AI safety can be a Pascal's mugging even if p(doom) is high

In a recent LessWrong post, Elliott Thornley (EJT) tackles a common debate in AI safety circles: whether AI safety advocacy constitutes a Pascal's mugging—a philosophical argument where a low-probability, high-stakes claim is used to justify extreme actions. Thornley argues that critics who dismiss this concern because p(doom) (probability of AI catastrophe) is high miss the point. He illustrates with a story: if God flips a coin at the end of time with a 50% chance of hell, and a stranger claims he can guarantee heaven for your wallet, it's still a mugging because the probability you make a difference is astronomically low (1 in a bajillion). The baseline risk doesn't change that.

However, Thornley diverges from the typical Pascal's mugging critique by asserting that AI safety is not a true mugging. He contends that p(you avert doom)—your personal ability to influence outcomes—is high enough to justify action. He notes that individuals are often just a few degrees of separation from key players like lab CEOs and policymakers, making personal impact plausible. Commenter ryan_greenblatt adds that the argument often applies to societal actions (e.g., U.S. policy) rather than personal choices, and high p(doom) defeats the mugging claim when societal intervention is feasible. Thornley's post reframes the debate from abstract risk to actionable influence, urging nuanced thinking about where individuals can actually make a difference.