Nature Is Weird
/ Desk lead
People trust AI more when the problems get harder, which is exactly when it’s most likely to be wrong.
This 'verification paradox' reveals a 46-point gap between perceived and actual AI accuracy. Because humans are susceptible to 'automation bias,' they mistake the confident, fluent tone of a large language model for correctness precisely when the complexity of the task has caused the model's actual performance to plummet.