Noah Smith just said the quiet part loud: AI safety isn't a technical problem, it's a fear distribution problem.
The Summary
- Noah Smith argues the AI safety debate is missing two critical pieces: making skeptics genuinely afraid of AI risk, and making China afraid enough to cooperate on safety measures
- Current safety discourse splits between doomers who can't convince skeptics and techno-optimists who dismiss existential risk entirely
- The real challenge isn't building alignment tech, it's building political will across adversarial nations before someone builds AGI without guardrails
The Signal
Smith's thesis cuts through the usual AI safety theater. He's not rehashing the Yudkowsky doom scenarios or the e/acc "just build" mantras. He's pointing at the gap between what AI safety researchers worry about in private and what actually moves policy or pauses development.
The first missing piece is visceral fear among skeptics. Not the abstract "paperclip maximizer" thought experiments that make rationalists nod and normies roll their eyes. Smith argues we need something like the 1983 TV movie "The Day After" did for nuclear war: make the threat real enough that people who currently think AI doom is nerd LARPing start taking it seriously. The problem is AI risk lacks the visual immediacy of mushroom clouds. A misaligned superintelligence doesn't photograph well.
"We need to make skeptics properly scared. And we need to make CHINA properly scared."
The second piece is harder and more dangerous to talk about: China. Smith points out that even if the U.S. solves alignment, coordinates industry safety measures, and gets regulation right, none of it matters if China (or any other major AI power) decides to race ahead without safety constraints. This is the geopolitical dimension that most AI safety writing politely ignores.
Here's where it gets tricky:
- International AI safety cooperation requires mutual fear of the technology itself
- China has shown zero public interest in AI pause advocacy or existential risk frameworks
- Any U.S. attempt to "make China scared" of AI risk could backfire as perceived weakness or manipulation
- The prisoner's dilemma is stark: whoever slows down for safety might lose the AGI race entirely
Smith doesn't offer clean solutions because there aren't any. He's diagnosing the problem most AI safety discourse won't touch: this isn't just about building better RLHF or interpretability tools. It's about convincing populations and rival governments to coordinate on slowing down the most economically and militarily valuable technology in human history. Good luck with that.
The timing matters. Major AI labs are reportedly approaching capabilities thresholds that could make this coordination window close fast. If we hit AGI before we hit consensus on safety, the technical alignment work becomes moot.
The Implication
If Smith is right, the AI safety community needs a strategy shift from technical papers to political psychology. How do you make AI risk feel real without creating panic that accelerates the race? How do you get China to the table without it looking like a ploy to slow their tech development?
Watch for more public intellectuals making this same pivot from "will AI doom us" to "how do we get everyone scared enough to cooperate." The meta-question of AI safety might be more urgent than the object-level alignment research.