The Interpreters of Uncertainty

When Everyone Can Predict, Who Decides What the Prediction Means?

#

Two numbers are sitting on Margaret’s kitchen table.

One is a risk estimate from her long-term care insurer: 34.7% probability of requiring extended residential care within ten years. The other is the monthly premium that number apparently justifies: $847.

Margaret stares at both for a while. She is not unintelligent. She managed her own business for twenty-two years, raised a daughter largely alone, and has been navigating a healthcare system that increasingly feels designed for someone else. But she cannot make sense of these numbers in any way that would help her decide what to do. Is 34.7% high? Compared to what? Should she pay $847 a month to insure against it? Compared to what else she could do with that money? What went into the calculation? Would the number change if she exercised more, or worried less, or stopped going to bridge on Thursdays?

Between those two numbers lives the entire world of uncertainty interpretation. Someone decided what the prediction means. Someone decided what the prediction is worth. Someone bears the risk of getting it wrong.

In the economy of 2031, who that someone is, and what their work looks like, is one of the more consequential questions being worked out in real time.

The Same Unbundling, Different Terrain

#

The diagnosticians in the previous essay faced a familiar version of this. AI absorbed the pattern recognition and left the judgment: what does this finding mean for this patient, and who is accountable for the call?

The uncertainty professions face the same unbundling, but on different terrain. Radiology’s judgment was about individual biology. Financial analysts, health economists, actuaries, economic forecasters: all of them were in the business of collective human behavior. And collective human behavior has a property that individual patient biology does not.

It is reflexive. The prediction changes the thing being predicted.

A forecast of inflation shapes the bond market, which shapes borrowing costs, which shapes the inflation being forecast. A risk estimate for an insurance portfolio, published widely, changes the behavior of the insured. A market prediction, believed strongly enough, becomes self-fulfilling. George Soros named this half a century ago and built a fortune on understanding it. The models have gotten better every year since then. The reflexivity has not gone anywhere.

Four professions, four different stories, depending on what the computation was concealing.

Raj and the Story Behind the Numbers

#

Raj Krishnamurthy manages a mid-cap equity fund from an office in Singapore. He has two young children he does not see enough, a habit of running long distances early in the morning before the city heats up, and an analyst’s instinct that developed over twenty years by doing work that no longer exists.

Five years ago, his team of twelve spent most of their time on what they called coverage: reading quarterly filings, building financial models, tracking industry trends, generating the quantitative scaffolding on which investment theses were built. The work was slow, detail-intensive, and required years of training to do competently.

AI does all of it now. Not approximately. Comprehensively. The system processes every SEC filing, every earnings transcript, every supply chain signal, every patent application, every shipping container tracked by satellite. It builds models incorporating more variables than Raj’s team could have handled in a month, and it does this in seconds.

His team is now four people. Not because eight were fired, though some were. Because the work that justified twelve no longer exists as human work.

What Raj does now is something he struggles to name. He calls it narrative sensing. What he means is this: markets have never been pure mechanisms for processing information into correct prices. They are mechanisms for processing stories into prices. The story that AI will transform healthcare. The story that a particular CEO has lost her board’s confidence. The story that this time, the housing market really is different.

AI is superb at processing information. It is mediocre at reading stories, because stories are about collective human psychology, and collective human psychology is reflexive in ways that resist modeling. The story that a stock will rise causes buying that makes the stock rise, which confirms the story, which causes more buying. You cannot model this in real time, because navigating it requires understanding what humans will believe next, and what they believe next depends partly on what AI predicts, which humans then absorb and react to.

Raj’s job is to read the room. Not the data. The room. What story is the market telling itself? Where is the story wrong? Where is it right but for the wrong reasons?

This is not a new skill. It is what the best analysts always did. The quantitative work was necessary but never sufficient. The analysts who outperformed were the ones who understood narrative, sentiment, fear, and greed alongside the models. AI removed the quantitative layer and left the narrative layer exposed. The profession did not shrink. It clarified.

That said, clarification may not be enough on its own. Raj is good at narrative sensing. Not everyone who did quantitative analysis is good at narrative sensing, or can become so. The profession that remains is smaller and demands a kind of judgment that cannot be taught the same way the old skills could. Whether there are enough Rajs is a question with real stakes.

What the Model Cannot Hold

#

Six thousand miles from Singapore, Dr. Amara Osei sits in a conference room in Accra, looking at a resource allocation model that would have taken her department a year to build. The AI built it overnight.

The model is, technically, beautiful. It maps every health intervention available to Ghana’s public health system against outcomes data from forty-three countries, adjusted for local demographics, disease burden, infrastructure constraints, and budget. It can tell Dr. Osei, with high precision, that investing one million cedis in maternal health screening will produce X quality-adjusted life years, while the same million in childhood vaccination will produce Y, and in diabetes management, Z.

The math is settled. The question is not.

Because the question was never the math. The question is what do we value. Is a quality-adjusted life year worth the same for a newborn and a seventy-year-old? Should efficiency drive the allocation, maximizing total health per cedi spent? Or should equity drive it, directing resources toward populations most underserved, even if the aggregate numbers look worse? The WHO publishes cost-effectiveness thresholds, but those thresholds embed assumptions about whose life-years count how much, and those assumptions are moral choices dressed as technical parameters.

Dr. Osei’s work used to be building the model and making the value judgments. The model-building consumed most of her time and most of her team’s capacity. Now the model arrives pre-built, updated daily, more sophisticated than anything her team could have produced. What remains is the value judgment. And the value judgment requires not a health economist but a moral philosopher who understands health economics, which is a rarer and harder thing to produce.

This is where the demand-supply reframe cuts sharpest. Ghana has a handful of health economists capable of this work. Nigeria has slightly more, spread across two hundred million people. The modeling bottleneck meant that resource allocation decisions in these countries were often made by default: continuing last year’s funding patterns, responding to the loudest current crisis, following whatever the international donor community prioritized for its own reasons.

AI dissolves the modeling bottleneck. The question now is whether there are enough people with the judgment to use the models wisely. The answer is no. And the shortage of judgment is harder to address than the shortage of computation, because judgment develops through experience, through wrestling with the consequences of past decisions, through learning what the models leave out.

The Name on the Report

#

Kenji Watanabe is an actuary in Tokyo who spent twenty years building risk models for a major insurer. He was good at it. Precise, careful, methodical. He understood the mathematics of mortality, morbidity, catastrophe, and the thousand small probabilities that determine what insurance costs.

He has a daughter starting university this fall, which he mentions with the particular combination of pride and bewilderment that tends to accompany that transition. He has also spent the last three years doing work that is nearly the opposite of what he spent twenty years training for.

AI builds better models than Kenji ever did. This is not a slight. Machine learning processes more variables, more cases, more correlations than any human actuary can hold in mind. The models are not slightly better. They are categorically different in scope and sensitivity.

Kenji’s job has inverted. He used to build models. Now he interrogates them.

The AI produces a risk assessment for a portfolio of life insurance policies. Kenji’s job is to ask: what assumptions are embedded in the training data? If the data is drawn primarily from wealthy nations, does the model underestimate risk for populations with different healthcare access? If it was trained on a decade of historically low interest rates, does it underestimate the impact of rate changes on reserves? If it was built on pre-pandemic mortality data, has it been appropriately updated, or is it carrying forward assumptions that may no longer hold?

These are not computational questions. They are judgment questions. And they carry a specific weight that distinguishes actuarial work from most other uncertainty professions: someone must sign off.

The actuarial certification on an insurance product is not a suggestion. It is a professional guarantee that the numbers are sound. That guarantee carries legal liability. When the model is wrong and the insurer cannot pay claims, someone is accountable.

That someone cannot be the AI. The parallel to pathology is exact. In both cases, the computational work is automatable, but the accountability is not, because accountability requires a moral agent who bears consequences. Kenji does not just check the AI’s work. He stands behind it. His name, his certification, his career are on the line. The AI has no name, no career, no capacity to be held responsible.

The profession transforms from building models to auditing them. It is smaller in headcount. It is more consequential in responsibility. Whether those two facts can coexist in a way that sustains a profession and attracts people to it is not yet clear.

What AI Revealed About Economic Forecasting

#

The most uncomfortable transformation belongs to the economic forecasters, because AI has surfaced a secret the profession had been keeping from itself.

Economic forecasting was never very good.

This is not a criticism of economists’ intelligence. It is a structural observation about reflexive systems. The Federal Reserve’s forecast of inflation affects bond markets, which affects borrowing costs, which affects investment, which affects the inflation the Fed was forecasting. The IMF’s growth projection for a developing nation affects investor confidence, which affects capital flows, which affects the growth the IMF was projecting. The prediction and the thing predicted are entangled in ways that more data and better models do not resolve.

AI makes this visible. When you give a machine learning system access to every economic variable on Earth and it still cannot reliably predict next quarter’s GDP, the problem is not the model. The problem is the system. Economies resist prediction not because of insufficient data but because of reflexivity, political intervention, collective psychology, and the sheer irreducibility of billions of people making decisions that interact with each other and with the forecasts being made about them.

The economist’s transformation is, paradoxically, toward honesty. The profession spent decades cultivating an air of scientific precision. Forecasts were published with decimal points. Models were presented with the authority of physics. AI strips that pretense away. The models are better than anything a human could build, and they are still mediocre at prediction, which proves that the mediocrity was never about the modeler.

What remains is the honest version of the profession. The economist who says: I cannot tell you what will happen. I can tell you what the models suggest, what the models miss, where the reflexive dynamics are likely to amplify or dampen the prediction, and what the range of outcomes looks like. I can help you think about uncertainty rather than pretending to resolve it.

This is more useful, not less. A profession that honestly navigates uncertainty serves decision-makers better than one that pretends to eliminate it. But the transformation is also a loss of prestige, and prestige is not a trivial thing. It shapes who enters a profession, what institutions trust them with, what authority they carry in the room where decisions are made. Whether the honest version of economic forecasting can command the institutional standing that the confident version used to is an open question I do not know how to answer.

The Pathway Problem

#

In the diagnosticians essay, the apprenticeship crisis was about the eye: how do you develop the reading instinct if AI reads the scans? The uncertainty professions face a deeper version of the same problem.

Raj developed his narrative sense by spending years building the quantitative models that his narrative sense eventually transcended. The modeling taught him what the numbers could and could not capture. Hours spent building discounted cash flow projections, tracking margin trends, stress-testing scenarios, gave him an intuitive feel for when a story had drifted from financial reality. The computation was not just a task. It was training for the judgment that the task eventually produced.

Dr. Osei developed her moral reasoning about resource allocation by struggling with the models’ limitations. What could not be captured, what was lost in aggregation, what assumptions had to be made: these taught her where the values questions lived. She did not arrive at the judgment abstractly. She arrived at it through the computational work.

Kenji developed his auditing instinct by building models. You cannot effectively interrogate a model you could not have built. The vulnerabilities, the assumption traps, the subtle ways training data can bias predictions: these are visible only to someone who has done the work from the inside.

In every case, the computational work was not separate from the judgment work. It was the pathway to it. AI automates the pathway and leaves the destination intact. But without the journey, how does anyone arrive?

We do not have a satisfying answer. This is the deepest version of the apprenticeship problem, and it applies far beyond these four professions. It applies everywhere the junior task was also the training for the senior judgment. Everywhere the easy work was also the developmental work. Everywhere automation removes not just labor but the slow accumulation that labor was secretly building.

What Prediction Was Always For

#

Margaret is still looking at those two numbers. She calls her daughter, who looks up the insurer online and finds mixed reviews. She calls her doctor, who says the 34.7% sounds about right but that the right question is what she values, not what the model calculated. She calls a financial advisor, who spends forty-five minutes talking not about the premium but about how Margaret thinks about the last decade of her life and what she wants it to look like.

None of this is computation. All of it is judgment. And all of it was always what the uncertainty professions were for.

The standard framing asks what happens to these professions when AI can predict better than humans. I keep coming back to a different question: what does AI’s superior prediction reveal about what these professions were always doing?

They were never in the prediction business. They were in the judgment business. Prediction was the part that was hard enough to require professionals, the visible and billable component. But the value, the reason clients paid and societies needed these professions, was always the judgment that prediction enabled.

When prediction becomes cheap, judgment becomes expensive. Not because judgment is rare, but because developing it requires the slow accumulation of experience that cannot be accelerated, the encounter with consequences that cannot be simulated, the moral wrestling that cannot be automated.

Margaret eventually decides not to buy the policy. Not because the number is wrong. Because the $847 a month, over ten years, would cost her the financial independence that matters more to her than the care the policy would fund. That is a judgment. The AI provided the input. The judgment was hers.

The interpreters of uncertainty are not disappearing. They are becoming, at last, what their name always implied: not producers of predictions, but people who help the rest of us understand what the predictions mean.

Whether we are producing enough of them is the question I cannot answer, and the one that keeps me up at night when I think about what happens to Margaret when the profession that should be helping her has been halved and the halving was called progress.

The Transformed is a series within The Approximate Mind examining how AI reshapes professional work across six arcs. The first essay found that AI unbundled pattern recognition from judgment in diagnostic medicine, revealing the human core. This essay finds the same unbundling in professions built on uncertainty, with a crucial difference: reflexivity means better prediction does not resolve uncertainty but exposes it. The series builds on Part 2 (When to Trust Hunches), Part 3 (The Irrational Quest), Part 49 (The Confluence of Influence), and Part 56 (The Space Between Yes and No). The next essay examines the digital builders.

References

#

Decision-Making Under Uncertainty

Kahneman, Daniel. Thinking, Fast and Slow. Farrar, Straus and Giroux, 2011.

Taleb, Nassim Nicholas. Antifragile: Things That Gain from Disorder. Random House, 2012.

Taleb, Nassim Nicholas. The Black Swan: The Impact of the Highly Improbable. Random House, 2007.

Tetlock, Philip E. Expert Political Judgment: How Good Is It? How Can We Know? Princeton University Press, 2005.

Reflexivity and Markets

Merton, Robert K. “The Self-Fulfilling Prophecy.” The Antioch Review, vol. 8, no. 2, 1948, pp. 193-210.

Shiller, Robert J. Narrative Economics: How Stories Go Viral and Drive Major Economic Events. Princeton University Press, 2019.

Soros, George. The Alchemy of Finance. Wiley, 1987.

Health Economics and Resource Allocation

Drummond, Michael F., et al. Methods for the Economic Evaluation of Health Care Programmes. 4th ed., Oxford University Press, 2015.

Sen, Amartya. Development as Freedom. Knopf, 1999.

World Health Organization. Choosing Interventions That Are Cost-Effective (WHO-CHOICE). WHO, 2023, www.who.int/choice.

Actuarial Practice and AI

Institute and Faculty of Actuaries. The Actuary in the Age of AI. IFoA, 2022.

O’Neil, Cathy. Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy. Crown, 2016.

Expertise and Professional Judgment

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.

Klein, Gary. Sources of Power: How People Make Decisions. MIT Press, 1998.