Eliminating data bias with fuzzy logic

AI offers insurers many benefits, but one downside could be unintended bias in how claims are assessed based on historical outcomes.

Simply removing variables concerning race, gender, socioeconomic, geodemographic or other attributes is not enough to eliminate historical bias. (Photo: ipopba/Adobe Stock)

Due to significant industry challenges which have been considerably accelerated over the last 18 months, such as emerging fraud and increasingly sophisticated schemes, AI has rapidly evolved into a strategic priority and is among the leading keys for success among insurers today.

There is no doubt AI is propelling companies to success and providing them with a competitive edge — this can be attributed to AI’s ability to swiftly compare and process millions of data points, like claim details. However, for many insurers, organizational challenges remain which constrain fully embracing the power of AI. Prominent among these challenges are ethical concerns, namely the issue of bias.

Though bias is a considerable worry for some AI systems, most notably predictive model-based AI systems, this does not ring true for all. When insurers leverage fuzzy logic-based AI systems that deliver explainable decisions, the result is that biased outcomes are effectively eliminated.

Why predictive models are highly biased to historical outcomes

As mentioned earlier, some AI systems are in danger of making biased decisions. This is one of the downsides of using predictive model-based AI, which is used in many AI tools. These systems are built solely on historical outcomes and, as such, are highly biased to historical decisions. When using these kinds of AI systems, bias can creep into the system in several ways:

Bias in historical data

Primarily, bias can be found in historical data for claim records, which hold indicators for fraud. These indicators include whether a claim was paid, an underwriting premium, etc. Any potential bias that existed within the organization is captured in these historical decisions — including bias resulting from human operators, unintentional or otherwise.

Furthermore, simply removing variables concerning race, gender, socioeconomic, geodemographic or other attributes is not enough to eliminate historical bias. This is because those variables correlate to other variables in your database.

Structural bias and organizational evolution

The second source of bias in predictive models is structural and is based on the evolution of an organization. For example, if a customer base in the past was biased to a certain geodemographic set, it will be difficult when adding new or different types of customers. This is because they will likely have attributes or characteristics that are unfamiliar to the system.

Moreover, the system may overemphasize certain ethnic groups or economic groups, because that is how historical decisions are biased. As a result, it will likely unfairly bias customer segments that haven’t previously been seen within the organization.

There is no bias inherent to the laws of physics

One thing is clear: Great care must be taken when training models and the technology used must be carefully chosen. There is no bias inherent to the laws of physics. Thus, utilizing AI systems grounded in the fundamentals of mathematics, such as fuzzy logic AI, means outcomes and decisions will be entirely unbiased.

What is fuzzy logic AI?

Fuzzy logic is a mathematical method of solving problems using qualitative data to produce quantitative outputs. Consider ambiguous terms like ‘most likely’ or ‘maybe.’  Terms such as these are referred to as ‘fuzzy’ and can be utilized by fuzzy logic systems to produce quantitative outputs or decisions.

In insurance, fraud and risk are outlying human behaviors that deviate from similar people and similar claims. Moreover, fraud and risk are outliers that are particularly difficult to identify because they arise from being different than one’s peers. Similar to the notions of ‘most likely’ or ‘maybe,’ the notion of being different is a fuzzy concept — a fraudulent claim or person can be very different, moderately different and more.

As such, the degrees of difference that exist in fraudulent claims and activity can be expressed mathematically using fuzzy logic. Fuzzy logic AI systems will utilize these mathematical terms and make millions of comparisons to assign a score for the degree to which any individual or claim is an outlier. This information will ultimately be used to determine which claims or individuals are fraudulent. A score that depicts a significant difference is a flag for fraud.

What causes unbiased outcomes in fuzzy logic systems?

Primarily, fuzzy logic systems are unbiased as they are not based on historical outcomes. These AI systems identify fraud by pinpointing outliers in claims data. There is no bias in being an outlier, even in the case of new or underrepresented groups.

For example, in an ethnically sensitive group, outlying individuals will be identified by being compared to similar individuals and claims within that ethnic group. This same process of identifying outliers can be applied to any underrepresented group, as individuals will only be compared to those that share the same attributes and characteristics. Doing so is completely without bias as the concept of being an outlier within one’s peer group has no bias inherent to it.

How explainability contributes to a lack of bias

Explainability works to end bias by making all outcomes transparent — there is no bias inherent to an explanation of the facts. Fuzzy logic AI systems deliver highly explainable decisions. Again, this is because these systems uncover fraud by identifying outliers.

For example, suppose a claim for vehicle damage caused by winter weather was flagged by the AI system as fraud. For this claim to have been considered fraudulent, it needed to have attributes that differed significantly from claims like it. In the case of the vehicle, the claim was deemed fraudulent because the repair costs far exceeded the typical costs for similar vehicles with similar damage.

This example demonstrates that what makes a claim fraudulent is easily identifiable and is a matter of fact. It arises from the attributes that are inherent to it and that make it different from others like it. There is no bias in this data point.

The presence of AI is increasing across the insurance space, which can be attributed to the race to solve pandemic-related challenges and the significant competitive advantages insurers have realized as a result of AI adoption. However, in many cases, there’s still apprehension to implement AI due to some of the challenges it poses — such as the risk associated with biased decisions and outcomes.

Though it may be true that there is a risk of biased decisions in some instances, like in the case of predictive models, this is not entirely the case for all AI systems.

When insurers use fuzzy logic AI systems, they will receive highly explainable decisions that are entirely without bias — aiding the swift adoption of AI technology and enabling insurers to focus on their continued success in the rapidly changing insurance space.

Gary Saarenvirta is the founder and CEO of Daisy Intelligence, a firm specializing in creating AI solutions for insurers. Saarenvirta previously served as former head of IBM Canada’s data mining and data warehousing practices. 

Related: