As artificial intelligence (AI) becomes more high profile, insurance practitioners are taking note of the potential for the technology, but also the potential for AI bias.
Insurance regulators have been concerned about the dangers posed by AI for some time. In January 2019 the New York Department of Financial Service (NYDFS) put out a circular on the: ‘Use of External Consumer Data and Information Sources in Underwriting for Life Insurance’.
The circular was issued following a departmental investigation which highlighted two major problems with insurers using AI to collate external data sources for their underwriting process: the potential for unlawful discrimination, and a lack of consumer transparency.
Pointing to two existing laws which prohibited insurers from discriminating against customers the NYDFS said: “Based on its investigation, the Department has determined that insurers’ use of external data sources in underwriting has the strong potential to mask the forms of discrimination prohibited by these laws.”
It is not just New York State which is looking to regulate insurers’ use of AI, in 2022 the Connecticut Insurance Department, and its California equivalent, inked bulletins on AI and racial bias, while the US National Association of Insurance Commissioners (NAIC) has formed a working group on Big Data and Artificial Intelligence.
The NAIC task force has yet to release any findings but individual states are already taking the initiative. In 2021 the Colorado legislature passed a law requiring that insurers establish a framework to ensure that industry’s use of AI is not discriminatory. In February this year a draft framework was issued, which critically moved beyond a principles-based supervisory approach.
It is easy to see why regulators are concerned about the potential for AI powered large language models (LLMs), such as ChatGPT, to inculcate ethnic, and other biases, into insurers’ policies if they are used to synthesise the huge amounts of historical data the industry holds.
The litigation risk posed by AI became real in December 2022 when US life and P&C insurer State Farm, found itself on the wrong end of a data-based class action lawsuit, which alleges that Black American householders’ claims were subject to greater scrutiny than their White peers — in other words exactly what the NYDFS had warned about three years earlier.
The heart of the problem is the disparity in data held by US life insurers on different ethnic groups and differences in the types of policies they were sold in the past which makes aggregating the data via LLMs problematic.
Karl Ricanek, CEO of Lapetus Solutions, which provides AI products such as facial recognition software to the insurance sector, says these disparities will inevitably result in biased results if AI is used to analyse it.
“If an insurer’s data underrepresents the number of African Americans, or Hispanics, then the primary data will invariably overshadow it. Unless firms take specific action to augment their existing data the results won’t be accurate. And then the question is, how do you get that data? And how are they utilising that data?”
Ricanek says one solution would be for insurers to pool their existing statistics in order to gain a more holistic output from LLMs. There are precedents, both globally, and in the US, for this type of action. Since 2005, ORIC International has been providing pooled and anonymised data to help insurers manage their operational risk exposure, while the Lapetus CEO points to the US’s MIB (previously known as the Medical Information Bureau) which operates on a similar model.
The MIB is a pooled resource which enables US life insurers to cross-check life insurance applications to detect potential fraud, and Ricanek says that a corresponding approach could help ameliorate the issue of AI-linked discrimination.
Adding the caveat that this is not a perfect solution, the CEO says that with each carrier serving a slightly different demographic pooling data will result in a more accurate picture.
“Each insurer has a population that their products resonate with so if you pull all this data together then it’s possible to capture the largest swathe of policyholders. If you want to understand how to really use this technology, you have to involve different data points from all the subpopulations.”
Ricanek’s comments are in line with Keith Raymond, Principal North America Analyst at Celent, who says the issue is not AI technology itself, but instead putting in place the correct guidelines to manage it.
Raymond points to the already extensive use of chatbots by insurers’ customer services departments as an example of the technology prevalence in the industry, and he says that the issue now is how to manage AI.
“Insurers are already using AI. It’s just a matter of: ‘This is the new shiny toy’, and ‘What are the implications of using it?’. Do we have to put additional guardrails in place? Because the whole issue about ethics and bias in AI has been around for a long time; it’s nothing new. Instead the question is does this add a level of risk that requires additional governance structures?”
According to Raymond, insurers need to develop the enterprise data architecture that will enable them to overcome the issue of bias inherent in their data sets and enable them to gain value from using LLMs.
“An insurer may have over a dozen policy administration systems that it wants to bring data from into an AI model. So having the right enterprise data architecture is a key component of success in moving forward with AI utilisation.”
The analyst says developing this kind of data architecture is a long term process and therefore there is no second mover advantage from insurers waiting to see how other firms tackle this issue. Instead he says that developing a successful data management system, which overcomes the bias issues associated with LLMs could provide significant business advantages.
“If an insurer already has the architecture in place, it will be easier to build an AI based LLM model using legacy data, making it possible to identify trends and risk patterns which can impact how future products are designed.”