Creating a social mandate for AI governance

1. Introduction: why do we need public engagement?

1.1 Public engagement is often forgotten, but critical for the response to AI safety

AI is moving faster than any previous general-purpose technology, and how well it goes depends on decisions being made now, at all levels of society. There is no off-the-shelf blueprint for governing it. Shifts in attitudes, norms and behaviour around AI are emerging in different ways in different places, and nobody can predict them with confidence.

What is clear is that governing AI well, with the consent and participation of the public, needs effective communication and active public engagement.

Global AI investment reached around 580 billion US dollars in 2025 (Stanford HAI, 2026 AI Index). The share going to safety is well under a tenth of a percent, and the fraction of that spent on public understanding is smaller still. The field has built technical research, governance work, funders and training. It has not built the layer that explains any of it to the people it affects.

At the time of writing (2026), public concern about AI is high in most countries surveyed. But concern is not the same as committed public support for governance. The rules that both present-day-harms and existential-risk communicators want, from transparency and liability to limits on the most capable systems, need a public that understands the case, sees itself in it, and keeps supporting it when the political wind changes.

In many countries a form of AI silence has set in: people use the tools daily, worry about them in the abstract, and rarely talk about the issue with anyone. Meanwhile the debate among those who do talk about it has split into several conversations, one about bias, others about surveillance and labour, even more others about catastrophic risk.

As a result politicians face little sustained pressure to act. When they do act, policies are vulnerable: the UK's AI Safety Institute became the AI Security Institute in 2025 and the US executive order on AI was revoked in January 2025 with little public reaction either way. High-stakes decisions about AI are being made without a public mandate in either direction.

AI advocates focus on policy change, shifting elite opinion through the media, and mobilising grassroots pressure. To succeed in the long term these strategies also need to engage the wider public and build the social mandate for durable governance.

Understanding how people talk and feel about AI is crucial to AI safety discourse.

1.2 The challenge of AI engagement

Generating strong and consistent public support for AI governance is hard for a wide range of reasons.

1.3 A social mandate is critical to underpinning AI safety

We are in a vicious cycle:

• Weak social mandate: a minority of the public identify AI as a priority issue, and few talk about it.
• Weak governance: no political mandate to create or hold the line on rules; commitments are voluntary and reversible.
• Companies not incentivised: no public pressure to trade capability for safety, and no penalty for deployment ahead of understanding.
• Passive public behaviour: people adopt tools without questions, and cannot see how to act on their unease.

A strong social mandate creates a virtuous circle:

• Strong social mandate: a majority identify AI governance as a priority and see themselves in the issue.
• Durable governance: rules driven and supported by the public, hard to reverse when the wind changes.
• Companies compelled: public pressure and social licence make safety a condition of operating, not a marketing line.
• Active public behaviour: people ask questions, choose, refuse, and join the local and national conversations that shape deployment.

1.4 Changing levels of concern: the lesson of 2023 to 2025

"Citizens should be empowered to push AI in a direction that can fulfill its promise as an information technology. But for that to happen, we will need a new narrative in the media, policymaking circles, and civil society, and much better regulations and policy responses. Governments can help to change the direction of AI rather than merely reacting to issues as they arise. But first policymakers must recognize the problem."Daron Acemoglu, Massachusetts Institute of Technology · Nobel Prize in Economics, 2024

Concern about AI rose sharply after ChatGPT's release in November 2022. The March 2023 open letter calling for a pause, resignation statements from senior researchers, and the Bletchley Park summit in November 2023 produced a surge of attention.

Polling in that period found majorities in the UK and US wanting AI regulated, and more Americans concerned than excited.

That attention led to real institutional change: AI safety institutes in the UK and US, the EU AI Act, and a run of voluntary commitments from the companies.

But it did not convert into commitment. By early 2025 the framing had moved decisively to opportunity. The UK published an AI Opportunities Action Plan in January 2025, the US revoked its AI executive order the same month, and the Paris summit in February 2025 dropped "safety" from its name. The public, still concerned in the polls, did not react, because concern had never been turned into an understanding of what governance was for or what its loss would mean.

This is the same pattern climate communicators experience in the UK after 2010: a spike in concern, a policy win, then a collapse in salience that undid much of the political buy-in.

A strategy that focuses on demands for policy change, with little attention to moving public engagement from concern to commitment, is vulnerable to changes in the political winds. For AI the cycle ran in two years rather than a decade.

The lesson is that backsliding is possible, and that holding ground needs an ongoing, broad-based and robust social mandate.

Converting high concern into social change needs evidence-based public engagement, which has not played a central role in AI campaigns or government communications so far.

1.5 Context analysis: what is happening in 2026?

• Concern without conversation: polling across countries shows high concern about AI, alongside low knowledge and low discussion. People are worried.
• Governance on several tracks: the EU AI Act is coming into force in stages, California's SB 53 has set frontier-model transparency rules, and the US federal government is pressing to pre-empt state laws. None of these was driven by visible public pressure.
• A field split: existential-risk and present-day-harms communicators work from different evidence, different funders and different vocabularies.
• Evidence that the dominant framings underperform: Seismic Foundation's 2025 message-testing study found the field's leading frames, including existential risk, performed worse than expected. Our own analysis of 1,534 comments on a viral MIRI reel found that the argument persuaded most people who replied, then left them resigned.
• Backlash forming locally: opposition to data centres over water, power and land is the first mass, cross-partisan public reaction to AI infrastructure, and it is happening without any connection to the safety debate.
• Populism and polarisation: AI is not yet a partisan issue in most countries. That window will not stay open. The US debate over state pre-emption is the first place it is starting to close.
• Harms arriving in daily life: chatbot harms to children, deepfakes, hiring and welfare algorithms, and job losses attributed to AI are making the technology concrete for people who had no view on it a year ago.
• Rising capability: each model generation makes claims that sounded like fiction sound less so, and gives sceptics less to point at. The capability objection is losing ground slowly, not the personification objection.

2. The core principles of public engagement

2.1 Values-based engagement and social mandate

The social science evidence is consistent: people do not form attitudes or change behaviour mainly by weighing expert information. They are moved by stories that feel right, told by people they trust, and made acceptable by the norms around them. This is as true for AI as for climate, and probably more so, because the technical case is harder for most people to evaluate.

People act when they understand the reasons and those reasons connect to their concerns. Explaining alignment or scaling laws to an audience worried about their child's homework or their job does not do that. Connecting AI governance to their child's safety, their livelihood or their say over their own community does.

Change of the kind AI governance needs will not come from information campaigns alone, the deficit model in which experts transmit facts to a receiving public. It comes from the interaction between people and the society around them.

Social norms decide what is acceptable to say and do, and people shape their behaviour by what their peers say and do.

A social mandate emerges when different communities are given a voice.

Building support while avoiding polarisation means enabling a broad range of public audiences to:

• see their values, identity and concerns in the AI story;
• know that people they identify with are engaging with AI in a way that looks authentic;
• be able to take actions that fit them, their concerns and their community.

Generating sustained, committed action from high concern therefore needs proper attention to people's values and the social contexts they live in. Engaging with diverse communal values, from family and fairness to security and freedom, is central to Common Signals' approach.

People act based on their values, beliefs, and social norms. AI safety comms today mostly starts at beliefs and skips the values that decide whether the argument is accepted.

2.2 A social licence to operate

Public engagement also bears on the structural shifts needed from the companies building AI. They need a social licence to operate. While they have it, they have little incentive to trade speed for safety, and voluntary commitments stay voluntary.

Companies struggle when they lose that licence, as tobacco did. Leaders of AI companies are already sensitive to this: the public framing of each model release, each safety incident and each data centre siting decision is managed carefully. The finding in our reel analysis that "blame the companies" was the one frame shared by believers and sceptics alike suggests the licence is weaker than it looks.

But with AI now embedded in work, education and daily life, the challenge of shifting perceptions durably is significant.

3. The elements of social transformation

Social transformation is complex and there is no blueprint. But recurring features link many rapid shifts in recent history: the Irish abortion referendum, changing attitudes to same-sex marriage, the spread of smoking bans, and the 2020 pandemic response, when governments changed the daily lives of whole populations in weeks.

Transformative change appears to need:

• A social mandate: a consistent majority of the public supporting significant change.
• Compelling narratives that relate to communal values: letting diverse communities engage in a way that fits their identity, rather than as a technical policy question.
• A supportive economic context: without a favourable economic wind, even popular rules may not be implemented or kept.

The third is the hardest for AI. The economic wind is blowing toward adoption at almost any cost, and governments see AI as their route to growth. That makes the first two more important, not less.

Where the economics cut against governance, only a mandate that is broad, cross-partisan and durable will hold.

3.1 Public consent and people-led change

People will need to be central to the AI transition. The choices being made about AI in workplaces, schools, hospitals and councils are choices about their lives, and rules made without their consent will be resented, evaded or reversed. Carbon pricing failed where it lacked buy-in; AI rules will fail the same way if they are experienced as something done to people rather than for them.

This is not unprecedented. Social pressure ended the slave trade, moved smoking out of public buildings, and won recognition for rights once thought impossible.

None of these came through technocratic change alone. Each needed widespread acceptance and a rapid shift in social and moral norms around the need to act.

Positions that once seemed radical became normal.

4. Generating a social mandate for AI governance

Common Signals believes three elements combine to generate a social mandate for AI governance: cross-societal concern, avoiding polarisation, and concern turned into action.

The indicators of success for each are set out below.

1. Cross-societal concern

• Polling consistently shows AI among the top five concerns for a population, and people across different communities are talking about it.
• Politicians report being pressured or supported on AI governance by their constituents.
• A majority of the public supports the delivery of specific governance measures.
• Audiences across society express support for advocates of AI governance and their campaigns.

2. Avoiding polarisation

• Parties across the spectrum recognise AI governance in their manifestos and speeches.
• Spokespeople associated with the right, with business and with faith communities speak out on AI governance.
• Polling shows support for governance holding across party lines, and decreasing tolerance for deployment ahead of safety.

3. Concern turned into action

• Populations report knowing what they can do about AI and doing it, from questioning use in their workplace or school to joining local decisions.
• Populations report knowing others in their community who care about AI and are acting.
• Communities are genuinely involved in decisions about AI deployment and hold decision-makers to account.
• Companies report demand from customers for safe and accountable AI.

These three elements combine: building and sustaining cross-societal concern in nations critical to AI governance, plus overcoming polarisation where it is starting to form, plus turning concern into action on behaviour, rules and companies, generates a social mandate for AI governance in key countries.

5. Critical themes for effective public engagement

Common Signals will focus on five themes that cut across its work.

5.1 Polarisation

AI is not yet a left-right issue in most countries. Concern spans the spectrum, and the strongest early local reaction, to data centres, is cross-partisan. That is an asset climate never had, and it is fragile. The US debate over federal pre-emption of state AI laws is the first sign of the issue sorting by party. Where an issue becomes identified with one section of society, action becomes sporadic, as climate showed in the US, Australia and Brazil.

Our approach: unite rather than divide, start from communal values, and recognise different motivations. Test messages and messengers with right-of-centre and non-metropolitan audiences before anyone else, because the field's default voice is neither.

5.2 Efficacy and resignation

The distinctive failure mode of AI risk communication is not disbelief but despair. In our reel analysis, "too late" was the second most-liked comment and fatalistic comments drew a fifth of all likes, while almost nobody responded to the call to act. Persuading people that the danger is real, and then leaving them with nowhere to put that belief, is worse than not persuading them.

Our approach: pair every threat message with a concrete, already-happening route to efficacy, and test which routes each audience finds credible.

5.3 Bridging present harms and future risk

The people worried about bias and surveillance and the people worried about catastrophe are describing the same technology, built by the same companies, governed by the same rules. To the public these are not two issues. To the field they are two tribes. The one frame our analysis found shared by believers and sceptics, that the companies are responsible, is also the one frame both groups of communicators can use.

Our approach: serve both groups from one evidence base, test frames that work across them, and treat the balance of our users as a standing health check on whether the bridge holds.

5.4 Capability and personification

The two objections that sceptics actually make are that current AI is too weak or too hyped for the danger to be real, and that AI is being personified as an agent when the companies building it are responsible. Neither is answered by the field's current framing, which tends to describe AI as a monster and its capability as a given.

Our approach: test messages that meet these objections rather than talk past them, and track how the capability objection shifts as models improve.

5.5 Trusted messengers

For most of the public, the people explaining AI are strangers: researchers, founders and campaigners they have never heard of. Trust is the scarce resource. Who the public believes about AI, segment by segment, is unknown, and the field's messenger choices are made by availability rather than evidence.

Our approach: build and maintain a map of who reaches which audiences, test messenger trust as part of every study, and support trusted voices from outside the field, from teachers and doctors to faith leaders and union reps, to tell their own AI stories.

6. Engaging with critical stakeholders

Common Signals will work with six types of stakeholder.

6.1 AI safety advocates

• Existential-risk organisations (ControlAI, PauseAI, MIRI, Future of Life Institute and others) and present-day-harms organisations (Ada Lovelace Institute, Foxglove, AI Now, Connected by Data and others) are the primary recipients of our research and tools.
• Funders on both sides increasingly recognise the gap in public engagement. The need for a body like this has been stated openly on the EA Forum.

6.2 Non-AI communities

• Supporting spokespeople from outside the field, in unions, schools, faith groups, health services and local government, to tell authentic stories of AI concern and action is a central aim.
• Communities already reacting to AI locally, especially over data centres, are engaged as sources of evidence and as messengers.

6.3 Social scientists

• Public-attitudes researchers at the Ada Lovelace Institute and Alan Turing Institute (Public Voices in AI), More in Common, and Seismic Foundation generate the evidence we build on. We build on their work rather than duplicate it, credit it clearly, and feed our findings back.
• A quantitative research partner delivers segmentation and message-test analysis, because the founder has not run a national survey and says so.

6.4 Media and creators

• Journalists covering AI need a neutral read on public opinion. From Year 4, a live public opinion tracker gives journalists a reliable source to contact.
• Creators who explain AI to mass audiences are mapped and, where willing, briefed. Our messenger map lists 69 individuals and organisations across existential-risk and present-day-harms voices.

6.5 Experts

• Working with AI researchers, including those who have left the companies, to speak up in ways that resonate beyond the field.
• Enabling other expert influencers, such as teachers, doctors and economists, to communicate what they know about AI to the widest possible audiences.

6.6 Decision-makers

• Building public pressure on politicians to govern AI in a way that resonates across the spectrum.
• Providing governments, AI safety institutes and parliaments with neutral, sourced evidence on what the public thinks and what moves them.
• Making the case to funders that public engagement infrastructure deserves investment.

7. Common Signals' contribution

7.1 Pathways to generating a social mandate

Guided by the five themes and working with the six stakeholder groups, Common Signals will build the social mandate for AI governance through three complementary approaches, in order of priority:

• 1

  Mobilising understanding

  How to engage key audiences with AI, and ensuring the research drives practice rather than sitting in a report.

• 2

  Motivating communicators

  Supporting organisations and trusted messengers on both sides of the field to engage their audiences through evidence rather than instinct.

• 3

  Promoting informed public engagement

  To decision-makers and funders, and the centrality of people-based approaches to governing AI well.

7.2 Key activities

• Values-based audience research segmenting the public by values and testing which messages move which segments, starting with a US pilot of around 2,000 adults in 2027.
• Rapid analysis of real public reaction, coding comment threads and public responses to incidents, videos and campaigns, published within weeks. See the first examples:
  • Bad, Bad, Not Good: How did the public respond to a caveman‑speak explainer of AI risk?
  • OpenAI and Shut: how METR's investigation of the OpenAI and Hugging Face incident closed the marketing-stunt argument
• Mapping and testing trusted messengers: who reaches which segments, and whom each segment believes.
• A free, open toolkit, translating the evidence into practical guidance per segment, with the method documented for others to replicate.
• Messenger training, piloted in Year 2 and delivered increasingly through partners and workshops from Year 3.
• Making the case, to funders and decision-makers, that public engagement infrastructure for AI is worth building.

Out of scope: running campaigns, lobbying, taking a policy position of our own, or selling closed advice. The moment we do any of these the neutrality that makes the evidence usable and trustworthy is lost.

7.3 Geographical focus

• Nations that host frontier AI development, where the debate is starting to polarise and where governance decisions set the pattern for everyone else. Mainly the US.
• Nations with high public concern and stated governance ambition, but where concern has not become commitment. The UK and much of the EU.
• Nations adopting AI rapidly with little public discourse about it, where the pattern is still forming. India and parts of Africa and Latin America, where our messenger map is currently empty.

8. Placing people at the heart of the AI transition

Common Signals puts people, their values and their experience, at the heart of public engagement on AI, to build the social mandate that governing it well will need.

Our role is as a catalyst of understanding and practice, not a campaigner. We support the organisations already doing this work to do it with evidence, and we serve both sides of a field that currently does not share a table.

8.1 Capacity

Common Signals is at pilot stage: one part-time founder, a contracted quantitative partner, and an advisory group of five to seven being formed with present-day-harms and existential-risk voices in balance.

The five-year strategy takes it to an independent institute of ten to twelve staff by 2031, on £1m to £1.5m a year, with no funder above a quarter of income and neither side of the debate a majority. It works with and through partners rather than delivering campaigns, and the staff ceiling is deliberate.

Each year ends with a published decision gate: named evidence that justifies the next stage, and a named fallback, including the conditions under which we would stop.

8.2 Added value

Our expertise is in generating, synthesising and translating public-engagement evidence into practice. We fill a gap nobody else occupies: adjacent organisations either test messages to advance a position, or study elite audiences. We test messages for the public, take no position, and give the results away.

An organisation whose whole case is that evidence beats instinct has to evidence itself. From Year 2 we publish an annual impact report with metrics defined in advance, adoption tracked at follow-up rather than assumed, cost per output reported, and null results and failures published alongside the rest.

We believe public involvement is essential for effective AI decisions, and we will work with funders and decision-makers to ensure that it gets built.

References

• Stanford HAI (2025) AI Index Report 2025, economy chapter.
• Seismic Foundation (2025) message-testing study, arXiv:2511.06525.
• Ada Lovelace Institute and Alan Turing Institute, Public Voices in AI (attitudestoai.uk).
• Common Signals (2026) Bad, Bad, Not Good: How did the public respond to a caveman‑speak explainer of AI risk?
• Social Change Lab (2022) research highlights on protest and public opinion.
• EA Forum, Ten AI safety projects I'd like people to work on (July 2025).
• Diagram based on Values Beliefs Norms Theory chain of causal influence in: Sovacool, B. and Hess, D. (2017), Ordering theories: Typologies and conceptual frameworks for sociotechnical change. Social Studies of Science, 47:5, pp.703-750. Available at: journals.sagepub.com/doi/10.1177/0306312717709363.
• Clarke, J., Webster, R. and Corner, A. (2020) Theory of change: creating a social mandate for climate action. Oxford: Climate Outreach.