Chapter 7 — What Each Side Cannot See
Risk does not look the same to a policyholder and an insurer. The policyholder encounters it through a personal history, practical constraints and the possible consequences of loss. The insurer estimates it from a portfolio shaped by accepted contracts, prices, refusals and departures. Neither view is complete.
For twenty-seven years, Marianne has lived about a hundred metres from a river. She has known warnings, closed roads and a damp cellar, but never a serious flood. When her premium rises sharply, she opens the shutters and looks at the garden. Nothing seems to have changed. The river still runs behind the trees. The new map appears to exaggerate a danger she has never experienced directly.
At the insurer’s head office, an analyst studies the same area on a screen. He sees coloured plots, insured values, claim dates and estimates of frequency. The company’s records cover thirty years and thousands of contracts. They point to growing flood exposure.
The history is still incomplete. It excludes houses the company declined to insure and owners who walked away when they saw the price. It also misses repairs made without a claim, sales completed before the next flood and households that never managed to enter the portfolio. The insurer observes a population already shaped by its underwriting rules and by policyholders’ decisions.
Marianne relies on her own experience, that of her neighbours and the apparent stability of the landscape. The insurer has a broader but selected history. Neither sees the whole risk.
Their errors do not have the same consequences. Marianne may underestimate her exposure and remain insufficiently protected. The insurer’s assessment may change premiums, access to credit and eventually the future of the neighbourhood.
The psychology of risk has studied the limits of individual judgement in great detail. A blind spot appears when bias is attributed only to policyholders, as though the organisation classifying them observed the world without its own selection effects, routines or constraints.
The Event That Never Happens
A house that has not burned for thirty years seems to offer daily proof of its safety. Someone who has never been seriously ill may postpone buying protection. An old warning loses force as time passes. An abstract danger struggles to compete with present expenses.
Optimism bias, present bias, the illusion of control and the availability heuristic name some of these regularities. A recent or spectacular event appears more likely. A catastrophe never experienced remains remote. After a flood, demand for insurance may rise, then decline as the memory fades [1, 2].
These mechanisms do not produce one universal response. An extreme event changes resources, emotions, beliefs, networks of assistance and relations with institutions at the same time. It may increase caution, strengthen some forms of solidarity or create the feeling that no protection can be trusted [2].
The absence of insurance does not necessarily reveal a poor understanding of probability. It may reflect an unaffordable price, a contract that is difficult to understand, uncertain payment or greater trust in relatives and neighbours than in a company. It may also come from a forced choice between an immediate premium and needs that cannot wait.
The explanation matters because it shapes the response. Describing a household as irrational leads to better information, simpler messages or stronger warnings. Recognising that the contract itself may appear risky requires attention to its price, quality, provider, payment timetable and claims process.
When Insurance Itself Appears Risky
Microinsurance makes this problem especially visible. Microinsurance refers to policies designed for households with low incomes, usually with small premiums and limited benefits. It may cover health, crops, death or catastrophe. The apparent simplicity of the product does not remove the difficulty of paying the premium, understanding the terms or trusting the provider.
Households that face illness, crop failure or natural disaster should, according to standard models, place a high value on protection. Yet many do not buy the products offered [3, 4].
An explanation based on bias is tempting. Households may misunderstand probabilities, underestimate catastrophe or give too much weight to current spending. These factors matter, but they do not exhaust the decision. Price, limited cash, contractual complexity and the timing of premium payments also play an important role [5–7].
For a poor household, insurance can itself look uncertain. The premium immediately reduces scarce resources. The benefit depends on a future event and then on an institution that must recognise the loss and apply the contract. The household is not only judging the probability of damage. It is also judging whether the promise will respond to what happens.
Index insurance shows the problem clearly. Payment depends on a measure fixed in advance, such as rainfall at a weather station or the average crop yield in an area. A farmer can suffer a poor harvest without receiving anything if the index does not cross the contractual threshold. The gap between the loss experienced and the payment triggered by the index is known as basis risk. It can make low demand entirely rational even when the product is correctly priced [8, 9].
Trust in the provider matters as well. Research in rural Cambodia found that hypothetical demand for microinsurance depended on price, previous experience of disaster, financial knowledge and the credibility of the organisation selling the policy [10]. Having suffered a loss did not automatically increase the desire to insure.
Formal insurance is only one way of coping with damage. Households may rely on family, neighbours, savings, credit or local forms of mutual aid. These networks can respond quickly and reduce the appeal of a contract. They become fragile when the same shock reaches the whole community at once [11, 12].
Buying insurance therefore involves judging two risks. The first is the possible loss. The second is the contract expected to carry it. The insurer classifies a person according to the risk they bring. The person also classifies the insurer according to the trust they place in its promise. Pricing and underwriting models rarely make this second assessment visible.
Two Views of the Same Future
In the early 1990s, David lives in San Francisco. He is thirty-six and has left a job in advertising. He holds a life insurance policy bought several years earlier. He has known for three years that he is HIV-positive. After a bout of pneumocystis pneumonia, the discussion no longer concerns HIV alone. It concerns AIDS.
Treatment is expensive, work is becoming harder and the rent still has to be paid. A broker offers to arrange a viatical settlement.1 David will receive part of the death benefit immediately. The purchaser will pay the remaining premiums and collect the insured sum when David dies [13, 14].
For David, the transaction turns a promise that would arrive too late into money available now. For the investor, it turns a life into a probable duration. The sooner David dies, the higher the return. Both parties know the same diagnosis, but they do not see the same object [15, 16].
David sees months of housing and treatment, along with some remaining autonomy. The investor sees an uncertain timetable. The doctor sees a disease and changing therapies. The insurer sees an obligation that will be triggered by death.
More effective antiretroviral treatments disrupt the market. Policyholders live longer. This medical progress is good news for them and bad news for investors who had priced an early death. The same information takes on a different value according to the actor receiving it and the decision it informs [17].
An estimate of life expectancy can guide treatment, release cash, set a price or organise an investment. Data do not contain their own use. Their meaning comes from a relationship and an institution.
This plurality does not mean that every viewpoint is equally powerful. The imbalance begins when one actor alone can convert its interpretation into a price and a decision. David knows his illness from within. The investor holds the capital and the contract that turn this knowledge into a market.
One Person’s Risk, Another’s Danger
The distinction between risk and danger depends on a person’s position in the decision [18]. A consequence appears as a risk when it is connected to a decision the actor controls. It appears as a danger when it is experienced as coming from outside.
An insurer chooses a model, deductible, underwriting threshold and portfolio composition. It may describe withdrawal as a prudent decision in response to a possible loss. A household does not choose the reclassification of its neighbourhood or the disappearance of available offers. What is a balance-sheet risk for one actor becomes a contractual, housing or financial danger for another.
This distinction prevents exposure from being confused with choice. A loss can be attributed to a personal decision only when a genuine alternative was available. The abstract possibility of moving home, changing employment or taking another route does not turn every vulnerability into a voluntary risk.
Both views may be rational without being equivalent. The actor making the decision can define a situation as a risk to manage. The person receiving its consequences must often adapt to a danger they did not choose.
The View from Above
Aerial and satellite images give insurers a perspective residents do not have. They allow a company to inspect roofs, vegetation, debris and the distance between buildings. This view covers a wider area than the homeowner’s daily experience. It remains fallible.
American regulators have received complaints about refusals and non-renewals based on images that were old, unclear or attached to the wrong building. West Virginia and Tennessee reminded insurers that an image should remain one piece of evidence among others. When its interpretation is uncertain, the insurer should seek more information, show the image to the policyholder and accept recent photographs or a physical inspection [19, 20].
A policyholder cannot correct information they do not know exists. A tarpaulin may be mistaken for a swimming pool, a structure on a neighbouring property may be assigned to the wrong parcel and old equipment may remain visible after it has been removed. The error stays silent until it produces a higher premium, a demand for work or a non-renewal.
The answer is not to oppose the resident’s eye to technology. Both can be wrong. The answer is to organise a process in which the interpretation can be challenged before it becomes a decision. The power to see from a distance should come with a right to see what the institution claims to have observed.
Homes can also produce data from within. Energy use, temperature, door openings, smoke detection, presence, water leaks and alarm activation can help prevent damage. They can also turn the home into a continuously observed space. The boundary between prevention and surveillance depends less on the sensor than on the uses authorised for the information it produces [21, 22].
The Portfolio Is Not the Population
An insurer may hold millions of observations. Such abundance creates an impression of completeness. Yet a portfolio is a filtered population.
It contains those whom the company accepted, those who accepted its price, those whose contracts were retained and those who reported a loss recognised by the claims process. It makes declined applicants, uninsured households, former customers, losses below the deductible and repairs paid by relatives less visible.
Credit offers a simple comparison. A lender observes repayment among the applicants to whom it granted loans. It does not directly observe what rejected applicants would have done. The earlier decision removed part of the information that a later model would need. Insurance faces the same problem. The results observed depend on the rules that governed entry into and exit from the portfolio [23].
Older categories also acquire a cumulative advantage. They have generated long series, stable forms and professional routines. Groups excluded or poorly described rarely return to contradict the insurer’s statistical history. A variable may survive because it is built into systems, not because it offers the best available account of risk [24].
Before saying that the data show a difference, we should ask whom they describe and after which refusals, deductibles and departures.
Data Are Produced
The word data suggests something already present and waiting to be collected. A trace becomes data only when an institution decides to retain it, format it, attach it to a unit and make it comparable [25–27].
There is no completely raw braking event. A device selects a recording frequency and an acceleration threshold. An application decides when a journey begins and how to treat a switched-off phone, a steep road or a stop forced by another driver. A claims database depends on reports, benefits, deductibles, assessments and administrative categories.
Calling data constructed does not make all measures equivalent. Some are more precise, more complete and better suited to a decision. It does require the institution to document their origin, transformations and blind spots. When a measure raises a price or closes access to a market, the insurer is responsible for the whole chain and not only for the final calculation.
When Data Change Function
Information does not necessarily remain where it was collected. Data gathered to prevent damage become a pricing factor. Information created for treatment enters underwriting. A map designed for emergency planning accompanies a non-renewal letter. Ordinary information can become sensitive when several sources are combined [21, 28].
Meaning depends on the context through which information travels [22]. Medical data do not perform the same function when they guide care and when they decide access to insurance. A climate map creates a different relationship when it opens financial assistance rather than justifying withdrawal. This gradual extension of use is often called function creep [29].
The change rarely arrives through one dramatic decision. The data already exist and reusing them appears inexpensive. Another department gives them an additional task. A prevention tool slowly becomes a selection tool without the people concerned having had a chance to discuss the change.
A trace does not carry the story that explains it. A purchase may indicate risky consumption or an ordinary errand. Presence in a neighbourhood may reflect work, a medical appointment or a family visit. A predictive model may establish that a variable improves an estimate without showing why the association exists [30].
Actuarial work cannot demand a complete causal demonstration for every variable. Insurance data are largely observational. A stable association may still provide useful information. It does not determine the meaning of the variable or the consequence that should follow.
Professional actuarial work in the United States has proposed requiring a rational explanation. The insurer should provide a plausible and understandable account linking the variable to circumstances that may contribute to the risk [31]. This test can exclude accidental correlations. It still does not decide whether the information should produce a premium, an underwriting condition or suspicion of fraud.
Losses That Never Become Claims
A loss enters an insurer’s data only after several steps. It must be noticed, understood as potentially covered, reported, documented, accepted and valued. At every stage, some damage disappears from the file.
A household may expect the amount to remain below the deductible. It may fear a future increase or non-renewal. It may repair with help from relatives, postpone the work or fail to realise that the policy could respond. A tenant may live with damp while having no control over the owner’s insurance. A small business may absorb an interruption to preserve a commercial relationship.
The practical threshold for making a claim may therefore be higher than the contractual deductible. The additional amount that the policyholder absorbs before deciding to report a loss has been called a pseudodeductible [32].
The absence of a claim is not the absence of loss. It may show that the contract left part of the cost with the person it was meant to protect.
The decision to claim also depends on resources. Faced with the same damage, a household with savings may pay for repair, avoid the procedure and preserve its claims history. A household without cash is more likely to use the cover it purchased.
American controversies over credit-based insurance scores reveal this ambiguity. These scores predict the number and cost of claims, but the association does not prove that people with lower scores cause more accidents. In the data examined by the Federal Trade Commission, much of the difference concerned claim frequency rather than average claim cost. No definitive causal explanation was established [33, 34].
One possible explanation is the ability to absorb a small loss without using insurance. The score may partly measure the need to claim rather than the tendency to cause damage. The hypothesis remains debated. It is enough to show that observed claim frequency combines exposure, reporting decisions and available resources.
Pricing can then charge more to those least able to do without insurance. Claims procedures also select what becomes a recognised expense. An assessment may exclude a cause, reduce an amount or demand evidence that is difficult to produce. A complex process retains mainly those who know their rights, have time and can bear the cost of a challenge. Abandoned files disappear more easily from statistics than from the lives they have weakened [35–37].
A public account of damage must therefore bring several sources together. Paid claims are essential, but they should be supplemented by household surveys, health data, appeals, municipal expenditure, postponed work and residential trajectories. Insurers possess valuable knowledge. They are not comprehensive accountants of social harm.
Organisations Have a Psychology Too
Institutions do not feel fear or hope in the way people do. They still produce regularities that resemble biases. A recent event receives more attention in committees and models. A long period without loss normalises exposure. Annual targets make immediate prevention spending more visible than future losses avoided. Teams favour available data, established methods and categories regulators already understand [38–40].
This short-sightedness comes partly from the division of work. The underwriter watches the composition of the portfolio.2 The actuary seeks a defensible price. Finance protects capital. Legal staff limit disputes. Claims handlers process and close files. Each person may perform a task properly while the combination of these local rationalities produces a decision no one has examined as a whole [41].
An organisation sees precisely what it has learned to count and much less of what crosses its boundaries. It knows the claims burden, but not all the damage retained by households. It measures the cost of prevention assistance, but not always the value of loss avoided. It records cancellation, but often does not know what happens to the person after departure. Categories and procedures select which information can travel to the places where decisions are made [40, 42, 43].
Technological announcements also direct attention. Connected objects will reveal conduct, artificial intelligence will personalise policies and decisions will become almost instantaneous. Such announcements do not always describe settled practices. They still influence budgets, recruitment, partnerships and expectations. An experiment can help create the market it predicts [44].
Hype around artificial intelligence does more than exaggerate performance. It presents adoption as inevitable. An organisation that does not automate appears to be falling behind. Discussion then moves away from whether the project is useful, which task it will automate and who will benefit. It turns instead to the speed of deployment [45].
The label artificial intelligence also groups together very different techniques and uses. Speaking more precisely about automation forces a simpler question. Which task will be transferred to the system, for whose benefit and with what possibility of reversing the decision? This precision resists the idea that a general technological advance requires the same response everywhere [45].
The announced future becomes a present constraint. A company builds a database, selects a provider, modifies procedures and trains staff before establishing the real value of the tool. Once these investments have been made, abandoning the project becomes harder. A provisional decision gradually acquires the durability of infrastructure.
The idea of continuous insurance gives this horizon a name. A contract was traditionally organised around distinct moments such as application, renewal and claim. Connected devices and external data make it possible to imagine a relationship in which exposure is observed throughout the contract. Industry writing sometimes calls this continuous underwriting [46, 47].
Continuous observation can support very different operations. Data may warn the policyholder, update an internal assessment or change the terms offered at renewal. Observation, intervention and conversion into price should not be confused. An early alert may prevent damage without every short-lived variation becoming a premium change.
Even a disappointing experiment may teach the company to gather traces, build an interface or coordinate partners. Research on behaviour-based motor insurance calls some of these projects happy failures. The actuarial proof may fail while the organisation still acquires useful infrastructure [48]. Experiments should therefore be examined early, while several uses remain possible and withdrawal is still realistic.
Absences, Departures and Displaced Losses
Unreported losses are not the only elements missing from a portfolio. Departures also change the population observed. After an increase, policyholders who switch company, reduce coverage or become uninsured may differ from those who remain. The results of the surviving portfolio can improve while protection deteriorates across the market [49, 50].
A territory may appear less risky in one insurer’s accounts after its hardest contracts have moved to a residual market or lost coverage. Residual mechanisms concentrate the risks the ordinary market did not retain. Their results reflect both the properties covered and the selection process that sent them there [51, 52].
Behavioural programmes face a similar difficulty. A sensor describes people who agree to be monitored and remain in the programme. It says less about those who decline, withdraw or leave the insurer. The observed population is also the result of a participation decision [23, 53].
No insurer sees all these trajectories alone. It knows that a policy was not renewed, but rarely what followed. Did the household find another insurer, reduce coverage, enter a residual market or give up insurance? Market-wide observation must connect premiums, deductibles, benefits, non-renewals and departures. Without that view, protection may shrink through the accumulation of private decisions, each of which appears defensible when considered separately [51, 52].
The Loop That Manufactures Its Evidence
A classification can predict the future and help produce it. It first changes what the institution observes. A territory inspected more often generates more recorded defects. A group exposed to stronger controls produces more alerts and cases. These observations then seem to confirm that the group deserved special attention. The defects may be real, but the probability of detecting them depends on the initial classification. Similar feedback loops have been studied in predictive policing [54, 55].
A decision can also alter the risk it claims only to measure. A higher premium leaves less money for maintenance. A larger deductible delays repair. A credit refusal prevents adaptation work. Insurer withdrawal affects property values and may accelerate the departure of households able to move.
A few years later, the neighbourhood may indeed be more vulnerable. The prediction has become performative. By changing the options available to people, it enters the causal chain of the outcome that it will later appear to have anticipated [56, 57].
Another loop appears when a measure becomes a management target. Goodhart’s law is commonly summarised as the idea that a measure loses reliability once it becomes a target [58, 59].
An insurer seeking to reduce reported claim frequency may raise deductibles, complicate procedures or remove highly exposed contracts. The indicator improves without any reduction in loss. Costs have been left with policyholders or moved outside the portfolio. A target for rapid closure may shorten the average life of files while weakening the review of complex cases. Staff do not need to manipulate figures. They only need to adapt their work to the criteria by which it will be judged [60].
These loops do not make every prediction false. They require a distinction between predicting an external condition and evaluating a situation the institution helps transform. Good performance on later data is no longer enough when the decision changes the population observed, the resources available or the events that will be recorded.
A portfolio is therefore an archive of losses and of the policies that shaped them. It records the effects of planning, credit, refusals, deductibles, claims procedures and management targets. Treating it as a natural mirror of the world erases the institutions already present in the data.
Data Do Not Speak Alone
The rhetoric of scoring often contrasts objective data with human judgement, which is presented as intuitive and biased. Yet an institution chooses the outcome to predict, the reference population, the information retained, the treatment of missing values and the consequence attached to the number. Data choose neither the question nor their use.
The same requirement applies to systems sold as artificial intelligence. Strong performance on a standard test does not show that the system is suitable for the precise task, population and consequences planned by the organisation [45].
Mechanical objectivity has its own history [61]. Standardised instruments were intended to limit the observer’s visible intervention. The absence of a hand at the moment of calculation does not abolish judgement. It moves judgement into the design of the instrument, the selection of cases and the definition of categories.
Brian Glenn describes an insurance version of this belief as the myth of the actuary. Risk selection presents an outer face of numbers and objective criteria. It retains a less visible face made of stories about prudence, character and policyholders’ responsibility [62, 63].
Frederick Hoffman’s work offers an important historical example. At the end of the nineteenth century, the Prudential statistician assembled extensive data to argue that African Americans formed a population that was difficult to insure. The volume of tables gave the argument an appearance of impartiality. W. E. B. Du Bois and other critics showed that the comparisons grouped people living under very different social and economic conditions. Race had been retained as the central explanation when a different organisation of the data would have made other mechanisms visible [64].
The numbers had not invented the story. An earlier story had determined which data should be compared and which cause should be sought. Contemporary procedures can make such choices harder to see, but they do not remove them [65, 66].
The alternative is not untidy human subjectivity on one side and automatic objectivity on the other. An objective practice states its choices, documents its limits and makes challenge possible. A score becomes dangerous when the procedure disappears and the number arrives alone, as though reality itself had chosen the question and the decision.
Errors That Do Not Carry the Same Power
It would be easy to conclude that policyholder and insurer each make their own kind of mistake. That symmetry would be false. A policyholder may underestimate a danger. An insurer may confuse an unreported loss with no loss, learn from a selected population or treat as external a vulnerability that its own decisions help strengthen.
These errors do not have the same reach. Marianne’s mistake first affects her own protection. The insurer’s mistake enters a model, underwriting rule or pricing procedure. It can reach everyone sharing the misunderstood characteristic, then be repeated by a bank, reinsurer, regulator or public authority.
The authority of an estimate depends on the quality of the measure and on the position of the actor able to attach a consequence to it. Searching for an institution with no blind spots would be futile. The practical task is to confront perspectives that do not observe the same things. Residents reveal losses absent from files. Associations connect refusals. Researchers study selection. Regulators compare portfolios and follow people who disappear from the ordinary market.
This confrontation does not require every decision to yield to personal testimony. It requires the insurer to state what its data support, what they do not see and how the policyholder can provide contrary evidence.
The partial view of risk then begins to serve another function. It no longer predicts expenditure alone. It starts to qualify the effort, choice and merit of the person being classified. The next chapter examines this movement from an observed trace to a judgement about behaviour.
References
In a viatical settlement, the holder of a life insurance policy sells the right to receive the death benefit to a third party. The purchaser pays an immediate sum, usually takes over future premiums and later receives the insured amount when the policyholder dies.↩︎
Here, an underwriter is the professional who examines risks and decides whether the insurer will accept them, decline them or cover them subject to particular conditions.↩︎