The existential AI safety community needs to take building a civic and social movement seriously as a core intervention. We believe this is a high-value, badly neglected approach to reducing catastrophic/x-risks from AI because it may significantly enhance the likelihood of governance efforts succeeding at keeping humanity safe.
Note: this post is about PauseAI, not PauseAI US, which is a distinct entity with a different leadership team and approach.
This post was written by Matilda da Rui and Maxime Fournes, with significant contributions from Benjamin Schmidt (PauseAI Germany co-lead).
Executive Summary
The existential AI safety community needs to take building a civic and social movement seriously as a core intervention. We believe this is a high-value, badly neglected approach to reducing catastrophic/x-risks from AI because it may significantly enhance the likelihood of governance efforts succeeding at keeping humanity safe. As far as we can tell, only one organisation is building this infrastructure across continents: PauseAI. This post lays out our reasoning and our track record, and makes the case that funding this work is one of the highest value-for-money contributions available to anyone looking to reduce AI risk.
Why don’t we already have a pause or strong controls on frontier AI? Multiple advocacy groups are communicating clear and convincing arguments for AI existential risk, and policy experts are putting forward comprehensive proposals. We need more of this work, but this work alone will not be enough, because one link is missing: what policymakers hear doesn’t align with what their incentives reward. They face no political constituency that cares about existential risk.
We saw this firsthand at our European Parliament conference, where an MEP told us that she agreed with us but could not act without public support. Building that constituency is what PauseAI offers to the AI safety ecosystem.
Why do we think that PauseAI can deliver this constituency? On remarkably thin resources ($600k since we started in 2023), we have built an organised presence in more than 15 countries and a community of thousands of supporters and volunteers. We organise civic engagement, build movement infrastructure, and generate media coverage to push the catastrophic risk frame, and a pause as a viable answer to it, into the mainstream. In the past six months, we’ve professionalised our operations, run our first campaign coordinated across 12 national groups, held a conference at the European Parliament, and run the largest AI safety protest to date (300 participants in London).
In a wide range of plausible timelines, we can organise a focused constituency fast enough to make a difference. And with access to sufficient resources (which would still amount to a tiny fraction of the current spending on AI safety overall), we can build a social movement with the power to reshape norms. But the capacity to convert a sudden window of opportunity into political change has to exist before the window opens. We are already part of the way there, but we are bottlenecked on funding, and our current runway ends in October 2026. If you’re interested in funding us, see Part IV.1, which includes a breakdown of funding scenarios.
Our goal, ask, and day-to-day work
Our ultimate goal is preventing catastrophe. Humanity should not build superintelligent systems until or unless we know how to keep them safe and under control; everything we do is in service of that outcome.
Our public ask—the demand our movement carries—is a pause on advanced general-purpose AI development, ultimately anchored in a binding international treaty. Of the levers currently available, we think it is the one that most directly buys humanity the time it needs to orient to this challenge, and it is simple and concrete enough to organise a movement around. But it is an instrument, not our end goal. If a better lever for preventing catastrophe emerged, we would back it. A movement, however, cannot rally behind a carefully hedged portfolio of policy options; it needs one clear, defensible demand.
Our day-to-day work instantiates our higher-level goal and ask: building the movement itself—the constituency and the community infrastructure that make any adequate policy politically reachable when it is proposed, and durable after it passes (we come to division of labour considerations in Part II.1). This is also why this post says comparatively little about the pause proposal itself and its merits.
We think the strongest objections to our work are these:
Public opinion is low-leverage on a decision like this.
We think this prior is miscalibrated. Visible, organised constituencies change what policymakers can afford to do, even when they’re already convinced on the merits. We make this case in Part II.2.
Building a movement powerful and aligned enough, on short timelines, is intractable.
We’ve done a meaningful share of it already, on very little money. Part III outlines our track record.
A mass movement will drift, polarise, or turn anti-AI, and end up net-negative.
This could happen if we didn’t plan accordingly. Much of our structure and strategy is set up to prevent it. We explain how in Parts II.4, II.5, and III.
A pause is not achievable
What we are fundamentally building is an organised constituency that cares about catastrophic AI risks, and that is net positive for any serious safety policy, not just a pause. The pause is the ask we organise around because a movement needs one clear, defensible demand. Besides, “a pause is impossible” is a self-fulfilling belief: the conviction that the race cannot be stopped is itself a major reason no one tries. We think this story can be broken. See Parts I.1, II.2, and II.4.
This is a long post, so we have prefaced each section with a summary of its contents in italics. We want this to be the beginning of an ongoing conversation. We welcome clarifying questions, constructive comments and peer input.
Introduction
A somewhat common view in the AI safety community seems to be that a social and civic movement for existential AI safety is not clearly useful, or that it is lower-EV than other interventions—something like “the value of getting the general public onboard with regulating advanced AI seems low” (either relatively or absolutely). A related vector of scepticism rests on the assumption that successfully building such a movement is unreasonably hard (“surely building a movement powerful, aligned and responsive enough to create sufficient political will to reach an enforceable international pause in the face of arguably unparalleled incentives for industry and states, on potentially short timelines and under various forms of uncertainty, with a general public that is only just starting to register the problem, all amounts to something humanity can’t pull off”).
We think both versions of this outside view are incorrect because their assumptions are miscalibrated. We claim that building an effective movement for a global pause on advanced AI development is both tractable and high-EV—and even probably decisive. This is why we set out to do this work. PauseAI’s mission can be summarised as follows: to create a global constituency capable of generating sufficient political will in order to obtain a robust pause on advanced general-purpose AI development by building scalable and resilient civic and social movement infrastructure.
To our knowledge, we are the only group building the worldwide infrastructure for a social movement supporting a focused political constituency and building deep public buy-in for a pause on frontier AI development. Everything we have achieved so far was accomplished on under $600k over 3 years of operation across more than 15 countries. To us, this is in itself a measure of how neglected this important intervention is.
We want to lay out the reasoning behind our claims and the bet we’re making. We need the wider AI safety ecosystem to support us in this—it’s a large and hard undertaking and we don’t have the resources to do this all alone at this point, despite the high efficiency we are capable of. We’ve built the foundations for this movement and we now need to shift gears to reach the next step in our growth.
This post provides an overview of our theory of change, approach, and track record so far, then outlines our role in the AI safety ecosystem and on the path to a pause, and finally indicates how you can increase our capacity such that we are more likely to succeed across plausible timelines. We also want to share with you the blueprints and foundations which we think make this approach tractable; for the sake of readability, we’ll be sharing those separately.
I. Our theory of change
Prologue
Today, the impact of frontier AI policy efforts is significantly constrained by a lack of the demand side (constituency interest) for the product they offer (policy briefings and proposals).
Imagine a policy expert from ControlAI goes to Brussels to brief a member of the European Parliament. The briefing is excellent: clear, well evidenced, convincing. The MEP listens, and by the end, she more or less believes it. Catastrophic risk from frontier AI is real, timelines are short, and something should be done.
And then she does nothing. Not because she was unconvinced, but because nothing in her world is pushing her to act, and a great deal is pushing the other way. Acting would mean spending political capital, picking a fight with a fast-growing industry her own government is courting for investment, and getting ahead of colleagues who would rather wait and see who ends up looking foolish. Against all that, the briefing was one good conversation. So she files it under “interesting, probably true,” and turns back to the issues her constituents do write to her about. If it really mattered, she tells herself, surely she would be hearing about it from the people who elect her. She isn’t, so she doesn’t.[1]
The gap we exist to close is right there, and it is not a gap in the argument. The argument was fine. What was missing was any immediate reason for her to act on it. Let’s now imagine the same meeting with one thing changed: for the past six months, this MEP has been getting steady, unmistakable signals that many of her constituents care about this, and care enough to remember it at an election. The same briefing now lands as something she can act on and defend to her colleagues, because acting is no longer politically expensive overall (there’s a tangible benefit to her acting on this issue). We didn’t need to improve the expert’s case, but we needed to change whether she could afford to say yes to it. We saw this first-hand at our Brussels conference in February, where MEP Saskia Bricmont agreed with our stance but asked for more public support.
1. The shape of the problem as we see it
For most actors, the race to AGI/ASI is a coordination problem that can be solved via a treaty that binds everyone, including those who wilfully or recklessly keep the race going without the world’s consent; this requires leveraging political power as a counterforce to the drivers of the race, and we are building this political power by catalysing a pause-focused worldwide civic and social movement.
Frontier labs are racing toward ASI without democratic buy-in, before anyone knows how to keep such systems under control, and before any governance regime adequate to manage the risks exists. [2]
We can break the issue down through a typology of actor profiles:
Some corporate and state actors, and most of the general population, are oblivious to catastrophic AI risk. The bottleneck with regards to those actors is awareness of the issue and of pausing development as a viable solution.
Some company leaders and policymakers think the race is inevitable and stay in it lest the outcome be worse if they dropped out of it—and we agree that pausing unilaterally is no long-term answer. Relatedly, the leverage which states hold over frontier labs is not fixed, and the more deeply those actors’ systems are enmeshed into public infrastructure, the costlier binding labs becomes. With this group, the ones that are stuck in coordination failure, the focus ought to be on showing to them that there is a way out of the race through cooperation, and on making that latter claim true by making a pause politically rewarding, putting in place the technical architecture of the pause, and preparing the ground for those leaders to sign up for it when their peers do. Likewise for members of the public who think there is nothing citizens can do: they need to see they can take meaningful action, especially when they organise effectively, and this requires building infrastructure to lower the friction for them to do so.
Other corporate and state actors are just reckless as to the risk and are prepared to gamble everyone’s lives for a shot at winning the race—these are the actors that our approach seeks to bind, since they can’t be expected to willingly desist from binding humanity into catastrophe.
We can also frame this in terms of the dynamics keeping the race running:
Political leaders and civil societies are largely not AGI/ASI-aware. If everyone believed what the frontier AI leaders believe, AI research would be shut down tomorrow. The public and its representatives cannot influence the course of history if they don’t know the race is on.
Labs underestimate the difficulty of aligning AGI/ASI or are reckless as to it, and do not worry enough about gradual disempowerment and concentration of power or what to align the AI to.
A widespread fatalism, the assumption that AI development cannot be stopped, stops people from acting. This dynamic is a powerful self-fulfilling prophecy. If people believe widely enough that the race can’t be stopped, that very inevitability becomes real. So one of our goals is to dispel it by showing people that they can take useful action and that organised civic agency can compound into enough political will to clear the bar for meaningful policy change. AI safety has lacked the deep societal buy-in and organised constituency that could enable good policy to be passed and robustly enforced.
2. Necessary conditions for reaching a pause
Pausing requires (1) a public mandate generating the political incentive to act towards a pause; (2) elite consensus that pausing is the way to go; (3) enough international political will to make defection unattractive; (4) a prepared technical policy framework and environment; and (5) focused, coordinated civic/social activation to make sure major events are converted into political momentum. We can provide (1), (3), and (5).
We think a specific set of conditions underpins the enactment of a pause. Concern, organised into a constituency which politicians can see and count[3], changes politicians’ incentive landscape, because contributing to tackling AI risk stops being politically costly. That shift enables sympathetic leaders to act, and lets some governments integrate AI safety into an established frame, such as national interest or national security. As more countries adopt governance measures, holding out grows more expensive and, provided the technical policy prep work gets done in parallel, by the time an opportune window opens, a binding international agreement is reachable. A single triggering event can set off what Mark and Paul Engler (2016) call a moment of the whirlwind: a sudden surge of mass participation, far beyond a movement’s standing base, that throws open a brief window in which the politically impossible turns achievable, but only for a movement already built and waiting to seize it. The end state we aim for is a treaty with teeth—a global moratorium on training the most capable systems, with verification mechanisms and a standing body to enforce it.
In other words, to be effective in scope, a pause needs to be an international agreement, and we think it requires five conditions:
(1) a public mandate visible enough that signing becomes politically rewarding rather than costly for the leaders who do it;
(2) sufficient elite consensus that pausing is the responsible position;
(3) enough political will in multiple major-power governments that non-cooperation is unattractive;[4]
(4) a technical and governance framework ready to execute;
(5) the capacity to convert a triggering window into political action, almost certainly event-driven, that breaks inertia.
Tuning the world’s political engines to deliver a pause
Our work primarily enables conditions 1, 3 and 5: building a coordinated, multi-jurisdiction public mandate, making the politics of acting favourable for governments, and maintaining the standing capacity to convert opportune windows into political momentum. Conditions 2 and 4 are mostly the lane of policy organisations and credible experts.[5] While we cannot deliver a pause unilaterally, we can build the substrate that makes a pause possible when the other conditions are met.[6]
II. Our role towards a global treaty and in the AI safety ecosystem
1. Our niche within the ecosystem
Our work complements that of AI safety experts sounding the alarm from a scientific perspective, communications specialists who raise awareness over media channels, policy organisations who brief politicians and prepare technical proposals, and radical activists that stretch the Overton window; we think all those niches need each other to be effective.
Our work is one part of a larger ecosystem, in which different organisations do different jobs that reinforce one another rather than compete. A useful frame for this division of labour comes from social movement strategist Bill Moyer, whose Movement Action Plan (2001) argues that durable movements need four distinct roles filled at once and tend to stall when one is missing or crowds out the rest. Mapped onto AI safety, the roles fall out roughly as in the diagram below.
The Citizen is the population that understands the problem and treats concern about it as a normal, legitimate thing to hold. Communicators and science educators (Rob Miles, Tarbell fellows, etc) build that understanding, and credible experts such as Hinton, Bengio and Russell supply the seed for the elite consensus (condition 2 of our theory of change). Another important signal is elite defection in the form of researchers walking out of frontier labs over safety concerns. Under Moyer’s model, both experts and communicators fall under the Citizen role.
The Change Agent organises Citizen actors. It educates, builds organising capacity via grassroots structures and coalitions, and moves the majority over time. This creates:
a visible public mandate and the deeper societal buy-in that underpins it,
the political will which this mandate and buy-in catalyse, and
readiness to channel windows of opportunity into institutional change.
We are the layer that helps the latent Citizen role cohere into a primary driver of the political process.
The Reformer works inside institutions, providing briefings, drafting policy and organising policymakers. This is the lane of policy organisations such as ControlAI and MIRI, which develop frontier AI governance proposals. We don’t typically write policy. Our role is to build the political conditions under which proposed policy can pass and survive. We sometimes brief politicians directly when this supports our objectives and is easy for us to deliver, but this is not the core of our role.
The fourth role, the Rebel, forces the issue onto the public agenda from the margins, through opposition and defiance. Most readers’ default assumption probably places PauseAI here, because we run protests, as do Rebels, but we use them as signalling devices for the size and seriousness of the constituency rather than as a sign of anti-establishment rebellion; our protests are solution-oriented and mainstream-coded. The distinction is becoming strategically sharper, because as parts of the industry work to paint data centre NIMBY groups as saboteurs or foreign plants, a Change Agent coded as Rebel would forfeit its own credibility and make the Citizen role harder for everyone else to inhabit. The Rebel, in our context, is entities building on general anti-AI sentiment. We take seriously the responsibility of helping public concern cohere around a drift-proof movement focused squarely on catastrophic AI risk and pausing as a solution, and we are wary of the inflammatory and occasionally violence-infused ethos of the anti-AI space. We will not ally with actors whom we expect to foster violence or to employ tactics that undermine our longer-term objectives.
PauseAI’s niche in the AI safety advocacy ecosystem
Moyer’s framework is a practitioner’s heuristic, not a law, but if it is even roughly right, the impact of the roles compound one another. We need an ecosystem in which each role tends to its own niche well, ideally coordinating with each other.
2. Policymakers need strong enough incentives to act
In order for policy efforts to bear fruit, decision makers’ incentives have to line up with their model of their own scope for action; a well organised constituency, buoyed over time by a vibrant pro-pause community spanning many societal segments, can clear the path and create the momentum for effective political action.
The story of the Brussels policymaker with which we opened this post highlights the gap which we aim to fill. A representative who privately agrees can still be pulled quietly back the moment industry lobbying leans on them, unless a visible movement is there to hold them to it. The frontier AI industry can afford dense, concentrated, well funded pressure. Left unorganised, widespread private concern constitutes virtually no pressure. In the civic engagement context, our Change Agent role is to focalise that diffuse beam into a laser-sharp demand.
A natural objection is that a movement large enough to register politically will inevitably contain many people who are not skilled advocates, and that a flood of unpolished letters may be unpersuasive. This misconstrues the role of a constituency. The argument is carried by the briefers, and the constituency demonstrates that voters are invested. Representatives don’t typically weigh mass constituent messages by their prose, but rather by the volume and level of engagement they signify, i.e. a mass of constituents who treat this issue seriously enough to take time out of their busy lives to identify whom they should write to and to think about what to say, and who will remember the issue at the next election. Our stance on this is based on social movement researcher Charles Tilly’s model of what makes a movement register with political leaders, namely a combination of worthiness, unity, numbers and commitment (2004). In empowering an organised constituency, the majority of the signal is carried by the following indicators: numbers, consistency around the problem-solution frame, and commitment. The “worthiness” component can be fulfilled by volunteers who speak to politicians directly, as they are typically seasoned and can make a more confident, articulate case than the average message sender. Further, lighter-touch tools let supporters register genuine concern in a minute without composing a case from scratch. The product of this work is a constituency politicians can actually see and count, to use Tilly’s framing.
Reputational incentives are also a constraint, and not one confined to elected representatives. The same economics weigh on decision makers and senior figures across influential professions, as we often hear from volunteers who work in fields like law and risk management. I (Maxime) observed this mechanism at close range when I worked in quantitative finance. A fund manager whose positions track their peers’ is oddly safe: when a bad year comes, it comes for everyone; “the whole market was down” is a defence clients accept, and the money stays put. A manager who is uncorrelated to the rest of the market lives under a harsher rule: if you win alone, you look brilliant, but if you lose alone, there’s nobody to share the blame, clients leave, and you go broke.
I remember a parable about picking a restaurant to impress a date: choose the place that’s packed but unremarkable over the one you’ve heard is wonderful but looks empty, because if the packed place disappoints, the crowd absorbs the blame, whereas if the empty place disappoints, you alone chose it. Being early and wrong costs far more than being conventional and wrong, even when the lonely call was the better one. That asymmetry is much of why private concern about AI stays private. This is valid for members of the public but a fortiori for politicians: being seen to hold a view while it still reads as unconventional is a risky bet. A visible, credible constituency changes that calculus; it fills the empty restaurant. Once enough serious people are openly concerned, voicing concern stops being the lonely, uncorrelated bet and becomes the safe, shared one, and aligning with an established movement carries far less risk than sticking your neck out alone.
3. The path to a treaty
International coordination rests on domestic political will, in enough countries and in at least one with the standing to convene the process; we’re working towards generating the public buy-in and the organised constituency (including in Asia, if we get the support we need to do our work) that will underpin the political will necessary for governments to sign such an agreement and to successfully call other countries to join them.
You can read our perspective on the technical governance mechanics of a pause treaty here. Pausing general-purpose AI development only works if systems are not simply trained outside the treaty’s jurisdiction, which requires international binding coordination.
The relevant wider context is probably already legible to most readers, so we’ll just cite a couple of elements as calibration data on our broader outlook. Some factors are likely going to make societal coordination for robust international governance increasingly hard the more we wait, for instance:
Many countries, including major powers, are deploying AI into public and critical infrastructure, defence included. If that trend continues, acting against the immediate interests of the companies supplying the systems that run a country’s backbone will become extremely difficult.
States and political parties are beginning to lock in polarising narratives, predicated for instance on expected economic windfalls from AI development and infrastructure.
The frontier AI industry lobby is already significantly influencing policymakers. The arrival of powerful AIs will concentrate power in the hands of AI companies in a way that will make them harder to oppose. We expect AIs to be used to shape the public narrative, spread FUD (Fear, Uncertainty and Doubt), create false flag narratives, etc.
Already progressing gradual disempowerment might lock us into a bad future timeline.
We think now is the time to leverage the fact that the general public and leaders alike are beginning to register the enormity of the problem, and that the growing awareness and tangible challenges the world is starting to experience are going to make our work easier. The leaders of frontier labs, the Vatican, high-level representatives in the US and beyond, and a growing share of the public have all been questioning the wisdom of the race in their own register. In the past six months alone, the dispute between Anthropic and the US Department of War over military red lines put a frontier lab’s safety commitments on front pages; Pope Leo XIV devoted his first encyclical, Magnifica Humanitas, to AI and human dignity; Anthropic publicly stated that “it would be good for the world to have the option to slow or temporarily pause frontier AI development”, and Mythos is still generating intense turbulence in various directions.
A classic objection to the pursuit of a pause claims that China is set on racing, but we should be careful when treating the race narrative as descriptive; it is partly a self-fulfilling story. US hawks have long used the “do you want China to win?” argument to justify America and its labs scrambling unchecked towards ASI, incentivising China to respond in kind defensively. In 2024, a CCP plenum called for “oversight systems to ensure the safety of artificial intelligence”.[7] In 2025, China’s AI Safety Governance Framework 2.0 added loss of control and CBRN misuse to the risks it tracks. As of this month, Beijing is reportedly preparing a $295 billion plan to build a data centre network. However, Chinese expert and official opinion is markedly more safety-conscious than the race framing assumes: leading Chinese scientists have repeatedly co-signed international statements warning of catastrophic and even existential risk.[8] The Chinese public, for its part, pairs high general optimism about AI with sharp concern about its specific catastrophic dimensions, even though headline sentiment may seem optimistic.[9] China also has its own reasons to prefer a world in which frontier development is verifiably constrained.
We think one of the more valuable things our global federation can do over the coming years is to build chapters and partnerships across Asia and to put public, civic weight behind the existing evidence that the people of the countries with a significant role to play in pausing GPAI development would rather their governments coordinate than gamble. Different jurisdictions present different opportunities and constraints for grassroots organising, but they each need to trust the tractability of the other players signing. By being visibly present across key jurisdictions, we aim to weaken the corrosive story whereby “they” will never stop so “we” cannot. We have not yet built an organised presence in Asia, though we are already building contacts and scoping out our expansion in that part of the world.
A federation operating across several continents is already the right shape for this task, since it can produce coherent, synchronised political pressure across jurisdictions. A global pause needs a government with the clout to convene the process (the UK’s role at the Bletchley Park AI Safety Summit is an example of what that looks like). And it needs enough other governments whose domestic politics have shifted so that convening it is worth their while. Previous international moratoria, such as the one on human cloning, were reached without a strong political constituency behind them. Here, though, the existing incentives for companies and governments to press on are so strong, and the stakes so high, that a constituency is needed to counter the friction those incentives create against reaching and enforcing a treaty.
While we’re operating under fearsome constraints, we also believe we are at a point where, with modest additional resources—especially by the lights of the resources invested in AI safety at large—we can deploy at scale and build the civic and social movement that may prove decisive for humanity to steer clear of AI x-risk.
4. How we can grow fast without breaking
Under a broad range of plausible timelines, we aim to scale as fast as possible while keeping the movement coherent and on-mission as it grows. This rests on two things: compounding growth, and a set of design choices that keep fast growth from breaking the movement. Separately, several broader dynamics we don’t control could accelerate it, provided we are ready for them. What follows is a high-level overview of these mechanisms. We hope to publish a follow-up post to lay this out in greater detail.
Given the uncertainty around ASI timelines, our goal is to scale as fast as possible while keeping the movement coherent and on-mission as it grows. But when we say “scale,” we don’t just mean a bigger headcount or a bigger turnout at protests. A successful movement needs to scale the way a body grows, i.e. not by piling on more identical cells, but by differentiating. People take on roles; groups form and the healthiest of them seed new groups; chapters coordinate with one another and with the centre; and the same basic unit, a few people who hold the core and can act on it, recurs at every level from a single volunteer up to the whole federation. For the reasons below, we expect that what lets growth turn into durable political power rather than a brief spike of attention is that investment in structure.[10]
The primary driver of our approach to growth is recursive organising: organisers who train the next organisers, so capacity multiplies rather than depending on central hiring. A chapter lead’s job is not to do the work for everyone, but to teach the people around them to do it and to help the best of them go on to lead in turn. This is an old idea in organising. Ella Baker built parts of the US civil rights movement this way. Under this principle, groups develop their members’ capacity, rather than just mobilising them for a one-off turnout. This can build deeper and longer-lasting power (Han, 2014). We can make that multiplication easy and rewarding by building infrastructure that lowers friction at every step: onboarding process, training, community spaces (online and in person), guides and toolkits, microgrants, high-quality knowledge repositories and coordination tools, mentoring by senior organisers. Such a process slows down as it approaches the total size of the pool it draws from, but we are nowhere near that point. Concern about AI is already widespread and almost entirely unorganised, so what limits us is not a shortage of willing people but our capacity to absorb and develop them.
Social movement research demonstrates that recruitment runs largely through existing relationships rather than cold outreach: people are far more likely to join when someone they already trust has (Snow, Zurcher and Ekland-Olson, 1980). This holds most strongly for higher-commitment participation. When Doug McAdam (1986) studied who actually showed up for Freedom Summer, a high-risk campaign everyone who applied already believed in, social ties to other participants and prior organising experience, rather than strength of conviction, were the primary differentiating factors between people who did join and those who withdrew after applying.[11] People brought in through recursive organising also tend to stay, because we give them real ownership of clearly bounded work and try to make sure they leave meetings and actions more capable and more connected than when they arrived, rather than handing them a problem with nothing to do about it.
Investing in recursive organising, as opposed to merely focusing on broadcasting, offers another benefit. Getting someone to take a costly or controversial public stance is not something mass exposure achieves on its own; that kind of behaviour change needs reinforcement through repeated contact across overlapping trusted ties, a well evidenced dynamic called complex contagion (Centola, 2018). Our volunteers carry the “catastrophic AI risk + pause” frame into their own networks as trusted messengers with shared language and context, which a broadcast campaign cannot do. Lighter, low-commitment actions can still spread the ordinary way, through weak ties and broad exposure; deeper public commitment relies on the trusted-tie route.
Fast growth, though, is a liability if the movement drifts, polarises, or fragments as it spreads. It also stops being fast if the replication mechanism degrades through generations. To protect the integrity of our movement’s core as it propagates, we rely on two structural choices. The first is the federation: chapters with real local autonomy, each bound to a small set of foundational documents they can build on but not contradict. A useful analogy is that of an octopus’s nervous system. Most of an octopus’s neurons sit in its arms rather than its central brain, so each arm can sense and act on its own, with a high degree of contextual precision and responsiveness, without routing every decision through the centre.[12] The global team is the central brain, each chapter is an arm that responds to its own local terrain, and each group is a segment within that arm. The structure balances top-down and bottom-up power: pure top-down organising has no grassroots legitimacy and is costly to run, while pure bottom-up organising lacks effectiveness and coherence and is easy to capture. We describe the concrete governance levers that instantiate this (revocable brand licence, charter, code of conduct…) in Part III. The second choice is nonpartisanship and breadth: reaching critical mass in two to three years requires a broad church, so we choose language, alliances and partnerships that hold across the mainstream political spectrum, around a frame we believe unites people: that humanity faces a coordination failure threatening its self-determination and freedom, driven by a handful of labs caught in a race to the bottom, with the answer being peaceful, democratic engagement grounded in good epistemics.
Beyond our own organising, several broader dynamics could accelerate the growth described above. We don’t control these, but by being prepared for their arising, we can benefit from each of them:
Latent concern becoming visible: a great deal of worry about AI is privately held but rarely voiced, because people who assume their view is unpopular keep it to themselves. As voicing the concern becomes visibly safe and normal, a self-reinforcing preference cascade can start, in which each person who speaks up makes it easier for the next one;[13]
Windows of opportunity: moments of political flux, the “moments of the whirlwind” of our theory of change,[14] which a prepared movement can convert into a step change in membership and pressure (only if the standing capacity to seize them is already in place).
A shift from allyship to self-interested movement: existential AI safety may currently read to much of the public as an abstract cause pursued on behalf of future generations (an ally movement[15]) but, as people come to see that it also protects their own immediate interests, we expect it to scale better, in line with the broad historical pattern whereby “self-interested” movements outperform ally movements;
We don’t expect every lever to fire, but we expect some workable combination of them to arise in most timelines, and we hold the whole model as a working version, to be upgraded as we learn and bring in deeper expertise. The next section outlines the failure modes which this still leaves open and our safeguards against them.
5. Failure modes
Our theory of change rests on a set of assumptions and levers, some workable combination of which needs to hold at least to some extent for the movement to be highly effective, and it faces a range of failure modes, listed in this section. We have designed our structure largely to prevent or mitigate those failure modes.
We plan to share more about the failure modes we’re tracking and our prevention/mitigation approaches in upcoming content. We generally think structure is a social movement’s most effective protection against failure. The shape of our infrastructure and culture largely sets the bounds of our movement as it develops and we aim to leverage this natural bounding function by design. Here are the core pitfalls we’re aware of and investing in avoiding.
Failure mode
How it happens
Leading indicator
Mitigation/tripwire
Reversibility
”West pauses, China wins”
Movement reads as unilateral and Western, feeds the race narrative
Our own materials quoted in service of the race frame
Our ask is a binding internationaltreaty, never a unilateral pause; Asia presence strategy; evidence that publics in key jurisdictions prefer coordination
Reversible
Luddite-coding
We get read as anti-tech/anti-progress rather than pro-safety
Coverage framing us as Luddites; supporters self-describing in anti-tech terms
Message discipline around a positive, solution-oriented ask; explicit non-association with general anti-AI sentiment
Reversible, but slow
Capture/drift
Movement is pulled off-mission or co-opted as it grows
Chapters or messaging diverging from core values; entryism
Federation charter, brand licence revocable by the centre, code of conduct, messaging routed up, supply of approved training material and messaging guidelines
Reversible by design
Violence association
Violent or illegal factions attach to us, destroying legitimacy
Any faction advocating or using violence under our banner or adjacent to us, including false flag operations/sabotage
Nonviolence as a hard principle: consistently peaceful, solution-oriented messaging and continuously making the ethical and strategic case for it so the norm is internalised. Tripwire: we publicly disassociate from any violence-advocating actor and refuse to ally with them.
Reversible if acted on fast; catastrophic if not, hence the hard line
Crowding out
We absorb funding/attention that would do more good elsewhere
Hard to observe directly (counterfactual). Proxies: whether funding we raise is drawn from the same pool as high value work; whether our cost per unit of impact holds up as we grow; whether we or trusted actors point us at a concern that a better intervention is relatively less well resourced than we are.
Prevention: we complement (vs substitute to) the rest of the field; we try to broaden the funding pool, drawing in support that wouldn’t otherwise reach AI safety; high cost-effectiveness standard; diverse funding sources; regular critical-path checks; trusted advisers. Mitigation: future funding can move; we can scale back, hand work to others, or merge with peers.
Reversible
III. Our path so far and where we’re headed
Despite the challenges of bootstrapping from minute resources to build a grassroots organisation, we’ve achieved meaningful outcomes and demonstrated skilful calibration on important bets. We’ve already done some of the work to deliver a social movement that can materially mitigate catastrophic AI risk. At this point, we’re building infrastructure and scaling up while trying to achieve escape velocity with regards to our capacity.
1. Bootstrap phase (2023-2025)
In its initial phase, PauseAI grew from a single founder to a global organisation, training generations of volunteers who are now training the next generations; each streamlining stage increased the probability of success of the next steps.
PauseAI began as a scrappy, founder-driven volunteer network. It is now a professionally run federation with the foundations for a robust, scalable infrastructure. The first years produced far more achievements than we have room to list here, from the French chapter running the largest French-speaking AI safety conference to date to the UK chapter issuing an open letter to Demis Hassabis which 60 policymakers signed and which was covered extensively inTIME.[16] We see this as evidence that latent organising capacity exists. Concerned citizens can sustainably deliver remarkable outputs if we make a modicum of coordination available to them.
In 2023, Founder Joep Meindertsma started building PauseAI via a Discord server. The global chapter started recruiting its first members online and running weekly group calls. Single volunteers started groups in their countries, flyering or reaching out to policymakers. Most volunteers were new to organising and advocacy. These local groups then started to grow into national chapters, starting once again with weekly calls and Discord servers. Concurrently, PauseAI received enough funding to hire two contractors, who helped deliver our first organising conference, PauseCon 2025. There, dozens of volunteers received advocacy and organising training from world-class experts, learnt about technical AI governance, went out flyering, held a protest, created social media posts and discussed strategy.
At that point, volunteers from different chapters (UK, France, Germany, Australia) had achieved various degrees of progress: some had already talked to policymakers, organised a protest and run a small organisation for half a year. Others had only started recently and still felt awkward even talking to their families about AI risk. As those volunteers skilled up, over the following months, new volunteers started their journeys. Leaders ran their first campaigns and trained the second and third generations of volunteers, who would go on to found, or were already running, other national chapters. Chapters continued to grow in the UK, France, Germany, Czechia and Australia; others started in Romania and Spain, launching their first or second meaningful initiatives.
As an example of this process, here’s how local groups are founded in Germany. A local leader emerges organically among volunteers or through outreach to EA/AI safety communities. A full-time local group management volunteer plans the first event with local volunteers, travels to their city, helps them with flyering and online outreach for the event; the event is run. The local group gets integrated into a monthly local group call and a KPI matrix. The national chapter ensures the local group is stable even if its leader has to leave. Topics and actions for meetings are provided for local groups if desired, though often local volunteers run their own experiments, which we can then repeat if successful.
Each of the steps leading to our current stage of development could have failed and do commonly fail, especially in volunteer-led organisations. The first people founding PauseAI could have given up after a few months. There needed to be a first, highly motivated volunteer or a few volunteers to start a chapter, and someone to guide them through the initial stages. As chapters or local groups started, one or two main volunteers being busy for a while or moving away could have ended them, and sometimes did. At each of these early stages, when small circumstances had a big impact, the growth process was a bit of a random walk. We’ve tried some things that haven’t worked out as hoped, such as our first team retreat, which didn’t generate clear outputs—we now run our events with more focus on being clear as to the value we seek to create and how to design the event to deliver that value optimally. We’ve learnt by doing and built our movement’s architecture from extensive contact with the terrain.
2. New leadership, professionalisation and federation
PauseAI has now professionalised its leadership and operations and adopted a federated structure that keeps mission, message and values coherent from the global team down to local volunteers, while leaving chapters the autonomy to act with attunement to their own context.
Over the past 6 months, PauseAI Global has transitioned from an organically grown network into a professional organisation: a paid core team with defined roles, a leadership structure and a board to which it reports. Joep Meindertsma, who founded PauseAI, sits on that board; Maxime Fournes (co-writing this) now leads the organisation as CEO, working with organising and communications directors to coordinate with national chapter leadership in several countries. Our current structure is publicly viewable here. The earlier structure, where a large leadership team of around twenty people, with varying subsets attending any given meeting, shaped direction, has been replaced by a small professional executive team and an advisory council through which senior volunteers and chapter leads can give input. Decision making has moved out of a sprawling Discord server into a database-first Notion workspace, under a simple operating rule: if it is not written down, it does not exist. A professional CRM is being set up; it will carry contact-and-response logging, allowing us to track valuable metrics.
The organisation is built as a federation: a network of local groups and chapters that share a common mission, a common set of values and a common message, while running their own work with local autonomy. Every chapter adopts a small set of foundational documents—our core values, a stable mission statement, our public positions, shared messaging guidance and a code of conduct—and is free to build on top of them in whatever way fits its context, but not to contradict them. The global team maintains those foundations, coordinates international campaigns and protects the brand; chapters decide almost everything else for themselves.
This structure affords us functional leverage to keep the movement cohered and on the critical path: chapters operate under a brand licence agreement the global organisation can revoke. We are setting up the federation backend such that the shared tools, channels and infrastructure, including the domains and social media channels chapters depend on become fully ours to extend or withdraw (this is not complete yet). To keep the movement’s messaging consistent, local groups route regional media engagement to their national chapter teams; likewise, important or national press requests get routed up to the Communications Director. We don’t hold that leverage to micromanage chapters but so the movement can grow quickly without its core values mutating as it spreads. We also make membership boundaries clear through a code of conduct that every member signs.
3. Recent outputs
Our foundations are now mostly in place and our recent achievements (the world’s largest AI safety protest, a conference at the European Parliament, a coordinated campaign across 12 national groups, growing our presence past 15 countries, etc.) suggest these foundations are fit for purpose; we now need to scale them by building more and bigger chapters, a fuller training pipeline, and enough standing capacity to be responsive to external events with the potential to generate the momentum for a shift in the public debate, and whether we can do so now comes down to resources.
As an overview of our infrastructural “backend” work since the leadership transition we underwent in December 2025:
We hired and onboarded an executive team and several country organisers, deliberately recruiting experienced movement builders;
We wrote and adopted ourFederation Charter, the backbone designed to keep the movement aligned as it grows;
We completely overhauled our knowledge and project management systems;
We deployed Activoice, a grassroots lobbying tool enabling citizens to send personalised emails to representatives in under a minute, with each chapter running its own list of target politicians. The first Activoice campaign gave us our first actual participation numbers; we previously had no way of tracking them.
We ranPauseCon Brussels, which sent around eighty people from fifteen countries home with stronger organising skills and a meaningful network. Within weeks, alumni had founded PauseAI Belgium, launched the largest AI safety events held in Czechia this year, and started a volunteer-lawyer initiative exploring litigation over governmental negligence on AI risk, and PauseAI Germany gained another full-time organiser;
We built the first version of our rapid-response system and fired it once, when Claude Mythos was released. Numerous volunteers commented that they’d found our rapid-response broadcast useful to understand and communicate the issue in their local and online networks.
We’re also tracking the potential to reap the benefits of resources built “on the ground” by individual chapters, then generalised if the proof of concept succeeds, as this would allow different parts of the federation to accelerate their capacity gains.[17]
The most visible “frontend” (public-facing) outputs of early 2026 were largely planned before the leadership transition and federation model rollout and delivered by chapters; the new organisation staffed, funded and connected them. Accordingly, the new backend’s potential as an impact multiplier is not yet traceable; the following frontend achievements are evidence of what we could deliver as the backend upgrades were underway:
The largest AI safety protest to date (London, February 2026), where hundreds marched from OpenAI’s offices to Google DeepMind and Meta.[18] This was co-organised with organisations focused on present-day AI harms. At the closing citizens’ assembly, attendees who had joined out of concern for current harms converged on the pause ask of their own accord. The march was covered by the BBC and MIT Technology Review. Numerous international media outlets used protest content featuring placards demanding a pause to illustrate articles reporting public concern over AI (Wall Street Journal, The Independent, The Guardian, Taipei Times, Politis etc.). Each such prominent association between the broad topic of public concern about AI and evidence of a large group of citizens specifically demanding a pause helps crystallise the mainstream narrative towards a pause as concerned citizens’ go-to ask;
A conference at the European Parliament (Brussels, February 2026), hosted with MEP Ondřej Kolář, with a keynote from Stuart Russell and panellists spanning four political families. Eight MEPs attended, as well as aides and journalists. As well as building legitimacy, relationships, substantive shareable content, and press coverage, the conference led Brando Benifei, co-rapporteur of the EU AI Act, to suggest extending the GPAI Code of Practice towards conditional restrictions on development beyond capability thresholds;
Our first coordinated federation campaign, which focused on the India AI Action Summit. It involved twelve chapters and yielded thousands of emails to representatives across fourteen countries and thousands of petition signatures. This was our first attempt at a coordinated campaign across the federation. We consider it a successful test of our infrastructure, which we’re continuing to build on;
Our second coordinated campaign, “AI is not just coming for your job”, currently running across chapters. It uses near-term employment concerns as an accessible entry point into the deeper conversation about catastrophic and existential risk, reaching beyond people already concerned about AI safety. France’s execution illustrates what a strong chapter can achieve: a public survey relayed by Le Monde, an op-ed published in Le Nouvel Obs, an estimated 1,000+ emails sent to targeted decision-makers via Activoice, three simultaneous demonstrations across French cities, six partner organisations brought on board, and over seventy testimonies collected and repurposed as content.
Growth: the federation now counts more than fifteen national groups, and around two hundred people apply to volunteer every month. PauseAI France, our first self-sustaining chapter, shows the trajectory a chapter can achieve. It has more than a hundred paying members, staff on its own payroll, and has launched ten new local groups since December. In the same period, on top of executing the two global campaigns, it ran a municipal elections campaign collecting AI safety pledges from candidates, which drew national press coverage, and independently ran a campaign around the G7 summit which it packaged as a turnkey kit for other chapters to reuse;
As well as briefing over 40 members of the Bundestag on existential AI risk, our German chapter autonomously issued a public statement calling on the government to support red lines, which 152 professors signed. The statement was covered in Die ZeitNewsletter, various online publications and several national radio mentions.
IV. Support us
1. Fund us if you can
We believe our effort to build a civic and social movement for existential AI safety is tractable and likely decisive to humanity’s survival past the next few years, and we’ve already delivered the foundations for that movement, so funding our work is one of the highest-EV contributions to AI safety, both in relative and absolute terms; we are currently bottlenecked on money and our runway ends in four months (October 2026).
We are building the world’s infrastructure for a civic constituency and social movement targeting a pause on advanced AI development. The foundations above are our evidence that, done carefully, this works. We need the resources to deliver the most efficacious, expansive version of this work, i.e. building a broad and agentic AI safety constituency and shifting societal norms away from tolerating continued frontier AI development by nurturing buy-in across volunteers’ local communities throughout the world.
The rest of the ecosystem cannot substitute its work for ours and vice versa. We need each other. Researchers and briefers can raise what is politically possible, but not manufacture the organised public that makes it politically necessary. Without that organised public, the concern rising worldwide is likely to find more diffuse or more hostile outlets, which may make caring about AI safety appear socially radioactive rather than reasonable. This risks compromising the entire existential AI safety movement by foreclosing the peaceful, coordinated path and leaving mostly chaotic ones, which is not a risk worth running on stakes this high.
People’s engagement with their representatives, and the trust and shared context that underpin their ability to elicit awareness and support across virtually their entire social network, cannot be bought. However, money does buy the infrastructure that supports that relational work, from local groups and community spaces to training and tools, and the chapters that carry this infrastructure. The binding constraint on that, right now, is clearly funding rather than demand. At the federation scale, newcomers are already arriving faster than we can take them in, though some chapters do currently have enough capacity to cope.
The core elements of our model are designed to yield compounding value at a sufficient pace to make a difference within a range of plausible timelines to AGI/ASI: each organiser we train can train others, each chapter can seed the next, and the relationships that carry the message are tended to by the citizens inside the movement rather than being “rented”.
The value in resourcing PauseAI is highest now, given that the standing capacity to transform a sudden moment of opportunity into a regulatory shift must be fully operational by the time the event arises. With support, we’re confident, based on what we’ve been able to produce so far, that we can deploy that capacity expansively in a matter of months and keep building from there so we can absorb and channel much higher citizen advocacy and community engagement energy levels.
Our view is that building a civic and social movement for existential AI safety is tractable and badly under-resourced. We believe we’re building capacity very efficiently. The outcome that is a social movement capable of increasing the range of timelines in which we get to a pause before hitting a point of no-return to s/x-risk and of protecting the enforcement of a pause by making the societal cost of violating it unattractive can’t be bought any other way.
So how much funding do we need, and how would we use it?
At $300k per year: survival. This is the lower limit at which we are able to keep the executive team running and continue operating, including organising two PauseCon conferences per year (although their capacity would not scale).
At ~$1M per year, we can scale according to our plan. We complete the team, hiring an operations director, two new regional organisers, a policy director to make our asks rigorous and orchestrate grassroots lobbying campaigns, and a development/fundraising director so that the organisation is never again one grant away from closing. PauseCon conferences scale up proportionally to our growth, and we can provide meaningful seed funding for new chapters and support for existing ones. We focus on the development of new chapters in Asia.
At ~$3.5M per year, we can go further on every front. We expand the team: five regional organisers instead of two, senior public affairs staff based in key jurisdictions like the UK and EU to carry a coordinated political ask in several countries at once, a proper content team, and a chief of staff whose job is to make sure the organisation doesn’t break as it grows this fast. We run our first paid campaigns, rather than relying on earned media alone, to reach the specific communities we want to target. Chapter seed funding and ongoing support grow roughly tenfold, and PauseCons grow into much larger events.
If you cannot support us substantially but want to make a smaller donation (one-off or recurring), please do so. We hope this post has convinced you that marginal contributions to PauseAI are high-EV; this is true if you have modest amounts of money to invest as well as if you have large sums at your disposal.
2. What you can do if you can’t fund us
You can also usefully contribute to our work in non-financial ways.
Since the most impactful way to help us right now is to fund us, the second most impactful thing you can do is to get the word out about us to potential funders.
If you have notable skill points in relevant areas and would like to help us refine various aspects of what we’re building, do reach out.[19] We’re a skilled, experienced specialist team, and the ceiling for what the organisation we’ve set out to build can achieve is very high, which means we could usefully absorb more expertise from external advisers.
Maxime here. I thought I would close this post on a more personal note.
In 2023, after a ~10-year career in deep learning, I became convinced that the race to build ASI was the most severe and urgent danger of our time. I spent about six months trying to work out where I could have the most impact, and I arrived at that answer: PauseAI. So I gave it two years, full-time and unpaid, most of it spent building PauseAI France and making a dent in France’s AI public discourse, which remains the work I am proudest of.
Three years on, I am more convinced than ever that it was the right call. We have tried in this post to lay out the reasoning behind that conviction. But the single most convincing reason is dead simple: PauseAI has operated, across the world, on a total budget of less than $600k since 2023. Let that sink in. Everything described in this post was executed on that budget.
I lack words to convey how neglected this path is. I am highly confident (80%+) that with a 10x increase in resources, we could build a large, aligned and effective movement in under two years. Whether that movement is decisive for reducing AI risk rests on the arguments we have made throughout this post. Conditional on the case being roughly right, this is, as far as I can tell, the highest value-for-money intervention available to anyone trying to reduce catastrophic risks from AI.
So, here is what I am asking, ordered by priority. If you can fund it, fund it. If you can help us find someone who can fund it, please do. And if you can’t, there are other ways to help, and we would be glad to hear from you.
Bibliography
Centola, Damon (2018). How Behavior Spreads: The Science of Complex Contagions. Princeton University Press.
Engler, Mark & Paul Engler (2016). This Is an Uprising: How Nonviolent Revolt Is Shaping the Twenty-First Century. Nation Books.
Han, Hahrie (2014). How Organizations Develop Activists: Civic Associations and Leadership in the 21st Century. Oxford University Press.
McAdam, Doug (1986). “Recruitment to High-Risk Activism: The Case of Freedom Summer.” American Journal of Sociology 92(1), 64–90.
McAdam, Doug & Ronnelle Paulsen (1993). “Specifying the Relationship Between Social Ties and Activism.” American Journal of Sociology 99(3), 640–667.
Moyer, Bill, with JoAnn McAllister, Mary Lou Finley & Steven Soifer (2001). Doing Democracy: The MAP Model for Organizing Social Movements. New Society Publishers.
Snow, David A., Louis A. Zurcher & Sheldon Ekland-Olson (1980). “Social Networks and Social Movements: A Microstructural Approach to Differential Recruitment.” American Sociological Review 45(5), 787–801.
Tilly, Charles (2004). Social Movements, 1768–2004. Paradigm Publishers.
Two psychological tendencies compound. The first is motivated reasoning, whereby we more easily reach conclusions we have an interest in reaching, and its corollary, motivated stopping, due to which, once we land somewhere comfortable to act on (“interesting, probably true, not my problem”), we stop looking for reasons to revise it (Kunda, “The Case for Motivated Reasoning”, Psychological Bulletin, 1990; Yudkowsky, “Motivated Stopping and Motivated Continuation”, 2007). The second tendency is reputational herding: deviating from one’s peers is punished asymmetrically, since an error made alone reads as incompetence while the same error made alongside everyone reads as bad luck, so the privately convinced stay publicly conventional (Scharfstein and Stein, “Herd Behavior and Investment”, American Economic Review, 1990; see also Keynes’s “beauty contest”, General Theory, 1936, ch. 12).
Given current uncertainty, we can model a few geopolitical scenarios unfolding. We’re describing this condition in broad enough terms to account for this range.
See the IDAIS “red lines” dialogues, whose Chinese signatories include Andrew Yao, Zhang Ya-Qin and Zhou Bowen. IDAIS consensus statements, https://idais.ai/ ; and Carnegie, “How Some of China’s Top AI Thinkers Built Their Own AI Safety Institute” (2025). Yao is a main convener of IDAIS, under which Chinese and Western experts have published a series of statements on existential risk from advanced AI. Carnegie Endowment for International Peace
On the optimism side: a November 2025 Edelman poll (reported by Al Jazeera) found 87% of people in China say they trust AI, against 32% in the US. On the specific-concern side: a small exploratory China–Germany study of AI-risk perception (Brauner et al., 2025, arXiv:2412.13841) found that even this comparatively optimistic Chinese sample rated “AI misused by criminals” at the maximum risk score — every respondent picked the top of the scale — and “AI more intelligent than humans” among its highest-rated risks. Broad enthusiasm about AI coexists with sharp concern about its catastrophic specifics.
We want to show you how we’re doing this at the level of mechanics, but we’ve had to accept that we can’t do that as well as give you the broad overview this post provides, because it would make this piece exceed most people’s context windows. So we hope to share more detailed blueprints in a separate post.
More specifically (McAdam and Paulsen, 1993), a social tie tends to lead someone to join a movement when joining supports their existing identity: a worried parent, a professional who takes the timelines seriously, or simply a sensible person acting on experts’ warnings rather than a crank or an alarmist. Because people hold many ties at once, some pulling the other way (e.g. colleagues who find the whole subject embarrassing), what matters is the net balance of affirming over discouraging ones. Our local groups and community support the identity-affirming side: a setting where caring about this is ordinary, which is what both recruits and keeps people. This is also why we try to stay mainstream-coded and locally respected, since the doomer caricature is an attempt to load the opposing side of that balance and raise the social cost of being known to care.
If you’re curious about octopus’ sensorimotor and cognitive configuration, we recommend Olson, C.S., Schulz, N.G. & Ragsdale, C.W. Neuronal segmentation in cephalopod arms. Nat Commun 16, 443 (2025). https://doi.org/10.1038/s41467-024-55475-5 and Flash T, Zullo L. Biomechanics, motor control and dynamic models of the soft limbs of the octopus and other cephalopods. J Exp Biol. 2023 Apr 25;226(Suppl_1):jeb245295. doi: 10.1242/jeb.245295. Epub 2023 Apr 21. PMID: 37083140.
An ally movement is a movement through which members use their higher capacity to stand up for the interests of third parties deemed marginalised/disempowered, as “allies” to them. This could be vulnerable members of their local community, non-human animals, future generations etc.
The article which broke the story established for the first time that Google DeepMind did not provide the UK AISI with pre-deployment access to Gemini 2.5 Pro. Notably, Google did provide AISI with pre-deployment access to Gemini 3 Pro a couple of months after the letter was published.
Our first example of this is a volunteer coordination platform which PauseAI UK built for itself and which we’re considering rolling out federation-wide. On this platform, volunteers can join or own projects (e.g. local events, research/content creation projects or lobbying visits), claiming clearly defined tasks and logging project updates. This tool:
allows volunteers to get involved based on availability, location and skills; to form teams; and to smoothly load the relevant context to be productive when they have time to contribute;
enables groups and chapters to do light project management and coordination in one place, which we expect to improve their effectiveness and resilience.
Examples: expertise in designing metrics and tracking methods that could be useful in our context; forecasting/modelling, especially if you have first-hand context on relevant parts of the world; press relations; public affairs; high-quality social movement research…
Existential AI safety needs an effective social movement. PauseAI is building it.
Link post
The existential AI safety community needs to take building a civic and social movement seriously as a core intervention. We believe this is a high-value, badly neglected approach to reducing catastrophic/x-risks from AI because it may significantly enhance the likelihood of governance efforts succeeding at keeping humanity safe.
Note: this post is about PauseAI, not PauseAI US, which is a distinct entity with a different leadership team and approach.
This post was written by Matilda da Rui and Maxime Fournes, with significant contributions from Benjamin Schmidt (PauseAI Germany co-lead).
Executive Summary
The existential AI safety community needs to take building a civic and social movement seriously as a core intervention. We believe this is a high-value, badly neglected approach to reducing catastrophic/x-risks from AI because it may significantly enhance the likelihood of governance efforts succeeding at keeping humanity safe. As far as we can tell, only one organisation is building this infrastructure across continents: PauseAI. This post lays out our reasoning and our track record, and makes the case that funding this work is one of the highest value-for-money contributions available to anyone looking to reduce AI risk.
Why don’t we already have a pause or strong controls on frontier AI? Multiple advocacy groups are communicating clear and convincing arguments for AI existential risk, and policy experts are putting forward comprehensive proposals. We need more of this work, but this work alone will not be enough, because one link is missing: what policymakers hear doesn’t align with what their incentives reward. They face no political constituency that cares about existential risk.
We saw this firsthand at our European Parliament conference, where an MEP told us that she agreed with us but could not act without public support. Building that constituency is what PauseAI offers to the AI safety ecosystem.
Why do we think that PauseAI can deliver this constituency? On remarkably thin resources ($600k since we started in 2023), we have built an organised presence in more than 15 countries and a community of thousands of supporters and volunteers. We organise civic engagement, build movement infrastructure, and generate media coverage to push the catastrophic risk frame, and a pause as a viable answer to it, into the mainstream. In the past six months, we’ve professionalised our operations, run our first campaign coordinated across 12 national groups, held a conference at the European Parliament, and run the largest AI safety protest to date (300 participants in London).
In a wide range of plausible timelines, we can organise a focused constituency fast enough to make a difference. And with access to sufficient resources (which would still amount to a tiny fraction of the current spending on AI safety overall), we can build a social movement with the power to reshape norms. But the capacity to convert a sudden window of opportunity into political change has to exist before the window opens. We are already part of the way there, but we are bottlenecked on funding, and our current runway ends in October 2026. If you’re interested in funding us, see Part IV.1, which includes a breakdown of funding scenarios.
Our goal, ask, and day-to-day work
Our ultimate goal is preventing catastrophe. Humanity should not build superintelligent systems until or unless we know how to keep them safe and under control; everything we do is in service of that outcome.
Our public ask—the demand our movement carries—is a pause on advanced general-purpose AI development, ultimately anchored in a binding international treaty. Of the levers currently available, we think it is the one that most directly buys humanity the time it needs to orient to this challenge, and it is simple and concrete enough to organise a movement around. But it is an instrument, not our end goal. If a better lever for preventing catastrophe emerged, we would back it. A movement, however, cannot rally behind a carefully hedged portfolio of policy options; it needs one clear, defensible demand.
Our day-to-day work instantiates our higher-level goal and ask: building the movement itself—the constituency and the community infrastructure that make any adequate policy politically reachable when it is proposed, and durable after it passes (we come to division of labour considerations in Part II.1). This is also why this post says comparatively little about the pause proposal itself and its merits.
We think the strongest objections to our work are these:
Public opinion is low-leverage on a decision like this.
We think this prior is miscalibrated. Visible, organised constituencies change what policymakers can afford to do, even when they’re already convinced on the merits. We make this case in Part II.2.
Building a movement powerful and aligned enough, on short timelines, is intractable.
We’ve done a meaningful share of it already, on very little money. Part III outlines our track record.
A mass movement will drift, polarise, or turn anti-AI, and end up net-negative.
This could happen if we didn’t plan accordingly. Much of our structure and strategy is set up to prevent it. We explain how in Parts II.4, II.5, and III.
A pause is not achievable
What we are fundamentally building is an organised constituency that cares about catastrophic AI risks, and that is net positive for any serious safety policy, not just a pause. The pause is the ask we organise around because a movement needs one clear, defensible demand. Besides, “a pause is impossible” is a self-fulfilling belief: the conviction that the race cannot be stopped is itself a major reason no one tries. We think this story can be broken. See Parts I.1, II.2, and II.4.
This is a long post, so we have prefaced each section with a summary of its contents in italics. We want this to be the beginning of an ongoing conversation. We welcome clarifying questions, constructive comments and peer input.
Introduction
A somewhat common view in the AI safety community seems to be that a social and civic movement for existential AI safety is not clearly useful, or that it is lower-EV than other interventions—something like “the value of getting the general public onboard with regulating advanced AI seems low” (either relatively or absolutely). A related vector of scepticism rests on the assumption that successfully building such a movement is unreasonably hard (“surely building a movement powerful, aligned and responsive enough to create sufficient political will to reach an enforceable international pause in the face of arguably unparalleled incentives for industry and states, on potentially short timelines and under various forms of uncertainty, with a general public that is only just starting to register the problem, all amounts to something humanity can’t pull off”).
We think both versions of this outside view are incorrect because their assumptions are miscalibrated. We claim that building an effective movement for a global pause on advanced AI development is both tractable and high-EV—and even probably decisive. This is why we set out to do this work. PauseAI’s mission can be summarised as follows: to create a global constituency capable of generating sufficient political will in order to obtain a robust pause on advanced general-purpose AI development by building scalable and resilient civic and social movement infrastructure.
To our knowledge, we are the only group building the worldwide infrastructure for a social movement supporting a focused political constituency and building deep public buy-in for a pause on frontier AI development. Everything we have achieved so far was accomplished on under $600k over 3 years of operation across more than 15 countries. To us, this is in itself a measure of how neglected this important intervention is.
We want to lay out the reasoning behind our claims and the bet we’re making. We need the wider AI safety ecosystem to support us in this—it’s a large and hard undertaking and we don’t have the resources to do this all alone at this point, despite the high efficiency we are capable of. We’ve built the foundations for this movement and we now need to shift gears to reach the next step in our growth.
This post provides an overview of our theory of change, approach, and track record so far, then outlines our role in the AI safety ecosystem and on the path to a pause, and finally indicates how you can increase our capacity such that we are more likely to succeed across plausible timelines. We also want to share with you the blueprints and foundations which we think make this approach tractable; for the sake of readability, we’ll be sharing those separately.
I. Our theory of change
Prologue
Today, the impact of frontier AI policy efforts is significantly constrained by a lack of the demand side (constituency interest) for the product they offer (policy briefings and proposals).
Imagine a policy expert from ControlAI goes to Brussels to brief a member of the European Parliament. The briefing is excellent: clear, well evidenced, convincing. The MEP listens, and by the end, she more or less believes it. Catastrophic risk from frontier AI is real, timelines are short, and something should be done.
And then she does nothing. Not because she was unconvinced, but because nothing in her world is pushing her to act, and a great deal is pushing the other way. Acting would mean spending political capital, picking a fight with a fast-growing industry her own government is courting for investment, and getting ahead of colleagues who would rather wait and see who ends up looking foolish. Against all that, the briefing was one good conversation. So she files it under “interesting, probably true,” and turns back to the issues her constituents do write to her about. If it really mattered, she tells herself, surely she would be hearing about it from the people who elect her. She isn’t, so she doesn’t.[1]
The gap we exist to close is right there, and it is not a gap in the argument. The argument was fine. What was missing was any immediate reason for her to act on it. Let’s now imagine the same meeting with one thing changed: for the past six months, this MEP has been getting steady, unmistakable signals that many of her constituents care about this, and care enough to remember it at an election. The same briefing now lands as something she can act on and defend to her colleagues, because acting is no longer politically expensive overall (there’s a tangible benefit to her acting on this issue). We didn’t need to improve the expert’s case, but we needed to change whether she could afford to say yes to it. We saw this first-hand at our Brussels conference in February, where MEP Saskia Bricmont agreed with our stance but asked for more public support.
1. The shape of the problem as we see it
For most actors, the race to AGI/ASI is a coordination problem that can be solved via a treaty that binds everyone, including those who wilfully or recklessly keep the race going without the world’s consent; this requires leveraging political power as a counterforce to the drivers of the race, and we are building this political power by catalysing a pause-focused worldwide civic and social movement.
Frontier labs are racing toward ASI without democratic buy-in, before anyone knows how to keep such systems under control, and before any governance regime adequate to manage the risks exists. [2]
We can break the issue down through a typology of actor profiles:
Some corporate and state actors, and most of the general population, are oblivious to catastrophic AI risk. The bottleneck with regards to those actors is awareness of the issue and of pausing development as a viable solution.
Some company leaders and policymakers think the race is inevitable and stay in it lest the outcome be worse if they dropped out of it—and we agree that pausing unilaterally is no long-term answer. Relatedly, the leverage which states hold over frontier labs is not fixed, and the more deeply those actors’ systems are enmeshed into public infrastructure, the costlier binding labs becomes. With this group, the ones that are stuck in coordination failure, the focus ought to be on showing to them that there is a way out of the race through cooperation, and on making that latter claim true by making a pause politically rewarding, putting in place the technical architecture of the pause, and preparing the ground for those leaders to sign up for it when their peers do. Likewise for members of the public who think there is nothing citizens can do: they need to see they can take meaningful action, especially when they organise effectively, and this requires building infrastructure to lower the friction for them to do so.
Other corporate and state actors are just reckless as to the risk and are prepared to gamble everyone’s lives for a shot at winning the race—these are the actors that our approach seeks to bind, since they can’t be expected to willingly desist from binding humanity into catastrophe.
We can also frame this in terms of the dynamics keeping the race running:
Political leaders and civil societies are largely not AGI/ASI-aware. If everyone believed what the frontier AI leaders believe, AI research would be shut down tomorrow. The public and its representatives cannot influence the course of history if they don’t know the race is on.
Labs underestimate the difficulty of aligning AGI/ASI or are reckless as to it, and do not worry enough about gradual disempowerment and concentration of power or what to align the AI to.
A widespread fatalism, the assumption that AI development cannot be stopped, stops people from acting. This dynamic is a powerful self-fulfilling prophecy. If people believe widely enough that the race can’t be stopped, that very inevitability becomes real. So one of our goals is to dispel it by showing people that they can take useful action and that organised civic agency can compound into enough political will to clear the bar for meaningful policy change. AI safety has lacked the deep societal buy-in and organised constituency that could enable good policy to be passed and robustly enforced.
2. Necessary conditions for reaching a pause
Pausing requires (1) a public mandate generating the political incentive to act towards a pause; (2) elite consensus that pausing is the way to go; (3) enough international political will to make defection unattractive; (4) a prepared technical policy framework and environment; and (5) focused, coordinated civic/social activation to make sure major events are converted into political momentum. We can provide (1), (3), and (5).
We think a specific set of conditions underpins the enactment of a pause. Concern, organised into a constituency which politicians can see and count[3], changes politicians’ incentive landscape, because contributing to tackling AI risk stops being politically costly. That shift enables sympathetic leaders to act, and lets some governments integrate AI safety into an established frame, such as national interest or national security. As more countries adopt governance measures, holding out grows more expensive and, provided the technical policy prep work gets done in parallel, by the time an opportune window opens, a binding international agreement is reachable. A single triggering event can set off what Mark and Paul Engler (2016) call a moment of the whirlwind: a sudden surge of mass participation, far beyond a movement’s standing base, that throws open a brief window in which the politically impossible turns achievable, but only for a movement already built and waiting to seize it. The end state we aim for is a treaty with teeth—a global moratorium on training the most capable systems, with verification mechanisms and a standing body to enforce it.
In other words, to be effective in scope, a pause needs to be an international agreement, and we think it requires five conditions:
(1) a public mandate visible enough that signing becomes politically rewarding rather than costly for the leaders who do it;
(2) sufficient elite consensus that pausing is the responsible position;
(3) enough political will in multiple major-power governments that non-cooperation is unattractive;[4]
(4) a technical and governance framework ready to execute;
(5) the capacity to convert a triggering window into political action, almost certainly event-driven, that breaks inertia.
Our work primarily enables conditions 1, 3 and 5: building a coordinated, multi-jurisdiction public mandate, making the politics of acting favourable for governments, and maintaining the standing capacity to convert opportune windows into political momentum. Conditions 2 and 4 are mostly the lane of policy organisations and credible experts.[5] While we cannot deliver a pause unilaterally, we can build the substrate that makes a pause possible when the other conditions are met.[6]
II. Our role towards a global treaty and in the AI safety ecosystem
1. Our niche within the ecosystem
Our work complements that of AI safety experts sounding the alarm from a scientific perspective, communications specialists who raise awareness over media channels, policy organisations who brief politicians and prepare technical proposals, and radical activists that stretch the Overton window; we think all those niches need each other to be effective.
Our work is one part of a larger ecosystem, in which different organisations do different jobs that reinforce one another rather than compete. A useful frame for this division of labour comes from social movement strategist Bill Moyer, whose Movement Action Plan (2001) argues that durable movements need four distinct roles filled at once and tend to stall when one is missing or crowds out the rest. Mapped onto AI safety, the roles fall out roughly as in the diagram below.
The Citizen is the population that understands the problem and treats concern about it as a normal, legitimate thing to hold. Communicators and science educators (Rob Miles, Tarbell fellows, etc) build that understanding, and credible experts such as Hinton, Bengio and Russell supply the seed for the elite consensus (condition 2 of our theory of change). Another important signal is elite defection in the form of researchers walking out of frontier labs over safety concerns. Under Moyer’s model, both experts and communicators fall under the Citizen role.
The Change Agent organises Citizen actors. It educates, builds organising capacity via grassroots structures and coalitions, and moves the majority over time. This creates:
a visible public mandate and the deeper societal buy-in that underpins it,
the political will which this mandate and buy-in catalyse, and
readiness to channel windows of opportunity into institutional change.
We are the layer that helps the latent Citizen role cohere into a primary driver of the political process.
The Reformer works inside institutions, providing briefings, drafting policy and organising policymakers. This is the lane of policy organisations such as ControlAI and MIRI, which develop frontier AI governance proposals. We don’t typically write policy. Our role is to build the political conditions under which proposed policy can pass and survive. We sometimes brief politicians directly when this supports our objectives and is easy for us to deliver, but this is not the core of our role.
The fourth role, the Rebel, forces the issue onto the public agenda from the margins, through opposition and defiance. Most readers’ default assumption probably places PauseAI here, because we run protests, as do Rebels, but we use them as signalling devices for the size and seriousness of the constituency rather than as a sign of anti-establishment rebellion; our protests are solution-oriented and mainstream-coded. The distinction is becoming strategically sharper, because as parts of the industry work to paint data centre NIMBY groups as saboteurs or foreign plants, a Change Agent coded as Rebel would forfeit its own credibility and make the Citizen role harder for everyone else to inhabit. The Rebel, in our context, is entities building on general anti-AI sentiment. We take seriously the responsibility of helping public concern cohere around a drift-proof movement focused squarely on catastrophic AI risk and pausing as a solution, and we are wary of the inflammatory and occasionally violence-infused ethos of the anti-AI space. We will not ally with actors whom we expect to foster violence or to employ tactics that undermine our longer-term objectives.
Moyer’s framework is a practitioner’s heuristic, not a law, but if it is even roughly right, the impact of the roles compound one another. We need an ecosystem in which each role tends to its own niche well, ideally coordinating with each other.
2. Policymakers need strong enough incentives to act
In order for policy efforts to bear fruit, decision makers’ incentives have to line up with their model of their own scope for action; a well organised constituency, buoyed over time by a vibrant pro-pause community spanning many societal segments, can clear the path and create the momentum for effective political action.
The story of the Brussels policymaker with which we opened this post highlights the gap which we aim to fill. A representative who privately agrees can still be pulled quietly back the moment industry lobbying leans on them, unless a visible movement is there to hold them to it. The frontier AI industry can afford dense, concentrated, well funded pressure. Left unorganised, widespread private concern constitutes virtually no pressure. In the civic engagement context, our Change Agent role is to focalise that diffuse beam into a laser-sharp demand.
A natural objection is that a movement large enough to register politically will inevitably contain many people who are not skilled advocates, and that a flood of unpolished letters may be unpersuasive. This misconstrues the role of a constituency. The argument is carried by the briefers, and the constituency demonstrates that voters are invested. Representatives don’t typically weigh mass constituent messages by their prose, but rather by the volume and level of engagement they signify, i.e. a mass of constituents who treat this issue seriously enough to take time out of their busy lives to identify whom they should write to and to think about what to say, and who will remember the issue at the next election. Our stance on this is based on social movement researcher Charles Tilly’s model of what makes a movement register with political leaders, namely a combination of worthiness, unity, numbers and commitment (2004). In empowering an organised constituency, the majority of the signal is carried by the following indicators: numbers, consistency around the problem-solution frame, and commitment. The “worthiness” component can be fulfilled by volunteers who speak to politicians directly, as they are typically seasoned and can make a more confident, articulate case than the average message sender. Further, lighter-touch tools let supporters register genuine concern in a minute without composing a case from scratch. The product of this work is a constituency politicians can actually see and count, to use Tilly’s framing.
Reputational incentives are also a constraint, and not one confined to elected representatives. The same economics weigh on decision makers and senior figures across influential professions, as we often hear from volunteers who work in fields like law and risk management. I (Maxime) observed this mechanism at close range when I worked in quantitative finance. A fund manager whose positions track their peers’ is oddly safe: when a bad year comes, it comes for everyone; “the whole market was down” is a defence clients accept, and the money stays put. A manager who is uncorrelated to the rest of the market lives under a harsher rule: if you win alone, you look brilliant, but if you lose alone, there’s nobody to share the blame, clients leave, and you go broke.
I remember a parable about picking a restaurant to impress a date: choose the place that’s packed but unremarkable over the one you’ve heard is wonderful but looks empty, because if the packed place disappoints, the crowd absorbs the blame, whereas if the empty place disappoints, you alone chose it. Being early and wrong costs far more than being conventional and wrong, even when the lonely call was the better one. That asymmetry is much of why private concern about AI stays private. This is valid for members of the public but a fortiori for politicians: being seen to hold a view while it still reads as unconventional is a risky bet. A visible, credible constituency changes that calculus; it fills the empty restaurant. Once enough serious people are openly concerned, voicing concern stops being the lonely, uncorrelated bet and becomes the safe, shared one, and aligning with an established movement carries far less risk than sticking your neck out alone.
3. The path to a treaty
International coordination rests on domestic political will, in enough countries and in at least one with the standing to convene the process; we’re working towards generating the public buy-in and the organised constituency (including in Asia, if we get the support we need to do our work) that will underpin the political will necessary for governments to sign such an agreement and to successfully call other countries to join them.
You can read our perspective on the technical governance mechanics of a pause treaty here. Pausing general-purpose AI development only works if systems are not simply trained outside the treaty’s jurisdiction, which requires international binding coordination.
The relevant wider context is probably already legible to most readers, so we’ll just cite a couple of elements as calibration data on our broader outlook. Some factors are likely going to make societal coordination for robust international governance increasingly hard the more we wait, for instance:
Many countries, including major powers, are deploying AI into public and critical infrastructure, defence included. If that trend continues, acting against the immediate interests of the companies supplying the systems that run a country’s backbone will become extremely difficult.
States and political parties are beginning to lock in polarising narratives, predicated for instance on expected economic windfalls from AI development and infrastructure.
The frontier AI industry lobby is already significantly influencing policymakers. The arrival of powerful AIs will concentrate power in the hands of AI companies in a way that will make them harder to oppose. We expect AIs to be used to shape the public narrative, spread FUD (Fear, Uncertainty and Doubt), create false flag narratives, etc.
Already progressing gradual disempowerment might lock us into a bad future timeline.
We think now is the time to leverage the fact that the general public and leaders alike are beginning to register the enormity of the problem, and that the growing awareness and tangible challenges the world is starting to experience are going to make our work easier. The leaders of frontier labs, the Vatican, high-level representatives in the US and beyond, and a growing share of the public have all been questioning the wisdom of the race in their own register. In the past six months alone, the dispute between Anthropic and the US Department of War over military red lines put a frontier lab’s safety commitments on front pages; Pope Leo XIV devoted his first encyclical, Magnifica Humanitas, to AI and human dignity; Anthropic publicly stated that “it would be good for the world to have the option to slow or temporarily pause frontier AI development”, and Mythos is still generating intense turbulence in various directions.
A classic objection to the pursuit of a pause claims that China is set on racing, but we should be careful when treating the race narrative as descriptive; it is partly a self-fulfilling story. US hawks have long used the “do you want China to win?” argument to justify America and its labs scrambling unchecked towards ASI, incentivising China to respond in kind defensively. In 2024, a CCP plenum called for “oversight systems to ensure the safety of artificial intelligence”.[7] In 2025, China’s AI Safety Governance Framework 2.0 added loss of control and CBRN misuse to the risks it tracks. As of this month, Beijing is reportedly preparing a $295 billion plan to build a data centre network. However, Chinese expert and official opinion is markedly more safety-conscious than the race framing assumes: leading Chinese scientists have repeatedly co-signed international statements warning of catastrophic and even existential risk.[8] The Chinese public, for its part, pairs high general optimism about AI with sharp concern about its specific catastrophic dimensions, even though headline sentiment may seem optimistic.[9] China also has its own reasons to prefer a world in which frontier development is verifiably constrained.
We think one of the more valuable things our global federation can do over the coming years is to build chapters and partnerships across Asia and to put public, civic weight behind the existing evidence that the people of the countries with a significant role to play in pausing GPAI development would rather their governments coordinate than gamble. Different jurisdictions present different opportunities and constraints for grassroots organising, but they each need to trust the tractability of the other players signing. By being visibly present across key jurisdictions, we aim to weaken the corrosive story whereby “they” will never stop so “we” cannot. We have not yet built an organised presence in Asia, though we are already building contacts and scoping out our expansion in that part of the world.
A federation operating across several continents is already the right shape for this task, since it can produce coherent, synchronised political pressure across jurisdictions. A global pause needs a government with the clout to convene the process (the UK’s role at the Bletchley Park AI Safety Summit is an example of what that looks like). And it needs enough other governments whose domestic politics have shifted so that convening it is worth their while. Previous international moratoria, such as the one on human cloning, were reached without a strong political constituency behind them. Here, though, the existing incentives for companies and governments to press on are so strong, and the stakes so high, that a constituency is needed to counter the friction those incentives create against reaching and enforcing a treaty.
While we’re operating under fearsome constraints, we also believe we are at a point where, with modest additional resources—especially by the lights of the resources invested in AI safety at large—we can deploy at scale and build the civic and social movement that may prove decisive for humanity to steer clear of AI x-risk.
4. How we can grow fast without breaking
Under a broad range of plausible timelines, we aim to scale as fast as possible while keeping the movement coherent and on-mission as it grows. This rests on two things: compounding growth, and a set of design choices that keep fast growth from breaking the movement. Separately, several broader dynamics we don’t control could accelerate it, provided we are ready for them. What follows is a high-level overview of these mechanisms. We hope to publish a follow-up post to lay this out in greater detail.
Given the uncertainty around ASI timelines, our goal is to scale as fast as possible while keeping the movement coherent and on-mission as it grows. But when we say “scale,” we don’t just mean a bigger headcount or a bigger turnout at protests. A successful movement needs to scale the way a body grows, i.e. not by piling on more identical cells, but by differentiating. People take on roles; groups form and the healthiest of them seed new groups; chapters coordinate with one another and with the centre; and the same basic unit, a few people who hold the core and can act on it, recurs at every level from a single volunteer up to the whole federation. For the reasons below, we expect that what lets growth turn into durable political power rather than a brief spike of attention is that investment in structure.[10]
The primary driver of our approach to growth is recursive organising: organisers who train the next organisers, so capacity multiplies rather than depending on central hiring. A chapter lead’s job is not to do the work for everyone, but to teach the people around them to do it and to help the best of them go on to lead in turn. This is an old idea in organising. Ella Baker built parts of the US civil rights movement this way. Under this principle, groups develop their members’ capacity, rather than just mobilising them for a one-off turnout. This can build deeper and longer-lasting power (Han, 2014). We can make that multiplication easy and rewarding by building infrastructure that lowers friction at every step: onboarding process, training, community spaces (online and in person), guides and toolkits, microgrants, high-quality knowledge repositories and coordination tools, mentoring by senior organisers. Such a process slows down as it approaches the total size of the pool it draws from, but we are nowhere near that point. Concern about AI is already widespread and almost entirely unorganised, so what limits us is not a shortage of willing people but our capacity to absorb and develop them.
Social movement research demonstrates that recruitment runs largely through existing relationships rather than cold outreach: people are far more likely to join when someone they already trust has (Snow, Zurcher and Ekland-Olson, 1980). This holds most strongly for higher-commitment participation. When Doug McAdam (1986) studied who actually showed up for Freedom Summer, a high-risk campaign everyone who applied already believed in, social ties to other participants and prior organising experience, rather than strength of conviction, were the primary differentiating factors between people who did join and those who withdrew after applying.[11] People brought in through recursive organising also tend to stay, because we give them real ownership of clearly bounded work and try to make sure they leave meetings and actions more capable and more connected than when they arrived, rather than handing them a problem with nothing to do about it.
Investing in recursive organising, as opposed to merely focusing on broadcasting, offers another benefit. Getting someone to take a costly or controversial public stance is not something mass exposure achieves on its own; that kind of behaviour change needs reinforcement through repeated contact across overlapping trusted ties, a well evidenced dynamic called complex contagion (Centola, 2018). Our volunteers carry the “catastrophic AI risk + pause” frame into their own networks as trusted messengers with shared language and context, which a broadcast campaign cannot do. Lighter, low-commitment actions can still spread the ordinary way, through weak ties and broad exposure; deeper public commitment relies on the trusted-tie route.
Fast growth, though, is a liability if the movement drifts, polarises, or fragments as it spreads. It also stops being fast if the replication mechanism degrades through generations. To protect the integrity of our movement’s core as it propagates, we rely on two structural choices. The first is the federation: chapters with real local autonomy, each bound to a small set of foundational documents they can build on but not contradict. A useful analogy is that of an octopus’s nervous system. Most of an octopus’s neurons sit in its arms rather than its central brain, so each arm can sense and act on its own, with a high degree of contextual precision and responsiveness, without routing every decision through the centre.[12] The global team is the central brain, each chapter is an arm that responds to its own local terrain, and each group is a segment within that arm. The structure balances top-down and bottom-up power: pure top-down organising has no grassroots legitimacy and is costly to run, while pure bottom-up organising lacks effectiveness and coherence and is easy to capture. We describe the concrete governance levers that instantiate this (revocable brand licence, charter, code of conduct…) in Part III. The second choice is nonpartisanship and breadth: reaching critical mass in two to three years requires a broad church, so we choose language, alliances and partnerships that hold across the mainstream political spectrum, around a frame we believe unites people: that humanity faces a coordination failure threatening its self-determination and freedom, driven by a handful of labs caught in a race to the bottom, with the answer being peaceful, democratic engagement grounded in good epistemics.
Beyond our own organising, several broader dynamics could accelerate the growth described above. We don’t control these, but by being prepared for their arising, we can benefit from each of them:
Latent concern becoming visible: a great deal of worry about AI is privately held but rarely voiced, because people who assume their view is unpopular keep it to themselves. As voicing the concern becomes visibly safe and normal, a self-reinforcing preference cascade can start, in which each person who speaks up makes it easier for the next one;[13]
Windows of opportunity: moments of political flux, the “moments of the whirlwind” of our theory of change,[14] which a prepared movement can convert into a step change in membership and pressure (only if the standing capacity to seize them is already in place).
A shift from allyship to self-interested movement: existential AI safety may currently read to much of the public as an abstract cause pursued on behalf of future generations (an ally movement[15]) but, as people come to see that it also protects their own immediate interests, we expect it to scale better, in line with the broad historical pattern whereby “self-interested” movements outperform ally movements;
We don’t expect every lever to fire, but we expect some workable combination of them to arise in most timelines, and we hold the whole model as a working version, to be upgraded as we learn and bring in deeper expertise. The next section outlines the failure modes which this still leaves open and our safeguards against them.
5. Failure modes
Our theory of change rests on a set of assumptions and levers, some workable combination of which needs to hold at least to some extent for the movement to be highly effective, and it faces a range of failure modes, listed in this section. We have designed our structure largely to prevent or mitigate those failure modes.
We plan to share more about the failure modes we’re tracking and our prevention/mitigation approaches in upcoming content. We generally think structure is a social movement’s most effective protection against failure. The shape of our infrastructure and culture largely sets the bounds of our movement as it develops and we aim to leverage this natural bounding function by design. Here are the core pitfalls we’re aware of and investing in avoiding.
III. Our path so far and where we’re headed
Despite the challenges of bootstrapping from minute resources to build a grassroots organisation, we’ve achieved meaningful outcomes and demonstrated skilful calibration on important bets. We’ve already done some of the work to deliver a social movement that can materially mitigate catastrophic AI risk. At this point, we’re building infrastructure and scaling up while trying to achieve escape velocity with regards to our capacity.
1. Bootstrap phase (2023-2025)
In its initial phase, PauseAI grew from a single founder to a global organisation, training generations of volunteers who are now training the next generations; each streamlining stage increased the probability of success of the next steps.
PauseAI began as a scrappy, founder-driven volunteer network. It is now a professionally run federation with the foundations for a robust, scalable infrastructure. The first years produced far more achievements than we have room to list here, from the French chapter running the largest French-speaking AI safety conference to date to the UK chapter issuing an open letter to Demis Hassabis which 60 policymakers signed and which was covered extensively in TIME.[16] We see this as evidence that latent organising capacity exists. Concerned citizens can sustainably deliver remarkable outputs if we make a modicum of coordination available to them.
In 2023, Founder Joep Meindertsma started building PauseAI via a Discord server. The global chapter started recruiting its first members online and running weekly group calls. Single volunteers started groups in their countries, flyering or reaching out to policymakers. Most volunteers were new to organising and advocacy. These local groups then started to grow into national chapters, starting once again with weekly calls and Discord servers. Concurrently, PauseAI received enough funding to hire two contractors, who helped deliver our first organising conference, PauseCon 2025. There, dozens of volunteers received advocacy and organising training from world-class experts, learnt about technical AI governance, went out flyering, held a protest, created social media posts and discussed strategy.
At that point, volunteers from different chapters (UK, France, Germany, Australia) had achieved various degrees of progress: some had already talked to policymakers, organised a protest and run a small organisation for half a year. Others had only started recently and still felt awkward even talking to their families about AI risk. As those volunteers skilled up, over the following months, new volunteers started their journeys. Leaders ran their first campaigns and trained the second and third generations of volunteers, who would go on to found, or were already running, other national chapters. Chapters continued to grow in the UK, France, Germany, Czechia and Australia; others started in Romania and Spain, launching their first or second meaningful initiatives.
As an example of this process, here’s how local groups are founded in Germany. A local leader emerges organically among volunteers or through outreach to EA/AI safety communities. A full-time local group management volunteer plans the first event with local volunteers, travels to their city, helps them with flyering and online outreach for the event; the event is run. The local group gets integrated into a monthly local group call and a KPI matrix. The national chapter ensures the local group is stable even if its leader has to leave. Topics and actions for meetings are provided for local groups if desired, though often local volunteers run their own experiments, which we can then repeat if successful.
Each of the steps leading to our current stage of development could have failed and do commonly fail, especially in volunteer-led organisations. The first people founding PauseAI could have given up after a few months. There needed to be a first, highly motivated volunteer or a few volunteers to start a chapter, and someone to guide them through the initial stages. As chapters or local groups started, one or two main volunteers being busy for a while or moving away could have ended them, and sometimes did. At each of these early stages, when small circumstances had a big impact, the growth process was a bit of a random walk. We’ve tried some things that haven’t worked out as hoped, such as our first team retreat, which didn’t generate clear outputs—we now run our events with more focus on being clear as to the value we seek to create and how to design the event to deliver that value optimally. We’ve learnt by doing and built our movement’s architecture from extensive contact with the terrain.
2. New leadership, professionalisation and federation
PauseAI has now professionalised its leadership and operations and adopted a federated structure that keeps mission, message and values coherent from the global team down to local volunteers, while leaving chapters the autonomy to act with attunement to their own context.
Over the past 6 months, PauseAI Global has transitioned from an organically grown network into a professional organisation: a paid core team with defined roles, a leadership structure and a board to which it reports. Joep Meindertsma, who founded PauseAI, sits on that board; Maxime Fournes (co-writing this) now leads the organisation as CEO, working with organising and communications directors to coordinate with national chapter leadership in several countries. Our current structure is publicly viewable here. The earlier structure, where a large leadership team of around twenty people, with varying subsets attending any given meeting, shaped direction, has been replaced by a small professional executive team and an advisory council through which senior volunteers and chapter leads can give input. Decision making has moved out of a sprawling Discord server into a database-first Notion workspace, under a simple operating rule: if it is not written down, it does not exist. A professional CRM is being set up; it will carry contact-and-response logging, allowing us to track valuable metrics.
The organisation is built as a federation: a network of local groups and chapters that share a common mission, a common set of values and a common message, while running their own work with local autonomy. Every chapter adopts a small set of foundational documents—our core values, a stable mission statement, our public positions, shared messaging guidance and a code of conduct—and is free to build on top of them in whatever way fits its context, but not to contradict them. The global team maintains those foundations, coordinates international campaigns and protects the brand; chapters decide almost everything else for themselves.
This structure affords us functional leverage to keep the movement cohered and on the critical path: chapters operate under a brand licence agreement the global organisation can revoke. We are setting up the federation backend such that the shared tools, channels and infrastructure, including the domains and social media channels chapters depend on become fully ours to extend or withdraw (this is not complete yet). To keep the movement’s messaging consistent, local groups route regional media engagement to their national chapter teams; likewise, important or national press requests get routed up to the Communications Director. We don’t hold that leverage to micromanage chapters but so the movement can grow quickly without its core values mutating as it spreads. We also make membership boundaries clear through a code of conduct that every member signs.
3. Recent outputs
Our foundations are now mostly in place and our recent achievements (the world’s largest AI safety protest, a conference at the European Parliament, a coordinated campaign across 12 national groups, growing our presence past 15 countries, etc.) suggest these foundations are fit for purpose; we now need to scale them by building more and bigger chapters, a fuller training pipeline, and enough standing capacity to be responsive to external events with the potential to generate the momentum for a shift in the public debate, and whether we can do so now comes down to resources.
As an overview of our infrastructural “backend” work since the leadership transition we underwent in December 2025:
We hired and onboarded an executive team and several country organisers, deliberately recruiting experienced movement builders;
We wrote and adopted our Federation Charter, the backbone designed to keep the movement aligned as it grows;
We completely overhauled our knowledge and project management systems;
We deployed Activoice, a grassroots lobbying tool enabling citizens to send personalised emails to representatives in under a minute, with each chapter running its own list of target politicians. The first Activoice campaign gave us our first actual participation numbers; we previously had no way of tracking them.
We ran PauseCon Brussels, which sent around eighty people from fifteen countries home with stronger organising skills and a meaningful network. Within weeks, alumni had founded PauseAI Belgium, launched the largest AI safety events held in Czechia this year, and started a volunteer-lawyer initiative exploring litigation over governmental negligence on AI risk, and PauseAI Germany gained another full-time organiser;
We built the first version of our rapid-response system and fired it once, when Claude Mythos was released. Numerous volunteers commented that they’d found our rapid-response broadcast useful to understand and communicate the issue in their local and online networks.
We’re also tracking the potential to reap the benefits of resources built “on the ground” by individual chapters, then generalised if the proof of concept succeeds, as this would allow different parts of the federation to accelerate their capacity gains.[17]
The most visible “frontend” (public-facing) outputs of early 2026 were largely planned before the leadership transition and federation model rollout and delivered by chapters; the new organisation staffed, funded and connected them. Accordingly, the new backend’s potential as an impact multiplier is not yet traceable; the following frontend achievements are evidence of what we could deliver as the backend upgrades were underway:
The largest AI safety protest to date (London, February 2026), where hundreds marched from OpenAI’s offices to Google DeepMind and Meta.[18] This was co-organised with organisations focused on present-day AI harms. At the closing citizens’ assembly, attendees who had joined out of concern for current harms converged on the pause ask of their own accord. The march was covered by the BBC and MIT Technology Review. Numerous international media outlets used protest content featuring placards demanding a pause to illustrate articles reporting public concern over AI (Wall Street Journal, The Independent, The Guardian, Taipei Times, Politis etc.). Each such prominent association between the broad topic of public concern about AI and evidence of a large group of citizens specifically demanding a pause helps crystallise the mainstream narrative towards a pause as concerned citizens’ go-to ask;
A conference at the European Parliament (Brussels, February 2026), hosted with MEP Ondřej Kolář, with a keynote from Stuart Russell and panellists spanning four political families. Eight MEPs attended, as well as aides and journalists. As well as building legitimacy, relationships, substantive shareable content, and press coverage, the conference led Brando Benifei, co-rapporteur of the EU AI Act, to suggest extending the GPAI Code of Practice towards conditional restrictions on development beyond capability thresholds;
Our first coordinated federation campaign, which focused on the India AI Action Summit. It involved twelve chapters and yielded thousands of emails to representatives across fourteen countries and thousands of petition signatures. This was our first attempt at a coordinated campaign across the federation. We consider it a successful test of our infrastructure, which we’re continuing to build on;
Our second coordinated campaign, “AI is not just coming for your job”, currently running across chapters. It uses near-term employment concerns as an accessible entry point into the deeper conversation about catastrophic and existential risk, reaching beyond people already concerned about AI safety. France’s execution illustrates what a strong chapter can achieve: a public survey relayed by Le Monde, an op-ed published in Le Nouvel Obs, an estimated 1,000+ emails sent to targeted decision-makers via Activoice, three simultaneous demonstrations across French cities, six partner organisations brought on board, and over seventy testimonies collected and repurposed as content.
Growth: the federation now counts more than fifteen national groups, and around two hundred people apply to volunteer every month. PauseAI France, our first self-sustaining chapter, shows the trajectory a chapter can achieve. It has more than a hundred paying members, staff on its own payroll, and has launched ten new local groups since December. In the same period, on top of executing the two global campaigns, it ran a municipal elections campaign collecting AI safety pledges from candidates, which drew national press coverage, and independently ran a campaign around the G7 summit which it packaged as a turnkey kit for other chapters to reuse;
As well as briefing over 40 members of the Bundestag on existential AI risk, our German chapter autonomously issued a public statement calling on the government to support red lines, which 152 professors signed. The statement was covered in Die Zeit Newsletter, various online publications and several national radio mentions.
IV. Support us
1. Fund us if you can
We believe our effort to build a civic and social movement for existential AI safety is tractable and likely decisive to humanity’s survival past the next few years, and we’ve already delivered the foundations for that movement, so funding our work is one of the highest-EV contributions to AI safety, both in relative and absolute terms; we are currently bottlenecked on money and our runway ends in four months (October 2026).
We are building the world’s infrastructure for a civic constituency and social movement targeting a pause on advanced AI development. The foundations above are our evidence that, done carefully, this works. We need the resources to deliver the most efficacious, expansive version of this work, i.e. building a broad and agentic AI safety constituency and shifting societal norms away from tolerating continued frontier AI development by nurturing buy-in across volunteers’ local communities throughout the world.
The rest of the ecosystem cannot substitute its work for ours and vice versa. We need each other. Researchers and briefers can raise what is politically possible, but not manufacture the organised public that makes it politically necessary. Without that organised public, the concern rising worldwide is likely to find more diffuse or more hostile outlets, which may make caring about AI safety appear socially radioactive rather than reasonable. This risks compromising the entire existential AI safety movement by foreclosing the peaceful, coordinated path and leaving mostly chaotic ones, which is not a risk worth running on stakes this high.
People’s engagement with their representatives, and the trust and shared context that underpin their ability to elicit awareness and support across virtually their entire social network, cannot be bought. However, money does buy the infrastructure that supports that relational work, from local groups and community spaces to training and tools, and the chapters that carry this infrastructure. The binding constraint on that, right now, is clearly funding rather than demand. At the federation scale, newcomers are already arriving faster than we can take them in, though some chapters do currently have enough capacity to cope.
The core elements of our model are designed to yield compounding value at a sufficient pace to make a difference within a range of plausible timelines to AGI/ASI: each organiser we train can train others, each chapter can seed the next, and the relationships that carry the message are tended to by the citizens inside the movement rather than being “rented”.
The value in resourcing PauseAI is highest now, given that the standing capacity to transform a sudden moment of opportunity into a regulatory shift must be fully operational by the time the event arises. With support, we’re confident, based on what we’ve been able to produce so far, that we can deploy that capacity expansively in a matter of months and keep building from there so we can absorb and channel much higher citizen advocacy and community engagement energy levels.
Our view is that building a civic and social movement for existential AI safety is tractable and badly under-resourced. We believe we’re building capacity very efficiently. The outcome that is a social movement capable of increasing the range of timelines in which we get to a pause before hitting a point of no-return to s/x-risk and of protecting the enforcement of a pause by making the societal cost of violating it unattractive can’t be bought any other way.
So how much funding do we need, and how would we use it?
At $300k per year: survival. This is the lower limit at which we are able to keep the executive team running and continue operating, including organising two PauseCon conferences per year (although their capacity would not scale).
At ~$1M per year, we can scale according to our plan. We complete the team, hiring an operations director, two new regional organisers, a policy director to make our asks rigorous and orchestrate grassroots lobbying campaigns, and a development/fundraising director so that the organisation is never again one grant away from closing. PauseCon conferences scale up proportionally to our growth, and we can provide meaningful seed funding for new chapters and support for existing ones. We focus on the development of new chapters in Asia.
At ~$3.5M per year, we can go further on every front. We expand the team: five regional organisers instead of two, senior public affairs staff based in key jurisdictions like the UK and EU to carry a coordinated political ask in several countries at once, a proper content team, and a chief of staff whose job is to make sure the organisation doesn’t break as it grows this fast. We run our first paid campaigns, rather than relying on earned media alone, to reach the specific communities we want to target. Chapter seed funding and ongoing support grow roughly tenfold, and PauseCons grow into much larger events.
To discuss funding, please e-mail contact@pauseai.info.
If you cannot support us substantially but want to make a smaller donation (one-off or recurring), please do so. We hope this post has convinced you that marginal contributions to PauseAI are high-EV; this is true if you have modest amounts of money to invest as well as if you have large sums at your disposal.
2. What you can do if you can’t fund us
You can also usefully contribute to our work in non-financial ways.
Since the most impactful way to help us right now is to fund us, the second most impactful thing you can do is to get the word out about us to potential funders.
If you have notable skill points in relevant areas and would like to help us refine various aspects of what we’re building, do reach out.[19] We’re a skilled, experienced specialist team, and the ceiling for what the organisation we’ve set out to build can achieve is very high, which means we could usefully absorb more expertise from external advisers.
And of course, no matter who you are, you can join us and volunteer with us.
Conclusion
Maxime here. I thought I would close this post on a more personal note.
In 2023, after a ~10-year career in deep learning, I became convinced that the race to build ASI was the most severe and urgent danger of our time. I spent about six months trying to work out where I could have the most impact, and I arrived at that answer: PauseAI. So I gave it two years, full-time and unpaid, most of it spent building PauseAI France and making a dent in France’s AI public discourse, which remains the work I am proudest of.
Three years on, I am more convinced than ever that it was the right call. We have tried in this post to lay out the reasoning behind that conviction. But the single most convincing reason is dead simple: PauseAI has operated, across the world, on a total budget of less than $600k since 2023. Let that sink in. Everything described in this post was executed on that budget.
I lack words to convey how neglected this path is. I am highly confident (80%+) that with a 10x increase in resources, we could build a large, aligned and effective movement in under two years. Whether that movement is decisive for reducing AI risk rests on the arguments we have made throughout this post. Conditional on the case being roughly right, this is, as far as I can tell, the highest value-for-money intervention available to anyone trying to reduce catastrophic risks from AI.
So, here is what I am asking, ordered by priority. If you can fund it, fund it. If you can help us find someone who can fund it, please do. And if you can’t, there are other ways to help, and we would be glad to hear from you.
Bibliography
Centola, Damon (2018). How Behavior Spreads: The Science of Complex Contagions. Princeton University Press.
Engler, Mark & Paul Engler (2016). This Is an Uprising: How Nonviolent Revolt Is Shaping the Twenty-First Century. Nation Books.
Han, Hahrie (2014). How Organizations Develop Activists: Civic Associations and Leadership in the 21st Century. Oxford University Press.
McAdam, Doug (1986). “Recruitment to High-Risk Activism: The Case of Freedom Summer.” American Journal of Sociology 92(1), 64–90.
McAdam, Doug & Ronnelle Paulsen (1993). “Specifying the Relationship Between Social Ties and Activism.” American Journal of Sociology 99(3), 640–667.
Moyer, Bill, with JoAnn McAllister, Mary Lou Finley & Steven Soifer (2001). Doing Democracy: The MAP Model for Organizing Social Movements. New Society Publishers.
Snow, David A., Louis A. Zurcher & Sheldon Ekland-Olson (1980). “Social Networks and Social Movements: A Microstructural Approach to Differential Recruitment.” American Sociological Review 45(5), 787–801.
Tilly, Charles (2004). Social Movements, 1768–2004. Paradigm Publishers.
Two psychological tendencies compound. The first is motivated reasoning, whereby we more easily reach conclusions we have an interest in reaching, and its corollary, motivated stopping, due to which, once we land somewhere comfortable to act on (“interesting, probably true, not my problem”), we stop looking for reasons to revise it (Kunda, “The Case for Motivated Reasoning”, Psychological Bulletin, 1990; Yudkowsky, “Motivated Stopping and Motivated Continuation”, 2007). The second tendency is reputational herding: deviating from one’s peers is punished asymmetrically, since an error made alone reads as incompetence while the same error made alongside everyone reads as bad luck, so the privately convinced stay publicly conventional (Scharfstein and Stein, “Herd Behavior and Investment”, American Economic Review, 1990; see also Keynes’s “beauty contest”, General Theory, 1936, ch. 12).
You can find our theory of change on our website.
We unpack “constituency politicians can see and count” in Part II.2.
Given current uncertainty, we can model a few geopolitical scenarios unfolding. We’re describing this condition in broad enough terms to account for this range.
See parts II.1.
A fuller version of this model is on our theory of change page.
Carnegie, “China’s Views on AI Safety Are Changing—Quickly” (Aug 2024) - the CCP’s July 2024 policy document called to create oversight systems to ensure the safety of AI. https://carnegieendowment.org/research/2024/08/china-artificial-intelligence-ai-safety-regulation Carnegie Endowment for International Peace
See the IDAIS “red lines” dialogues, whose Chinese signatories include Andrew Yao, Zhang Ya-Qin and Zhou Bowen. IDAIS consensus statements, https://idais.ai/ ; and Carnegie, “How Some of China’s Top AI Thinkers Built Their Own AI Safety Institute” (2025). Yao is a main convener of IDAIS, under which Chinese and Western experts have published a series of statements on existential risk from advanced AI. Carnegie Endowment for International Peace
On the optimism side: a November 2025 Edelman poll (reported by Al Jazeera) found 87% of people in China say they trust AI, against 32% in the US. On the specific-concern side: a small exploratory China–Germany study of AI-risk perception (Brauner et al., 2025, arXiv:2412.13841) found that even this comparatively optimistic Chinese sample rated “AI misused by criminals” at the maximum risk score — every respondent picked the top of the scale — and “AI more intelligent than humans” among its highest-rated risks. Broad enthusiasm about AI coexists with sharp concern about its catastrophic specifics.
We want to show you how we’re doing this at the level of mechanics, but we’ve had to accept that we can’t do that as well as give you the broad overview this post provides, because it would make this piece exceed most people’s context windows. So we hope to share more detailed blueprints in a separate post.
More specifically (McAdam and Paulsen, 1993), a social tie tends to lead someone to join a movement when joining supports their existing identity: a worried parent, a professional who takes the timelines seriously, or simply a sensible person acting on experts’ warnings rather than a crank or an alarmist. Because people hold many ties at once, some pulling the other way (e.g. colleagues who find the whole subject embarrassing), what matters is the net balance of affirming over discouraging ones. Our local groups and community support the identity-affirming side: a setting where caring about this is ordinary, which is what both recruits and keeps people. This is also why we try to stay mainstream-coded and locally respected, since the doomer caricature is an attempt to load the opposing side of that balance and raise the social cost of being known to care.
If you’re curious about octopus’ sensorimotor and cognitive configuration, we recommend Olson, C.S., Schulz, N.G. & Ragsdale, C.W. Neuronal segmentation in cephalopod arms. Nat Commun 16, 443 (2025). https://doi.org/10.1038/s41467-024-55475-5 and Flash T, Zullo L. Biomechanics, motor control and dynamic models of the soft limbs of the octopus and other cephalopods. J Exp Biol. 2023 Apr 25;226(Suppl_1):jeb245295. doi: 10.1242/jeb.245295. Epub 2023 Apr 21. PMID: 37083140.
We hope to describe this in an upcoming post.
See condition 5 of our theory of change, Part I.2.
An ally movement is a movement through which members use their higher capacity to stand up for the interests of third parties deemed marginalised/disempowered, as “allies” to them. This could be vulnerable members of their local community, non-human animals, future generations etc.
The article which broke the story established for the first time that Google DeepMind did not provide the UK AISI with pre-deployment access to Gemini 2.5 Pro. Notably, Google did provide AISI with pre-deployment access to Gemini 3 Pro a couple of months after the letter was published.
Our first example of this is a volunteer coordination platform which PauseAI UK built for itself and which we’re considering rolling out federation-wide. On this platform, volunteers can join or own projects (e.g. local events, research/content creation projects or lobbying visits), claiming clearly defined tasks and logging project updates. This tool:
allows volunteers to get involved based on availability, location and skills; to form teams; and to smoothly load the relevant context to be productive when they have time to contribute;
enables groups and chapters to do light project management and coordination in one place, which we expect to improve their effectiveness and resilience.
By contrast, only a handful of people participated in PauseAI’s first protest, in May 2023.
Examples: expertise in designing metrics and tracking methods that could be useful in our context; forecasting/modelling, especially if you have first-hand context on relevant parts of the world; press relations; public affairs; high-quality social movement research…