What AI-accelerated biomedicine lacks is the testing of existing conclusions

AI companies plan to compress decades of biomedical progress into a few years, and they treat the poor reproducibility of the biomedical literature as a data problem to be solved by generating new data. This essay argues that the contamination lies in the conclusions, which rest on surrogate measures that fields stopped updating, and that the patient-derived omics data already in the public domain are adequate for testing those conclusions. Such tests fall outside routine research: researchers inside a field have no reason to run them, researchers outside it could not afford them, and AI models trained on the literature reproduce its paradigm. AI has lowered the cost of testing, but the doubt that starts a test still comes from a person. The record contains the essay in English and in the original Chinese (AI 加速生物医学,缺少的是对已有结论的检验).