Abstract: This article compares the European Union Artificial Intelligence Act (EU AI Act) and California’s Transparency in Frontier Artificial Intelligence Act (TFAIA), arguing that the EU framework is more effective in regulating the risks posed by current AI systems. Assessed against four criteria (risk coverage, fundamental rights protection, responsible AI innovation, and clear compliance structures), the EU AI Act’s comprehensive, deployment-focused approach proves to be the stronger framework compared to the TFAIA’s narrower regulation of frontier models. Moreover, these frameworks reflect two fundamentally different theories of what AI harms are and when the law should intervene to address them: the EU’s primary regulatory concern is fundamental rights harms arising from routine AI deployment, while California’s framework operates only where risks reach catastrophic scale. This divergence suggests that AI governance may develop globally either as a tool primarily oriented toward protecting individuals or as one oriented toward enabling technological growth.
Introduction
Artificial Intelligence (AI) is rapidly being integrated into decision-making systems across public, private, and commercial sectors. This expansion has intensified longstanding concerns about the risks that these systems pose, including algorithmic discrimination, opacity and lack of accountability, and intrusive data practices. As these risks become more prominent, different regulatory responses have been adopted across jurisdictions, particularly regarding which AI risks should be prioritised, when regulatory intervention should occur, and how responsibility for preventing harm should be allocated. Among the most influential contemporary approaches are the European Union Artificial Intelligence Act (EU AI Act), which came into force in August 2024, and California’s Transparency in Frontier Artificial Intelligence Act (TFAIA), which is effective from January 2026. With the EU AI Act representing the first comprehensive horizontal AI regulatory framework in the world and in the absence of federal AI legislation in the United States—as well as the fact that the headquarters of the majority of the top 50 leading AI companies including OpenAI and Meta are located in California—the regulatory approaches of these two jurisdictions have emerged as the most significant legislative models currently influencing global debates on AI governance.1 2
Despite their prominence, the EU and Californian frameworks have fundamentally different priorities and regulatory philosophies. The EU AI Act focuses on the protection of consumer welfare and human rights across a wide range of public and private sectors, using a risk-tiered system to categorise risk levels. By contrast, the TFAIA replaces this deployment-focused risk hierarchy with a capability threshold, regulating the development of frontier AI systems, while seeking to balance safety considerations with rapid technological innovation. A further key distinction is that the EU model adopts a pre-deployment compliance focus, while the TFAIA targets the development of AI models and post-deployment accountability.3 These two distinct philosophies reflect deeper theoretical debates about the function of legal intervention in the governance of emerging AI technologies.4
This article contends that the EU AI Act is more effective than California’s TFAIA in regulating the risks posed by current AI systems, when assessed against four foundational criteria: risk coverage, fundamental rights protection, responsible AI innovation, and clear compliance structures. It is argued that the EU’s comprehensive ex-ante obligations, combined with ongoing post-market monitoring, are more effective because they cover a substantially wider range of AI systems and directly target the real-world harms that arise when AI is deployed in critical situations. While the TFAIA offers an important contribution to AI regulation by focusing narrowly on frontier model development, which enables it to successfully manage the low probability, catastrophic AI risks, it leaves everyday rights-based harms—such as discriminatory hiring systems and biased decision-making tools—largely unregulated. As a result, the EU AI Act offers a more substantively reliable legislative model for mitigating AI risks at the current stage of technological development.
Although the TFAIA’s purpose is to specifically regulate the development of frontier AI models, this comparison is justified because the TFAIA currently operates as the primary AI governance framework in California in the absence of federal AI legislation. It must therefore be evaluated against the full range of harms that AI systems currently produce, not only those it was intended to address.
In this article, “effective” regulation refers to legal frameworks that: cover a broad spectrum of AI-related risks, protect individuals’ fundamental rights, support innovation and technological advancement responsibly, and establish clear enforcement mechanisms. These criteria are consistent with the foundational principles identified in international standards such as the OECD Recommendation on Artificial Intelligence.5 The article will demonstrate why each criterion constitutes a necessary condition of effective AI regulation before applying it to each framework.
A. Scope and Coverage of AI Risks
The effectiveness of AI regulation depends not only on how severe the risks addressed are, but also on whether the framework regulates the types of harms that are most prevalent in current AI deployment. Therefore, one must consider the scope of each regulatory approach when evaluating its effectiveness. A model that fails to regulate routine harms posed by AI systems cannot be said to fulfil one of the basic functions of the law, which is to protect individuals from discriminatory outcomes in sectors such as employment, healthcare and criminal justice.6
When examining the scope of risks that each framework governs, it is evident that the EU approach adopts a far more expansive regulatory model than California’s TFAIA. Firstly, the EU AI Act’s scope is extensive because it deliberately adopts broad foundational definitions that enable the legislation to reach a wide range of AI systems across public and private sectors. An AI system is defined in Article 3(1) as “a machine-based system designed to operate with varying levels of autonomy,” which generates outputs—such as predictions or decisions—from input data in ways that influence physical or virtual environments.7 Risk is likewise framed broadly to apply to different probabilities and degrees of severity of harm.8 These definitions reinforce the Act’s sector-based regulatory approach that applies not only to providers, but also to deployers, importers, distributors and product manufacturers.9 As a result, regulatory obligations are determined by the contexts in which AI systems are used, rather than by a model’s technical capability alone.
Secondly, the EU AI Act’s expansive scope is illustrated through its risk-tiered classification system, which categorises AI systems into unacceptable risk, high-risk, limited risk, and minimal risk. Unacceptable-risk systems—such as social scoring, biometric categorisation, and certain forms of biometric identification—are prohibited due to their inherent incompatibility with fundamental rights. High-risk systems encompass AI deployed in contexts where errors or algorithmic bias can materially affect individuals in certain sectors, including employment, education, credit scoring, healthcare, welfare administration, migration control, and law enforcement. These systems are subject to extensive obligations, including conformity assessments, risk-management procedures, human oversight mechanisms and post-market monitoring.10 By contrast, limited-risk systems face certain transparency obligations aimed at preventing deception, while minimal-risk models remain largely unregulated to avoid unnecessary regulatory interference with low-impact AI uses. This structure allows regulatory intervention to be adjusted to the severity of risk, ensuring that the EU framework captures both routine risks and more exceptional forms of harm arising from AI systems.
By contrast, the scope of the TFAIA is defined narrowly through a number of statutory requirements that significantly limit the range of AI systems that are subject to regulation. Firstly, the TFAIA adopts a technology-based approach of regulating only “frontier models,” a special category of the most powerful AI systems, where the foundational model is trained with at least 10^26 floating-point operations per second of computing power.11 Only the developers of such models—classified as ‘frontier developers’—particularly ‘large frontier developers’ that have more than $500 million in annual gross revenues, are directly regulated by the TFAIA.12 Even within this narrow category, regulatory obligations arise only when a system poses a catastrophic risk, classified in the TFAIA as a “foreseeable and material risk” that may cause over $1 billion in damages or “materially contribute” to causing the death or serious injury of more than fifty people.13 These definitions illustrate that the TFAIA is designed to target a very small proportion of AI systems, specifically the models that are capable of causing extreme, low-probability harms, such as AI-enabled development of CBRN weapons and situations resulting in large-scale human casualties. Consequently, this legislative framework is unlikely in practice to substantially impact the operations of the vast majority of AI applications, since the most common, low-risk social and economic harms remain outside its regulatory scope. From a regulatory perspective, this model’s scope is difficult to justify because the decision to focus only on frontier models illustrates the TFAIA’s prioritisation of technological development over the protection of individuals from harms that are already occurring.
Legal scholars have criticised that focusing on frontier AI systems is harmful, as it takes attention away from the uses that are most prominent in the real-world context at the moment.14 This is a compelling argument when one examines the scope of the EU AI Act in comparison to the TFAIA because the fundamental issue with the scope of the Californian legislation is that it appears misaligned with the current technological reality of AI harms. The harms that are not classified as frontier models under the TFAIA can still pose considerable risks. One pertinent example of this is that AI systems which are classified as having unacceptable risks under the EU AI Act, such as biometric identification in public spaces, biometric categorisation of individuals and social scoring systems, are not directly covered by the TFAIA regulations, despite being banned with very limited exceptions in the EU jurisdiction.15
Overall, what makes the EU framework more effective than the TFAIA is that it is both horizontally and vertically comprehensive. Horizontally, it applies across numerous sectors and different stages of model development and use. Vertically, the EU addresses harms of varying severity, from unacceptable risks to minimal transparency risks. Together, these elements reinforce the Act’s ability to directly regulate the systems that are most prevalent in the current technological field, rather than focusing narrowly on a small subset of frontier models.
B. Protection of Fundamental Rights
The protection of fundamental rights is a central criterion for assessing the effectiveness of AI regulation because artificial intelligence increasingly influences decisions that directly affect crucial elements of public life, such as employment, welfare administration, healthcare and policing. In these sectors, algorithmic bias can result in systemic discrimination, as well as violations of fairness and privacy rights.16 There have been numerous instances of discrimination in recent years, such as AI recruitment tools discriminating against women in the employment sector and predictive policing algorithms disproportionately targeting Black communities in certain US jurisdictions.17 18 These examples highlight that the issue of rights-based harms is currently highly prevalent in numerous social and economic contexts, making it crucial for current AI regulatory frameworks to directly respond to these everyday rights violations, alongside addressing catastrophic AI risks. It follows that the regulatory effectiveness of an AI system cannot be determined without assessing whether the legal framework successfully restricts AI models to ensure that they are compatible with fundamental rights—including the rights to non-discrimination and protection of personal data under the Charter of Fundamental Rights of the EU.19
Firstly, a central feature of the EU AI Act’s rights-focused approach is the introduction of a mandatory Fundamental Rights Impact Assessment (FRIA) for deployers of high-risk AI systems. The FRIA requires deployers to identify, assess, and mitigate foreseeable risks to fundamental rights prior to deployment, and to document the measures adopted to address those risks.20 This ensures that high-risk AI systems are conforming with EU human rights and data protection standards, but more importantly, this obligation extends beyond any previous EU legislative attempt to protect fundamental rights.21 As Mantelero has argued, the FRIA is significantly stronger than the Data Protection Impact Assessment (DPIA) under the EU’s GDPR: while the DPIA is confined to privacy and personal data protection, the FRIA adopts a more comprehensive approach by including a broader range of rights, such as non-discrimination, equality and freedom of expression.22
Nevertheless, the EU AI Act’s framework has faced intense scrutiny by various international non-governmental organizations, such as Amnesty International. It may be argued that the EU’s legislative framework does not go far enough to protect fundamental rights, by allowing exemptions from the FRIA for law enforcement and migration authorities—which is particularly significant in sectors like biometric surveillance and emotion recognition.23 While deployers of high-risk systems are generally required to publish summaries of FRIA outcomes, this requirement does not apply in the areas where upholding the rule of law and human rights principles seems to be the most crucial. Moreover, further exemptions justified on the grounds of ‘exceptional reasons of national security’ allow authorities to withhold even basic information about the use of high-risk AI systems, severely limiting the transparency and public accountability efforts of the EU AI Act.24 These exemptions are a clear weakness of the EU AI Act because a regulatory framework that treats fundamental rights as a central legislative aim should ensure that they are protected in precisely the sectors where the consequences of algorithmic bias are the most severe. In the contexts where the ground of ‘exceptional reasons of national security’ is available, such as law enforcement and border control, the consequences of AI discrimination are the most severe specifically because the affected individuals are the ones that are the least able to challenge the authorities and seek redress.25
Despite these limitations, the EU AI Act constitutes a substantially stronger framework than the TFAIA on the grounds of fundamental rights protections. The Californian framework does not require deployers to conduct any equivalent human-rights-based assessment, nor does it prohibit AI practices that are incompatible with universal fundamental rights, particularly biometric categorisation or social scoring.26 Instead, the TFAIA’s primary mechanism to protect individual and social interests is a requirement on frontier developers to publish a transparency report, while large frontier developers are further obliged to include catastrophic risk assessments.27 These two instruments that the TFAIA adopts are not focused on protecting fundamental rights, since the transparency report discloses information about a model’s capabilities and limitations, while the catastrophic risk assessment addresses very significant harms rather than individual rights violations. This approach stands in direct contrast to the FRIA, which directly requires deployers to identify and mitigate foreseeable risks to specific fundamental rights, including non-discrimination, equality and freedom of expression.28 While the TFAIA’s approach reinforces the principle of accountability to the public for frontier developers, it does not adequately address the rights-based harms arising from routine AI deployment. A framework that leaves algorithmic discrimination in sectors like employment, welfare and policing unaddressed in an effort to advance innovation is fundamentally incompatible with a key principle underlying fundamental rights: that individuals are entitled to legal protection from state and private power. Therefore, when evaluated against the criterion of fundamental rights protection, the EU AI Act emerges as the more effective framework out of the two, due to its adoption of a binding framework of obligations to ensure high-risk AI systems conform with EU fundamental rights standards.
C. Responsible AI Innovation
While the EU’s risk-based framework is fundamental in minimising AI harms across the different risk levels, legal scholars have emphasised that the compliance requirements may disproportionately burden certain firms, especially start-ups and small and medium-sized enterprises (SMEs) which will have to provide a larger share of their revenue for additional compliance costs.29 30 This argument is especially compelling for AI models in the high-risk category, where the obligations for deployers are extensive—requiring the implementation of risk management and human oversight systems, as well as post-market monitoring and continuous compliance testing.31 These obligations inevitably increase operational costs, as firms must allocate additional time, resources, and hire legal experts to ensure regulatory compliance. The European Commission’s ‘Study to Support an Impact Assessment of Regulatory Requirements for Artificial Intelligence in Europe’ supports these concerns, estimating that the total average compliance cost for a single AI unit will be €29,277 annually.32 The issue becomes even more consequential for start-up firms, which may face average costs of up to €23,000 to obtain certification for an AI system, accounting for up to 10-14 percent of total development costs.33 This may lead to stifling nthe EU, particularly as SMEs and start-up firms may become motivated to move their operations to alternative jurisdictions with more lenient regulatory frameworks, such as the United States. At the aggregate market level, the European Commission projects that in 2025, total global AI compliance costs range from €1.6 billion to €3.3 billion for systems categorised as high-risk under the EU AI Act.34 These costs may be especially concerning for the healthcare industry, where many AI applications are classified as high-risk, potentially slowing the deployment of socially valuable innovations.35 Moreover, the uneven distribution of compliance burdens risks reinforcing market concentration, as large, well-resourced firms are better positioned to deal with regulatory costs than smaller competitors with limited budgets—potentially stifling competition and long-term innovation in certain industries.36
By contrast, the Californian approach highly prioritises boosting innovation in the AI sector over any other regulatory concerns. This is illustrated by Governor Newsom’s rejection of the earlier version of the TFAIA, bill SB1047, on the grounds that it was too restrictive on AI developers with models valued at over 100 million dollars.37 The TFAIA’s structure is deliberately designed to avoid imposing regulations on smaller companies and startups, as it applies only to large frontier models and companies exceeding $500 million in annual gross revenue. In addition, the California Department of Technology is authorised to update statutory definitions as technological capabilities develop, allowing the framework to adapt without imposing rigid compliance structures.38 Thus, this regulatory framework ultimately minimises barriers to entry and preserves California’s competitive advantage in AI development.
However, the divergence between the EU AI Act and the TFAIA reflects a deeper regulatory disagreement over what responsible AI innovation ought to look like. International governance frameworks, most notably the OECD Recommendation on Artificial Intelligence, emphasise that AI progress must be pursued only within enforceable safeguards that protect fundamental principles such as human rights, democratic values, fairness and privacy.39 From this perspective, AI innovation that is permitted to develop without legal limitations risks enabling AI systems to increase discrimination against individuals in crucial public sectors such as employment and healthcare.40
The EU AI Act aligns closely with this regulatory philosophy by embedding ethical and rights-based considerations directly into legally enforceable obligations, particularly for high-risk systems. Innovation and regulation are thus treated as complementary rather than antagonistic. However, it must be acknowledged that the EU AI Act’s compliance obligations do increase operational costs and may in practice slow the development of certain socially valuable innovations, particularly in the healthcare industry. The question therefore is whether the costs that this regulation imposes are justified by the social benefits it creates. Regulatory limitations are justified, since compliance costs are a foreseeable consequence of operating in sectors where algorithmic errors directly affect individuals’ fundamental rights, and cannot be treated as equivalent in weight to those harms.
The TFAIA by contrast adopts an approach that aims to reinforce California’s commercial dominance in frontier AI development, without imposing any limitations on the grounds of ethical or public interests.41 As a result, although California’s framework may better enable rapid technological development in the short term, it fails to adopt the concept of socially sustainable long-term innovation. Therefore, when assessed against the criterion of responsible AI innovation, the EU model, despite its greater demands, is ultimately more effective in integrating innovation with enforceable social interest protections. This article argues that this is the stronger approach to responsible AI innovation because it weighs the social benefits of protecting individuals from harm caused by AI systems above the prospect of incrementally slowing innovation.
D. Clear Enforcement Mechanism
The fourth criterion for assessing the effectiveness of the EU AI Act and TFAIA is the implementation of a clear enforcement mechanism. Regulatory obligations are only valuable to the extent that they can be successfully monitored, enforced, and sanctioned. This is particularly significant in the context of AI regulation, where the technical complexity of the systems involved and the significant commercial interests at stake create strong incentives for non-compliance. Therefore, this section evaluates which framework provides a stronger compliance structure in practice.
As an EU regulation, the EU AI Act applies directly across all Member States following the end of its transitional period in 2027.42 Penalties for violations of the EU AI Act range from 7.5 million (or 1% of the firm’s total annual turnover, where applicable) to 35 million (or 7% of total annual turnover), with the highest tier reserved for violations involving systems in the unacceptable risk category.43 Crucially, these sanctions form part of a wider enforcement structure that involves continuous post-market compliance obligations and institutional oversight by the European Artificial Intelligence Board and the AI office within the European Commission.44 Together, these mechanisms provide information and advice to national authorities, enhancing the Act’s practical enforceability across Member States.
However, there are certain challenges to the EU enforcement model. Legal scholars have argued that due to difficulties in coordination between EU regulatory bodies and national authorities, there may be enforcement inconsistencies when it comes to complex, demanding frameworks such as the EU AI Act.45 For example, in certain cases, it may be difficult to establish which category in the risk-based system a particular model falls under. A study by the appliedAI Institute for Europe found that in approximately 40% of cases, it could not be determined whether an AI system fell within the high-risk category.46 Such ambiguity not only increases administrative burdens for both AI providers and regulatory agencies, but also raises the issue of potential strategic misclassification by firms seeking to avoid compliance obligations.
Nonetheless, these implementation challenges do not substantially affect the strength of the EU’s enforcement structure in comparison to the Californian framework. Under the TFAIA, the California Attorney General may impose civil penalties of up to $1 million per violation.47 These sanctions are significantly lower than those available under the EU AI Act, particularly when considered in light of the fact that the TFAIA applies only to firms with annual gross revenues of at least $500 million. As a result, such financial penalties are highly unlikely to operate as a meaningful deterrent for large frontier AI developers. Moreover, the TFAIA does not establish a dedicated supervisory body, mandate systematic post-market monitoring, or require independent audits.48 Instead, the framework relies heavily on self-reporting, leaving regulators with limited ways to identify, investigate and prevent non-compliance with the TFAIA.49
Overall, despite certain implementation difficulties, the EU’s multi-level enforceability approach reduces reliance on voluntary industry self-regulation, setting this regulatory framework apart from comparatively lighter enforcement models like the TFAIA. Therefore, when assessed against the criterion of clear enforcement mechanisms, the EU AI Act provides a substantially more effective regulatory model than the TFAIA.
Conclusion
This article has demonstrated that the EU AI Act and California’s TFAIA do not merely adopt different regulatory philosophies, but reflect fundamentally incompatible theories of what AI harm is and when the law should intervene to address it. The TFAIA is premised on the theory that the law should act only when the AI risk may reach a catastrophic scale and when powerful frontier models are involved, particularly the largest technology firms such as Google and OpenAI.50 This is a significant contribution in the context of the United States, where federal AI regulation remains absent. However, its limited scope, weak enforcement mechanisms, and lack of fundamental rights protections restrict its effectiveness when it is evaluated in comparison to a more comprehensive legislative framework like the EU AI Act.51 Ultimately, this comparison illustrates that regulatory effectiveness in the context of AI systems largely depends on limiting the pervasive, real-world risks currently influencing the fundamental rights and interests of individuals and the wider society. The EU AI Act therefore presently constitutes the more effective regulatory model for governing AI because the rights-based harms that individuals currently experience from AI deployment demand a regulatory response that matches their scope and severity. Underlying this is the EU AI Act’s theory of harm, that the law must intervene before harm occurs, wherever AI systems are capable of affecting individuals’ fundamental rights.
This distinction is significant beyond the article’s comparison of the two regulatory models. As jurisdictions globally begin designing domestic AI regulations, the choice between adopting the EU or the Californian regulatory philosophy will determine whether AI governance develops primarily to protect the rights of individuals or to encourage technological growth. This article contends that protecting individuals from foreseeable harm is the stronger foundation for AI governance, and that the EU AI Act presently constitutes the more compelling model for how the law should respond to the challenges posed by artificial intelligence.
- European Parliament, ‘EU AI Act: first regulation on artificial intelligence’ (European Parliament, 8 June 2023)<https://www.europarl.europa.eu/topics/en/article/20230601STO93804/eu-ai-act-first-regulation-on-artificial-intelligence> accessed 12 January 2026 ↩︎
- Governor of California, ‘Governor Newsom signs SB 53, advancing California’s world-leading artificial intelligence industry’ (Gov.ca.gov, 29 September 2025) <https://www.gov.ca.gov/2025/09/29/governor-newsom-signs-sb-53-advancing-californias-world-leading-artificial-intelligence-industry/>acc accessed 12 January 2026 ↩︎
- Connor Attridge, ‘California’s Frontier AI Act: National Security Implications for the UK’ (CETaS Expert Analysis, 14 October 2025) <https://cetas.turing.ac.uk/publications/californias-frontier-ai-act-national-security-implications-for-the-uk> accessed 16 January 2026 ↩︎
- Michael Veale, Kira Matus and Robert Gorwa, ‘AI and Global Governance: Modalities, Rationales, Tensions’ (2023) 19 Annual Review of Law and Social Science 255–275 <https://doi.org/10.1146/annurev-lawsocsci-020223-040749> accessed 12 January 2026 ↩︎
- OECD, Recommendation of the Council on Artificial Intelligence (OECD/LEGAL/0449, adopted 22 May 2019) <https://legalinstruments.oecd.org/en/instruments/OECD-LEGAL-0449> accessed 12 January 2026 ↩︎
- Rashida Richardson, Jason M Schultz and Kate Crawford, ‘Dirty Data, Bad Predictions: How Civil Rights Violations Impact Police Data, Predictive Policing Systems, and Justice’ (2019) 94 New York University Law Review Online 15 ↩︎
- Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act) Art 3(1) ↩︎
- Ibid Art 3(2) ↩︎
- Ibid Art 2(1) ↩︎
- ‘High-level summary of the EU AI Act’ (ArtificialIntelligenceAct.eu, 27 February 2024) <https://artificialintelligenceact.eu/high-level-summary/> accessed 18 January 2026 ↩︎
- Senate Bill 53 (Transparency in Frontier Artificial Intelligence Act) (Cal 2025) s 2(i)(1) ↩︎
- Ibid s2(h), (j) ↩︎
- Ibid s2(c) ↩︎
- Gina Helfrich, ‘The harms of terminology: why we should reject so-called “frontier AI”’ (2024) 4 AI and Ethics 699 ↩︎
- European Parliament, ‘EU AI Act: first regulation on artificial intelligence’ (European Parliament Topics, 2023) <https://www.europarl.europa.eu/topics/en/article/20230601STO93804/eu-ai-act-first-regulation-on-artificial-intelligence> accessed 12 January 2026 ↩︎
- European Network of National Human Rights Institutions, ‘Key Human Rights Challenges Related to Artificial Intelligence’ (ENNHRI) <https://ennhri.org/ai-resource/key-human-rights-challenges/> accessed 17 January 2026 ↩︎
- Jeffrey Dastin, ‘Amazon Scraps Secret AI Recruiting Tool That Showed Bias against Women’ Reuters (10 October 2018) <https://www.reuters.com/article/us-amazon-com-jobs-automation-insight-idUSKCN1MK08G> accessed 27 March 2026 ↩︎
- Rashida Richardson, Jason M Schultz and Kate Crawford, ‘Dirty Data, Bad Predictions: How Civil Rights Violations Impact Police Data, Predictive Policing Systems, and Justice’ (2019) 94 NYU L Rev Online 15 ↩︎
- Charter of Fundamental Rights of the European Union [2012] OJ C326/391 (‘Charter’) ↩︎
- Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act) Art 27 ↩︎
- Athina Thomaidou and Konstantinos Limniotis, ‘Navigating Through Human Rights in AI: Exploring the Interplay Between GDPR and Fundamental Rights Impact Assessment’ (2025) 5(1) J Cybersecurity and Privacy 7 ↩︎
- Alessandro Mantelero, ‘The Fundamental Rights Impact Assessment (FRIA) in the AI Act: Roots, Legal Obligations and Key Elements for a Model Template’ (2024) 54 Comput L & Sec Rev 106020 ↩︎
- Amnesty International EU Office, ‘EU’s AI Act Fails to Set Gold Standard for Human Rights’ (Amnesty, 3 April 2024) <https://www.amnesty.eu/wp-content/uploads/2024/04/EUs-AI-Act-fails-to-set-gold-standard-for-human-rights.pdf> accessed 13 January 2026> ↩︎
- Ibid ↩︎
- Forster C, Refugee Protection in the Artificial Intelligence Era: A Test Case for Rights (Chatham House, September 2022) <https://www.chathamhouse.org/2022/09/refugee-protection-artificial-intelligence-era> accessed 27 March 2026 ↩︎
- Senate Bill 53 (Transparency in Frontier Artificial Intelligence Act) (Cal 2025) ↩︎
- ‘California Enacts Landmark AI Safety Law, but with Very Narrow Applicability’ (Cleary CyberWatch and Privacy Watch, 13 October 2025) <https://www.clearycyberwatch.com/2025/10/california-enacts-landmark-ai-safety-law-but-with-very-narrow-applicability/> accessed 18 January 2026 ↩︎
- Athina Thomaidou and Konstantinos Limniotis, ‘Navigating Through Human Rights in AI: Exploring the Interplay Between GDPR and Fundamental Rights Impact Assessment’ (2025) 5(1) J Cybersecurity and Privacy 7 ↩︎
- Amir Al-Maamari, ‘Between Innovation and Oversight: A Cross-Regional Study of AI Risk Management Frameworks in the EU, U.S., UK, and China’ (2025) arXiv 2503.05773, https://doi.org/10.48550/arXiv.2503.05773 ↩︎
- Philipp Hacker, ‘Comments on the Final Trilogue Version of the AI Act’ (European New School, 23 January 2024) <https://www.europeannewschool.eu/images/chairs/hacker/Comments%20on%20the%20AI%20Act.pdf> accessed 18 January 2026 ↩︎
- Pascal Yves Schroeder and others, ‘Zooming in on AI – #10: EU AI Act – What Are the Obligations for High-Risk AI Systems?’ (AO Shearman, 29 October 2024) <https://www.aoshearman.com/en/insights/ao-shearman-on-tech/zooming-in-on-ai-10-eu-ai-act-what-are-the-obligations-for-high-risk-ai-systems> accessed 12 January 2026 ↩︎
- Andrea Renda and others, Study to Support an Impact Assessment of Regulatory Requirements for Artificial Intelligence in Europe (Publications Office of the European Union 2021) ↩︎
- Ibid ↩︎
- Ibid ↩︎
- Elena Bignami and others, ‘Balancing Innovation and Control: The European Union AI Act in an Era of Global Uncertainty’ (2025) 4(1) JMIR AI e75527 ↩︎
- Francesco Filippucci and others, The Impact of Artificial Intelligence on Productivity, Distribution and Growth: Key Mechanisms, Initial Evidence and Policy Challenges (OECD Artificial Intelligence Papers No 15, OECD Publishing 2024) ↩︎
- Pascal Yves Schroeder and others, ‘Zooming in on AI – #10: EU AI Act – What Are the Obligations for High-Risk AI Systems?’ (AO Shearman, 29 October 2024) https://www.aoshearman.com/en/insights/ao-shearman-on-tech/zooming-in-on-ai-10-eu-ai-act-what-are-the-obligations-for-high-risk-ai-systems accessed 12 January 2026 ↩︎
- Connor Attridge, ‘California’s Frontier AI Act: National Security Implications for the UK’ (CETaS Expert Analysis, 14 October 2025) <https://cetas.turing.ac.uk/publications/californias-frontier-ai-act-national-security-implications-for-the-uk> accessed 16 January 2026 ↩︎
- OECD, ‘Recommendation of the Council on Artificial Intelligence’, OECD/LEGAL/0449 (adopted 22 May 2019) <https://legalinstruments.oecd.org/en/instruments/OECD-LEGAL-0449> accessed 12 January 2026 ↩︎
- Rashida Richardson, Jason M Schultz and Kate Crawford, ‘Dirty Data, Bad Predictions: How Civil Rights Violations Impact Police Data, Predictive Policing Systems, and Justice’ (2019) 94 NYU L Rev Online 15 ↩︎
- Connor Attridge, ‘California’s Frontier AI Act: National Security Implications for the UK’ (CETaS Expert Analysis, 14 October 2025)<https://cetas.turing.ac.uk/publications/californias-frontier-ai-act-national-security-implications-for-the-uk> accessed 16 January 2026 ↩︎
- ‘Long-Awaited EU AI Act Becomes Law after Publication in EU’s Official Journal’ (White & Case, 16 July 2024) <https://www.whitecase.com/insight-alert/long-awaited-eu-ai-act-becomes-law-after-publication-eus-official-journal> accessed 16 January 2026 ↩︎
- Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act) Art 99 ↩︎
- European Commission, ‘European AI Office and European Artificial Intelligence Board’ (European Commission) <https://digital-strategy.ec.europa.eu/en/policies/ai-office> accessed 15 January 2026 ↩︎
- Melanie Smith, ‘Challenges in the Implementation of EU Law at National Level’ (European Parliament Think Tank, 15 November 2018)<https://www.europarl.europa.eu/thinktank/en/document/IPOL_BRI(2018)608841> accessed 15 January 2026 ↩︎
- Initiative for Applied Artificial Intelligence, ‘AI Act: Risk Classification of AI Systems from a Practical Perspective’ (initiative for applied artificial intelligence, March 2023) <https://www.appliedai.de/uploads/files/Whitepapers/appliedAI-Whitepaper-Risk-Classification-of-AI-Systems-from-a-Practical-Perspective-2023-EN.pdf> accessed 18 January 2026
↩︎ - Transparency in Frontier Artificial Intelligence Act (Cal 2025), Cal Bus & Prof Code s 22757.15(a) ↩︎
- Connor Attridge, ‘California’s Frontier AI Act: National Security Implications for the UK’ (CETaS Expert Analysis, 14 October 2025) <https://cetas.turing.ac.uk/publications/californias-frontier-ai-act-national-security-implications-for-the-uk> accessed 16 January 2026 ↩︎
- Ibid ↩︎
- Amir Al-Maamari, ‘Between Innovation and Oversight: A Cross-Regional Study of AI Risk Management Frameworks in the EU, U.S., UK, and China’ (2025) arXiv 2503.05773 <https://doi.org/10.48550/arXiv.2503.05773> ↩︎
- Ibid ↩︎
Marilena Triantafillidis
LLB (LSE) ’28 and Notes Editor of the LSE Law Review 2025-27
