Just two months after its launch in November 2022, the consumer generative AI ChatGPT was estimated to have reached 100 million monthly active users, making it the fast-growing consumer application in history. But what’s behind its ongoing rapid advancement and how can organisations overcome concern about the risk management implications of these tools?
In conversation with Insurance Business, Michael Berger (pictured), head of Insure AI at Munich Re, noted how quickly generative AI (GenAI) reached its wide audience and how it’s viewed as one of the largest disruptors in recent years.
“It is able to increase productivity, efficiency, decision quality and even creativity to a new level and innovate beyond traditional business models or demands,” he said. “GenAI helps to democratise technology at high speed – for the good and evil, as one can see in the cyber threat landscape. All of this is true via all sectors and regions at once and will further develop.”
Munich Re’s Insure AI team works to structure insurance solutions for AI risks, he said, including the risk of wrong predictions, discrimination or hallucination. Its first insurance solution for an AI application was created in 2018, covering the financial losses stemming from AI underperformance, followed in 2019 by the debut cover for a Large Language Model.
“Insuring generative AI is the current development field for the Munich Re AI team, exploring solutions to cover the various ways AI potentially and randomly goes wrong,” he said. “Munich Re provides protection to AI model developers as well as AI model users, this with limits from €5 million up to €50 million per AI model.”
Given its capacity to produce assets like images or text, make unstructured data accessible, enable AI access for a layman, unlock new business opportunities and drive advancements across the organisation, he said, both GenAI’s power, but also the inherent risks, need to be examined, including risk transfer solutions.
“We believe that insurance will become vital for a smooth and widespread adoption of GenAI and management of emerging AI risks over the years to come,” he said. “Collaboration of the insurance and tech industries can aid in unlocking the tremendous potential of GenAI for all.”
Looking across the wider market, Berger noted that insurers are still in the early stages when it comes to fully embracing the possibilities of advanced AI technologies and AI models. Munich Re is one of the few insurance companies to actually offer a dedicated AI insurance product.
“There are certainly other players who are looking into this area and will be familiar with our initial products,” he said. “We see that the interest in participating in this attractive growth area is increasing, just as the demand for risk transfer will continue to grow enormously.
“For Munich Re at least, this is an exciting business opportunity. And my team is able to insure AI error risk across lines and industry sectors and build a business process as an insurer or reinsurer by basically covering this type of risk.”
Acknowledgement of the opportunities presented by GenAI is being met with increased recognition of the risks it could pose, and Berger identified some of the “enormous new risks”. For instance, he said, GenAI produces misinformation or inaccurate output, a phenomenon widely discussed as "hallucination". And even the best GenAI models will occasionally produce these types of errors.
Another risk is copyright infringement in the sense that the output is too close to a copyrighted asset that was used in the model’s training data. There is also the risk of discrimination against certain groups in the results, e.g. by over-representing a group with negative associations. There are also other risks, he said, such as environmental concerns related to the need for huge computing resources to train and update the model, as well as the impact in terms of high energy and water consumption.
Effective risk management strategies will be critical in mitigating these risks, he said, as it is vital for model firms and model users to be aware of the special risks accompanied with GenAI, as well as if it is of priority for management to know how to mitigate and transfer them.
"Both have to rely on the results of the AI model they are producing or adapting,” he said. “That is where insurers come into play with their risk management offering which provides protection in the event of defined errors. So how does Munich Re tackle the challenge? In principle, every AI system, including every generative AI system, is a probabilistic system.
“It is technically unavoidable that even if you build the most perfect AI or generative AI model, there will always be a certain probability that the AI will make mistakes. This residual risk, this residual error probability, is basically the risk we actually take after diligently analysing it. This is, in a nutshell, what our solution approach is all about.”
