The Skeptic

TAM-INS.01 · The Insufficient · The Approximate Mind

Dr. Meera Chandran has been a rheumatologist in Pune for twenty-two years. She collects brass figurines of Nataraja, the dancing Shiva, which she arranges along her office windowsill in no particular order. When patients ask about them, she says she likes to watch something hold still and move at the same time.

Last March, a forty-six-year-old woman came in with joint pain, fatigue, and intermittent low-grade fevers. The AI diagnostic system processed her bloodwork, her history, her age, her occupation as an agricultural laborer in Satara district. It generated a differential: rheumatoid arthritis, lupus, reactive arthritis, fibromyalgia. It ranked rheumatoid arthritis highest based on inflammatory markers and symptom presentation.

Dr. Chandran looked at the differential and then looked at the woman. Something was wrong with the list. Not in the logic. The logic was sound. Each diagnosis on the differential was consistent with the available data. The ranking was defensible. The system had done exactly what it was designed to do.

But the list felt closed in a way that bothered her. She could not have articulated why at the time. What she said to her resident was: “The system answered a question. I am not sure it answered the right one.”

The Closed Question

#

Every AI diagnostic system works the same way at its core. It receives inputs: symptoms, lab values, demographics, history. It matches those inputs against patterns it has learned from training data. It generates a set of possible explanations, ranked by probability.

This process contains an assumption so fundamental that it is invisible. The assumption is that the correct explanation exists within the ontology the system was built to search.

The differential diagnosis 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 patient’s presentation. If the correct explanation is a condition the system has never encountered, or a mechanism that operates across categories the system treats as separate, or a causal pathway that does not exist in any published study because the population in which it manifests was never studied, the differential will not contain it. Not because the system failed. Because the system succeeded, perfectly, within a frame that was too small.

The question was not “what is wrong with this patient.” The question was “which of the diagnoses in our existing catalog best matches this patient’s presentation.” Those are different questions, and the system cannot tell the difference.

This matters beyond medicine. Every optimization, every recommendation engine, every policy model, every research pipeline begins with a frame. The frame determines what the system can see. Nothing inside the system can question the frame, because questioning the frame requires standing outside it, and the system is the frame.

What an Interrogator Cannot Do

#

The Approximate Mind has argued, in earlier essays, for an epistemic AI: a system that interrogates the objective function, that asks what the optimizer cannot see. That argument is real and necessary. But it contains a limit that this essay needs to name.

The epistemic interrogator accepts the domain and probes its edges. It asks: within this domain, what knowledge traditions are you missing? Which populations are underrepresented? What second-order consequences are invisible? These are important questions. They extend the frame. They do not question whether the frame itself is the right one.

A red team attacks the system’s conclusions. A devil’s advocate argues the other side. An adversarial reviewer tests the methodology. Each of these accepts that the problem as stated is a real problem and works within its terms.

None of them does the thing that Dr. Chandran did instinctively in her office. She did not question the differential. She questioned whether the act of generating a differential, from these categories, trained on this data, applied to this woman, was the right move.

That is a different intellectual operation. It is not interrogation. It is not adversarial testing. It is something more radical and more disorienting.

It is the refusal to believe that the problem as stated is the problem.

The Pyrrhonian Posture

#

There is a philosophical tradition for this, and it is not the one most people think of when they think of skepticism.

Descartes doubted everything in order to find the one thing he could not doubt. His skepticism was a method for arriving at certainty. The doubt was instrumental. It served a goal. Once the goal was reached, the cogito was established, the doubting stopped.

Two thousand years before Descartes, a different tradition emerged. Pyrrho of Elis, and later Sextus Empiricus, practiced a skepticism that did not resolve. They suspended judgment as a permanent condition, not a stage on the way to certainty. The Pyrrhonist does not doubt in order to arrive at truth. The Pyrrhonist holds the doubt itself as the practice.

This sounds paralytic. In philosophy, for an individual human being, it may be. But as a design principle for a specific component of a multi-agent AI architecture, it is the most practical idea in this essay.

An AI system whose resting state is non-belief. Not an AI that attacks conclusions. Not an AI that argues alternatives. An AI that receives a specification, any specification, and responds: I do not believe any of this. Show me why I should.

Its only output is a list. Not a list of errors. Not a list of alternatives. A list of things the specification assumes to be true that have not been independently established.

“Patient” is a classification, not a fact. You have not established that treating this person as a patient, rather than as a household, a water-access situation, a labor arrangement, a position in a caste economy, is the right unit of analysis.

“Crop yield per hectare” assumes both the crop and the hectare are meaningful units. You have not established that a hectare captures anything real about a polyculture farm where different crops occupy different vertical layers and different seasonal windows.

“Student performance” assumes that performance is a property of the student rather than a property of the relationship between the student and the institution. You have not established that.

The skeptic does not say these assumptions are wrong. It says they are unexamined. The difference matters.

What It Does to a Pipeline

#

The skeptic sits upstream of everything else. Upstream of the optimizer, upstream of the epistemic interrogator, upstream of the consequence modeler. It gets the first pass at every specification. Its function is to slow down one specific moment: the moment when a specification becomes operational.

That moment is where the frame locks. Once the specification is accepted, everything downstream works within it. The optimizer optimizes within it. The interrogator interrogates within it. The consequences are modeled within it. If the frame was wrong, everything downstream is wrong in a way that nothing downstream can detect.

The skeptic’s job is to hold the frame open for one additional step before it closes. Not permanently. Not indefinitely. Long enough for the unstated assumptions to become visible to the humans who will decide whether to proceed.

This is the asymmetry that prevents paralysis. The skeptic participates in problem definition. It does not participate in solution. Its output feeds downstream systems that can still act. The optimizer still optimizes. The interrogator still interrogates. The pipeline still runs. But it runs on a specification that has been examined at a level the pipeline itself cannot examine, because the pipeline is the specification made operational.

The Trained Skeptic Problem

#

Here is the problem this essay cannot solve, and needs to name honestly.

A skeptic built from machine learning is a skeptic trained on some body of knowledge. Its skepticism has a shape. It has learned, from whatever corpus it was trained on, which categories tend to be constructed, which units of analysis tend to be misleading, which assumptions tend to be dangerous. It will doubt the kinds of things its training taught it to doubt.

And it will accept the kinds of things its training did not flag.

A skeptic with blind spots is worse than no skeptic at all. It produces false confidence. The team using the system believes the assumptions have been checked. They have been checked by a system whose checking had a shape the team cannot see.

I wonder whether the honest response to this problem is that the skeptic can never be a single system, that it must always be plural, and that the plurality itself must be designed to contain perspectives that disagree with each other about what constitutes an assumption worth questioning.

This is where the next essay begins.

The Index That Already Doubts

#

There is a system that does something close to what this essay describes. It was not built by philosophers or AI researchers. It was built by a father and son who had spent years watching healthcare systems process people whose lives exceeded the system’s categories.

The Intersectional Systemic Harm Index measures how barriers compound for a single person. It refuses to accept that any single barrier is the barrier. It treats the interaction between barriers, transportation plus digital divide plus economic strain plus social isolation plus language, as the real unit, not the individual components.

When the compounding score exceeds what the individual barrier scores would predict, conventional assessment treats the excess as noise. The index treats it as signal. Something is operating in the compound that the decomposed view cannot see.

This is the skeptic’s move, performed operationally before it was named philosophically. The system refuses to believe the decomposed version of the problem. It insists that the categories the conventional assessment uses, individual barriers treated as isolable variables, are not the categories that describe reality. It does not argue against the individual assessments. It holds that they are insufficient.

The index was built from practice, not from theory. The theory came later, which is itself instructive. The people who understood that the categories were insufficient were the people who had spent years watching the categories fail, not the people who had built the categories in the first place.

Dr. Chandran’s Nataraja

#

The woman from Satara did not have rheumatoid arthritis. She did not have lupus, reactive arthritis, or fibromyalgia.

What she had was a body that had been carrying water for four kilometers a day for twenty years, through a monsoon pattern that had shifted enough to extend the dry season by six weeks, in a household where she was the only adult performing physical labor because her husband’s chronic illness, itself misdiagnosed for years, had made him unable to work. Her joint pain was not a disease. It was the cumulative consequence of a life the system’s categories could not hold.

Dr. Chandran did not arrive at this understanding through diagnosis. She arrived at it through conversation. Forty minutes in which the system’s categories fell away one by one and the actual structure of the woman’s life became visible. The fatigue was not a symptom. The joint pain was not a presentation. They were the body’s honest accounting of what had been asked of it.

The AI system had performed perfectly. The differential was well-reasoned, well-ranked, and wrong in a way that no amount of better data or better algorithms could have corrected. The categories themselves were the problem. The system’s ontology did not contain the explanation because the explanation was not a medical entity. It was a life.

Dr. Chandran looked at her Natarajas afterward. The dancing Shiva holds still and moves at the same time. Destruction and creation in a single gesture. She wonders sometimes whether the practice of medicine is learning when to hold the categories still and when to let them dance.

The woman went home with no diagnosis. She went home with something Dr. Chandran thinks may have been more useful: the experience of having been seen as a whole life rather than processed as a symptom cluster. Whether that experience changes anything material about the four-kilometer walk or the shifted monsoon or the household’s single laboring body, Dr. Chandran does not know.

She suspects the system could have flagged the moment. Not the diagnosis. The moment when the categories stopped fitting. The moment when the specification became insufficient for the life it was trying to describe. A system that could say, clearly and without judgment: this encounter has exceeded my ontology, and someone who sees differently should be in the room.

That would require a system that does not believe its own categories. That holds them provisionally. That treats every classification as a hypothesis rather than a fact.

A system that believes nothing.

This is the first essay in The Insufficient, a four-essay sub-series of The Approximate Mind examining what lies beneath the empirical record that AI systems are built to search. The Ungoverned Frontier asked what the knowledge map is missing. The Insufficient asks whether the map’s projection system is distorting the territory. This essay introduces the skeptic architecture: an AI component whose resting state is non-belief, designed to identify unstated assumptions in any specification before the pipeline makes them operational. The second essay, “The Traditions,” populates this architecture with seven philosophical operations drawn from traditions the AI ecosystem was not built to see.

References

#

Pyrrhonian Skepticism

Sextus Empiricus. Outlines of Pyrrhonism. Translated by R.G. Bury. Harvard University Press, 1933.

Striker, Gisela. “Scepticism as a Kind of Philosophy.” Archiv für Geschichte der Philosophie, vol. 83, no. 2, 2001, pp. 113-129.

Category Construction and Classification

Bowker, Geoffrey C., and Susan Leigh Star. Sorting Things Out: Classification and Its Consequences. MIT Press, 1999.

Hacking, Ian. The Social Construction of What? Harvard University Press, 1999.

Lakatos, Imre. Proofs and Refutations: The Logic of Mathematical Discovery. Cambridge University Press, 1976.

Critical Realism and Ontological Depth

Bhaskar, Roy. A Realist Theory of Science. Verso, 1975.

Danermark, Berth, et al. Explaining Society: Critical Realism in the Social Sciences. Routledge, 2002.

Diagnostic Epistemology

Montgomery, Kathryn. How Doctors Think: Clinical Judgment and the Practice of Medicine. Oxford University Press, 2006.

Groopman, Jerome. How Doctors Think. Houghton Mifflin, 2007.

Optimization Failures

Scott, James C. Seeing Like a State: How Certain Schemes to Improve the Human Condition Have Failed. Yale University Press, 1998.

Muller, Jerry Z. The Tyranny of Metrics. Princeton University Press, 2018.