An Algorithm Denied a Man's Epidural Three Times. Nobody Told Him It Wasn't Human. Bureaucratic Opacity, Now Automated.
A patient named Magnuson was denied an epidural steroid injection for back pain by Medicare three separate times despite his doctors submitting additional documentation. Only on the third denial did Magnuson learn the decision was made by AI. The system requires a human clinician at the tech company to manually review denials only after the AI has already rejected the treatment. This is part of a pilot programme operating across six US states aimed at reducing so-called wasteful medical procedures.
This illustrates what scholars of algorithmic governance call automated adverse selection. The mechanism is straightforward: when a cost-optimization model is placed between a patient and a clinician, the burden of proof shifts from the institution to the individual. You cannot appeal a decision effectively if you do not understand who, or what, made it. The lesson for any AI-adjacent system is that opacity plus authority equals unaccountability, and that is not a technical flaw. It is a design choice.
Medicare is running the pilot programme across six US states. The AI system requires a human clinician at the tech company to review denials only after rejection occurs.
- Open ChatGPT or any free large language model and ask it to evaluate a mock insurance appeal letter you write about a denied procedure. Observe how it identifies gaps in reasoning. Expected outcome: you will see how an AI surfaces logical weaknesses that a human reviewer might also flag.
- Write a second appeal letter addressing every weakness the AI found. Resubmit it to the same AI and ask for a denial or approval decision. Expected outcome: the model will either accept the revised letter or find new objections, demonstrating how automated review systems iterate.
- Ask the AI to explain its decision in plain language. Expected outcome: you will receive a rationale that may or may not match clinical reality, illustrating why human oversight in automated medical decisions remains essential.