Chapter 8 — The False Merit of Behaviour
Behavioural data seem to offer an answer to the limits of older classifications. Instead of charging a person for what they are, insurers can claim to take account of what they do. The idea is easy to understand. Careful driving, home maintenance and other preventive efforts can reduce losses. The policyholder should receive part of the benefit.
This form of personalisation can correct a category that is too broad. It can also turn protection into a continuing test. Avoiding an accident is no longer enough. The policyholder must produce the right traces, satisfy the indicator and show that they deserve to contribute less. Several different steps are folded into one another. A trace becomes the description of a behaviour. The behaviour is treated as a choice. The choice becomes a responsibility. That responsibility is then used to justify a price [1, 2].
In 1982, the National Organization for Women, widely known as NOW, launched its Insurance Project against the use of sex in American insurance pricing. Motor insurance raised a particular difficulty. Under the rating systems challenged by the organisation, women under twenty-five generally paid less than men of the same age. Beyond that age, men and women were often placed in the same classes [3, 4].
Insurers could therefore present sex classification as discrimination in women’s favour. NOW activists replied that the advantage was limited. Adult women drove fewer miles on average than adult men, but that difference in exposure was not always reflected in premiums. A woman who used her car very little could pay as much as a man who drove much farther [4, 5].
NOW did not simply ask insurers to remove the lower rate for young women and apply the male rate to everyone. In proceedings in Pennsylvania, the organisation argued that classes based on sex should not disappear until pricing took greater account of mileage. The disputed variable was to be replaced by a more direct measure of vehicle use [4].
The demand sought to loosen the hold of an assigned identity. A woman should not be treated as the representative of a group average when her own use of the vehicle could be observed. The challenge nevertheless helped establish another view of fairness. A price would be fairer when it came closer to the conduct of each person [3, 6].
A few decades later, telematics devices1 would record much more than distance. They could collect information on time of day, speed, acceleration, braking, routes and sometimes location. A demand for individual treatment, made in order to escape a social category, opened the way to much closer observation of driving [7].
NOW was not asking for continuous surveillance. It was challenging a classification it considered unjust. The history shows how a criticism can be absorbed by the system it seeks to change. Removing one variable does not necessarily change the idea of justice behind the price. It may simply send insurers looking for more data, often closer to private life.
Observing behaviour is not the issue in itself. A measure closer to action can recognise a real reduction in risk. Difficulty arises when every gain in precision is assumed to require a more individual price.
Claire and Nadia agree to install the application offered by their motor insurer. For three months, it records their journeys before calculating a score. The programme is not presented as compulsory. It promises a discount to careful drivers.
Claire teaches at a secondary school fifteen minutes from home. She drives mainly in the morning and late afternoon, rarely travels at night and parks in her driveway. The application records low mileage, few night journeys and few instances of sudden braking. At the end of the observation period, her annual premium falls from €680 to €520.
Claire finds the result fair. She drives carefully and sees no reason to pay for people who take more risks.
Nadia works as a care assistant at a residential facility on the edge of town. Her shifts change from week to week. Sometimes she starts before dawn. At other times she returns after ten in the evening. She drives farther on roads used by lorries, with roundabouts and congestion that sometimes require rapid braking.
Nadia has never had an accident or received a fine. Her score nevertheless records night driving, high mileage, repeated braking and street parking. Her premium rises from €720 to €960.
The application records several real differences between their journeys. It does not show why those differences exist. It does not bring Nadia’s home closer to her workplace, change her shifts or repair the roads. It classifies traces produced by two very different arrangements of daily life.
The discount given to Claire does not by itself create new resources. When total losses remain unchanged, asking less from some policyholders means asking more elsewhere. The score decides which differences will remain shared and which constraints will be returned to the people who bear them.
Claire is not inherently more prudent than Nadia. She has more room to arrange her life around what the application rewards. Behavioural pricing can be defended when the measured action really reduces risk. The action must be reasonably accessible, and the policyholder must receive part of the benefit. False merit appears when these conditions are assumed rather than established.
From Category to Behaviour
Traditional segmentation uses age, address, vehicle, occupation or claims history. Behavioural pricing claims to move closer to action. It observes driving, steps, sleep, home equipment or photographs requested by the insurer. It does more than estimate a future loss. It also tries to change what may cause that loss [1, 8].
This aim is not foreign to insurance. A contract can affect the risk it covers. A deductible may encourage care or discourage small claims. A maintenance requirement may reduce some forms of damage. A prevention visit may identify unsafe wiring or a leaking pipe. Insurers estimate future claim frequency while also trying to influence its causes [8, 9].
Connected tools extend that logic. They make conduct more visible and allow the insurer to intervene during the life of the contract. Research on motor insurance based on vehicle use suggests that some drivers change their behaviour when they receive feedback or know that they are being observed. The effect varies with the design of the programme, the variables recorded and the way rewards are calculated [10, 11].
The same technology can serve different purposes. It can inform a driver, help explain a particular journey, recognise a measurable reduction in risk or identify policyholders who are already less costly. These are not equivalent uses. An application that gives advice does not exercise the same power as a score that raises a premium or prepares a non-renewal [12, 13].
The language of reward tends to blur the distinction. The good policyholder is said to be recognised at last. The person who pays more is invited to change. Producing favourable data, adapting a home, buying a newer vehicle or changing working hours requires time, money and alternatives. Those resources are not distributed equally.
An incentive also carries a judgement. It changes the cost of an action and signals which conduct the institution regards as normal, responsible or deserving [14]. A discount for affordable safety work may recognise a contribution to common protection. An opaque score may instead turn insurance into a contest in which policyholders are ranked by their ability to conform [15].
From Trace to Merit
A sensor never observes the whole situation. It records a change in speed, a location, a time or a number of steps. The trace must then be interpreted as a behaviour. Sudden braking may follow from distraction, a child crossing the road, poor road design or dense traffic. The sensor measures the movement. It does not record the story of the act.
The behaviour is then treated as a choice. Driving at night may be leisure or a hospital shift. Walking little may reflect preference, disability, sedentary work or the absence of a safe public space. High heating use may come from a chosen temperature or from poor insulation that a tenant cannot change.
The choice is finally turned into responsibility. Even when a person can act, they are not the sole cause of the risk. Driving depends on roads and working hours. Health also depends on housing, income, environment and access to care. The condition of a house depends on its owner, but also on builders, local authorities and surrounding infrastructure.
Responsibility for conduct is not the same as responsibility for the loss that occurs. Driving farther can increase the probability of an accident. It does not mean that the driver chose the particular accident that happened. The part of risk connected to a decision must remain distinct from bad luck [16].
The last step gives this attribution a financial consequence. The person who receives a good score appears to have earned the right to contribute less. The person with a poor score appears to owe the cost revealed by their conduct. The classification no longer measures exposure alone. It also distributes approval and suspicion [17, 18].
The history of credit helps explain the change. Before automated scoring, lenders explicitly judged an applicant’s character, reputation, stability and presumed willingness to keep promises. The score made this judgement less visible without removing it. The morality of the reliable borrower moved into the language of risk and regularity [19].
Disputes become sharper when the statistical division between good and bad risks does not match ordinary ideas of responsibility. A credit-based insurance score may predict claims without describing how a person drives. Illness, separation or job loss can damage a financial record and later raise the price of insurance. Predictive value alone does not justify the use [20].
A technology can be closer to action than a postcode without being closer to responsibility. Between the trace and the price, the institution constructs an account of what the person could have done and what they must now bear.
Prediction Is Not Attribution
A variable is predictive when it improves an estimate of future loss. This property does not explain why the relationship exists. An address may be associated with fire risk without the act of living at that address causing the fire. A time of day may be associated with accidents without explaining all the danger of the journey [21].
Causation asks a different question. Would changing the factor really change the probability or severity of the damage? A statistical association may reflect other features that have not been measured. A factor that directly contributes to loss may also be difficult to observe.
The distinction matters for prevention. Telling a driver that some roads or hours are more dangerous may help reduce exposure. An association between vehicle colour and accidents would not mean that repainting the car would make the journey safer. A useful predictor is not automatically a useful lever.
Even when a factor contributes to risk, the person may not have the means to change it. Housing, working hours, infrastructure and the money required for repairs are rarely matters of free choice. Conduct that can be changed in theory is not necessarily an option available in practice.
The next question concerns attribution. What part can reasonably be assigned to the person? The relevant comparison is not with a life from which all risk has been removed. It is with the situation the person could have reached by taking an accessible precaution at a bearable cost [16].
One decision still remains. Even when a difference is predictive, causal and partly controllable, should it become a premium, deductible, restriction or refusal? The answer depends on the good being protected, the size of the consequence and the alternatives left to the person. An accessible precaution may justify assistance, a limited requirement or a modest price difference. It does not follow that the policyholder should bear the whole loss when bad luck occurs.
Prediction, assistance, attribution and pricing are separate operations. Putting them into one score makes it appear that a statistical measure already contains the rule that should govern its use.
When a Body Becomes Behaviour
Body weight shows how a physical characteristic can be read as the trace of conduct. Body mass index, usually called BMI, relates weight to height. It is simple to calculate and allows comparisons across populations. A BMI of 30 or more is commonly used to define obesity in adults [22].
The simplicity of the measure explains its wide use. It also marks its limits. BMI does not directly measure the quantity or distribution of body fat. It does not describe the effects on organs, physical ability or daily life. Two people on opposite sides of the threshold may be more similar than two people placed in the same category.
A threshold does not necessarily reveal a natural break. It marks the point at which an institution has decided to act. It may trigger an examination, open access to treatment, determine eligibility or produce an administrative consequence. Its value depends on the decision attached to it [23, 24].
An international commission recently proposed separating screening by BMI from clinical diagnosis. It uses the term clinical obesity when excess body fat is already accompanied by impaired organ function or a serious limitation in everyday activities. It uses preclinical obesity when these effects are not present but future risk is higher [25].
This distinction reduces the risk of treating an indicator as a complete diagnosis. It also shows why weight cannot by itself tell the story of a behaviour. Weight is shaped by biological mechanisms, sleep, some medicines, work, income and the food environment. Common forms of obesity are associated with many genetic variations, most of which have small effects and interact with living conditions [26].
Recognising this complexity can counter the simple story of insufficient willpower. Medical classification can still create new boundaries. A more precise category may open care and also become a condition of eligibility. One person is recognised as ill and receives treatment. Another is told to change their way of life.
Two changes are then confused. An approximate indicator becomes a complete description of the body. A partly modifiable characteristic becomes evidence of a free choice. The same number appears to measure health, reveal conduct and justify a consequence.
Medical knowledge does not authorise every use. BMI may open a discussion or support further examination without acquiring the power to set a premium, select an employee or close access to protection. A category that is defensible in a clinical setting should not travel alone into every institution interested in ranking the person.
Data Do Not Determine Their Use
The word personalisation covers very different operations. An application can adapt advice, award points, offer a discount, decide entry into a contract or affect a claim. The person is treated individually in each case. The data do not create the same value or the same power.
Health programmes such as Vitality show this range. Steps, activities and reported behaviours can lead to rewards and maintain a continuing relationship with the brand without directly recalculating each medical risk. The data then have commercial and relational value as well as possible actuarial value [27].
Uses also differ across contracts, lines of insurance and countries. Information that changes a motor premium may remain excluded from health pricing. A programme bought by an individual does not operate in the same way as one arranged through an employer. An experiment may be designed mainly to retain customers or sell other services [10, 13].
Personalised data do not always describe insured risk. They may estimate whether a customer will buy, compare offers, cancel or accept an increase. The price then reflects the customer’s value as a buyer as well as the expected cost of claims. A reward described as recognition of prudence may work mainly as a tool of retention.
Information collected for advice may later be used for selection. This change in purpose does not require the data to be false. It changes the relationship in which the data act [12]. A step count may help someone follow a personal goal without deciding access to insurance. A photograph of a roof may support repairs without automatically becoming a reason for non-renewal.
This development is part of a broader movement. Households are asked to manage health, housing, retirement and mobility as a portfolio of risks. A walk or a journey acquires monetary value when an application converts it into points. Security becomes continuing work on the self [28–30].
The institution has not disappeared. It still defines the indicators, rewards and consequences. It asks each person to produce evidence of prudence and to answer for the result.
The Discount and Its Other Side
Behavioural programmes often begin with an invitation. Accepting the sensor brings a discount, points or services. Refusing is said to leave the contract unchanged. This neutrality may not last if the reference price rises or if people who refuse monitoring are gradually treated as less transparent.
When expected losses remain the same, a discount does not create new resources. It must be offset by a higher contribution elsewhere, a larger deductible, narrower cover or a reduction in the insurer’s margin. A bonus also determines the burden left to others [2, 31].
The calculation changes when the programme really reduces the frequency or severity of claims. The discount can then share part of the benefit created by prevention [9, 11]. This reduction must be distinguished from three other outcomes. The programme may simply identify policyholders who were already less costly. It may discourage some claims. It may move risks to another contract or to people who have left the portfolio.
The score should therefore be judged by losses actually avoided and not only by the improvement of its indicator. Once a measure becomes a target, people and organisations learn to satisfy it, sometimes without achieving the purpose it was meant to serve. This problem is commonly associated with Goodhart’s law and Campbell’s law [32–34].
A driver may learn to avoid the braking events recorded by the application without becoming more attentive in every situation. A policyholder may decide not to report a problem in order to preserve a score. A company may increase enrolment in a prevention programme without reaching the people who need help most. The measure improves faster than safety.
Personalisation can also remove from the common price those who know how to produce the conduct the system rewards. The group left behind then contains more night work, long commutes, chronic illness, disability, older vehicles and poorly equipped neighbourhoods. Its average claims cost rises and appears to confirm that it always contained the expensive risks.
A prevention policy should not begin by measuring the people who already have the means to succeed. It should first distribute the means to act, then decide carefully what should be recognised in the price.
Taking Behaviour into Account without Converting Everything
Rejecting behavioural merit does not mean ignoring conduct or making every premium identical. Some actions increase avoidable harm. Some forms of prevention work. Pooling does not require indifference to what people do [15, 17].
It does require the institution to separate operations that behavioural programmes often merge. The insurer can help reduce risk. It can consider the part reasonably attributable to an available decision. It must then share the loss that remains and decide whether the observed difference may alter the price without closing access to protection.
The result is a need for zones of non-conversion. Information may be used for advice, care or prevention without gaining an automatic right to enter selection [2, 12]. Genetic information, an old diagnosis, a climate map and a driving score are not the same. They raise a common question. What consequence may an institution attach to what it knows?
Any answer must consider whether the person can understand the rule, act on the information and challenge an error. It must also consider what options remain after an adverse decision. Insurance often controls access to housing, credit, health or work. In those settings, a statistical difference requires a strong justification before it becomes an increase, exclusion or refusal.
Older categories are not a safe alternative. Sex, social origin and place of birth do not become fair criteria because individual tracking is intrusive. The choice lies between these extremes. It concerns permitted uses, available assistance and the part of risk that the collective will continue to carry.
A trace never possesses the force of a sanction by itself. An organisation must still choose a threshold, a procedure and a consequence. The next chapter begins when the score stops being information about a person. It becomes a decision that they must accept or challenge.
References
In motor insurance, telematics refers to a device or application that records selected information about journeys and driving. The data may be used to advise the policyholder or adjust the contract.↩︎