Which audience is most neglected in AI safety communications?
By benrmatthews @ 2026-08-19T11:51 (+2)
A first-cut look at the public, policymakers and the labs, through scale, neglectedness and tractability.
Summary
- A question I keep hearing among people who work on AI communications is which audience to prioritise: the general public, policymakers, or the people inside the labs. It matters for comms strategy and for career choice, and there does not seem to be a clean analysis of it.
- Run through scale, neglectedness and tractability, my tentative read is that the public is the most neglected of the three and probably the most underrated, that policymaker-facing work is high-value but less neglected and access-gated, and that lab-directed communication is the least neglected per head and the least tractable through information alone.
- The bottom line is weaker than that sounds. The audiences are complementary, personal fit dominates at the individual level, and the weakest link in the public case, whether moving opinion actually moves outcomes, is precisely the thing nobody has measured.
- Conflict of interest: I am building a public-facing research organisation, so I gain if the "public is neglected" conclusion holds. I have tried to argue the other side fairly, and I would most value pushback on the public's tractability.
What I'm asking
Corrections and pointers, in roughly this order:
- Is there an audience-level analysis of AI comms I have missed?
- Is my read of the labs as saturated and incentive-bound fair, or am I underrating normative pressure?
- How much does moving public opinion on a technology actually move policy on it?
The question
Most discussion of AI safety communications is about what to say. This piece is about who to say it to. The three audiences people usually have in mind are the general public, policymakers and regulators, and the technical and executive staff at frontier AI labs.
The natural tool is the framework EA already uses for causes: importance or scale, neglectedness, and tractability (the person who prompted this used "controllability," which I read as the same thing). It is usually applied to whole causes; using it to compare audiences inside a single cause is less common, and I have not found anyone who has done it properly for AI comms.
One caveat before the ratings. What follows is judgement. I will use high, medium and low, but the reasoning matters more than the label, and I have tried not to dress a guess up as a score.
The public
Scale: high, but mediated. The public is enormous and shapes the political weather that everything else operates in. Its effect on AI outcomes, though, runs mostly through policymakers, so the influence is real but indirect.
Neglectedness: high, the most neglected of the three. Overall AI safety resources are tiny against the technology: global AI investment reached around 250 billion US dollars in 2024, while safety takes well under a tenth of a percent, and the slice of that going to evidence-based public engagement is smaller again. The field has almost no shared, neutral public-comms infrastructure, and the demand for one has been stated openly.
Tractability: uncertain, and this is the crux. Public attitudes are movable, and message-testing such as the Seismic Foundation's 2025 study shows some framings work far better than others, including the finding that the field's favoured existential-risk framing underperforms. Appetite exists: Public Voices in AI finds broad public support for regulation. But whether moved opinion moves outcomes is contested. Gilens and Page find economic elites dominate US policy; Burstein finds opinion does move policy, more so as salience rises. And the effect sizes public campaigns achieve are modest. So the public is easy to inform and hard to convert into outcomes, and the size of that gap is unknown.
Policymakers and regulators
Scale: high and direct. The numbers are small and the per-person leverage is large, which is the opposite shape to the public.
Neglectedness: medium, and less neglected than the public. A substantial and growing AI governance ecosystem already targets this audience: policy institutes, technical governance groups, fellowships routing talent into government, and dedicated briefing work. It is not saturated, but it is comparatively resourced, and the existing field maps skew toward this end and the technical end.
Tractability: medium to high, conditional on access. The mechanism is well understood, evidence, briefings, relationships, and it works when you are in the room. The constraints are that access is gated, the space is crowded with sophisticated actors including industry,.
People at the AI labs
Scale: very high per person. A few hundred people make the decisions that most shape frontier development. If communication could move them, the leverage would be enormous.
Neglectedness: low, relative to the size of the audience. This is the counterintuitive one. The labs are tiny as an audience but heavily engaged already, through large internal safety teams, external technical collaboration, and constant discourse aimed squarely at them. Per head, they may be the least neglected audience in the field.
Tractability: low, through communication specifically. The blunt version, which I heard put well recently, is that the people at the labs already know the arguments; they do not act on them because of competitive and commercial incentives, not because they are underinformed. Communication can shift norms and what is sayable, which is not nothing, but it rarely changes a lab's trajectory directly.
The comparison
| Audience | Scale / importance | Neglectedness | Tractability |
|---|---|---|---|
| The public | High, but mediated through policy | High, the most neglected | Uncertain: easy to inform, unclear whether that moves outcomes |
| Policymakers and regulators | High and direct | Medium, comparatively resourced | Medium to high, if you have access |
| People at the labs | Very high per person | Low, saturated relative to its size | Low via information; the bottleneck is incentives |
What I think this adds up to
Tentatively: the public is the most neglected audience and, I suspect, the most underrated, because the neglect is severe and the first half of the job, moving attitudes, is tractable and improvable with better evidence. The weakness is the second half, the link from attitudes to outcomes, which is contested and unmeasured. That weakness is not a reason to walk away; it is the most valuable thing to research, because a field cannot rationally allocate between these audiences while the public's leverage is a question mark.
Policymaker-facing work is high-value and I would not talk anyone out of it, but it is less neglected and gated by access, so the marginal contribution is lower than the raw leverage implies. Lab-directed communication is the one I would be most cautious about as a primary focus. The people matter; the trouble is that they are already saturated, and the lever comms pulls is weak against their incentives.
Has anyone done this yet?
ThI went looking for a rigorous importance-neglectedness-tractability analysis by audience for AI safety comms, and did not find one. What would make this rigorous:
- A map of communications effort and spend by audience, so "the public is neglected" stops being an impression and becomes a number.
- Effect-size estimates per audience, so tractability is compared on evidence rather than vibes.
- A model of the mediation, since the public's value depends almost entirely on how reliably public opinion transmits to policy, and that transmission is exactly what the literature disputes.
None of this is necessarily expensive: It is desk research, a spend map, and a handful of expert interviews, and it would improve a lot of career and strategy decisions that are currently made on instinct. I think someone should do it.
What I'm asking
Corrections and pointers, in roughly this order:
- Is there an audience-level analysis of AI comms I have missed?
- Is my read of the labs as saturated and incentive-bound fair, or am I underrating normative pressure?
- How much does moving public opinion on a technology actually move policy on it?
Sources
- Stanford HAI, 2025 AI Index, economy chapter
- Overview of the AI safety funding situation
- Ten AI-safety projects people want built (EA Forum)
- Seismic Foundation message-testing (arXiv:2511.06525)
- Public Voices in AI
- Gilens and Page, Testing Theories of American Politics (2014)
- Burstein, The Impact of Public Opinion on Public Policy (2003)
- Social Change Lab, research highlights