral and Litigation-Driven
3.1. No Single Federal AI Code
In contrast, the United States does not have a comprehensive federal AI statute comparable to the EU AI Act. Instead, AI governance is:
- Fragmented across sectors (healthcare, finance, transportation, employment),
- Guided by a mixture of agency guidance, soft law, and existing statutes,
- Strongly shaped by litigation and common-law reasoning in tort and contract.
AI is, by default, treated as another technology or product: if it causes harm, courts ask whether existing doctrines on negligence, product defect, professional malpractice or cybersecurity can address it.
3.2. Tort Law and Product Liability as Primary Tools
US liability for AI-induced harm typically runs through:
- Negligence law – Did the developer, deployer or professional user exercise reasonable care?
- Product liability – Is the AI-enabled product unreasonably dangerous or defective in design, warning, or manufacture?
- Contract and warranty – What did the parties allocate by agreement?
Rather than imposing ex ante, risk-based duties at the system level, the US model relies more on ex post adjudication: courts examine facts case-by-case, and doctrines evolve through precedent. Criminal liability may attach in egregious situations (e.g. reckless disregard for safety, fraud), but there is no distinct AI criminal code. The result is a model that is flexible and innovation-friendly, but usually less predictable ex ante for global providers looking for uniform compliance strategies.
4. Asian Approaches: Growth-Oriented, State-Led, and Divergent
“Asia” is far from uniform; nonetheless, some common threads and contrasts can be sketched for China, Japan, South Korea and Singapore, which are often cited in AI governance debates.
4.1. China: Strong State Control and Platform Responsibility
China’s emerging AI governance regime combines:
- Top-down regulation of recommendation algorithms, deep synthesis (deepfakes), and generative AI services,
- Platform and provider responsibilities for content moderation, user identification and security,
- A strong emphasis on social stability, ideological control and data governance.
Liability is framed not only in terms of private-law compensation, but also in terms of administrative and quasi-criminal sanctions for providers and platforms that fail to meet state-imposed duties. AI-related harm is often treated through the lens of online platform governance and cybersecurity, rather than classic tort alone.
4.2. Japan and South Korea: Soft-Law Guidance and Incremental Reform
Japan and South Korea have generally preferred:
- Soft-law frameworks (ethics guidelines, best practices),
- Gradual adaptation of existing laws (consumer protection, data protection, product safety),
- Promoting AI innovation while recognizing the need for accountability and transparency.
Liability for AI-induced harm typically passes through existing civil and administrative law, with increasing discussion about how to adjust doctrines of negligence, product liability and corporate responsibility to reflect AI-specific risks.
4.3. Singapore: Experimental and Regulatory-Sandbox Approach
Singapore has positioned itself as a testbed for trustworthy AI, using:
- Non-binding frameworks such as its Model AI Governance Framework,
- Regulatory sandboxes in finance and other sectors,
- A strong focus on responsible AI as a competitive advantage.
Liability is still largely anchored in traditional doctrines, but there is an explicit policy interest in clarifying accountability chains for AI systems, especially in financial services and critical infrastructure.
Overall, leading Asian jurisdictions tend to blend innovation promotion with targeted state control, with liability developing through a mix of administrative enforcement and private law, rather than publicly codified, AI-specific civil liability regimes.
5. Criminal Liability: A Narrow but Symbolic Space
In all three regions, criminal liability for AI-related harm is relatively rare and usually indirect:
- The EU entertains corporate and individual criminal liability where gross negligence or intentional misconduct in AI deployment violates safety, data protection or anti-discrimination norms.
- The US may prosecute actors for fraud, cybercrime or willful safety violations where AI is used as a tool, but does not treat AI as a special source of criminal liability.
- Asian jurisdictions use criminal or quasi-criminal sanctions mainly to enforce content, cybersecurity and public-order rules, with AI treated as a medium or instrument.
In none of these regions is there serious movement towards recognizing AI systems themselves as criminal actors. The debates on “electronic personhood” remain largely academic. Responsibility is consistently traced back to human or corporate decision-makers.
6. Convergence and Divergence: What Global Providers Must Watch
For global AI providers and large deployers, the comparative picture can be summarized as follows:
- Convergence
- AI is treated as a tool; liability attaches to humans and corporations.
- Negligence and product liability doctrines remain central.
- There is growing attention to data quality, transparency and human oversight.
- Divergence
- The EU relies on strong ex ante regulation (AI Act) and structured civil liability reforms, emphasizing fundamental rights.
- The US relies on ex post litigation and sectoral rules, emphasizing innovation and private ordering.
- Asian jurisdictions mix state-led governance, platform responsibility and incremental legal adaptation, with particular focus on social stability and industrial policy.
This means that AI providers must develop multi-layered compliance strategies: EU-style regulatory compliance, US-style litigation and risk management, and Asia-specific attention to platform rules, cybersecurity and state priorities.
Conclusion: Towards a Pluralistic Map of AI Liability
There is no single, global model for AI liability. Instead, we see a pluralistic landscape:
- The EU’s risk-based, rights-oriented and regulation-heavy model,
- The US’s litigation-driven, sectoral and innovation-focused model,
- A diverse set of Asian approaches that blend growth strategies with varying degrees of state control and platform accountability.
For scholars and practitioners concerned with AI criminal and civil liability, this diversity presents both challenges and opportunities. On the one hand, it complicates cross-border compliance and raises the risk of forum shopping and regulatory arbitrage. On the other hand, it creates a laboratory of legal ideas, where different models can be compared, refined and, over time, partially harmonized.
In the near future, the key task will not be to impose a single liability model worldwide, but to ensure that different regional regimes remain interoperable enough to protect individuals from AI-induced harm, while still allowing innovation to cross borders. AI may be global, but liability will remain, for the foreseeable future, profoundly local.
Yanıt yok