Thanks for this! Like I said above, I spent years in advertising and market research and the finding that people feel first and rationalise afterwards is strong. So, as you say, we need to get to that visceral response as the first step.
But I don't think we need to oversimplify or overstate to be effective. If we did that, there would also be a cost because credibility is one of the few advantages the AI safety side actually has. What I'd argue instead is that we already have accurate material that is alarming but simply badly delivered. Making that content more effective is the challenge.
On targeting. I agree that some orgs deliberately target elites and this is sometimes correct. My argument is about the balance. When a large share of the field's output goes to audiences that already agree with the AI safety ‘thesis’, all that produces is good discussion and little movement. If I had to reallocate, I'd shift toward the public since public pressure is often what makes policymakers act in the first place.
The data centre backlash in the US is the clearest recent example. Both parties are now legislating on it, not because anyone made a careful technical case to legislators but because people got angry about something in their own neighbourhood. Some of the specific claims about water usage and energy prices are contested but the anger moved policy anyway.
On abandoning epistemics. That isn't what I argued at all. My request is for professional communications teams, messages tested against outside audiences, and spokespeople who can talk to a journalist or a legislator. None of that requires exaggeration.
One of the things I learned from working in advertising research is that emotion comes before reason in how people process arguments. That doesn’t mean we need to shock people. A better option could be to lead with what people already feel rather than inventing things they should feel.
The main point here is that the opposition understands that we’re in a political fight. The AI safety field largely still treats this as an academic debate, where better reasoning is expected to win on its own. It won’t.
On money and scale. I'm not proposing taking money from AMF or any other cause. My argument is about allocation within AI safety funding, which is already substantial. For example, the grantmaking.ai database has research entries outnumbering comms entries by more than five to one. Bring that ratio down, not necessarily to 1:1, would be a good start.
I don't think it's actually about polish versus volume since both models currently have the same problem, i.e. they don't seem built to end in a political ask.
AI in Context gets millions of views but is more focused on converting those into newsletter signups rather than political outcomes. And then there's PDKU, which went the other way with dozens of lower-context creators given minimal message discipline and told fellows not to post about AI at all if they didn't want to. But Celia Ford's reporting found the "AI will kill us" narrative reproduced itself anyway, right down to protest posters reading "EVEN THE CEOs WILL DIE": https://www.transformernews.ai/p/inside-ai-safety-influencer-bootcamp-plz-dont-kill-us
So volume didn't diversify the message the way it was meant to. It just found the same message through less curated channels. PDKU's own organisers measured success as "views per dollar," which is the same theory-of-change gap I write about, just in a different aesthetic.
Thanks for this! Like I said above, I spent years in advertising and market research and the finding that people feel first and rationalise afterwards is strong. So, as you say, we need to get to that visceral response as the first step.
But I don't think we need to oversimplify or overstate to be effective. If we did that, there would also be a cost because credibility is one of the few advantages the AI safety side actually has. What I'd argue instead is that we already have accurate material that is alarming but simply badly delivered. Making that content more effective is the challenge.
In response to the three points:
On targeting. I agree that some orgs deliberately target elites and this is sometimes correct. My argument is about the balance. When a large share of the field's output goes to audiences that already agree with the AI safety ‘thesis’, all that produces is good discussion and little movement. If I had to reallocate, I'd shift toward the public since public pressure is often what makes policymakers act in the first place.
The data centre backlash in the US is the clearest recent example. Both parties are now legislating on it, not because anyone made a careful technical case to legislators but because people got angry about something in their own neighbourhood. Some of the specific claims about water usage and energy prices are contested but the anger moved policy anyway.
On abandoning epistemics. That isn't what I argued at all. My request is for professional communications teams, messages tested against outside audiences, and spokespeople who can talk to a journalist or a legislator. None of that requires exaggeration.
One of the things I learned from working in advertising research is that emotion comes before reason in how people process arguments. That doesn’t mean we need to shock people. A better option could be to lead with what people already feel rather than inventing things they should feel.
The main point here is that the opposition understands that we’re in a political fight. The AI safety field largely still treats this as an academic debate, where better reasoning is expected to win on its own. It won’t.
On money and scale. I'm not proposing taking money from AMF or any other cause. My argument is about allocation within AI safety funding, which is already substantial. For example, the grantmaking.ai database has research entries outnumbering comms entries by more than five to one. Bring that ratio down, not necessarily to 1:1, would be a good start.
I don't think it's actually about polish versus volume since both models currently have the same problem, i.e. they don't seem built to end in a political ask.
AI in Context gets millions of views but is more focused on converting those into newsletter signups rather than political outcomes. And then there's PDKU, which went the other way with dozens of lower-context creators given minimal message discipline and told fellows not to post about AI at all if they didn't want to. But Celia Ford's reporting found the "AI will kill us" narrative reproduced itself anyway, right down to protest posters reading "EVEN THE CEOs WILL DIE": https://www.transformernews.ai/p/inside-ai-safety-influencer-bootcamp-plz-dont-kill-us
So volume didn't diversify the message the way it was meant to. It just found the same message through less curated channels. PDKU's own organisers measured success as "views per dollar," which is the same theory-of-change gap I write about, just in a different aesthetic.