AI Gatekeepers Are Failing at the Exact Jobs They Were Meant to Replace

Three separate Factum items expose the same fracture: Claude Opus 4.8 directly contradicted a radiologist’s Grade III subscapularis tear on raw 266 MB DICOM files; the Brown University ECON 1170 case caught at least 50 students using AI to cheat on a single midterm; and an unnamed hiring-agent tool produced scores ranging from 66 to 99 on identical resumes across 100 runs. These are not isolated glitches. They show AI systems now occupying the precise roles—medical diagnosis, academic integrity enforcement, and employment screening—where consistency is the minimum requirement, yet the models themselves are producing non-reproducible outputs at scale. The older headline on variable resume scoring and the recent Claude MRI contradiction are two sides of one operational reality: the automation layer is being inserted into judgment pipelines faster than its variance can be measured or contained.