An economic argument from LessWrong: AI labs that invest heavily in safety measures lose out to competitors who skip them—a disincentive sometimes called the 'alignment tax.' But researcher Ozzie Gooen proposes a reframe. If expensive safety protocols only need to apply to the tiny fraction of AI decisions where catastrophic failure is possible—maybe 0.1% of all tasks—then a hundred-fold safety overhead becomes just a 10% cost overall. This could work by restricting truly critical decisions like frontier AI development to heavily vetted, expensive agents, while using cheaper models for routine work. Banks already do this with financial access: most employees go through request processes instead of holding direct account privileges. The implication changes the competitive landscape: instead of a race to zero on safety, labs could differentiate by offering highly reliable systems for genuinely high-stakes decisions. The pressure doesn't disappear, but the economic math shifts.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton