The Interrogator

TAM-074 · The Approximate Mind

A man I know has been looking at job postings at AI research institutions for the past year. He has thirty years of experience across healthcare, technology, and policy. He has worked in systems that served populations of millions. He has watched, from the inside, what happens when institutions optimize for the wrong thing: the metric that looked clean while the community it measured deteriorated, the efficiency gain that erased the relationship that was carrying the real load, the policy that simplified beautifully on paper and destroyed compromises that had taken decades to negotiate.

He keeps a notebook. Not digital. A leather-bound thing his son gave him two years ago, already half full. In it, he writes down the questions that the systems he encounters are not asking. Not the answers they are getting wrong. The questions they are not asking at all.

He has looked at hundreds of job postings. Machine learning engineer. Alignment researcher. Prompt engineer. Product manager, AI safety. Research scientist, foundation models. Every role is about building systems that answer questions better. Not one is about building systems that question the questions.

He is not sure whether the role he wants exists. He is increasingly sure it needs to.

What Every Optimizer Misses

#

There is a pattern in the history of consequential optimization failures, and it is not the pattern most people think.

The Green Revolution optimized Indian agriculture for yield per hectare. It succeeded. Caloric output increased dramatically. It also depleted soil health across entire regions, collapsed groundwater tables, created pesticide dependencies, pushed millions of smallholder farmers into debt cycles, and contributed to a suicide crisis in the cotton belt that is still ongoing decades later. The optimization worked. The objective function was the catastrophe.

Structural adjustment programs optimized developing economies for macroeconomic stability. They succeeded by their metrics. GDP growth, inflation control, trade liberalization. They also devastated public health systems, education infrastructure, and social safety nets in the countries that adopted them, producing a generation of outcomes that the metrics were not designed to measure.

Health systems that optimize for DALYs averted per dollar spent produce a rational allocation that systematically defunds mental health, chronic pain management, disability support, and elder care. Conditions that are high-suffering but low-mortality. That matter enormously to the people experiencing them. That barely register in the framework.

The pattern is not that the optimizer gets the wrong answer. It is that the optimizer answers the wrong question, perfectly.

Every objective function is a decision about what counts. Yield per hectare counts. Soil microbiome health does not. GDP growth counts. The knowledge carried by the community health worker who was defunded does not. DALYs averted counts. The quality of a life lived with chronic pain does not. These are not oversights. They are structural features of optimization itself. You cannot optimize for everything. The act of choosing what to optimize for is the act of choosing what to make invisible.

And nobody is building the system that makes the invisible visible again.

The Job That Doesn’t Exist

#

The man with the notebook can see what the optimizer misses. Not because he is smarter than the people who build optimizers. Because he has spent three decades on the other side of the optimization: in the communities where the policy landed, in the hospitals where the metric was met and the patient was not helped, in the government offices where the simplified process erased the accommodation that a specific population needed.

He has a particular kind of knowledge. It is not technical. It is not academic. It is the knowledge of someone who has watched systems interact with human lives for long enough to recognize the gap between what the system measures and what the person experiences. He can feel when a question is incomplete. He cannot always articulate what is missing, but he can point at the silence in the data and say: something should be here, and it is not, and its absence will produce consequences the model cannot predict.

There are thousands of people like him. Retired public health officials. Experienced social workers. Agricultural extension officers who have spent decades walking between the policy and the field. Educators who have watched three generations of reform and can tell you what each reform could not see. They carry institutional memory. They carry the knowledge of what went wrong and why, knowledge that lives in experience rather than in publications.

They have no role in the AI transition. There is no job posting that says: “Wanted: someone who has spent decades watching optimization fail and can articulate what the objective function is not seeing.” There is no department of epistemological interrogation. No budget line for adversarial questioning of the questions being asked.

The people most qualified to identify what AI systems are missing are the people the AI industry has no role for.

What Would Need to Exist

#

An AI system whose purpose is not to optimize but to interrogate. Not to answer questions but to question answers. A system that takes a specification, an objective function, a policy proposal, a research agenda, and asks, with structural rigor, four things:

What knowledge exists in this domain that your model cannot see? Not “what data are you missing,” which implies the existing framework is correct and merely incomplete. But “what ways of knowing are relevant here that your architecture cannot represent?” The farmer in Odisha whose intercropping practice maintains soil-health interactions that no published paper documents. The health worker in Rajasthan whose body has learned to read other bodies in distress. The pharmacist who noticed what the algorithm could not notice because noticing required a relationship over time. Their knowledge is real. It is valid. It is invisible to any system trained on published, digitized, propositional text.

Who is affected by this optimization, and whose experience is absent from the model? Not a demographic checklist. A genuine accounting of whose lives will change and whose situations the training data does not represent. The rural woman in Bihar whose interaction with the legal system looks nothing like the urban, English-speaking, digitally connected profile the system was trained on. The elderly patient whose disease presents differently than the clinical trial cohort. The community whose compromise was encoded in the complexity the system is simplifying away.

What consequences does the objective function render invisible? Not first-order effects, which the optimizer can model. Second and third-order effects that cascade through systems the optimizer was not designed to see. Epistemological consequences: what knowledge is lost when a practice is displaced? Social consequences: what relationships change when an institution is restructured? Political consequences: what compromises are erased when a law is simplified? Cultural consequences: what practices are disrupted when an optimization reshapes how people live?

What is being implicitly valued and what is being implicitly discounted? Every objective function embeds a value judgment. Yield over resilience. Efficiency over recognition. Speed over deliberation. Aggregate welfare over individual dignity. These judgments are usually invisible, made silently in the design of the function rather than explicitly by a human decision-maker. The interrogator makes them visible. Not to resolve the value conflict. To ensure that the humans making the decision know the conflict exists.

Why It Doesn’t Require a Trillion Parameters

#

This is the part that matters most for the man with the notebook. And for the thousands of people like him.

The interrogator does not need to be a frontier model. It does not need to cost billions. Its functions are specialized, not general. A domain-specific model trained deeply on Indian agricultural knowledge, including documented traditional practices, gray literature, extension service reports, costs tens of thousands of dollars to train, not tens of billions. A values analysis system built on structured representations of ethical traditions is a knowledge engineering project, not a machine learning scaling problem. A consequence modeling system for a specific domain requires causal reasoning about that domain, not universal intelligence.

The interrogator can be built from small, specialized models at a fraction of the cost of the systems it would interrogate.

This is the equity argument in technical form. If the epistemic function required frontier scale, it would belong to the same three to five institutions that own the frontier models, and it would serve their priorities. If it can be built from small models, it can be built by state universities, by development banks, by agricultural cooperatives, by the communities whose knowledge the frontier models cannot see. The architecture choice is the equity choice.

A pilot in Indian agriculture. A domain-specific model trained on the available literature plus documented traditional knowledge. The four interrogation modes built for this specific domain. Three to five active AI-driven agricultural optimization projects evaluated. Twelve to eighteen months. Under two million dollars. That is what it would take to find out whether this concept works in practice or only in essays.

The Role

#

The man with the notebook does not want to build another optimizer. He wants to build the thing that questions the optimizer. And he wants to be able to feed his family while doing it.

This is not a trivial point. The AI industry has created enormous economic value by building systems that answer questions. It has created no economic value, and no institutional home, for the work of questioning the questions. The people who can do this work, the ones with decades of experience watching systems interact with human lives, the ones who can feel the silence in the data, are either retired, or working in roles that do not use this capacity, or trying to explain to hiring committees why their thirty years of institutional experience is relevant to AI development when they do not have a machine learning degree.

They are relevant. Their experience is the training data for the interrogator. Not in the technical sense. In the human sense. They know what objective functions miss because they have lived on the receiving end of the missing. They know what questions are not being asked because they have spent careers watching the consequences of the unasked questions.

The role that needs to exist is something like: epistemological architect. Someone who designs the interrogation, who specifies what the adversarial layer should look for, who bridges between the communities whose knowledge is invisible and the systems that need to see it. This is not a technical role in the way the AI industry currently defines technical. It is a role that requires deep domain experience, philosophical clarity, and the particular kind of pattern recognition that develops only through decades of watching systems interact with human complexity.

There will be more people like him. As AI optimization extends into agriculture, healthcare, education, governance, urban planning, the demand for people who can question the questions will grow. Not because anyone plans for it. Because the consequences of unquestioned optimization will become visible, the way the Green Revolution’s consequences became visible, the way structural adjustment’s consequences became visible, and someone will need to do the work of asking what went wrong, and the answer will always be the same: we optimized for the wrong thing, and nobody was asking whether it was the right thing.

The question is whether we build the role before the consequences or after. We always build it after. This essay is an argument for building it before, for once.

I wonder whether the institutions that most need this work will recognize the need in time, or whether the recognition will come only after the optimization has run, and the consequences have compounded, and someone with a notebook full of unasked questions will be invited to explain what went wrong.

The Notebook

#

The man is still writing in it. This morning he wrote down a question about a health system AI that a state government is piloting. The system optimizes appointment scheduling for patient throughput. It is good at this. Wait times have decreased by 30%. Patient satisfaction scores are up.

He wrote: “What happens to the patients who used the wait? The ones who talked to each other in the waiting room. The ones who got information from the person next to them. The ones for whom the trip to the clinic was the only time they left the house that week. The system is measuring throughput. Nobody is measuring what the waiting room was carrying.”

He does not know if anyone will read the notebook. He does not know if the role he is describing will exist in time for him to fill it. He knows that the question he wrote down this morning is not in any AI system’s objective function, and that the people affected by its absence are the people who were sitting in the waiting room that is now empty.

He turns the page. There are questions on every line. The questions are not answers. They are something older and, he suspects, more necessary.

He is fifty-three. He does not have time for another career after this one. He would like this to be it.

This is Part 74 of The Approximate Mind, a series exploring how artificial intelligence reshapes what it means to think, create, work, and live. Previous essays examined the toll booth economy (Part 52), the invisible tiers (Part 57), the threshold of robotic convergence (Part 65), and the wrong question (Part 67). This essay argues that the most important AI system we can build is not a better optimizer but a better interrogator: one that questions what the objective function cannot see. The detailed design specification for this system, including ontological, epistemological, methodological, and axiological architecture, is available as a companion document in Part 75, “The Epistemic Framework.”

References

#

Optimization Failures and Their Consequences

Shiva, Vandana. The Violence of the Green Revolution: Third World Agriculture, Ecology, and Politics. Zed Books, 1991.

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.

Knowledge Systems and Epistemological Justice

Santos, Boaventura de Sousa. Epistemologies of the South: Justice Against Epistemicide. Routledge, 2014.

Chambers, Robert. Whose Reality Counts? Putting the First Last. Intermediate Technology Publications, 1997.

Polanyi, Michael. The Tacit Dimension. University of Chicago Press, 1966.

AI, Equity, and Institutional Design

Crawford, Kate. Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. Yale University Press, 2021.

Mohamed, Shakir, Marie-Therese Png, and William Isaac. “Decolonial Artificial Intelligence: Decolonial Theory as Sociotechnical Foresight in Artificial Intelligence.” Philosophy & Technology, vol. 33, 2020, pp. 659-684.

Mazzucato, Mariana. Mission Economy: A Moonshot Guide to Changing Capitalism. Harper Business, 2021.

Global Health and Development

Farmer, Paul. Pathologies of Power: Health, Human Rights, and the New War on the Poor. University of California Press, 2003.

Sen, Amartya. Development as Freedom. Anchor Books, 1999.

Tacit Knowledge in Professional Practice

Dreyfus, Hubert L., and Stuart E. Dreyfus. Mind Over Machine: The Power of Human Intuition and Expertise in the Era of the Computer. Free Press, 1986.

Collins, Harry, and Robert Evans. Rethinking Expertise. University of Chicago Press, 2007.