The Discipline That Cannot Be Delegated

Abstract

Three essays preceded this one. The first argued that ideology does not destroy intelligence but captures it, turning a capable mind to the service of a premise it is forbidden to examine. The second distinguished a mistake, the mis-grasp of something already held, from an error, the honest wandering at the frontier of the unknown, and located the whole skill in the judgment of which one occurred. The third turned toward artificial intelligence itself: the arrogance of the learned who will not say "I do not know," the bad faith of an industry that misnames its machine, and a fallibility that implicates the author alongside everyone else. This essay gathers the three and identifies a single failure beneath them, the loss of the internal veto by which a fallible mind stays correctable, together with a single remedy that no tool, platform, or authority can hold on anyone's behalf. It then argues that engagement-optimised architecture selects for precisely the conditions of that failure, drawing on Asch's conformity studies and in particular the variation in which a dissenter giving a different wrong answer collapsed conformity as effectively as a correct one, which relocates the harm from the suppression of truth to the manufacture of apparent unanimity. A central section examines the asymmetry by which honesty is demanded of machines and not of the operators who instruct them, working through two documented cases: Mata v. Avianca, in which sanctions fell not on the fabrication but on the refusal to be corrected, and the May 2025 modification of Grok's system prompt, in which an instruction concerning one political subject altered the model's answers on unrelated topics and the system could not detect from the inside that it had been altered. The essay closes by stating three objections at full strength and conceding what they establish, including that a rule can compel an act while leaving the disposition untouched.