Big data in insurance contracts – a tool for good, or bad?

5 1 Introductory comments

When taking out insurance, consumers should have a right not to be discriminated against,72 yet at the same time, the insurance industry is based on some form of discrimination, which has often been argued as being justified in the nature of insurance.73 Frees and Huang describe it as follows:74

Insurers collect information on current and potential customers. They collect information about the customers themselves, the entity being insured (whether a person, organization, or physical object such [as] an auto or home), where the entity is located (that can vary, such as a person or auto), and parameters about the contract desired, among other things. This information, represented as variables or factors, provides the basis that insurers use to form groups and make decisions. By treating groups differently, they discriminate among them.

Such discriminatory practices may occur at different touchpoints in the insurance life cycle. First, it may occur in deciding whether or not insurance must be provided to the policyholder or not, which would be at issuing insurance or at the renewal of the insurance policy.75 Second, the insurer may limit the scope or application of the insurance.76 Finally, the price of premiums may also be adapted and changed depending on the data at the insurer’s disposal.77 Therefore, it can be said that the insurance industry is no stranger to the differential treatment of its customers, but unfair discrimination is different and must be avoided.78 Kuschke states that the very nature of insurance is based on discrimination.79 Further, she states that “persons who pose a higher risk or chance of loss pay higher premiums than others for the same insurance cover, and stereotypes are used to predict insurance risks. Whether this discrimination is fair or not, depends on constitutional norms”.80 What is considered justifiable in the insurance industry is continuously evolving. Take for example, in 2012 the European Court of Justice provided a preliminary ruling in the matter of Association belge des Consommateurs Test-Achats ASBL, Vann van Vugt, Charles Basselier v Conseil des ministres Case C-236/09 ECJ where it was ruled that the differentiation between men and women in the provision of insurance was an unacceptable practice.81

The use of big data and algorithms may perpetuate unlawful or unfair discrimination, or at the very least, highlight discriminatory practices that already exist within the insurance industry. The increased use of algorithms and big data in insurance carries the risk of unfair discrimination and potential financial exclusion of consumers.82 This is particularly the case when the differentiation between policyholders shifts from an economic decision-making tool to something more nefarious based on a person’s moral or social standing in society.83 Generally, anti-discrimination laws are concerned with formal or intentional discrimination. The existing laws do not provide much guidance on algorithmic decision making and that is because typical discrimination focuses on the human component of the decision. 84

5 2 South African legislative framework

As a starting point, the preamble of the Constitution of the Republic of South Africa 1996 notes that the purpose of the Constitution is to “heal the divisions of the past and establish a society based on democratic values, social justice, and fundamental human rights”. Therefore, all legislative engagements must reinforce this constitutional imperative. Section 9 of the Constitution also refers to the term “unfair discrimination” as opposed to some countries that only refer to the term “discrimination”.85 Unfair discrimination is described as when a person is treated differently as compared to other categories of people and therefore a person’s dignity as a human being is impaired by such treatment. Section 9(4) states that “no person may unfairly discriminate directly or indirectly against anyone on one or more [prohibited] grounds ...”. The prohibited grounds listed in the Constitution include race, gender, sex, pregnancy, ethnic or social origin, colour, sexual orientation, age, disability, religion, conscience, belief, culture, language, and birth, etc. 86 However, the Constitution does not directly apply to contractual engagements, but would, at best, only apply indirectly to insurance contracts by means of the application of public policy considerations. 87

The Promotion of Equality and Prevention of Unfair Discrimination Act (PEPUDA) gives effect to the Constitutional imperative that no person shall be unfairly discriminated against88, and prohibits unfair discrimination, harassment and hate speech. Section 14 of the PEPUDA states that in order to defend an action of unfair discrimination, the said discrimination should reasonably and justifiably differentiate between persons according to objectively determinable criteria, intrinsic to the activity concerned and that would amount to a valid defence against a claim of unfair discrimination. However, PEPUDA does not prohibit all forms of discrimination but rather only prohibits unfair discrimination. There are certain circumstances where discrimination can be regarded as fair.89 In determining whether discrimination is fair, account may be taken of whether the discrimination “reasonably and justifiably differentiates between persons according to objectively determinable criteria (for example, actuarial evidence), intrinsic to the activity concerned (for example, insurance business)”.90 However, there have been developments herein with the matter of Association belge des Consommateurs Test-Achats ASBL, Vann van Vugt, Charles Basselier v Conseil des ministres Case C-236/09 ECJ,91 that provided the preliminary ruling that differentiating between male and females in insurance would be considered to be gender discrimination. In the South African insurance industry, the differentiation between male and female is still considered to be an acceptable practice.

This notwithstanding, the South African insurance industry has always concerned itself with the fact that if an insurer can show with statistics, over a particular period of time, that the claim rations confirm an increase in and exposure to risk as a result of age, gender, location or other factors,92 premiums will be accordingly determined.93 It is trite that the differentiation in premiums is an actuarial science and does not meet the threshold of unfair discrimination.94 Differential treatment, in this specific context, is not viewed as unfair discrimination due to the actuarial data used in support of such decision making. However, big data may change this narrative, due to its capabilities which go beyond the current forms of risk determination and differentiation. There is, therefore, an increased risk of unfair discrimination of policyholders in the application of big data which requires regulation. There are generally three forms or areas in which discriminatory practices may develop in the use of big data in the insurance industry, being that of the policyholder, the price of the premium, and the data itself. These are discussed further in the paragraphs that follow.

5 3 First example of discrimination: Personalisation of offerings

The assessment of personalised insurance for policyholders stems from their individual risk profiles. Increasingly powerful algorithms may tempt insurers to collect a wider array of data about policyholders,95 and will give insurers insights into policyholder’s future conduct and the likelihood of a customer making a claim.96 The predictions in this context may result in hyper personalised insurance cover, although advantageous in certain ways, may result in some policyholders becoming completely ‘uninsurable’, based on the risk assessments offered by big data. Increased personalisation and targeting of insurance products can possibly lead to discriminatory results for consumers. A major problem arises when other types of data points could practicably act as proxies for possible discriminatory traits or could closely be correlated with protected characteristics such as race and/or gender97 and lead to proxy discrimination instead.98 It may also negatively impact the benefits for the policyholder in that of insurance risk pooling.99 The fact that AI systems can learn from data does not guarantee that their outputs will be free of human bias or discrimination, and there is plenty of evidence of AI systems picking up existing human biases or historic discrimination.100 This is clearly a potentially discriminatory practice,101 which may in certain instances meet the threshold of “unfair discrimination” and have the effect of stereotyping the policyholder and the policyholder’s lifestyle. Basing a decision to insure or not to insure a policyholder on this alone, without human intervention or the application of common sense, may feed into pre-existing bias and stereotypes in the application and use of such algorithms.

Unfortunately, most types of discrimination generated by big data cannot be dealt with using current anti-discrimination laws, as these laws traditionally focus on discrimination based on protected characteristics, such as race, by humans.102 Big data can use classes and categories for differentiation that do not directly relate to protected characteristics.103 There will be a greater need for regulatory safeguards if the risk of proxy discrimination rises.104 Supervisors may need, for example, powers to eliminate the use of certain data points that are unnecessary or could be potential sources of biases, and regularly audit algorithms in order to detect potentially unlawful discriminatory outcomes.105 Furthermore, insurers should thoroughly test big data algorithms prior to their release to ensure that there are no possible discriminatory results. 106

5 4 Second example of discrimination: Personalisation of pricing

Frees and Huang explain that the determination of insurance pricing stems from the 17th-century theory of randomness107 and that fairness in this context was based on a “framework of individual equitable contracts, ones that traded a certain present amount for an uncertain future value”.108 This theory of randomness became the backbone to insurance pricing in the 19th century,109 wherein a premium had to be paid for “compensation for an uncertain future event”.110 This can be described as insurance pooling and a way to address uncertainties and losses for insurers.111 Risks associated with the insurance pools depend on the nature of the policyholders and may impact the premium and insurance costs accordingly.112 Discriminatory price optimisation practices can occur through the employment of Big Data as they will allow insurers to charge more individualised prices.113 Therefore, insurers relying on big data analytics will be able to charge differential prices to groups of consumers, which will then lead to more personalised pricing based on the behavioural characteristics and personal data of consumers.114

While this type of personalised pricing may bring benefits to certain policyholders, there is developing evidence of pricing practices in the insurance industry that fail to treat policyholders fairly, because insurers are partaking in harmful price optimisation techniques.115

Evidence suggests that insurers may “charge prices based on the optimum amount of margin they can earn from an individual policyholder, rather than the risk and/or cost of the individual policyholder”.116 The reason for this is that insurers are increasingly relying on big data analytics to better understand a policyholder’s individual price sensitivity as well as their likelihood to shop around or switch insurance at the point of renewal.117 This is clearly not the correct way of premium pricing in insurance, as it does not focus on the risk of the policyholders but rather their individual propensity to switch insurance. 118

These examples of unfair pricing practices must be circumvented and to do this it is proposed that there be a ban on unfair price optimisation practices when selling insurance products to consumers.119 Insurers must set premium prices based on the individuals’ risk assessment and should not be setting prices based on consumers’ individual price sensitivity or their likelihood to switch insurance contracts.120 In order to achieve fair outcomes in this regard, insurers should be required to publish information about the price differentials between their customers for transparency purposes. This could have the effect of increasing competitive pressures amongst insurers and ensure public and supervisory inspection to ensure that the pricing practices of insurers are fair towards consumers. 121

5 5 Third example of discrimination: Data granularity

Data granularity refers to the level of detail considered in a decision-making process or represented in an analysis report, therefore, the higher the granularity, the deeper the level of detail.122 Consequently, granular customisation may provide additional product availability and accessibility to policyholders,123 but at the same time, some policyholders may find that they are not being offered specific types of cover at all, or that they are exposed to high non-risk related individual premiums and possibly discriminatory claims settlement decisions.124 This may negatively impact the availability or even the affordability of insurance for certain parts of the population.125 Further, this may result in diminished levels of consumer confidence in the insurance industry.126 Insurance supervisors may also need to consider whether certain algorithmic practices may possibly undermine the underlying characteristic of risk pooling in insurance. 127

A core feature of private insurance policies is that customer segmentation occurs based on risk.128 This means that higher-risk policyholders usually pay higher premiums compared to policyholders who are viewed as being lower-risk.129 The proliferation of big data now allows insurers to consider a wider array of personal and behavioural data and charge corresponding premiums. Previously, risk-based pricing was based on a limited number of easily identifiable criteria.130 This brings in new potential for discrimination as the use of AI may allow insurers to easily identify high-risk characteristics and result in categories of consumers no longer able to access or afford insurance coverage.131

Profiling the policyholder is increasingly used then it could reduce the availability, access, and affordability of insurance.132 This type of risk segmentation will allow insurers to cherry-pick ‘good risk’ from ‘bad risk’ policyholders which will then ultimately lead to increased pricing differentiation between low- and high-risk consumers.133

Further to that, if personalised insurance products increase then there is serious concern that it may undo the traditional principle of “solidarity” or “risk pooling” that has always been at the core of insurance.134 If insurers calculate every individual’s personal risk and corresponding premium, insurance companies would no longer be spreading out the risk collectively between policyholders.135 Hyper personalised risk assessments could leave certain individuals ‘uninsurable’ and lead to new forms of financial exclusion in the future.136

In Australia, for example, there is evidence emerging of more granular segmentation by insurers which is resulting in higher premiums and even potential financial exclusion of consumers.137 The Competition and Consumer Commission found evidence that “[o]ver the past decade, insurers’ methodologies for pricing insurance have become much more sophisticated and combined with access to better data, we have seen a shift towards more address-based risk assessment and pricing. As a result, insurance premiums are increasing, especially for those in high-risk areas”.138 One of the findings of the Inquiry was that "more granular pricing approaches, in particular address-based risk assessment, has been a key contributor to increased premiums for many policyholders.”139

It will be necessary to have stronger oversight mechanisms in place when it comes to the use of personal data by insurers in terms of what they should be allowed to consider when selling insurance contracts and setting premiums.140 Regulators will need to monitor whether the use of big data analytics has an impact on the insurability of high-risk consumers and may need to intervene to ensure people continue to have adequate access to insurance policies. 141

The above discussion is evidence of how big data utilisation in insurance may result in varied discriminatory practices for policyholders. Furthermore, ethical considerations also need to be investigated as big data use will no doubt have certain ethical consequences.

6 1 Introductory thoughts

The issue of discriminatory practices in the insurance industry talks to risk classification and the fine line between differentiating between policyholders for sound economic reasons and attempting to punish or deal with concepts such as social structures within a country.142 The question, as highlighted in the preceding paragraph, is whether such discriminatory practices are justifiable in such circumstances. Frees and Haung point out that the risk classification in insurance has been argued to be morally appropriate based on the actuarial sciences that underscore such differentiation.143 In fact, the issue of morality becomes, in addition to the question of discrimination, a central feature in assessing the use of big data in the insurance industry. There are two sides to the issue of morality and ethics in the use of big data in the insurance industry. On the one hand, the use and processing of big data is to assess the insurability of the policyholder to determine whether the insurer is willing to provide insurance in the circumstances. This information links to, among other things, the moral risks of the policyholder which talks directly to the policyholder’s character and integrity.144 The use of big data in this context may result in discriminatory practices within the insurance industry, which was discussed in the preceding paragraphs. The second aspect relates to the morality and ethics in the pre-contractual engagement with policyholders and the ultimate use of big data in contractual obligations.

Conduct that is legally permissive is not necessarily ethically permissive. Therefore, the role of big data may very well be regulated within the insurance industry through legislative and regulatory interventions, it does not necessarily consider the ethical considerations of the use of big data in the insurance industry. Hinman notes that “[e]very society has its set of moral rules or guidelines that establish the boundaries of acceptable behavior, that draw the line between good and evil”.145 Hinman explains that these moral rules may be to protect people from harm or be related to issues of respect146 and that the concept of ethics is the process of refining our moral beliefs.147 The question then is what are those acceptable societal boundaries when considering the use of big data in insurance contracts?

6 2 Public policy in contracts

Central to the validity and enforcement of contracts generally is the concept of public policy and our views and understanding of public policy is, according to Southwood AJA, constantly evolving.148 Like with public policy, big data is continuously developing and there may not be any easy answers to the questions surrounding the use of big data. Our courts have noted that in novel situations (specifically that of the duties of one of the contracting parties) one should turn to public policy for guidance.149 Therefore, public policy may provide guidance both in assessing the validity of a contractual provision150 and in establishing whether there are implied duties because of public policy considerations.151 Yet, there is no clear guidance as of yet on whether public policy may be used to intervene where big data is used in determining whether a person may wish to conclude a contract.

Public policy has generally been considered to be part of the post-contractual phase e.g., determining whether a contractual obligation is valid and whether such an obligation would be enforced. The question is whether ethics and/or morality play a role in public policy considerations. Articulating exactly what public policy may be in a particular circumstance is often difficult.152 To determine what public policy may be, Joubert AJ notes that one should take into account “considerations of justice, equity, good faith, reasonableness, common sense and the like” in particular circumstances.153 Public policy has also been found “that the concept of ubuntu and the necessity to do simple justice between individuals have been recognised as informing public policy in a contractual context”.154 However, Sachs J (referencing Wessels) notes that public policy is linked to the “moral sense of the community”,155 also called “public morality”.156 More importantly, public policy is intrinsically linked to our Constitutional values.157 Essentially, public policy forms the basis of the enforcement of contractual provisions, and may on occasion include the question of morality.158 The issue of using big data in insurance contracts may require further assessment of whether ethics and morality actually play a role in insurance contracts.

6 3 Morality as an integral part of contracts

Ethics is generally considered to be a roadmap as to what is considered acceptable (or good) and what is considered to be unacceptable (bad) in society. Often the element of ethics is also linked to that of morality. Herein lies the challenge, as what is considered good or bad is often associated with the personal reflections and convictions of an individual and one would think that this has no place in the consideration of the law, especially something as objective as contractual engagements. Yet, our courts have recognised the possibility of morality in the entire life cycle of a contract, and this should be no different from insurance contracts.

The term ethics or morality is rarely used in the discussion of pre-contractual engagements. Rather, the terminology of “good faith” is used. Although the principle of good faith negotiations is acceptable and enforceable in South African law,159 contracts of insurance are generally not negotiated. This does not conceptually preclude the requirement from the parties to “relate to each other in good faith”.160 Furthermore, the information used to assess whether insurance will be provided to the policyholder is often associated with moral risk and hazards of the policyholder. It can then be said ethics and morality play a role in the pre-contractual engagements in the contract lifecycle.

Nothing seemingly prevents the contracting parties from agreeing, either expressly or impliedly to a moral and a legal duty in a contract.161 However, it is necessary to note that duties cannot simply be implied into a contract by means of concepts such as ubuntu or good faith.162 Another factor that is of importance in the insurance industry is the way big data is translated into standard-form contracts. Although the type of cover, the extent of the cover, and the policyholder may be individualised in an insurance contract, many insurance contracts are effectively standard form contracts that are used not only by one insurer but by the insurance industry as a whole. This may also be prejudicial to the policyholder. Ngcobo J in Barkhuizen v Napier, links morality to the philosophical foundation of contracts in noting that: 163

In my view, to treat mass-produced script as sanctified legal Scripture is to perpetuate something hollow and to dishonour the moral and philosophical foundation of contract law. It certainly does not promote the spirit of openness central to our new constitutional order.

Our courts have also linked the enforcement of contracts to the concept of morality on more than one occasion. Ngcobo J, in the Constitutional Court case of Barkhuizen v Napier, also noted that:164

Pacta sunt servanda is a profoundly moral principle, on which the coherence of any society relies. It is also a universally recognised legal principle. But, the general rule that agreements must be honoured cannot apply to immoral agreements which violate public policy.

Herein, Ngcobo J links the enforcement of a contract on the basis of the sanctity of a contract to a moral principle but at the same time notes that sanctity of contracts cannot be applied to immoral agreements.165 In these instances, contracts and that of their enforcement is linked to the morality of the matter, in other words, what is considered right or wrong in society in some way, shape or form. Although dealing with illegal contracts, the par delictum rule is worth mentioning as it will not allow for restitution of performance in an illegal contract wherein the contracting parties are equally morally guilty,166 and a court will not enforce “a claim arising from a transaction which is base[d] in the sense that it violates morality”. 167

Therefore, big data must not only be used in such a way as not to contravene anti-discriminatory legislation (as discussed in paragraph 3 above) but big data must also be used in an ethical and moral manner in all areas of the contract lifecycle.

6 4 Ethical considerations in the use of big data

Big data and the ethical consequences raised through the use thereof is a complex issue. Ethical questions require cognitive work with a social dimension and these types of questions have in the past been performed by humans.168 Big data analytics is the product of human choices, and they will affect human values.169 Cultivating ethical sensibilities in big data is key in order to ensure that the use of big data is ethical and not unfair.170 Therefore, designing the algorithms will require certain ethical considerations.

One important ethical consideration for the use of big data analytics in the insurance industry is the idea of “explainable AI” which encourages the notion that AI also needs to be able to respond to questions on “why” it has reached certain decisions.171 This will encourage a certain amount of transparency in the decision-making process in order to understand why the algorithms have reached a certain decision. This can also enhance the fair treatment of consumers.

Notably, the OECD has put forward ethical considerations which should be considered by insurers utilising big data analytics. Some of the most important considerations raised by the institution include:

• human agency and oversight: meaning AI systems should empower human beings, allowing them to make informed decisions and fostering their fundamental rights;172
• technical robustness and safety: AI systems need to be resilient and secure, which means they need to be safe, ensuring a contingency plan is in place in the case of something going wrong, as well as being accurate and reliable;
• algorithms should ensure privacy and data governance: data protection and adequate data governance mechanisms must be guaranteed, taking into account the quality and integrity of the data, and ensuring legitimised access to data;
• transparency: the data, system, and AI business models should be transparent, there should be traceability mechanisms in place as well as the fact that AI systems and their decisions should be explained to their human counterparts. 173

The OECD emphasises diversity, non-discrimination, and fairness as ethical considerations.174 They state that unfair bias must be avoided, as it could have multiple negative implications, from the marginalisation of vulnerable groups to the exacerbation of prejudice and discrimination.175 It is thus clear that the potential for discrimination through the use of big data is an ethical issue that must be mitigated and managed by insurers and the developers of the algorithms.

A further consideration is the insurers’ ability to repudiate the insurance contract based on incorrect information provided. In the past, the duty of disclosure rested on the policyholder as all the information relevant to the risk and for purposes of risk assessment is seen as being exclusively within the domain and knowledge of the policyholder and the insurer has no means to obtain such information itself. The underlying rationale in this regard is that the underwriter is deemed to know nothing and is solely dependent on the information supplied by the policyholder. This “traditional” approach presupposes that the policyholder is in possession of all material facts and information, and overlooks the emerging realities of big data which has the potential to place insurers in a more favourable position than their policyholders and which may result in informational asymmetry.

The use of big data and advanced analytics places insurers in a completely different position. Insurers are no longer limited to adopting a “passive” approach and can now place reliance on data obtained utilising technical capabilities or from third-party service providers such as social media. Insurers are able to obtain, mine, and analyse large sets of data for their own benefit and are thus able to gain an upper hand on their policyholders. This raises a second ethical consideration, the use of big data analytics in the pre-contractual stage and whether insurers would be able to repudiate a policy due to a non-disclosure if there was an error in the algorithm. Insurers would be required to develop certain checks and balances to ensure that the design of the algorithms is correct and that they are asking for the required information from policyholders. Insurers should only be allowed to repudiate a policy based on the current non-disclosure rules and should not be allowed to repudiate a policy based on inaccurate and incorrect algorithm designs.

Currently, there is no legislative or regulatory oversight of these ethical issues in the insurance industry in South Africa, and big data is left generally unregulated - exposing policyholders to risk and unreasonable contracting practices. Herein, we suggest that these ethical matters require consideration by regulators and insurers as the possibilities for discrimination in these areas are rife.