The use of AI in insurance has the potential to yield economic and societal benefits that go beyond insurers and their customers by improving risk pooling and enhancing risk reduction, mitigation and prevention. In order to foster the adoption of AI systems and realise these benefits, insurers need to earn the trust of customers by using new technologies responsibly. The Geneva Association’s Benno Keller provides useful insights.
Since the end of the ‘AI winter’, the period from the 1970s to the end of the 1990s characterised by setbacks and disappointment, AI has made remarkable progress. Today, many insurers are rolling out intelligent systems that automate routine tasks or assist human decision-making along the entire insurance value chain. Such systems combine new types of learning algorithms with the analysis of data from new sources, such as online media and the internet of things (IoT).
Intelligent systems excel in detecting patterns and correlations in complex data in ways that humans find difficult or even impossible. Such identified patterns are the basis for analytical tasks such as classification, regression and clustering that play an important role in the insurance business model. Compared to traditional modelling approaches used in insurance (such as generalised linear models), intelligent systems have the potential to provide much more accurate predictions, because they can learn complex non-linear relationships between variables.
Further progress in learning algorithms will enable these systems autonomously to make standardised decisions under human supervision in a growing number of areas.
In order to foster the adoption of AI systems and realise the benefits, insurers need to earn the trust of their customers by using the new technology responsibly. AI can also help insurers to enhance their role in risk reduction, mitigation and prevention.
Principles for responsible AI
Over the past few years, an intense and ongoing debate has emerged on what the responsible use of AI entails, and various governmental and non-governmental actors have issued guidelines for its ethical use. An analysis of these guidelines reveals a global convergence towards five core principles for responsible AI: Transparency and explainability, fairness, safety, accountability and privacy.
Considerable uncertainty remains, however, regarding how ethical principles and guidelines should be implemented in a specific context.
In insurance, these core principles have long played an important role. As a matter of fact, various laws and regulations - including insurance law, privacy and data protection laws, anti-discrimination laws and supervisory requirements - govern the fair, transparent and accountable behaviour of insurers as well as the protection of privacy.
Nevertheless, the use of AI raises some intricate questions for insurers. Which trade-offs arise with the implementation of core principles? How can insurers foster and demonstrate adherence to such principles? What changes, if any, are necessary to existing governance mechanisms and risk management frameworks for this purpose? Insurance regulators, too, are increasingly considering these questions.
The Geneva Association’s report, ‘Promoting responsible artificial intelligence in insurance’, analyses important ethical guidelines for the use of AI and explores how to implement them in the context of insurance, taking into account current developments in computer science on fair machine learning. A definitive answer to these questions will remain elusive.
Transparency and explainability
Transparency and explainability are important for building trust with customers and other stakeholders. Some guidelines mention the importance of transparency and explainability to enable individuals to seek redress against decisions affecting them. Providing an explanation is particularly important when a decision has a significant impact on the affected individual. Therefore, the degree to which explainability is needed highly depends on the context and severity of the consequences of an erroneous or otherwise inaccurate output. Interpretability of algorithmic outcomes is also an indispensable prerequisite for the assessment of the performance of AI systems and their continuous improvement and thus for sound data science.
Providing meaningful explanations is a challenge, as some ‘black box’ algorithms are by nature complex - the price to pay for better accuracy - and therefore difficult to interpret and explain. In recent years, considerable efforts have been undertaken in computer science to overcome the challenges of interpreting and explaining ‘black box’ algorithms. Reverse engineering approaches, for example, consist of building interpretable algorithmic surrogates, a recent technique which needs to be understood better. Design approaches rely on imposing certain constraints on the predictions.
When it is not possible to provide an explanation of the role of different variables in a given decision, other approaches, such as certification of the AI system by an independent body, may be used to foster trust. Nevertheless, the implementation of interpretable models should be encouraged, in particular if their outcomes have a significant impact on customers. When used for risk selection and pricing, trust in AI systems can be fostered by using data sources that are related to the insured risk in a way which is intuitively understandable to customers. The benefits of overly complex models may not always justify their reduced interpretability.
How to ensure fairness is particularly intricate as there are several mutually exclusive definitions of a fair decision. Fairness is usually associated with many different values such as freedom, dignity, autonomy, privacy, non-discrimination, equality and diversity, among others. These values often need to be interpreted in context, including the cultural context. It is therefore impossible to provide a universal standard of fairness.
Decisions should be fair in the sense that they do not discriminate and disadvantage individuals or groups of individuals. To eliminate or minimise unfair bias in AI-driven decisions is thus a core requirement in ethical guidelines.
Humans are not free from bias, and the use of AI may actually enhance fairness of decisions in some circumstances. However, if deployed at scale, even a minimal bias in an AI system can affect large numbers of individuals. For data scientists, the challenge is therefore to identify, measure and mitigate potential bias that could put certain groups at a systematic disadvantage.
In order to promote the responsible use of AI within their organisations, insurers should consider the following three actions:
Establish internal guidelines and policies for the use of AI
Internal guidelines and policies play an important role in raising awareness of the benefit–risk trade-offs in the use of AI in insurance. Insurers should therefore develop and adopt relevant guidelines and policies that include principles for dealing with issues of transparency and explainability and of fairness. In particular, guidelines should help to clarify how the benefits and risks of using AI should be assessed on a case-by-case basis. Actuaries, risk managers, data scientists and data protection officers should closely cooperate in the development and implementation of such guidelines and policies.
In doing so, insurers may adopt a risk-based approach to the governance of AI, implying a special focus on the uses of AI systems that may have a significant impact on individuals. The significance of impact refers to the consequences decisions have on affected individuals and depends on the specific circumstance of the use of AI.
Adopt appropriate governance structures
The responsible use of AI in insurance requires governance structures that hold decision makers - individuals, committees and departments - accountable. Decision makers should have the necessary competencies, skills and expertise, and organisations should have in place effective processes, including triggers and escalation.
With many different governance models possible, each with its own advantages and disadvantages, the optimal organisational model will depend on a company’s structure and culture.
Develop and roll out internal training programmes
Finally, ensuring a responsible use of AI requires awareness of related benefits and risks across different functions and managerial levels. In order to raise awareness, insurers should consider developing and rolling out comprehensive training programmes on the benefits and risks of AI and their respective internal guidelines and policies. Such training programmes would ideally target employees at different management levels and from various internal functions involved in decision-making, including agents and other customer-facing employees. A
Mr Benno Keller is special adviser, digital and innovation for the Geneva Association.