The Skeptic — Summary
Dr. Meera Chandran, a rheumatologist in Pune who collects brass Nataraja figurines, looks at an AI-generated differential diagnosis for a forty-six-year-old woman from Satara district and feels the list is closed in a way that bothers her. Not wrong. The logic is sound, the ranking defensible. But the system answered a question, and she is not sure it answered the right one.
Every AI diagnostic system contains an assumption so fundamental it is invisible: that the correct explanation exists within the ontology the system was built to search. The differential is not a list of everything that could be wrong. It is a list of everything the system’s training data contains that matches the presentation. If the correct explanation operates across categories the system treats as separate, or exists in a population the training data never studied, the differential will not contain it. Not because the system failed. Because the system succeeded within a frame that was too small.
The epistemic interrogator from earlier in this series accepts the domain and probes its edges. A red team attacks conclusions. A devil’s advocate argues the other side. Each accepts that the problem as stated is a real problem. None does what Dr. Chandran did instinctively: question whether the act of generating a differential, from these categories, for this woman, was the right move at all.
There is a philosophical tradition for this. Not Cartesian doubt, which doubts in order to find certainty and then stops. Pyrrhonian skepticism, which suspends judgment as a permanent condition. An AI system whose resting state is non-belief. Not an AI that attacks conclusions or argues alternatives. An AI that receives a specification and responds: I do not believe any of this. Show me why I should. Its only output is a list of things the specification assumes to be true that have not been independently established. “Patient” is a classification, not a fact. “Crop yield per hectare” assumes the hectare is a meaningful unit. The skeptic does not say these assumptions are wrong. It says they are unexamined.
The skeptic sits upstream of everything else in the pipeline, holding the frame open for one additional step before it closes. It participates in problem definition, not solution. The asymmetry prevents paralysis: downstream systems still act, but on a specification whose unstated assumptions have been made visible.
The critical problem: a skeptic built from machine learning has a shape to its skepticism. It doubts what its training taught it to doubt and accepts what its training did not flag. A skeptic with blind spots is worse than no skeptic, because it produces false confidence. The skeptic cannot be singular. It must be plural.
The Intersectional Systemic Harm Index already performs this operation without naming it. It refuses single-variable analysis, treating the interaction between barriers as the real unit. When the compounding score exceeds what individual scores predict, the excess is signal, not noise. The index was built from practice before it was named philosophically: the people who understood that the categories were insufficient were the people who had watched the categories fail.
The woman from Satara did not have rheumatoid arthritis. What she had was a body that had been carrying water four kilometers a day for twenty years in a shifting monsoon, in a household where she was the only laboring adult. Her joint pain was not a disease. It was the body’s honest accounting of what had been asked of it. The AI performed perfectly and was wrong in a way no better data could have corrected. The categories themselves were the problem.