1. You have conducted a study into car damage fraud with AI-generated fake images. Why?
"The simple answer is: because it's a super interesting subject. Many parties are mainly concerned with the positive side of AI, but of course there are also risks. For example, AI can be used to commit fraud, for example by submitting a fake claim to an insurer. At the same time, we wanted to use our research to show that such fake damage is much more difficult to recognise than you initially think."
2. Is that also one of the most important findings of your research: that fake damage is so difficult to recognise?
"As far as I'm concerned, the most important outcome is that on the one hand it is very easy to make a fake image of a car damage, while on the other hand it is very difficult to recognise such an image. We have focused our research mainly on recognising fake damages in a claim, but the reverse can of course also occur. A fraudster can make a car with serious damage look like new 'for a while', then apply for all-risk coverage and claim the damage later. Perhaps that acceptance risk in the Dutch market is an even greater risk. In any case, we always emphasise that insurers (and other fraud detection agencies) must set up a complete framework. This consists not only of transparent processes, but also of the right tooling. One tool does not recognise everything, as the research has shown. Moreover, AI is and remains a tool. In other words, you can't blindly trust it. A good example is a photo of a broken side mirror of a car. In the reflection you can see that the mirror is still intact. This is relatively easy to see for a human, but it may be more difficult for AI tooling. This applies to more things, by the way: some signals people recognise easily, while AI tooling has more difficulty with them. Conversely, AI can recognise some patterns that we as humans do not see. So it is mainly the combination that makes the difference."
Summer series!
AI is hot. Insurers are experimenting with all kinds of AI. But, so do their customers. For example, when submitting a claim for damages. Milliman consultants have investigated car damage fraud with AI-generated fake images. Their conclusion? Making such images is "very easy".
In this first part of our summer series on AI: a curse or a blessing, Niels van der Laan, Partner at Milliman Benelux and speaker at the Association Day of the Dutch Association of Insurers, discusses the most important results of the research. In addition, he gives tips on how to prevent fraud with AI.
3. In the survey, you made a distinction between young and old(er) people and also between people with a lot or little/no experience with AI and fraud detection. Does that still result in major differences?
"No, we didn't see any significant differences in how well or badly people took the quiz. All participants were shown twenty images and had to indicate which were fake and which were real. You have to think of images of a flat tire, hail damage and a broken bumper to collision damage and a broken mirror. The average score (of correct answers) is 46 percent. That is even lower than you would expect based on guessing (50-50)."
4. 'Human detection is inadequate', you conclude in the research report. What do you mean by that? Should AI be used to combat fraud with AI?
"Fraud investigators can certainly do that, but in addition. Just as we emphasise that you can't just rely on one tool, you also need people to detect fraud. In fact, the role of humans remains essential. If you ask ChatGPT a question, you always have to check whether the answer is correct. So you have to keep thinking consciously. I recommend that here, in the complete fraud process, too. The human in the loop is essential. AI certainly has added value, but humans must always be aware of the risks and make the final decisions."
5. Fraud with car insurance is hopefully still limited, because an insured person is often referred to a specific garage by his insurer. What about travel and home contents insurers, for example? How can they arm themselves against AI fraud?
"The chain is indeed different with travel and home insurers. Sometimes uploading a photo or a receipt is sufficient for a damage report. Then a fake image is not only easy to generate, but also simple to use in the process. Insurers can train an AI agent to recognise such fake images as well as possible. They can also purchase existing tools. A lot is already possible, but insurers must above all think consciously about their processes. This new form of fraud is here. The question is: how do you set up your processes in such a way that the risk of fraud is minimised? That can sometimes be very simple. For example, in the event of damage, ask for a video. Nowadays, you can even use Telematics to read (with data from the car) how damage occurred. You can also create an app that allows policyholders to take a photo in real time . There are many tools and tricks to reduce the risk, but the most important message is: the risk is there, be aware of it and think about how you (re)organise the chain."
Photo Niels van der Laan: Ivar Pel