Last year, the super PAC Leading the Future raised $125 million to shape how Americans think about AI regulation. Its affiliates paid influencers $5,000 a video and ran a news operation staffed by AI agents posing as journalists. Contrast that to what happened earlier this year, when a few hundred people gathered outside an AI lab holding handmade signs that couldn't agree on what they were protesting.
As someone who moved into AI safety from journalism, this dichotomy is hard to ignore. The safety side has the Nobel laureates, the scientific consensus, and a public already anxious about AI. But it cannot convert any of that into political consequences because it never built the machinery to do so. Indeed, AI safety is treating a political fight like a research debate and, until it builds a professional communications operation, being right will keep losing to being organised.
Celia Ford wrote that AI safety's insularity "may otherwise become a liability". I believe it already is. And while she covered the problem as being about who the field talks to, the failures run deeper by focusing on the wrong message, the wrong channel, and the wrong theory of change.
How AI safety is getting it wrong
First, the field leads with the concern that its own audience ranks last. According to Anthropic's 81,000-person study, the largest qualitative AI study ever conducted, existential risk was cited by 6.7 per cent of participants, ranking it last of the 13 concerns. The top four were unreliability, jobs, autonomy, and cognitive atrophy. Some will say that this proves existential risk is neglected and deserves more attention, not less. But neglectedness tells you what to talk about, not how to do so. And the 'how' is still broken.
This brings us to the second failure, which is that even the right message fails if nobody finds it. The field publishes to LessWrong, the Alignment Forum, the EA Forum and arXiv. But the public it needs to move is on mainstream television, podcasts, and social platforms. So while these forums may coordinate the community, they will never persuade anyone outside it.
Now, I'm aware of the irony of publishing this article here. But the distinction matters. This argument is addressed to the AI safety field and this forum is where the field gathers. Indeed, these forums are good at what they are for, i.e. coordinating the people already inside the community. The failure comes from publishing what should be outward-facing work through these platforms and then wondering why nobody who isn't already part of the AI safety bubble reads it.
Both of those failures rest on a third one, which is what the field believes will actually change things. When it faces a policy fight, AI safety's response is more research, more papers, more roundtables, and more convenings. But that is how ideas circulate inside expert communities, not how policy gets made. Meanwhile, the opposition runs political organising, media relationships, lobbying, and dark money advocacy. Indeed, the results are visible in who gets through the door. Corporate Europe Observatory found that of 97 senior European Commission meetings on AI in 2023, 84 were with industry and just twelve with civil society.
The opportunity that almost worked
The field does occasionally break through, which is why the obvious objection is AI 2027. It went viral, JD Vance discussed it, and it sparked copycats like Europe 2031. Isn't that reach?
Reach, yes. A campaign, no.
The report anchored its headline on a specific year, which many readers took as a prediction. The caveat that 2027 was the authors' "modal estimate", a term few readers would even understand, sat in a footnote. Only in November 2025 did they move the clarification into the text itself. Then in December, lead author Daniel Kokotajlo revised the timeline outright: "around 2030, lots of uncertainty though".
A year on, the authors and independent trackers still argue over whether the scenario ended up with 51% of predictions on track or two-thirds instead. But regardless of the grading, the walk-back is now a talking point opponents will reach for whenever the field's credibility comes up.
Why communications is the fix
The gap between reach and results is the whole problem; better communications is the solution. Yes, research matters and nobody is arguing otherwise. But communication is what makes research actionable. Indeed, every safety outcome the field wants, from evaluation standards to enforceable regulation, runs through this bottleneck. It's also where the community has invested the least.
The public anxiety Anthropic documented will stay raw until someone builds the operation that converts it into legislation and enforcement. For example, a finding on model deception can only become an evaluation standard that a regulator demands if someone has translated it for that regulator. And a policymaker who understands why frontier training runs need oversight will legislate for it but only if someone has explained it well.
The resources clearly exist. Funders have put hundreds of millions of dollars into AI safety over the years and the field now has its own dedicated funding database at grantmaking.ai. Yet look at how its roughly 6,900 entries are tagged: 1,416 carry the research tag, while comms appears 267 times, education 222, and advocacy 221. These are tags rather than budgets so they measure where effort clusters and not necessarily where every dollar flows. But the staffing tells the same story: one AI governance advocate has estimated the field runs at roughly three researchers for every advocate. The field, in other words, is not prioritising communications.
Of course nobody expects nonprofits to match lab salaries, which can reach upwards of $400,000. But the under-investment speaks to a field that has not made communications part of its core infrastructure. And that is the ask. We need professional teams who are paid fairly, messages tested on outside audiences, and spokespeople who can face a journalist or a legislator in a language they understand.
Ultimately, the accelerationist side never won the argument because it built the operation instead. Until the AI safety field does the same, handmade signs that can't agree with each other won't amount to much at all.