About
AI governance: the efforts under way
International institutions
UNESCO – Recommendation on the Ethics of AI
Adopted at UNESCO in 2021 and endorsed in 2023 by 193 member states, this text sets out principles (human rights, transparency, impact assessment, accountability) and calls for consistent national frameworks. It serves as a reference, without imposing limits on technical capabilities or providing mechanisms to shut down frontier AI systems.
UN – Global Dialogue on AI Governance
In 2024, a “United Nations System White Paper on AI Governance” coordinated by UNESCO and the ITU mapped out possible institutional options and stressed coordination and risk monitoring. In 2025, the Summit of the Future led to the creation of the Global Dialogue on AI Governance, an intergovernmental, multi-stakeholder forum to bring existing initiatives together and prepare possible more formal agreements, with no global moratorium mechanism.
Office of the High Commissioner for Human Rights – Targeted moratoriums
In 2021, High Commissioner Michelle Bachelet called for a moratorium on the sale and use of certain AI technologies, notably remote biometric recognition in public spaces, until adequate human rights safeguards are in place. The logic is a targeted pause on certain uses, not a halt to frontier AI systems in general.
ICRC / UN – Position on autonomous weapon systems
The International Committee of the Red Cross and the UN Secretary-General have for several years advocated a legally binding instrument banning or restricting certain categories of autonomous weapons, to preserve “meaningful human control” over the use of force. It amounts to a sectoral moratorium or ban limited to military AI.
G7 – Hiroshima Process and statements on AI
From 2023, the G7 Hiroshima AI Process produced principles and a code of conduct for generative AI and foundation models : risk assessment, transparency, accountability, information sharing with authorities. The French G7 presidency in 2026 announced its intention to strengthen this convergence around “safe” and protective AI, without calling into question the continued development of frontier AI systems.
NGOs and think tanks
Future of Life Institute (FLI)
”Pause Giant AI Experiments” letter (2023)
Published in March 2023, this letter calls for a pause of at least six months on training systems more powerful than GPT-4, the frontier models of the time. The pause must be public and verifiable ; if self-regulation fails, the text asks governments to “step in and institute a moratorium”. Despite wide media coverage, no major lab put a coordinated pause in place.
Statement on Superintelligence (2025)
In 2025, FLI coordinated a new letter, this time calling for a long-term prohibition on developing superintelligent AI until there are sufficient safety guarantees and clear democratic support. The NGO then came out in favour of a pause on frontier AI experiments and of a moratorium of at least fifteen years on certain forms of general AI or superintelligence.
Machine Intelligence Research Institute (MIRI)
”AI Governance to Avoid Extinction” report (2025)
Published in May 2025, this report sets out several possible paths for AI governance, including a central “Off Switch and Halt” scenario. It starts from the diagnosis that frontier AI systems could, in the short term, surpass human experts in many fields and create a risk of extinction (loss of control, wars, authoritarian lock-in). The Halt scenario involves building a global capacity to restrict or stop the development and deployment of these systems, then maintaining a global moratorium, potentially for decades, while alignment problems are solved.
Corresponding draft treaty (2025)
Building on this, MIRI outlines an international treaty setting compute thresholds (FLOP) above which training frontier AI systems requires a licence, along with tracking of chips and data centres, international inspections and a sanctions regime. The architecture draws on arms control regimes, with a US–China coalition envisaged as its core.
ControlAI
”A Narrow Path” (2024) and ”Direct Institutional Plan” (2025)
ControlAI, founded in the mid-2020 s, published its “A Narrow Path” plan between 2024 and 2025, focused on preventing uncontrolled superintelligence. It proposes banning, for at least twenty years, the development of superintelligence and of certain extreme capabilities (escape, self-improvement, strategic uses), while allowing other frontier AI systems deemed less dangerous under licence. It also recommends compute governance measures (power caps, licences), stronger safety requirements and, eventually, the creation of an international AI safety agency.
To implement this framework in democracies, ControlAI then developed the “Direct Institutional Plan” (DIP), a two-step action plan (drafting policies targeting superintelligence and its precursors, then systematically engaging institutional actors) aimed at securing national bans and, ultimately, a lasting international agreement against the development of superintelligence.
Existential Risk Observatory (ERO)
”Conditional AI Safety Treaty” (2025)
From 2024–2025, the Existential Risk Observatory has proposed a “Conditional AI Safety Treaty” intended to be more politically acceptable than an immediate, general moratorium. Rather than decreeing a halt to frontier AI systems right away, it provides a mechanism : states commit to stopping certain training runs as soon as independent international AI safety institutes conclude that a project or capability level comes too close to a loss of control. The length of the halt is not set in advance ; it depends on how risk assessments and progress on alignment evolve.
PauseAI (global)
”Building the Pause Button” (2025)
The PauseAI movement proposed a hardware-based approach to the pause at the 2025 AI Safety Camp. Training frontier LLM-type models requires large amounts of highly specialised and expensive hardware (chips). The supply chain for this hardware relies on a very small number of highly specialised companies, such as ASML (Netherlands) and SMEE (China) for the lithography machines needed to make the chips, or TSMC (Taiwan), Samsung (South Korea) and SMIC (China) for the fabrication plants. Through the AI chip supply chain, governments or other regulators can know about, and physically control, the activity of the companies developing AI.
