Researchers identify nine distinct categories of hard-to-verify AI tasks, challenging the oversimplified binary view of task verification difficulty

Easy-to-verify tasks have efficient verification programs that score solutions quickly without substantial resource costs or side effects

Hard-to-verify tasks aren't a single category but rather a negative classification encompassing multiple distinct types with different underlying reasons

One category of hard-to-verify tasks requires expensive AI inference, where verification is theoretically possible but computationally prohibitive to run repeatedly

The article suggests task verification difficulty exists on a spectrum rather than as a simple binary classification like 'ravens vs non-ravens'