I think that's the right intuition. Legal AI feels especially dangerous because the output can look competent while hiding jurisdiction-specific footguns