Many-analysts results don’t prove bias; they often show the effect is under identified and analyst choices flip sign. Best fix: preregistration + multiverse/sensitivity reporting. Public preferences ≠ measured welfare-support effects.