The People Most at Risk Are Often Missing From the Spreadsheet

What if the first victims of artificial intelligence are not the workers whose jobs machines can perform, but the workers our measurements fail to see?

That possibility creates an unsettling connection between two seemingly separate problems. Efforts to predict which occupations are most exposed to AI often focus on the cognitive complexity of tasks. Efforts to understand education systems often focus on enrollment figures, teacher credentials, and learning outcomes. Both appear to be exercises in careful measurement.

Yet both can produce the same distortion: they make visible the people and institutions that are already documented, regulated, and legible, while hiding those with the weakest protections. The result is a dangerous confusion between what can be measured, what is formally recognized, and what is actually vulnerable.

A software developer may receive a high AI exposure score because software can generate code. A small private school may barely appear in national education statistics because no reliable database records its students or teachers. In the first case, visibility can exaggerate risk. In the second, invisibility can conceal it. Together, they reveal a broader principle:

The map of a system is often a map of its records, not a map of its risks.

The measurement trap: exposure is not vulnerability

Imagine a city preparing for a flood. Engineers inspect the buildings with detailed blueprints, recent safety certifications, and known occupancy levels. They conclude that the tallest office towers are most exposed because they contain the greatest number of rooms with expensive electrical systems.

Meanwhile, an informal settlement near the river is left out of the model. Its buildings have no reliable plans, its residents are not fully counted, and its infrastructure is poorly documented. The model does not show that the settlement is in greater danger. It shows only that the engineers know more about the towers.

This is close to what happens when AI risk is assessed through occupational exposure scores. Researchers examine the tasks associated with an occupation and estimate how readily an AI system might assist with or replace them. Occupations involving writing, analysis, coding, diagnosis, or legal reasoning tend to score highly because their tasks are easier to describe in the language of information processing.

But an occupation's technical exposure is not the same as a worker's economic vulnerability. A financial analyst may perform tasks that an AI system can accelerate, yet work inside a highly protected institution, possess specialized credentials, belong to a professional network, and have enough bargaining power to redirect the technology toward augmentation rather than displacement.

A low status worker may face the opposite situation. Their tasks may be too irregular, physical, fragmented, or poorly documented to register as highly exposed. Yet they may have no contract, no union, little savings, weak legal protection, and limited access to retraining. The occupation looks safer in the model precisely because the model has trouble seeing it.

This produces what we might call the protection paradox: the workers whose tasks are easiest to model may be better positioned to survive technological change, while those whose work is harder to classify may be less able to survive any change at all.

The crucial question is therefore not simply, “Can AI perform this task?” It is also, “Who absorbs the consequences when the task changes?”

The same blindness appears in education

Consider a country with a large and diverse private education sector. Provincial departments may maintain registration lists, but those lists can be incomplete, outdated, or disconnected from one another. There may be no annual student census, no centralized database of institutions, and no standardized system for evaluating performance.

At first glance, this looks like a bureaucratic inconvenience. In reality, it changes the political status of the people inside the system.

If officials cannot reliably say how many students attend private schools, they cannot accurately estimate the number of classrooms needed. If they do not know how many teachers work there, they cannot assess the scale of training needs. If learning outcomes are not collected consistently, they cannot distinguish successful schools from institutions that are failing children at scale.

The absence of data does not mean the absence of an education system. It means the system exists without full public visibility.

This is the educational version of the protection paradox. Schools with strong reporting systems may appear more accountable because they generate more evidence. Schools operating outside coherent oversight may appear less significant because their failures are not aggregated into official knowledge. A documented problem attracts scrutiny. An undocumented problem is often treated as a minor or isolated one.

There is a profound difference between being unmeasured and being unaffected. A child does not receive a better education because the state lacks a record of the school. A teacher does not gain security because their credentials are absent from a database. A family does not become less exposed to poor instruction because its school is statistically invisible.

This is why measurement is never merely descriptive. It allocates attention, resources, and responsibility. What enters the database can become a policy target. What remains outside it can be left to private improvisation.

The hidden variable is institutional power

The connection between AI labor forecasts and invisible education systems becomes clearer when we add a third variable to the usual analysis: institutional power.

Most risk models focus on the object being changed. In labor markets, that object is the task. In education, it may be the school or student. But the consequences of change depend heavily on the surrounding institutions: contracts, professional norms, regulation, unions, data systems, enforcement capacity, and social status.

A useful framework is to separate four dimensions:

1. Technical exposure: How easily can a technology perform or alter the activity?
2. Observability: How accurately is the activity recorded and classified?
3. Protection: What legal, financial, or institutional safeguards surround the person doing it?
4. Adaptation capacity: Can the person or institution gain new skills, negotiate new arrangements, or move to a better position?

These dimensions are often collapsed into one headline number, but they should not be. A job can have high technical exposure and low vulnerability if protection and adaptation capacity are strong. Another job can have modest technical exposure and high vulnerability if protection and adaptation capacity are weak.

The same matrix applies to schools. A school may be visible but poorly governed. It may be invisible but effective. It may be visible, regulated, and well funded, or invisible, unregulated, and unable to improve. Enrollment counts alone cannot reveal these differences.

The central mistake is treating visibility as a neutral property. Visibility is produced by institutions, and institutions are unevenly distributed.

Measurement does not simply reveal inequality. It can reproduce inequality by making some people easier to govern than others.

A professional employee in a regulated sector leaves behind abundant traces: qualifications, payroll records, performance reviews, tax documents, and industry classifications. An informal worker may leave behind fewer traces even while facing greater instability. Similarly, a large school chain may produce reports, websites, financial statements, and examination data, while a small neighborhood school may exist in a provincial file that has not been updated for years.

The first group becomes easier to model. The second becomes easier to ignore.

Why prediction fails when categories become targets

There is another danger. Once a category becomes influential, organizations begin to behave according to it.

Suppose policymakers identify software developers as highly exposed to AI. Universities may reduce investment in computing education, employers may use the classification to justify hiring fewer junior developers, and students may avoid the field. The prediction then changes the environment it was meant to describe.

Now suppose private schools are excluded from reliable national data. Funding formulas may overlook them, teacher training programs may not reach them, and accountability systems may focus elsewhere. Their invisibility becomes self reinforcing. Because they are not measured, they receive less institutional attention. Because they receive less attention, their quality and conditions remain unknown.

This is a form of administrative feedback. A measurement system does not merely observe a social world. It helps distribute opportunity within that world.

The problem is especially acute when categories are built around occupational labels or institutional registration. A job title can conceal radically different realities. “Teacher” might describe a tenured professional in a well funded public school, a temporary instructor in a private academy, or an uncredentialed worker paid by the day. “Analyst” might describe a protected employee at a major bank or a contractor whose income disappears when a client adopts automation.

Likewise, “school” can refer to institutions with radically different levels of staffing, oversight, curriculum, and student support. A category may be administratively convenient while being socially misleading.

Better analysis therefore requires moving from labels to conditions. Instead of asking only which jobs are exposed, we should ask which workers have the least ability to negotiate technological change. Instead of asking only how many schools exist, we should ask which children are enrolled in institutions that have no reliable route to public accountability.

This shift changes the purpose of data. Data should not merely rank objects by exposure. It should identify where power is weakest and where correction is least likely to occur without intervention.

From exposure maps to vulnerability maps

A better approach would combine technical exposure with institutional context. Think of it as a vulnerability map, not an exposure map.

For workers, the map could include:

• The share of income dependent on a single employer or platform.
• Access to contracts, benefits, legal remedies, and collective bargaining.
• The cost and availability of retraining.
• Whether AI adoption is negotiated with workers or imposed unilaterally.
• The worker's ability to move into tasks that technology complements rather than removes.

For schools, the equivalent map could include:

• Verified enrollment and attendance.
• Teacher qualifications, pay, turnover, and employment status.
• Student learning outcomes measured through comparable assessments.
• The existence of inspection, complaint, and improvement mechanisms.
• Access to public support, training, infrastructure, and transparent funding.

Notice what this framework does. It does not discard technical analysis or institutional records. It places them in a wider model. It asks not only what a system can do, but who has the power to respond when it does it.

This also suggests a practical rule for public policy: the less visible a population is, the more cautious we should be about interpreting its apparent level of risk. Low measured exposure may reflect genuine resilience. It may also reflect poor classification, informal work, missing records, or institutional neglect.

The first policy response should not always be a new prediction. Sometimes it should be an audit of the blind spots that make prediction unreliable.

That audit can begin with simple questions:

• Who is absent from the database?
• Which categories combine people with very different protections?
• Which institutions report regularly, and which are merely registered?
• Where do workers or families have no channel to contest the official picture?
• What consequences follow when an occupation or school is labeled high risk or low priority?

These questions are not technical decorations. They are tests of whether measurement is serving the public or merely serving the convenience of administrators.

Key Takeaways

1. Separate exposure from vulnerability. When evaluating AI's effect on a job, assess not only whether the tasks can be automated, but also the worker's legal protection, bargaining power, financial cushion, and access to retraining.

2. Treat missing data as a warning signal. An absent record does not indicate a small problem. It may indicate that a population is informal, weakly protected, or overlooked by the institutions responsible for support.

3. Replace broad labels with local conditions. “Software developer,” “teacher,” and “private school” are not sufficiently precise categories for predicting lived outcomes. Examine contracts, resources, autonomy, oversight, and pathways for adaptation.

4. Measure institutional response capacity. A useful risk assessment should ask who can negotiate change, appeal a decision, obtain help, or improve performance after conditions shift.

5. Audit the model before acting on its rankings. Before directing funds or making workforce predictions, identify who is excluded, how classifications were created, and whether the measurement itself could reinforce existing inequalities.

The deepest lesson is not that data is useless. It is that data becomes dangerous when it is mistaken for reality rather than understood as a record of what institutions have chosen, or managed, to notice.

An AI system may be excellent at generating code while being unable to tell us which developer can survive losing half their tasks. A government may count registered schools while remaining unaware of the educational lives of thousands of children. In both cases, the failure is not primarily a failure of intelligence. It is a failure of attention organized through institutions.

The future will be shaped not only by what machines can do, but by who is visible when machines change the rules. The people most in need of protection may be precisely those who appear least exposed, because the spreadsheet has mistaken their absence for safety.

The question we should ask before trusting any forecast is therefore simple: Who is missing from the picture, and what happens to them when the picture becomes policy?