Concentrating AI safety on 0.1% of tasks could make safety tax affordable

A new analysis by Ozziegooen on LessWrong challenges the standard objection that AI safety taxes are economically unviable because they make systems uncompetitive. The key insight: the tax doesn't have to be uniform. By applying expensive safety measures only to the small fraction of tasks that carry catastrophic risk (<1% of actions), the blended overhead becomes manageable. For instance, a 100x cost multiplier on 0.1% of tasks translates to roughly a 10% increase on the entire system. This concentrated approach could make even extremely costly safety techniques — like inference-time monitoring, specialized transparent models, or heavily vetted agents — practical for high-stakes operations.

The post identifies two main strategies: first, improving risk assessment to identify which tasks are most critical (e.g., frontier LLM development vs. routine queries); second, designing processes that isolate critical work into a narrow computational channel, similar to how companies restrict bank account access. This could lead to a tiered market: inexpensive general-purpose agents for everyday tasks and expensive, highly reliable agents for sensitive decisions. The author suggests this makes AI safety work on high-cost systems highly valuable, as long as we can ensure those systems are used for the top fraction of risky tasks. Objections include cases where risk is too distributed to concentrate, but the author argues this is unlikely with good design.