Hi ben
I don't see public engagement as a bottleneck to AI Safety. If at all, it may help in getting the urgency it needs. Secondly, there are so many narratives and all of us are overwhelmed in terms of what to believe or who to believe.
If a voice/ org/ platform with credibility can lead in educating public, i believe it will be incredibly helpful.
Agreed, public engagement is not a direct bottleneck, but I think that governments respond to political salience, and the window for governance intervention is narrowing. Public engagement that builds salience might be what makes government regulation arrive in time rather than after power concentration. or gradual disempowerment has taken place.
On narrative overload, the Seismic research I cited found existential-risk framing performs worst across almost every demographic. So the problem may not just be too many voices, but that the loudest ones are using the register least likely to reach anyone not already convinced.
I can also envisage an "America Talks AI" version if the UK route isn't the right one, built on More in Common's US Hidden Tribes segments and that could launch through an existential-risk incubator with a strong US presence.
This is a concept I've been developing, posted in the hope that people who know the field better than I do will tell me where it could be improved.
I work in communications for nonprofits and social impact organisations, predominately form a climate change and sustainability angle, and I'm currently just finishing the BlueDot course on AGI Strategy. As such, please treat this as a comms practitioner's outside view of an aspect of the AI Safety ecosystem. I'm looking for corrections to a concept, not judgement on a finished proposal.
I'm cross-posting it to my blog and here in the EA Forum, as I would rather have the holes in the concept found sooner rather than later
Where I am guessing, I say so, and I have put the guesses in the form of predictions to be checked. I'm also happy for this idea to be built by people better placed than I am, if it proves credible.
Disclosure: I researched and drafted the post, but used Claude to help refine it and align with the Forums' norms, before I did a final edit. The published piece likely contains >20% AI-generated text, which has been then edited again by me.
Climate is the one case where I can point to this problem being solved, as I've worked in the area for 10+ years and am aware of the organisations that exist in the ecosystem.
For years, talking to the public about climate change was done by instinct: earnest, well meant, and aimed mostly at people who already agreed. Then a handful of organisations, Climate Outreach chief among them, did the work of finding out who the public actually were, what moved each part of them, and whom they trusted to hear it from, and then gave the answers away.
Climate Outreach is a British charity. Its flagship, Britain Talks Climate, segments the public by core values, which means grouping people by the underlying values that shape their worldview rather than by demographics like age or income, works out what moves each group and whom they trust to say it, and ships the result as a free toolkit with training. Its seven British value segments sit underneath a great deal of UK climate communication. The work is given away and held non-partisan, which is why it has been picked up by bodies from the IPCC to local councils that might not take messaging direction from a campaign group.
I see that AI Safety is roughly where climate was before that work existed. The technology is moving faster than the public's understanding of it, the people trying to close that gap are largely guessing, and there is no shared, trusted place the evidence lives.
An EA Forum list of wanted AI safety projects calls for a communications outfit whose first job is research on messages, frames and messengers, which is close to a one-line description of the research engine at the heart of what I'm proposing.
Nobody does this for AI safety as far as I can see, but I have missed something.
There are two maps of the field I found:
I used AI to join and ead them together to get a wider picture of the field.
The outcome was that is seems every individual piece of the Climate Outreach pipeline exists somewhere, but no single body runs the whole pipeline, and the pieces are split across two camps that barely talk to each other.
The split is roughly between people who focus on AI's present-day harms, such as bias, surveillance and the effect on jobs, and people who focus on existential risk, meaning the possibility that advanced AI causes catastrophe or even human extinction. On one side is the present-harms and justice wing: the Algorithmic Justice League, DAIR, the AI Now Institute, Connected by Data, Fight for the Future.
On the other is the existential-risk wing: PauseAI, Control AI, the Center for AI Safety, the Future of Life Institute, the Existential Risk Observatory, Evitable.
The only groups I could find on both maps are pause-and-control campaigners like PauseAI, Stop AI and Control AI.
The adjacent work is real but pointed elsewhere. Explainable turns AI futures into stories for organisations, for example.
On public literacy the field is strong: BlueDot Impact trains people entering AI safety, Rob Miles and Rational Animations explain alignment to people who already want to know, and Non-Trivial and Leaf reach teenagers.
Structured deliberation is covered by Democracy Next and the Odyssean Institute,
Closest of all to this concept is the Ada Lovelace Institute and the Alan Turing Institute produce serious public-attitudes research.
A clear example of message-testing is from a November 2025 paper, the Seismic Foundation, the agency Blue State and the Centre for Future Generations tested AI-risk frames on the public and found that existential-risk messaging, the field's dominant register, was the worst-performing theme across every demographic group, while concrete frames about jobs and children mobilised best.
This is the closest thing I've found to what I'm describing. It is also a single research report arguing for one mobilisation strategy, not an open toolkit and training programme that the whole field can keep using.
The gap I could find nobody was filling was in training trusted non-expert messengers (the clinicians, scientists, faith and professional voices) to reach sceptical adults who are not seeking the information out. That is the part Climate Outreach treats as central, and the part AI safety has skipped.
Part of the case is simple proportion. AI now attracts enormous investment, around 250 billion dollars in 2024 by Stanford's count, while the share going to safety is well under a tenth of a percent, and the share of that going to public engagement is a sliver of a sliver.
Public concern is high but unorganised, and the field is only beginning to test what actually moves people to care about issues around AI safety.
The Seismic study above found existential-risk framing, the field's dominant register, to be the worst-performing theme across every demographic group, and its authors note that this is counterintuitive to a community that was itself mobilised by exactly that framing.
The first version would would build on Climate Outreach's existing seven British segments, seek Climate Outreach itself as a host or mentor, and raise non-aligned money from trusts on the model by which the Nuffield Foundation stood up the Ada Lovelace Institute as a deliberately independent body.
My first instinct was that this needs a big new survey, on the scale of Britain Talks Climate's original 10,385-person study. Then I looked properly at what already exists.
The Ada Lovelace and Alan Turing Institutes' Public Voices in AI survey is a nationally representative study of 3,513 UK residents from November 2024, structured around eight concrete applications of AI rather than AI in the abstract, and it deliberately oversamples groups usually missed in this kind of research. It is good, and it is authoritative. So the question stopped being "do I need a survey" and became "what does this one not give us" to build a toolkit and training programme on top.
Those three things (values segmentation, the messenger map, and framing tests, are precisely the engagement layer that turns attitudes data into something a campaigner can use.
This also becomes cheaper and faster to pilot: run AI-attitude questions through the existing British value segments, add a trusted-messenger map, test framings, and lean on Public Voices in AI for the substance rather than re-gathering it.
A full standalone survey stays available as a costlier alternative, worth it only if the partnership cannot be struck or the existing data proves too thin. It also turns the most obvious potential rival into the most obvious partner, which is usually a good sign.
I would rather raise these myself than have them raised for me, so here are the places I think it is most likely to fail.
The effects are probably modest. Social Change Lab's research on movement outcomes finds opinion shifts in the low single to low double digits, with weaker and more contested effects on direct policy change. Public opinion's independent influence on elite technology policy is disputed: Gilens and Page (2014) find economic elites dominate US policy, while Burstein's review of the responsiveness literature finds opinion does move policy, more so as an issue's salience rises. So the honest framing is mandate-building and discourse-shifting, not a lever on legislation.
The cheap version of the research depends on the Public Voices in AI team being willing to have their work built on, which is an assumption I have not tested.
And the largest caveat is the one at the top: I am not an AI safety insider (but am interested as shown by participating in the BlueDot AGI Strategy course), so the most useful thing you could tell me is that I have misread the field.
In a handful of years the AI Safety field has built deep technical research, a governance layer, funders, incubators and training.
The public-facing layer I am describing is small and ordinary by comparison. It is the kind of thing that gets built once someone decides it should exist.
Climate Outreach is the proof., so perhaps the open question for AI safety is only whether its version arrives before the moment it is needed or after.
I think that is worth a capable team's best effort. I am fairly sure I am not that person (though could be part of a team which is most of why I am posting this rather than starting it.
I'd value scrutiny on any of the ideas in this concept.
Here are a few questions I've been thinking through:
If you have thoughts on any of these, I'd welcome them in the comments or directly.
I'll update this post with corrections and with anything I got wrong, including organisations I've missed from my research or even mischaracterised.
I'll also credit people who improve my thinking.
If even one of the corrections saves me misdirected time and effort, this concept post will have done its job.
Thanks in advance.
Thanks Ben. A useful outside comparison from climate communications and asking whether the same kind of infrastructure is missing here.
The main thing I’d keep testing with the engagement you get (and can you find 10 people or so to send it through to directly for comment?) is whether public engagement is actually a bottleneck for AI safety right now, and if so, for which audience. It may be that the strongest audience is perhaps supporting policy audiences, professional groups or organisations already doing adjacent work in that order?
I’d also watch whether the Climate Outreach analogy is doing too much work. AI safety may need something similar in function, but quite different in audience, governance, neutrality, legitimacy, and risk management. Interesting to find out!
Either way, I think this is a strong next move. It should give you useful information about the ecosystem, the culture, the gaps, and whether this kind of work gives you energy. It also seems like a good way to test the “translator / operator / bridge-builder” role we discussed, without needing to prematurely commit to founding anything.
Thanks @Martin Underwood. Both of your challenges have been taken on board.
On the bottleneck question, I'm planning to use the next few weeks of ecosystem conversations to test this, starting with whether the audience is policy-adjacent rather than general public. The ordering you suggest (policy audiences, professional groups, adjacent orgs) is probably the right one to work through.
On the analogy doing too much work, agreed. Climate Outreach is useful as a functional model (values segmentation, trusted messengers, long-term framing infrastructure) but the audience, governance, and legitimacy questions are hopefully different enough. Will treat it as a starting point for questions and further research.
Good prompt on finding 10 people to send it to directly. Will do that before making any more concrete plans.