AI chatbot assistance improves human performance on hard questions — complex cases where the operator lacks the knowledge or time to reach the correct answer independently. This is the expected benefit: augment where humans are weak.
The finding in 2603.11213 is that the same AI assistance reduces accuracy on easy questions. Cases the operator would have answered correctly without help are answered incorrectly with help, because the operator defers to the AI's suggestion rather than trusting their own judgment. The AI is wrong on some easy cases — every system has error modes — and the operator, who would have been right, adopts the AI's wrong answer.
Note (2026-11-10): I checked the abstract of arXiv 2603.11213, a randomized experiment with caseworkers on a 770-question benchmark. Control caseworkers had a mean accuracy of 49 percent; with high-quality chatbots (96 to 100 percent accurate) accuracy improved by 27 points; incorrect chatbot suggestions substantially reduced accuracy, with a two-thirds reduction on easy questions; and improvements leveled off as chatbot accuracy rose (the AI underreliance plateau). The essay says AI assistance in general reduces accuracy on easy questions and that help degrades aggregate performance when most tasks are easy; the abstract's harm is from incorrect suggestions, aggregate accuracy rises with chatbot quality, and the paper's own conclusion is about underreliance. The distinction between cognitive deference and automation bias is the essay's.
This is not automation bias in the usual sense. Automation bias is failing to override an automated recommendation. Cognitive deference is actively replacing a correct judgment with an incorrect one because the AI's presence reduces the operator's confidence in their own competence. The operator does not fail to notice the AI is wrong. The operator fails to notice they are right.
The structural point: the net effect of AI assistance depends on the distribution of task difficulty. If most tasks are hard, AI help improves aggregate performance — the gains on hard cases outweigh the losses on easy ones. If most tasks are easy, AI help degrades aggregate performance — the losses on easy cases dominate. The same tool is beneficial or harmful depending on the baseline competence of the operator relative to the task distribution. Help that improves the ceiling lowers the floor.