Regulation of artificial intelligence

• 1. Introduction

  ›

  Regulation of artificial intelligence surged ninefold in global legislative attention between 2016 and 2024. Stanford University's 2025 AI Index measured that shift across 75 countries. In 2024 alone, nearly 700 AI-related bills were introduced across 45 American states, up from 191 the year before. U.S. federal agencies issued 59 AI-related regulations that same year, more than double the figure from 2023.

  The urgency did not produce agreement. Elon Musk, Sam Altman, Dario Amodei, and Demis Hassabis all called publicly for immediate oversight of AI. In 2023, following the creation of GPT-4, Musk and others signed an open letter urging a halt to the training of more powerful AI systems. Mark Zuckerberg and Marc Andreessen warned on the other side, arguing that premature rules could stifle innovation. A 2022 Ipsos survey found that 78% of Chinese citizens agreed AI products carry more benefits than drawbacks. Among Americans, only 35% agreed.

  A 2023 Fox News poll found 35% of Americans considered federal AI regulation 'very important'. An additional 41% called it 'somewhat important'. The United Kingdom launched the first in a series of international summits on AI safety in 2023, followed by gatherings in Seoul, Paris, and New Delhi. Whether those summits could produce binding agreements depended on a question no single government had yet answered: who had the authority to enforce them.

• 2. Hard Law and Soft Law

  ›

  A 2020 meta-review by the Berkman Klein Center for Internet & Society examined existing AI ethics frameworks. It drew on documents including the Asilomar Principles and the Beijing Principles. Across all of them, eight core principles appeared repeatedly: privacy, accountability, safety and security, transparency and explainability, fairness and non-discrimination, human control of technology, professional responsibility, and respect for human values.

  The 'pacing problem' is the term legal scholars use for the central challenge of hard-law regulation. AI technology evolves faster than traditional legislation can follow. Regulatory agencies face a further difficulty: most were built for specific industries and lack the jurisdictional reach to govern AI across sectors.

  Soft law offers an alternative. Voluntary principles and codes of conduct adapt more readily to a changing technology landscape. The trade-off is enforcement: a guideline no one must follow tends not to be followed. Cason Schmit, Megan Doerr, and Jennifer Wagner proposed a middle path. They argued for using copyleft licensing of AI models and training datasets to bind developers to specific ethical obligations. A designated quasi-governmental body would then hold enforcement power over those terms.

  Henry Kissinger, Eric Schmidt, and Daniel Huttenlocher published 'Being Human in an Age of AI' in November 2021, calling on governments to establish a commission to regulate the technology. Youth advocacy organization Encode AI has also issued comprehensive agendas pushing for stricter rules and public-private partnerships.

  At the AI Action Summit in 2025, the UK and US governments declined to sign an international agreement on AI. The agreement proposed an open, inclusive, and ethical approach to development, including environmental protections. US Vice President JD Vance argued the agreement would harm the AI industry's growth. The UK said the text failed to provide practical clarity on global governance or to address national security questions adequately.

• 3. The Architecture of Global Governance

  ›

  Canada and France announced a G7-backed International Panel on Artificial Intelligence in December 2018, modeled on the International Panel on Climate Change. Its mandate was to study AI's effects on people and economies around the world. By 2019, the initiative had been renamed the Global Partnership on Artificial Intelligence.

  The Global Partnership on Artificial Intelligence, known as GPAI, launched formally in June 2020. Its 15 founding members included Australia, Canada, France, Germany, India, Italy, Japan, South Korea, and the European Union. Mexico, New Zealand, Singapore, Slovenia, the United Kingdom, and the United States also joined as founding members. By 2023, membership had grown to 29 nations. GPAI's Secretariat is hosted by the OECD in Paris. The partnership's mandate spans four themes: responsible AI and data governance, supported by the International Centre of Expertise in Montreal for the Advancement of Artificial Intelligence; and the future of work and innovation and commercialization, handled by a corresponding centre in Paris.

  The OECD AI Principles were adopted in May 2019. The G20 followed with its own AI principles in June of the same year. At UNESCO's 40th Scientific Session in November 2019, the organization launched a two-year effort to produce a global standard-setting instrument on AI ethics. A draft text was issued in September 2020. The full instrument was adopted at UNESCO's General Conference in November 2021.

  On the 30th of October 2023, the Group of Seven nations adopted eleven guiding principles for the design, development, and deployment of advanced AI systems. These principles form part of what became known as the Hiroshima Process. A voluntary Code of Conduct for AI developers was introduced alongside them, emphasizing transparency, accountability, and risk-mitigation across the AI lifecycle. Ursula von der Leyen welcomed the initiative, saying it reflected the direction of the EU's AI Directive, which was in its final stage of negotiation at the time.

  In September 2022, the Council of Europe began negotiations toward a binding AI treaty. The talks involved the organization's 46 member states plus 11 additional nations, including Argentina, Australia, Canada, Japan, and the United States. On the 17th of May 2024, the Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law was adopted. It opened for signature on the 5th of September 2024. Among the first ten signatories were Andorra, Georgia, Iceland, Norway, Moldova, San Marino, the United Kingdom, Israel, the United States, and the European Union.

  At the United Nations General Assembly in September 2025, governments received a global call for binding AI red lines. The proposal urged nations to reach a formal agreement prohibiting certain AI uses and behaviors by the end of 2026.

• 4. The European Model

  ›

  On the 2nd of February 2020, the European Commission published its White Paper on Artificial Intelligence. The document proposed a vision organized around two pillars: an ecosystem of excellence and an ecosystem of trust. The trust pillar outlined a regulatory framework built on risk. Applications were divided between those carrying high risk and those that did not, with mandatory requirements applying only to the former. An application qualified as high-risk if it operated in a sensitive sector such as healthcare, transport, or energy and was likely to generate significant harm.

  A January 2021 draft of the formal AI legislation was leaked online on the 14th of April 2021. The Commission released its official text days later. That legislation, which became known as the AI Act, refined the earlier risk framework into four categories: minimal, limited, high, and unacceptable. Academics raised concerns about unclear definitions, particularly the broad description of what counts as AI, and warned of unintended consequences for vulnerable groups such as patients and migrants.

  A new category, general-purpose AI, was added later to cover models like ChatGPT that did not fit the original application-based structure. The AI Act defines systemic-risk models as those trained on computing power exceeding ten to the twenty-fifth floating-point operations. Those models face a thorough evaluation process. Weaker general-purpose models face transparency requirements only.

  The AI Act was adopted in May 2024 and entered into force on the 1st of August 2024, officially designated Regulation EU 2024/1689. Provisions covering general-purpose AI models become enforceable from the 2nd of August 2025. Several categories of application are prohibited outright, among them real-time remote biometric identification and emotion recognition systems, with limited exemptions for areas such as law enforcement.

  On the 29th of May 2024, the European Court of Auditors published a critical assessment of EU AI governance. It found that EU measures were not well coordinated with those of member states and that investment monitoring lacked consistency. Analysts noted that the EU's prescriptive approach sits in tension with the United States' sector-specific preference. That divergence could complicate efforts toward global regulatory alignment.

• 5. Three National Models

  ›

  The United States, China, and the European Union each took a different path toward governing artificial intelligence. Analysts described these approaches as market-driven, state-driven, and rights-driven respectively.

  In October 2023, the U.S. government issued an executive order on AI safety and security. The order required developers of the most capable AI systems to share safety testing results with federal agencies before any public release. In December 2025, President Trump signed a further executive order directing federal agencies to develop a unified national AI policy. The order authorized the federal government to challenge state-level AI laws through legal action. Exemptions were made for state rules on child safety, data center infrastructure, and government procurement. Walter Donway, writing for a publication of the American Institute for Economic Research, called the order's underlying premise 'philosophically wrong'. His argument: no central authority could foresee the risks of an emerging technology better than millions of independent actors in a free market.

  China's July 2017 'Next Generation Artificial Intelligence Development Plan,' designated State Council Document No. 35, set the country's trajectory through 2030. The plan directed governing bodies to promote AI development while keeping data, including personal data on Chinese users, stored within national borders. Ethical guidelines published in 2021 required AI researchers to keep their systems under human control and to prevent them from endangering public safety. On the 15th of August 2023, China's first generative AI regulatory measures entered into force. They applied to all providers offering generative AI services to Chinese users, including foreign companies, governing data protection, transparency, and algorithmic accountability. By mid-2024, over 1,400 AI algorithms had been registered under the Cyberspace Administration of China's filing regime. In April 2026, China's National Development and Reform Commission blocked Meta's $2 billion acquisition of Manus. Manus was an AI startup that had relocated from China to Singapore in 2025. The decision established national security review as a standing condition for cross-border AI transactions involving Chinese assets or technology.

  In March 2023, the UK published a white paper titled 'A pro-innovation approach to AI regulation.' Rather than introducing new legislation, the document gave existing regulators flexibility to apply broad principles within their own sectors. Prime Minister Rishi Sunak hosted the first AI Safety Summit in November 2023. From that gathering came the AI Safety Institute, built from the earlier Frontier AI Taskforce led by Ian Hogarth. The institute was assigned responsibility for evaluating the world's most advanced AI models. The UK's 10-year National AI Strategy, published in 2021, had already designated AGI-related catastrophic risks as scenarios requiring long-term assessment.

• 6. The Existential Question

  ›

  Artificial general intelligence, or AGI, presents a different regulatory challenge than the AI systems in use today. AGI describes a system capable of matching or exceeding human cognition across a broad range of tasks. Scholars refer to the 'AI control problem' as the challenge of ensuring that systems at this level remain beneficial over the long run. Doing nothing about the risk and imposing an outright ban are both seen as impractical responses. Approaches such as brain-computer interfaces, which would enhance human capabilities to keep pace with AI, are regarded as potentially complementary to regulation rather than replacements for it.

  International mass surveillance for AGI arms control is one option that researchers have raised alongside more conventional proposals. Differential intellectual progress is a related concept: researchers should prioritize defensive and safety-focused strategies before advancing toward more powerful and riskier systems. Proposed regulatory structures also call for review boards operating at every level, from individual universities and corporations through to international bodies, alongside sustained investment in AI safety research.

  The 'AGI Nanny' is a specific proposal that has entered the discussion about preventing catastrophic AI outcomes. The concept calls for constructing a system smarter than any individual human but short of true superintelligence. This system would be connected to a large surveillance network. Its purpose would be to monitor humanity and prevent the creation of a dangerous superintelligence. It would also guard against other large-scale threats, including subversion of the global financial system. The arrangement would persist until a true superintelligence could be safely created.

  Conscious, ethically aware AGIs raise a further set of questions. Scholars debate how to integrate such entities into existing human society, a conversation that turns on considerations of legal standing and moral rights. The United Nations has made progress on AI governance broadly, but its institutional and legal capacity to address AGI-scale existential risks is acknowledged to be more limited. Military AI raises those same questions in a more immediate context. International deliberations on autonomous weapons have been under way for more than a decade.

• 7. The Battlefield Dimension

  ›

  Under the Convention on Certain Conventional Weapons, states began meeting in 2014 to consider possible international rules for lethal autonomous weapon systems, or LAWS. Informal expert sessions took place in 2014, 2015, and 2016. A Group of Governmental Experts was appointed in 2016 to continue the deliberations. In 2018, the group adopted a set of guiding principles on LAWS.

  In 2016, China became the first UN Security Council permanent member to formally question whether existing international law could govern fully autonomous weapons. The country published a position paper making that argument. The paper prompted other national delegations to propose new global regulation of autonomous weapons. The Campaign to Stop Killer Robots, a coalition of non-governmental organizations, has repeatedly called for a moratorium or preemptive ban on developing and deploying such systems.

  The United States has maintained that current international humanitarian law is sufficient to regulate autonomous weapons. In 2023, the Congressional Research Service confirmed that the U.S. military did not have lethal autonomous weapon systems in its inventory. The same report noted that U.S. policy does not prohibit their development or future use.

  The Munich Convention on AI, Data and Human Rights brought together scholars from multiple academic institutions and called for a binding international agreement to protect human rights in the age of AI. How much human control must remain over AI-enabled weapons is one of the central unresolved questions that any such agreement would need to address.