Civil society and AI safety
Movement building, field building, power building, in-fighting in the computationalization of our institutions
2024-09-30 — 2025-10-15
Wherein the nascent civil‑society mobilisation around AI safety is depicted as a movement ecology under strain, with typical fracture lines and coalition‑dynamics, and some relevant social‑movement theories are sketched.
2 Theories of Change
The Social Change Lab’s blog post frames the AI safety problem as a massive imbalance: the capabilities and resources of AI development (private labs, venture capital, state actors) far exceed the organising capacity of the public, NGOs, or activist infrastructure. Their implicit theory of change is:
- Make currently opaque AI risks legible to broader publics (i.e. make harms, probabilistic risks and governance failures visible).
- Mobilize legitimacy, moral pressure, narrative and political leverage so labs and regulators respond.
- Build durable institutions and counterweights (watchdogs, civic oversight, public participation) to channel that pressure over time.
They don’t lay out a detailed stage model (e.g. recruitment → disruption → policy uptake), but the categories of strategy (protest, narrative, field-building, deliberation) echo classic tactics in social movement theory. So let’s name-check some interesting theories that might help design movements for AI safety.
2.1 Resource Mobilization / Entrepreneurial Theory
Pioneered by McCarthy & Zald (McCarthy and Zald 1977), this approach treats movements as rational actors needing resources (money, personnel, legitimacy, networks) to act. It shifts the focus from grievances or moral outrage to capacity: we may have a compelling cause, but without infrastructure, we can’t sustain collective action.
- Advantages: Explains why well-funded, well-networked groups tend to persist and scale.
- Critiques: Tends to underplay culture, identity, or grassroots agency; can over-emphasize professionalization. Some movements succeed with lean resources (e.g. decentralized digital campaigns).
In the AI context, the mapping highlights how thin movement infrastructure is. If we overestimate public altruism or underinvest in connective tissue (platforms, staff, coordination), we risk collapse.
2.2 Political Process / Political Opportunity
This school (McAdam 1982; Tilly 2004) adds the dimension of opportunity windows: elites may split, crises may open legitimacy gaps, or institutional thresholds may shift. Movements succeed not just by having resources, but by acting when constraints momentarily loosen. In the AI domain, we might see moments when a lab’s failure or scandal breaks public trust, or when a policy window opens (a legislative hearing, a regulatory mandate). But those windows are narrow and unpredictable.
2.3 Framing Theory / Cultural Work
Snow and Benford (1988) and others argue that movements succeed or fail depending on how well they frame problems (diagnosis, moral attribution, solution). Even if we have resources and opportunities, if our narrative doesn’t resonate—or worse, alienates—we stall. Applied to AI: “AI is inevitable and unstoppable” is the default narrative. That framing supports inertia. If the movement reframes AI as a political choice rather than a destiny, it may shift how we engage. But catastrophic framings (extinction risk) can backfire by raising fatalism or distrust, as some experimental research in climate messaging suggests.
2.4 Movement Ecology / Ecosystem Models
In recent years, activists and scholars have talked of ecologies: a diversity of groups with different roles (advocacy, litigation, protest, research, local organising) coexisting and interacting. The idea is that no one strategy dominates; the strength lies in the networked interplay. (See Ayni Institute / Open Philanthropy discussion on movement ecology)
The ecological metaphor has critics: boundaries blur (who is inside the movement?), agency becomes diffused, and coordination is nontrivial (see “boundary,” “agency,” “interaction” problems). In AI safety, we might already see an incipient ecology: protest groups, policy institutes, public education projects, watchdog labs. The trick is whether they mutually reinforce one another rather than fragment or compete.
2.5 Organizer Models
Marshall Ganz’s “organizing” perspective emphasizes leadership, narrative, relational ties, and distributed leadership over hierarchical control. He sees movements as building power, not just contesting it. (See interview “Are You Building Something?”).
In AI, grassroots buy-in will require narrative coherence and relational organizing, not just top-down statements from experts. We may need “organizers” who can help connect technical researchers with civic actors, not simply more memos.
3 Careers in AI Safety
There’s a lot to say. For now, see AI safety careers.
4 Fracture lines and coalition dynamics
In practice, movements for social change are as much about keeping a coalition together as about winning specific reforms. Observationally speaking we’ve seen this kind of thing before: Consider anthropogenic climate change. What began as a scientific concern gradually became a cultural signifier. Political entrepreneurs and media ecosystems recognized the mobilizing power of the issue. Fossil fuel companies seeking to avoid regulation often funded scepticism, which conservative media amplified. Over time, belief in anthropogenic climate change became less correlated with scientific literacy and more correlated with political affiliation. The issue was reframed from a collective environmental challenge to a zero-sum conflict between economic liberty and government overreach, which solidified the partisan divide.
COVID-19 policy seems to have followed a similar trajectory, with public health measures becoming entangled in broader cultural and political identities, although the question of “who benefits” is more complex in that case.
Some claim such sectarian fracture lines are often engineered by the opponents of change, who recognize that a divided movement is less effective. This might be so, but I suspect they can also emerge organically, from a genuine divergence of priorities, or from the complexities of coalition politics, or from the narcissism of small differences.
One such split in the AI safety ecosystem is between “AI Ethics” communities and “AI Safety” communities.
AFAICT, the median person in each group worries about similar things (e.g. algorithmic prejudice, labour displacement, surveillance, resource consumption, corporate control, cementing authoritarian power, weaponized AI), which makes them more similar to each other than to the general public.
Within this broad let’s-not-hurt-people-with-AI consensus we can situate those tribes. People who claim an allegiance to the latter tribe (AI Safety) also worry about autonomous AI systems going rogue or becoming superintelligent. Partisans of the former tribe (AI Ethics) tend to dismiss such AI concerns as too speculative or otherwise unimportant, and thus diluting the effort that should be spent on short-term issues like algorithmic prejudice. Aspirational firebrand substackers spend effort trying to prise these communities apart by assigning such differences in priorities to moral failings or to being in the pocket of some large vested interest, and thus painting the other side as not just wrong but evil.
What follows is a worked example of that.
4.1 Anti-TESCREALism Case Study
Coverage of a recent attempt to start a new battle front in the forever culture war. You might want to skip reading about it unless you enjoy culture wars or are caught up in one of the touch-points of this one. Otherwise I would avoid feeding the attention/engagement cycle monster.
Public debate over AI risk has recently been punctuated by the emergence of the acronym TESCREAL (Transhumanism, Extropianism, Singularitarianism, Cosmism, Rationalism, Effective Altruism, Longtermism), which was coined by Timnit Gebru and Émile P. Torres. Gebru and Torres argued that these movements form an overlapping “bundle”, rooted in techno‑utopian thought, that uses existential‑risk narratives to justify speculative, high‑stakes interventions.
Because few outside those inner circles of “rationalist / EA / futurist” discourse encountered the term, the TESCREAL label functions partly as a sorting device — a quick tag for “those who endorse extreme futurism and technocratic risk arguments” and argues that they are in opposition to more justice‑ or ethics‑oriented critics.
The aspirational wedge here portrays TESCREAL as a kind of techno‑libertarian cult—indifferent to social justice and dismissive of near‑term harms—and tars much AI safety work with that brush. It also positions anti‑TESCREALists as affiliated with an AI ethics side of the proto‑culture‑war (culture beef?)
4.2 What TESCREAL does
This is a neat example of classic movement splitting. Notable features:
- “Othering” TESCREAL often acts less as a precise critique and more as a boundary marker: it signals “this is not us” in debates over AI (much like how woke, SJW, cultural Marxist are used in broader culture‑war settings).
- Simplification of pluralism. It flattens a diverse spectrum of positions (e.g., an EA working on malaria has little in common with a singularitarian) into a single antagonistic “side.”
- Incentives of performative polarization. Once an issue is sorted along identity lines, winning the battle becomes more valuable than resolving the technical problem, or negotiating trade‑offs.
In other words: TESCREAL illustrates the standard mechanics of movement fragmentation and divide‑and‑rule.
Under polarization, concerns about ASI and existential risk (X‑risk) get assigned respectively to a “TESCREAL” or an “anti‑TESCREAList” side. Conversely, concerns about algorithmic bias, energy consumption, labour displacement, and corporate accountability are claimed by the anti-TESCREAList side, and any expression of concerns about such matters by someone classified as TESCREAL‑aligned is treated as suspect or hypocritical. If this division hardens, it would reduce researchers’ ability to collaborate across domains and policymakers’ ability to address the full spectrum of AI risks in a coherent manner.
4.3 Connecting to movement theory
In social movement scholarship, cohesion and fragmentation are recurring hazards as movements scale. Collective identity, shared framing, and ritualized boundary work are essential to prevent implosion (or being co-opted) (Peters 2018).
But “coherence” is fragile. As coalitions incorporate more actors, strategic and ideological tensions accumulate. That’s when labels, purism tests, or rhetorical bundles (like TESCREAL) tend to emerge. Many movements in history have died of internal splits long before achieving institutional change (e.g. the decline of the U.S. civil rights coalition post-1968, the splintering of 20th-century feminist waves, or schisms in radical environmentalism).
Within framing theory, a movement’s success depends on maintaining a narrative that is inclusive enough to hold alliances but sharp enough to mobilize. The introduction of TESCREAL is a framing move that cuts a movement into two camps and sets them against each other.
Finally, from a processual view (stages of social movements), fragmentation is often the fourth phase: incipiency → coalescence → institutionalization → fragmentation (Hiller 1975). The attempt to build a divide around Anti-TESCREALists might be a symptom that the AI safety ecosystem is entering that danger zone of internal sorting and conflict.
4.4 Risks and lessons for AI safety
- Shrinking the coalition. If “TESCREAL vs. anti-TESCREAL” becomes the dominant axis, actors who have cross-cutting interests, for example opposing over-acceleration but still care about X-risk, may be forced to pick sides — or be silenced by association.
- Reductionism of policy space. Complex trade-offs (bias, labour, local harms, long-term alignment) get sorted into camps and then simplified into binary slogans.
- Overvaluing theatrics. The polarization dynamic rewards sharp calls, ridicule, exclusion and other engagement-pumping behaviour — rather than careful synthesis or cross-cutting engagement.
So what should we do, in a pluralist field where critique is necessary but collapse of the coalition would be fatal?
- Prefer connective critique to sweeping rejections. Resist the temptation to brand entire clusters as “evil.”
- Emphasize shared purpose and modular trust. Build bridges over specific projects or domains, leaving space for disagreement elsewhere.
- Metacognitive boundary work. Be explicit about when we’re naming a critique of a policy position versus stamping identity tags.
- Guard narrative pluralism. Encourage multiple frames (immediate harm, existential risk, institutional resilience) to coexist rather than battle prematurely.
If TESCREAL is a sorting device more than an analytical lens, then the fight is not merely about correctness but about preserving the possibility of a unified movement. The worst outcome would be that our internal battles drown out the external crisis.
5 Some Emerging AI-Safety Civil Society Actors
Mostly recycled from the Social Change Lab map, but I plan to extend it.
5.1 Protest, disruption, mobilisation
Groups using direct action and mass visibility to pressure labs and governments.
- PauseAI: Decentralized network demanding a pause on frontier-model training until enforceable safety standards exist; noted for disciplined, media-friendly protests.
- Stop AI: Civil-resistance organization calling for an outright ban on artificial general intelligence; embraces disruptive tactics to dramatize existential risk.
- People vs Big Tech: Youth-driven campaign linking AI harms to social-media dysfunction, surveillance, and concentration of tech power.
- Control AI: UK-based initiative urging public oversight of “frontier” AI; has gathered tens of thousands of signatories on open safety pledges.
5.2 Narrative-shaping and watchdogs
Actors exposing current harms and contesting industry framing.
- Algorithmic Justice League: Joy Buolamwini’s research-storytelling project documenting racial and gender bias in facial recognition.
- DAIR Institute: Timnit Gebru’s independent lab producing community-rooted empirical research on AI’s social impact.
- Fight for the Future: A veteran digital-rights NGO deploying rapid online mobilization; expanding its scope from surveillance to generative-AI governance.
- AI Now Institute: Academic centre analysing corporate AI power and advocating for democratic accountability in tech governance.
5.3 Public literacy and democratic engagement
Groups focused on public understanding, participation, and deliberation.
- We and AI: Volunteer UK collective running workshops and partnerships to build inclusive AI literacy.
- Connected by Data: Policy nonprofit reframing data and AI governance as democratic—not purely technical—questions.
- CivAI: U.S. nonprofit using hands-on demos (deepfake creation, phishing simulations) to teach critical awareness of AI misuse.
- Democracy Next / Citizens’ Assemblies on AI: Facilitators of deliberative mini-publics producing policy recommendations on AI oversight.
- Global Citizens Assemblies: Prototype for transnational deliberation treating AI governance as a global commons issue.
5.4 Infrastructure, field-building, and lobbying
Institutions providing connective tissue, research capacity, and policy leverage.
- Ada Lovelace Institute: Independent think-tank convening cross-sector dialogue and generating policy research on equitable AI.
- AI Incident Database: Open repository of real-world AI failures enabling transparency and pattern analysis.
- Data & Society / Future of Life Institute / Humane Intelligence: Research and convening hubs connecting academic, activist, and governance communities around AI risk.
- Stop Killer Robots: Global coalition campaigning for binding international law against autonomous weapons; a model of issue-specific global coordination.
6 Incoming
- The “TESCREAL” Bungle
- Segmentation faults: how machine learning trains us to appear insane to one another.
- Semi-counterpoint: Doctorow on Big Tech narratives.
- Steven Buss, Politics for Software Engineers, Part 1, Part 3

1 Social licence for AI safety
I think it’s amazingly under-regarded, probably because many of the more technical roots of the first movers in AI safety and the concomitant tendency to forget the human angle in systems.
If AI safety is to become more than a niche technical concern and instead a society-wide governance programme, we need to view it through the lens of movements, not just research or mechanism design. In democratic states, persuading people is part of persuading governments.
Social movement theory tells us movements need resources, narrative work, opportunistic timing, and ecosystem-building. This sets up the challenges of AI safety organising: public awareness is low, the technology seems exotic and distant, regulatory inertia is baked in, and commercial interests run deep.
What would it mean for a mass movement to intervene effectively in this particular policy issue — and how plausible is it? Below is a sketch of how social movement theory might frame an “AI safety movement,” and what that suggests for strategy.
There are some interesting theories of change that might be relevant. Some campaigns may fail spectacularly or fragment; that’s the nature of movement building in domains with high uncertainty, complexity and low salience (see analogies in early climate activism, vaccine policy, pollution regulation).
The Social Change Lab map offers a useful breakdown of actors and risk zones. I will defer to them for the details.