Artificial Intelligence is now a familiar concept but has only recently gained widespread public attention, mainly because of such software as ChatGPT, the chatbot developed by OpenAI.
Three developments have led to the impressive performance of modern AI systems. First, the amount of data and its complexity have increased as part of the seemingly inevitable and endless digital transformation.
Unsupervised learning can go further and even detect previously unknown fraud patterns.More and more complex data is now digitally available, which requires algorithms that can utilise this data effectively.
Accordingly, and second, more complex algorithms have been developed.
Third, computational power has increased substantially, enhancing the speed of these algorithms and fostering their widespread implementation.
An award-winning dissertation by Jörn Debener of the University of Münster in Germany has investigated how financial institutions can benefit from these incredibly powerful, modern AI systems, and come up with some significant findings.
The high cost of fraud
The aspect of the dissertation that is most accessible and directly relevant to practice is a specific research project on how AI algorithms can help to identify insurance claim fraud.
Such fraud is a serious, often alarming, and in some ways insidious problem for insurance companies. Because many fraudulent claims remain undetected, it is difficult to estimate the total damages.
A rough approximation is provided by Insurance Europe, the European insurance and reinsurance federation, which estimated the total damages caused by both detected and undetected claim fraud in Europe in 2017 at approximately €13bn (£11.1bn). More recent figures from the US confirm the huge sums associated with such activities.
Consequently, insurance companies benefit greatly from detecting claim fraud. However, such detection generally requires painstaking investigations by fraud experts, which is very time-consuming and cost-intensive.
Enabling fraud experts to focus their investigations only on highly suspicious claims therefore helps insurance companies substantially in dealing with the problem.
And this is where AI, and specifically machine learning, come into play, because these methods have the potential to efficiently identify suspicious claims that should be investigated in detail.
Various methodological machine-learning options are available for identifying fraudulent claims. Some algorithms “learn” fraud patterns from previously identified fraudulent claims. These algorithms are from the domain of supervised machine learning.
By contrast, other algorithms learn to identify anomalous claims without requiring a categorisation of past claims into fraudulent and non-fraudulent. These algorithms are from the domain of unsupervised machine learning.
To date, unsupervised machine learning has mostly been overlooked for fraud detection, but in fact offers an important advantage.
Since it does not rely on the categorisation of past claims – a task that requires a lot of resources and is prone to errors – it is less susceptible to relearning fraud patterns that are already known by the insurance company, and can thus potentially detect fraud patterns that have remained undetected in the past. This is clearly an enormous potential gain.
To test whether and to what extent both supervised and unsupervised machine learning approaches can identify claim fraud, Debener and his co-authors first analysed the capability of several algorithms to detect fraud in categorised past automobile insurance claims.
This is well known as an insurance segment in which fraud is particularly common.
Both approaches yielded promising results, indicating that AI in the form of machine learning is generally capable of identifying fraudulent claims, at minimum those that had been detected in the past.
In a second step, the authors then conducted a field experiment with an insurance company to examine the even more intriguing question of whether machine learning can identify previously undetected fraud patterns.
To this end, the insurance companies’ fraud experts (real people) investigated past claims that had been categorised as non-fraudulent by the existing (non-machine-learning) fraud detection mechanisms of the insurance company, but had nonetheless been marked as anomalous by the supervised and unsupervised machine-learning algorithms.
Interestingly, both supervised and unsupervised machine-learning algorithms managed to find a substantial number of claims that had not been investigated before settlement, but were judged as highly suspicious by the fraud experts after reinvestigation by the insurance company.
A powerful incentive
The finding that AI is able not only to detect insurance fraud but also to detect fraudulent claims that have been overlooked by existing, more conventional fraud detection mechanisms, thus certainly provides a powerful incentive for insurance companies to make use of modern AI systems.
A closer look at the identified highly suspicious claims then revealed that the claims identified by unsupervised machine learning differed to a meaningful extent from those that had been identified by supervised machine learning.
Debener and his co-authors suspect that supervised machine learning detects known fraud patterns more rigorously than the existing more conventional fraud detection mechanisms, so that these claims have a lower chance of slipping through, but unsupervised learning can go further and even detect previously unknown fraud patterns, which is a major additional benefit.
In short, the investigation suggests supervised and unsupervised machine learning algorithms act as complements rather than substitutes, and a combination should be considered for detecting insurance fraud.
The two together may well usher in a new era of fraud detection, which should provide a range of benefits for the industry and its legitimate clients.
Brian Bloch is a freelance journalist based in Germany
