This is my first post here. My leaning into AI safety is still in it's formative phase but I come at it as a tech founder and builder rather than a researcher. I've spent a while trying to work out why so much good safety work isn't becoming regulation and surviving contact with the industry's lobby. The argument below in a nutshell is: the AI safety field leads comms with its least mobilising frame, extinction, when the harms that actually move people to act are instead concrete and proximate, and an organised public might be the missing half that keeps any safeguard and regulation alive. If that's true, then I share some ideas on what we do about it . I'm learning in public so I'd love any comments, counter-arguments, clarifications and additional perspectives.
As a spring dawn emerged over the desert, residents of Doña Ana County were waking to find their letterboxes strewn with flyers of a smiling Hispanic woman promising hundreds of millions in investment and high-paying careers in return for their approval of Project Jupiter. Against that backdrop Daisy Maldonado, a first-time candidate, ran for local office with the endorsement of Bernie Sanders, pitting her candidacy directly against a $165 billion data centre.
In rural Utah, thousands have similarly voiced concern, and growing anger. Across the USA, over 800 community groups in 49 states are organising against data centre projects they believe will drain water, spike electricity bills and benefit shareholders thousands of miles away. That number more than doubled in the first three months of 2026 alone. In 2025, community opposition stalled more than $156 billion of planned construction, and the first quarter of 2026 blocked or delayed a further $130 billion. A Gallup poll this May found seven in ten Americans oppose data centres near their communities. You don't need epistemically aligned allies to create strategically useful ones.
Nonetheless, none of them are protesting artificial intelligence in the abstract, and I should be clear that neither am I. They're protesting their water and power bills, the incessant hum, the ruined landscapes and planning decisions that enrich others at their expense. And whether they realise it or not, they're protesting the reality that AI will have a profound impact on all reaches of humanity, and they don't believe the writers of their future existence should sit exclusively in the hands of corporations and investors. But is intermittent political pressure durable enough to help shape broad, effective AI safeguards?
Eighty-nine percent of Brits want an independent AI regulator with real enforcement powers, yet only 15% think there's currently enough regulation. Eighty-four percent worry the government cares more about the tech industry than keeping the public safe. Hardly news nor peculiarly British sentiment. An October 2025 Gallup Poll found 80% of Americans want rules for AI safety and data security, even if it means developing capabilities at a slower rate. And yet, recent US federal policy has cleared industry barriers rather than impose safeguards, with lobbying shelving a voluntary disclosure regime. In the UK, the Treasury Committee has warned that a wait-and-see approach risks serious harm, while the AI Bill stalls. In the EU, lawmakers have delayed and slimmed down the AI Act's key provisions.
The political and legislative response is slow, and in many places moving backwards. Why does public concern not match legislative actions? And why, when it has emerged, does it soften or face delays? Most critically, is there anything that can be done about it?
My journey into AI Safety is still in its formative phase but my perception is that much external communication to date has centred on catastrophe. Loss of control. Extinction. Well versed people making meritworthy arguments, many make me nervous, and few I object to. But evidence on whether or how that framing moves those outside the AI Safety space is consequential.
The Social Change Lab's (SCL) March 2026 RCT trial is worth perusing. Across 3,467 British adults, randomised and pre-registered, it measured 'willingness to act' across different AI harms once they'd spent time reading about each. Human extinction was the least mobilising and resonant frame, scoring just 14/100.
These findings match reality. PauseAI, the largest international group campaigning on AI extinction risk, held its biggest ever protest in London in February 2026. Around 300 people came, which organisers called a major step forward. In the same period, a single small-town petition in Pennsylvania against a proposed data centre collected over 1,800 signatures and killed the project outright. Town halls on data centres across the country have had to move meetings outdoors to fit everyone in. The gap between the world's largest AI safety protest and what one data centre dispute produces in a small town is the SCL finding written in behaviour.
There are good reasons for this, three to be precise. Trippenbach and colleagues (in "From Catastrophic to Concrete") identified three psychological barriers that make the extinction frame difficult for many to act on:
And concerningly, the trend is heading in the wrong direction. The Seismic Foundation, surveying 10,000+ people across five countries, found public concern about AI loss of control halved in two years, from 67% to 36%, despite an increase in existential comms.
Does this diminish the risk? No! And this is whereI want to be incredibly precise about what I'm arguing (and what I’m not). I'm not saying catastrophic AI risk isn't important, and I'm definitely not saying people should stop working on it. The opposite, in fact. ControlAI's three-person team used it to persuade over a hundred cross-party parliamentarians to publicly back AI safety measures, and efforts like that are critical. But extinction is the wrong frame for building the public pressure that shifts political will. If you want the public to act, organise, vote, write letters, sign petitions and make politicians feel democratic pressure, you have to meet them where they are. And some good news, from Hoes and Gilardi's experiments across 10,800 US and UK participants, is that exposure to existential-risk narratives increases concern about catastrophe without reducing concern about more immediate harms.
So here is my argument, laid out plainly:
And beneath the instrumental case sits a simpler one. A technology that will reach every corner of humanity is having its regulations shaped by the corporations and investors involved. A bargain for their agency has been struck without their consent, whether they realise it yet or not, and the public is starting to recognise that the terms of their future should not be set that way. There's nothing inherently anti-AI in that. It's the oldest demand in democratic politics. Decisions that bind everyone must be made with everyone in the room.
The stalled bill in London, the slimmed-down Act in Brussels, the shelved disclosure rules in Washington. One side turns up organised at every stage, and it hasn’t been the public, yet.
Even those building the technology acknowledge where it's heading. OpenAI's charter promises AI that benefits all of humanity. Anthropic's mission is AI for humanity's long-term benefit. Yet, Dario Amodei warned in January that without action AI risks creating extreme levels of inequality, and more recently Anthropic argued that the world should have the option to slow or temporarily pause frontier AI development, to let society and safety research keep up. BlackRock CEO, Larry Fink, told shareholders in March that AI threatens to concentrate wealth on an even larger scale. Investigating the role of race dynamics and capitalism are for another day, but trillions of dollars of capital investment makes it unlikely that we’re about to pivot towards the common good, regardless of what a charter says. A change of direction will only happen when a powerful external force redirects it.
And that capital becomes political power. Namely, a $125 million super PAC aligned with OpenAI and Andreessen Horowitz alongside 3570 registered federal AI lobbyists. The regulations are being written by the people who stand to gain from there being as few of them as possible.
So who's sitting on the other side of this table? Our governments? The regulators? Civil society? All of them, to a point. Anthropic donated $20m to Public First Action to help counter the OpenAI & A16Z super PAC. In the UK Ofcom is enforcing one of the most ambitious online-safety regimes globally. The AI Security Institute tests frontier models before release, the Ada Lovelace Institute is measuring public AI sentiment, and campaigners have carried AI risk onto the streets and parliamentary debates. They do exceptional work, but they are outgunned and outmaneuvered. AI companies and their allies held over 530 ministerial-level meetings in the UK and outnumbered every other voice by 8.7:1. The US situation is a worrying echo. And the public, a force that could rebalance the table? It exists, the stalled data centre billions prove it, but it organises sporadically and without intentional focus or coalition. It’s intermittent pressure against relentless, organised and disproportionately well funded opposition.
Unfortunately the political science backs the cynics here, and the game appears rigged. The most-cited study of US policy responsiveness, covering nearly 1,800 policy questions over two decades, found that what average citizens want has almost no independent effect on what the government does. Rather it's organised interests, lobbies and elites shaping public policy. But that study describes politics in its quiet state, and it comes with one stark exception. Money buys lobbyists, advertising and access, and all three are at their most effective while policy stays technical, low-profile and relatively uncontested, the kind that never reaches your dinner table. Once an issue starts being raised in local council meetings, deciding votes and elections, and lands in the grip of the media, the calculus starts to change. A capital war chest becomes less useful to a politician on the wrong side of something their constituents are willing to act on. Money regularly wins quiet fights but votes win louder ones and that's the theory of power everything I’m proposing rests on.
This is the pattern I’ll lean on and validate. Child safety crossed that threshold in Britain years ago, and it's why the Online Safety Act keeps surviving everything thrown at it. Immigration crossed it across much of the Western world and dragged the policies of major parties with it, no super PAC required. A data privacy fight started in California and became law across 20 states. The data-centre resistance is becoming a midterm election question. AI governance issues must be technically unmasked and become an electoral question candidates get asked, something representatives feel the weight of in their town halls and inboxes, until quietly repealing a safeguard costs more votes than it buys in donations.
Are there cases where public pressure doesn't produce regulation? Of course. Many. Gun control is the obvious one. Enormous, sustained pressure, Parkland, March for Our Lives, and still no federal legislation. So, why would AI be any different?
Well, the dynamics are starkly different. Gun control is a deeply partisan, constitutional issue with entrenched, well-funded opposition on both sides. However, AI is sharpening an economic inequality story, and I believe that changes the politics. Elon Musk just became the world's first trillionaire. The ten richest people on earth now have roughly $3 trillion between them, nine of them American, eight in tech, having added $550 billion to their collective wealth in 2025 alone. As Larry Fink told his shareholders, a dollar invested in the US stock market since 1989 has increased 15x more than a dollar tied to median wages. AI is now the leading reason companies give for cutting jobs in America and firms adopting AI have reduced junior hiring by roughly 13%. The standard response is that every technological revolution displaces before it creates, and the net forecasts still point that way. Maybe they're right. But even if AI eventually creates more jobs than it replaces, its wealth generation is concentrating at the very top with historic speed while the costs are disproportionately felt by those seeing the least benefit, and that asymmetry crosses political lines in a way guns never could.
In June, Donald Trump and Bernie Sanders both proposed that the American public should hold equity stakes in AI companies, and Sam Altman flew to Washington to discuss it with them. The underlying argument behind this rebalancing is that the models are trained on data that humanity collectively produced, and yet profits flow almost entirely to people who packaged it. Whether that argument leads to a sovereign wealth fund or something else entirely is unclear but the political energy beneath it is bipartisan. The data-centre resistance spans regions, demographics and parties. Coalitions built across political divides tend to last, and a bipartisan coalition is one thing even this much capital might struggle to split. This opportunity mustn't be squandered.
The SCL trial gives the clearest current answer in the UK (there is no representative causal mobilisation study in the USA). Environmental harms topped the index at 100. Human extinction propped up the table at 14. By the end of the study, over a quarter of participants had signed a real petition for AI regulation, so this wasn't hypothetical hand-wringing.
The study found three things move people from concern to action:
This might explain why data centres tick the box, as does an algorithm inadvertently filtering women or people of colour out of job interviews. A hypothetical superintelligence in 2030 ticks none!
There’s an interesting exception not covered in the SCL trial though. Children and AI companions. In the SCL index, ‘relationship harms’ scored a modest 38 but add children to that and the picture changes. 72% of American teenagers have used an AI companion, one in three has confided something serious to one instead of a person, and 69% of parents across five countries worry about their child forming a romantic relationship with an AI. Few harms combine stronger emotional proximity to someone worth protecting, and when a harm touches children the political response tends to follow quickly.
Notably, in many global surveys, job loss is the AI harm mentioned first. The Seismic Foundation found it ranked second as a concern, after AI replacing human relationships. But it doesn’t mobilise people as effectively as other harms, ranking a surprisingly low 58/100 according to SCL. What does this tell us? That caring or worrying doesn't always lead to action and using US concern polls as a proxy for mobilisation may be ineffective. The harms that will move people to act tend to be specific, local, fightable and immediate. This is an opportunity that needs exploring in the USA.
I’ll admit to generalising for the sake of brevity so far, but there isn't just one public. Seismic identified five distinct publics, roughly 100 million people, moved by different things, though some concerns cut across them all. 60% worry about AI replacing human relationships, 70% want humans kept in control of decisions. A message that resonates with one leaves another cold, and the most mobilising harms will continually evolve, thus, so must the communication.
This data should be treated as a moving signal, measured continually. The SCL is a snapshot, but its authors noticed that, in the week Grok's nudification scandal was loudest, child-safety anger surged up the open responses. A single study will always be but a photograph of a moving target which means treating this as a live instrument, tracking which harms are rising, in which publics, and testing the language continuously against real actions, petition clicks, emails to representatives.
I'd also wager that an engaged public is a teachable one. The parent who first pays attention to AI companions may begin to understand why we can't ever reliably make these systems completely safe for children, and start to comprehend what that means in other implementations, and more broadly still. The entry point is proximate but the destination is a fuller, more informed picture.
Imagine that the public does mobilise around the harms that move them, that pressure could misfire in a couple of ways. It could demand the wrong thing, and it could split down party lines.
Take the first. The public shouldn’t be writing the regulations. It’s a practiced craft and getting it wrong does damage. A case in point is a recent UK AI Bill, well-intentioned but drafted with loose, emotive definitions by authors who didn't appear to understand how regulation works in practice, as expert scrutiny made clear. The work of building pressure should split in two. The public must make issues impossible for politicians to ignore and experts must turn that pressure into regulations that function. Neither is likely to be effective without the other, but it’s critical that they focus on the same outcomes to amplify efforts.
A deeper objection might be that, however well-aimed, public pressure won't matter because even a serious and visible AI disaster, wouldn’t produce sensible regulation, as exemplified by COVID. The most expensive warning shot in history, produced little implemented reform. But there’s a distinction in that COVID was a catalyst for the Pandemic Agreement and strengthened International Health Regulations but, as life returned to normal, political will evaporated and budgets shifted before any durable implementation. However, the AI risk calculus is fundamentally different. Unlike a virus that burns through a population and retreats, AI's growing integration into the economic and financial backbone of societies means exposure will not diminish but is more likely to compound. And right now the AI industry is far better prepared for that moment than the public is.
A second prospective misfire is partisanship. The moment AI safety becomes one party's cause, the other turns against it, and the cross-party coalition, that was its greatest strength, falls apart. American climate politics spent two decades illustrating this. And avoiding politics won’t help! AI is already a political issue, with significant capital working to define it on the industry's terms. Stepping back just cedes the pen. But keeping the issue cross-party will take deliberate effort; different people, in different parties, in different jurisdictions are moved by different harms; a parent isn't a farmer, isn't a developer. So you would need to speak in the terms each recognises, keeping a close eye on how this changes over time. It’s important to note that none of this requires whipping up outrage and anger, whilst mobilising, can lead to the indefensible. What it will require is informed pressure through the ballot box and the legislative process; a public aware of the risks, who understands what they're asking for, not a mob baying for blood.
I can hear your mind whirring. Data centres, children, jobs are all a long way from the frontier governance and safeguards that are actually needed. I don't disagree. But the pattern has a precedent which we should study closely.
When Molly Russell died, and her father went public about Instagram's algorithm pushing harmful content, the coverage focused on children and social media. When the coroner recorded her death as partly from "the negative effects of online content," the narrative shifted, from bad content to bad design, engineered by the platform, and intertwined with politically sensitive child safety. It shifted from an individual, lifestyle question (of personal discipline) to a power question (who designs the platform, with what incentives).
What followed was far broader than the original trigger. And truthfully, far broader than anyone predicted. The UK's Online Safety Act brought in platform accountability, content moderation duties, age verification and an expandable system of oversight, built as a living regulatory framework. Ministers continue to add new priority offences and Ofcom extends its codes without fresh legislation and, once a framework like that exists, it can become an evolving vehicle for far more than the harm that created it. That’s the lesson here.
Political scientists called these ‘focusing events,’ with 40 years of literature validating their effectiveness. A specific, proximate harm forces political attention which broadens when campaigners, voters and reformers persuade institutions that the visible harm is symptomatic of larger governance failure rather than isolated abuse.
The UK example is far from a one-off. It’s played out in the tobacco and pharmaceutical industries, and more recently California's privacy law grew from years of diffuse unease about data extraction. The Cambridge Analytica scandal opened the door, one ballot initiative gave legislators a deadline, and the CCPA passed in a week to keep a stricter version off the ballot. Voters expanded it two years later, and some twenty states have since copied it into a de facto national standard. Is it perfect? No. But it is evolving. #MeToo similarly produced federal statutes banning forced arbitration and the NDAs that silenced harassment claims, without any single catastrophe trigger.
This same machinery is powering up in AI, but, as of yet, broadly unaccompanied by a mobilised public. The UK's All-Party Parliamentary Group on Children's Online Safety has opened an inquiry into AI harms and children to feed the forthcoming AI Bill. Character.AI, the biggest companion platform, faced a wave of lawsuits tied to teen deaths and barred under-18s from open-ended chat. California passed SB 243, the first US state law on AI companion chatbots. Senators Hawley and Blumenthal introduced a bipartisan federal bill to ban them for minors outright. And the UK extended the Online Safety Act to cover AI chatbots, after realising existing laws weren’t fit for purpose once Grok began to be used to generate intimate, deepfake images on X without a clear obligation to remove them. So far, so good, perhaps, but it’s far from enough.
It’s important to recognise that the Online Safety Act wasn’t a ‘clean’ win given its age checks are dodged with a VPN, a petition to repeal it passed half a million signatures, Wikipedia took the government to court and Reform UK calls it censorship. And yet Parliament debated repeal and rejected it, and the government has spent 2026 tightening the Act including an amendment requiring platforms to block known intimate-image abuse including deepfakes (keep the amendment link here), a 48-hour takedown rule and, in March, a new power to restrict children's access to entire categories of service. A law this contested, clumsily implemented and easy to mock is growing despite the big tech lobby. Why? Because child safety has a cross-party public standing behind it. The proximate harm ignites the regulation but an organised public is what keeps it alive long enough to evolve and broaden. Ignition is the phase everyone studies yet the far more critical survival phase is often overlooked.
It would be remiss of me not to acknowledge that regulation is distinct in the UK and USA. Britain has a single Act and an empowered regulator. America has neither. Legislation runs through courts, the states and Congress. In March a California jury found Meta and YouTube liable for designing addictive products that harmed a child through defective design. The award won't trouble these tech titans, but the finding will, with thousands of similar cases queued behind it. It isn't settled, the verdict will be appealed and causation was fiercely contested. Even so, the American system is reaching through its courts what Britain layered through Parliament.
And California's legislature has begun to go further. In the 2025 session that produced the AI chatbot regulation, it also passed SB 53, the first US law targeting frontier AI developers directly. The frontier model providers, OpenAI, Anthropic, Google and the rest, must now publish how they assess and manage catastrophic risks, report serious safety incidents and protect employees who raise concerns and Anthropic has already published its compliance framework. A stronger bill requiring pre-deployment safety testing had been vetoed the previous year. But the template is travelling with New York signing its own frontier safety law in December 2025, and in early July Illinois signed the Artificial Intelligence Safety Measures Act near unanimously, adding the nation's first requirement for annual independent third-party audits. Three states covering roughly 40% of the US AI market now run frontier transparency rules that Congress won't pass. It’s clear that the legislative machinery is evolving, but it still stops short of the requirement that matters most, mandatory independent evaluation before a model is released.
A survival test of this state driven legislation has already begun. A December executive order created a Justice Department task force to sue states over these new AI laws and tied federal funding to dropping them, and any bipartisan draft in Congress would freeze state rules on model development for three years anyway. Though, the last time Washington tried an outright freeze the Senate stripped it out 99 to 1 after a bipartisan revolt.
Taking a step back from the federal vs state battle, much of the safeguarding and regulatory machinery that other proximate harms might require (testing, incident reporting, transparency, independent oversight, independent risk assessment before deployment) are shared with frontier governance safeguards, and a standing public could keep them in place long enough to evolve effectively. Will a narrow trigger expand into the broad infrastructure required to govern frontier AI capabilities? I believe the evidence shows it can, but it's unproven, and it's the claim most worth testing. And if the bridge proves narrower than hoped, the wins along the way are still invaluable from safer products for children, transparency and incident reporting to communities with a say over their own infrastructure. Truly effective frontier governance is the upside but it’s not the whole case.
I suspect we can all agree that effective, independent regulation and safeguards alone won't be enough and nor will treaties and global coordination. What an informed and mobilised public provides are the conditions under which other forms of AI safety and accountability measures can work most effectively. A regulator enforcing independent pre-deployment evaluations and adversarial testing needs public support to survive pushback or softening. A journalist investigating an AI company's practices needs readers who care enough to make the story consequential. And a pension fund investor demanding independent audits and mandatory incident reporting at a shareholder meeting needs to know the public is watching and cares. None of these people can act effectively in a vacuum and each of them depends on an informed public that pays attention and is willing to act. That's the fertile ground, and right now, for generative AI, it needs urgent expansion.
Normalising and accepting the impact of AI is a very real danger and accelerates my concerns about a loss of agency. I’d define that as the point at which the public believes it has no choice, treats AI systems as beyond question and comes to see the future as something imposed rather than shaped. They must not become more-adaptable individuals, but rather more-demanding societies. So what might it mean for the public to become more demanding and mobilise?
The most direct route is electoral. Constituents write to their MP or representative about something specific like mandatory safety standards for AI companions, or incident reporting for AI hiring bias and discrimination. They respond to government consultations on AI regulation. They keep showing up to local planning decisions on data centres, as communities in 49 states already are, and they make AI governance a question candidates get asked. None of this is hypothetical given Daisy Maldonado won her primary in June and is on November's ballot, and this year's US midterms is likely to be the first at-scale test of whether this type of pressure sways votes. The difference between 300 people at an extinction risk march and hundreds of thousands organising against local data centres tells you everything about the kind of pressure politicians actually feel.
Additionally, individual letters and consultation responses are easy to ignore in isolation but they’re harder to dismiss when they arrive as a coordinated signal. Thousands of people responding to the same consultation in the same week, open letters timed to a parliamentary vote, collective positions on specific policy questions that show representatives the depth and breadth of feeling in their constituency. This ensures individual concern becomes the kind of aggregate pressure that I’m arguing for - visible, measurable and tied to something a politician can count.
There's a third layer beyond regulation in the form of market forces. Think ethical AI standards for business, a B Corp parallel, or procurement that favours credible governance practices but they will only gain traction when enough people demand them and someone has defined what credible looks like. Consumer choices, investor questions at shareholder meetings and advertiser pressure are all levers that exist in theory but lie dormant because there isn't yet a public informed or active enough to kickstart them.
The research to start segmenting public AI concerns and building targeted public awareness is well underway but it’s static. The Social Change Lab, the Seismic Foundation and the Ada Lovelace Institute have begun measuring what the public thinks about AI and what moves them to act, and uncovered the distinction between the two. But this needs to go further, akin to how you scale high impact growth marketing because it will change in real time. Think A/B testing headlines, understanding topic and narrative click through rates and the actions they drive, segmenting by audience and jurisdiction, (and reading the signal closer to real time).
In addition to this audience understanding, detailed, specific regulatory proposals already exist, drafted by aligned technical and regulatory experts. Many point in the right direction from independent regulators, pre-deployment safety evaluations, mandatory incident reporting, transparency requirements and liability frameworks. And the organisations pushing for safeguards are doing incredible work. As I previously mentioned, ControlAI, with three full-time staff, persuaded over a hundred cross-party parliamentarians to publicly back AI safety measures. Foundations like Seismic have mapped five distinct publics and what moves each of them. PauseAI (however you feel about them) has put AI risk protests on the streets, despite the struggle to build a broad base with their extinction messaging. The Future of Life Institute has spent millions putting AI risk in front of large audiences through paid media. But each of these efforts covers part of the picture and they need co-ordination.
I've spent a 15 year career founding and scaling regulated tech companies, building network effects, understanding behavioural psychology and growing and segmenting audiences and user bases, a background recognisably distinct from most working in AI safety. Over these last few months haven't yet found anyone running the whole loop above, tracking which harms are rising in which audiences, building a large mainstream audience and targeted comms around those specific harms, turning that audience into coordinated action aimed at specific safeguards. And lastly, but most critically, feeding these real time evolving insights with the people writing regulatory proposals, running campaigns and receiving their invaluable intelligence in return, which safeguards are viable, which frameworks legislators are considering, what's surfacing from the labs, so all partners are aligned and the messaging stays focused on what's achievable as well as what resonates in a compounding loop. Maybe someone is already building this and I haven't found them, in which case I'd genuinely like to meet them. From where I stand, it's a consequential gap in the field.
Public mobilisation around AI risks and safety is still in its infancy, yet the polling is emphatic that the public care deeply. However, concern that never mobilises is concern a politician will never have to feel. I believe the cases above show that the bridge between concern and political pressure can be built. So far, it has mostly been built by accident and until it's built purposefully and in a targeted manner, the best work in AI safety, the pre-deployment evaluations, the interpretability research, the incident-reporting frameworks, the carefully negotiated governance proposals, will stay voluntary. Something the industry can take or leave, safeguards a future administration can and likely will quietly shelve. It’s clear that the public concern exists but the political pressure doesn't, yet. Closing that gap is a challenge that must be addressed before it’s too late.
