AI raised the floor, but grading rubrics still penalize the ceiling

AI 抬高了下限，评分标准还在惩罚上限

Two large-scale RCTs show the same pattern: AI lifts the floor of student work while present, but once removed, performance drops, and traditional rubrics actively penalize deeper reasoning. In a Turkish high school math experiment, ChatGPT-assisted practice scores jumped 48%, yet closed-book exam scores fell 17% below the control group. In a Milan business writing study, students who spelled out failure conditions and causal mechanisms received systematically lower grades. The floor is borrowed from external compute; the ceiling only grows when rubrics reward it.

Why it matters: Two large-scale RCTs with hard numbers expose the illusion of AI-assisted learning: practice scores soar but closed-book tests drop, and students copy answers without reasoning. Strong HKR, but it's a synthesis piece rather than a primary research release, so it stays below 85.