PauseAI Proposal

The PauseAI Proposal is the living web policy document maintained by PauseAI on the organisation's own site that sets out the substantive demand structure of the international pause-on-frontier-AI movement PauseAI has anchored since its May 2023 founding in Utrecht. The proposal opens with the demand to "implement a temporary pause on the training of the most powerful general AI systems" and assembles a four-element treaty plan, a two-element "Other Measures" demand package, and a long-term-policy track around that opening demand. Originally published alongside PauseAI's founding in 2023 and carried into the October 2023 international protest cycle as the organisation's substantive policy reference, the proposal is currently marked "Version: April 5th, 2026" — making it a living manifesto whose successive revisions track the international AI-safety-summit cycle that PauseAI's protest and lobbying programme is built around.

The proposal is not named after a single human author; it is institutionally authored by PauseAI as an organisation, with Joep Meindertsma as its founding articulator and Maxime Fournes as the CEO under whom the current 2026 revision was published. Within the corpus it is the publication-side anchor of PauseAI's organising work — the document that the federation of national chapters co-ordinates around, that the international-protest calendar is timed against, and that the organisation's policymaker meetings substantively draw on.

What the proposal asks for

The body of the proposal is organised in four named sections: Treaty Measures (containing the International AI Safety Agency and the coalition-veto demand), Effects of a Treaty (clarifying scope), Other Measures (the copyright and liability demands), and Long-term Policy (the algorithm-publication and hardware-advance demands).

The proposal's central institutional demand is for an International AI Safety Agency (AISA) modelled explicitly on the International Atomic Energy Agency, with four named responsibilities: approving major deployments and training runs of AI systems judged not to pose a threat to humanity; granting incremental approval for more powerful AI once sufficient safety-research progress has been made; specifying the oversight requirements for permitted training runs; and verifying compliance through GPU tracking, energy monitoring, data-centre inspections, semiconductor-facility inspections, and chip-location tracking. The verification track is explicitly framed by reference to the existing nuclear-non-proliferation regime — the feasibility analysis that accompanies the proposal argues that GPUs can be tracked "as we track elements used in the development of nuclear weapons" because specialised-GPU production is concentrated among a small number of supply-chain chokepoints (NVIDIA, TSMC, ASML).

Alongside the AISA, the Treaty Measures section asks for a coalition veto: a "sufficiently large coalition" of opposing countries should be able to block the deployment of any AI system above the superhuman threshold, and the model weights and access to such systems should be kept "under the stewardship of the AISA" until inter-state agreement on deployment has been reached. The treaty's stated scope is general-purpose AI systems — language models and agentic systems — and would, on the proposal's own framing, "usually not affect narrow AI systems, like image recognition," with most existing AI products and services continuing to operate legally under the regime.

The Other Measures section adds two demands sitting outside the treaty mechanism but which the proposal asks national governments to adopt regardless: a ban on training AI systems on copyrighted material, and the imposition of criminal-acts liability on AI model creators when their systems are used to commit crimes. The Long-term Policy section names two further extensions — limiting publication of training algorithms and runtime-efficiency improvements, and considering limits on hardware-capability advances — as measures the proposal asks be held under consideration for the longer arc of frontier AI governance.

The proposal cites 64–69% U.S. public support for a pause on frontier AI development as the political-feasibility ground for the demand structure, and links forward to the Future of Life Institute's Pause Giant AI Experiments open letter of March 2023 as the prior-art public-discourse anchor under which PauseAI's specific demand framework operates.

How the proposal has been carried

The proposal is the substantive reference document of PauseAI's organising programme. The opening demand — that AI training above the most-powerful general-AI threshold be paused — is what the international protest calendar mobilises around. The proposal's four-section structure is what the federation's chapters cite when meeting with national policymakers, what the spokespeople carry into press appearances on outlets including NBC, Bloomberg, KQED, the New York Times, and Fortune (Wikipedia), and what the organisation's theory of change treats as the substantive content of the international treaty its summit-track strategy is built around.

The proposal's protest-track trajectory runs through the four-summit AI-safety-summit cycle that PauseAI's summit page treats as the institutional vehicle for the treaty: the November 2023 AI Safety Summit at Bletchley Park, the May 2024 AI Seoul Summit, the February 2025 Paris AI Action Summit, and the February 2026 India AI Impact Summit. The protest cycle PauseAI has run against this calendar — small Brussels Microsoft action May 2023, the October 21 2023 international protest, Bletchley Park November 2023, the thirteen-country May 13 2024 mobilisation ahead of Seoul, the November 2024 Anthropic-and-others international wave, the February 2025 Paris mobilisation, and the 2025 Google DeepMind London protest — has carried the proposal's demand framework into each summit's public discourse without yet converting it into binding treaty language at the institutional level. Each successive revision of the proposal (the current version stamped April 5 2026 sits in the wake of the New Delhi summit) reads as PauseAI's adjustment of the demand package to the political reality of the summit cycle's evolving framing from "safety" (Bletchley, Seoul) to "action" and "impact" (Paris, New Delhi).

The proposal also operates as PauseAI's training document for chapters and as the institutional reference for the organisation's companion materials — the feasibility analysis on GPU tracking and supply-chain chokepoints, the theory-of-change page on summit selection, and the FAQ explaining the scope of what would and would not be paused. It is in this sense the publication-side fixed point that PauseAI's entire programme of protests, petitions, public-comment campaigns, candidate questionnaires, and policymaker meetings rotates around.

Position in the corpus

Within the corpus this is the first PauseAI-side publication-class entity and the foundational manifesto of the pause-movement publication slate. It sits with A Narrow Path: How to Secure Our Future as one of the two corpus-anchoring pause-movement policy documents — the two distinguished by the organisational shape they come from (A Narrow Path is ControlAI's detailed three-phase policy plan, with named authorship and an inside-game UK-parliamentary track; the PauseAI Proposal is the protest-and-treaty grassroots federation's policy reference, institutionally authored and carried by the international-protest track). Together they bracket the corpus's representation of the pause-movement publication register: ControlAI's analytic policy plan and PauseAI's mass-political treaty demand, addressing related but distinct theories of change against the same diagnosis.

As a publication type it fills the manifesto slot for the AI-safety / pause register — structurally distinct from the corpus's other manifesto-type entry Civil society calls on the EU to put fundamental rights first in the AI Act, which is a one-off coalition statement on a specific legislative file rather than a living organisation-anchored policy document, and from the corpus's only other AI-pause publication-side artefact Pause Giant AI Experiments — the Future of Life Institute's March 2023 open letter, which is anchored in the corpus as a Message rather than a Publication because the FLI is a think-tank outside the corpus's grassroots-organising remit and because the open letter's role in the corpus is primarily as a propagating framing rather than as an organisational policy document.

In the corpus's mapping of the make-AI-good movement, the PauseAI Proposal is the document the publication slate points to for "what does the pause movement actually ask for" — the standing answer that PauseAI's chapters, protests, and policymaker meetings all reference, and the publication-side fixed point against which the movement's progress on the international-treaty front can be tracked.