Generative AI and LLMs in Insurance: Common Risks and Proven Mitigation Tactics

Generative AI and LLMs in Insurance: Common Risks and Proven Mitigation Tactics
Le Nguyen

by: Le Nguyen, Digital Marketing Manager at Zelros

Artificial Intelligence (AI) has taken center stage as a driving force behind innovation. One particular branch of AI that has garnered significant attention is Generative AI, a technology capable of generating highly realistic and complex content, such as images, text, and even videos. 

Generative AI is a broad category for a type of AI, referring to any artificial intelligence that can create original content. The text-generating part of Generative AI is built on a large language model (LLM).

While Generative AI (GenAI) holds immense promise, it also comes with its fair share of risks and challenges that must be carefully navigated. This blog post will delve into the risks associated with implementing Generative AI/LLMs within the insurance industry. Here is a rundown list of what to expect

Risk of implementing Generative AI in general. 

Generative AI, a subset of artificial intelligence, involves training models to produce creative content often indistinguishable from human-generated content or sometimes even better. Several risks that you might already be aware of: 

1. Hallucinations

Generative AI can create convincing fake content, such as deep fake videos and fabricated news articles. This poses a significant risk to information integrity and can potentially spread misinformation on a massive scale. Generative AI hallucinations typically occur due to limitations in the training data or the architecture of the AI model. These hallucinations are unintended and can be seen as a form of “creative” output from the AI system, though they are not typically useful for practical applications.

2. Privacy Concerns

Generative AI models can inadvertently expose private and sensitive information from the training data they were built on. This could lead to breaches of personal privacy and data security. A clear example was what happened at Samsung. 

3. Amplification of Biases

If the training data used for Generative AI contains biases, the AI model can inadvertently perpetuate and amplify these biases in the generated content. This poses ethical concerns and can reinforce existing societal inequalities.

4. Intellectual Property Infringement

Generating content resembling existing copyrighted material could lead to intellectual property disputes and copyright infringement issues.

Risk of implementing Gen Ai/ pre-trained LLMs in Insurance.

The majority of communication between policyholders and their insurance companies is conducted through written contracts. Therefore, a solid grasp of the language used in these documents, as well as an understanding of how information is presented, is crucial for insurance literacy.

As mentioned above, the text-generating part of Generative AI is built on a large language model (LLM). There are a lot LLMs use cases in the Insurance context.  The pressing question is by utilizing Large Language Models (LLMs) to enhance insurance operations, what are the risks of implementing LLMs? 

1. Underwriting Ambiguity

Pre-trained LLMs can generate intricate narratives, potentially leading to inaccuracy or bias in claims and underwriting cases.

2. Regulatory Compliance

The insurance industry operates under strict regulatory frameworks. Integrating LLMs without proper oversight could lead to non-compliance with regulations, resulting in legal and financial repercussions.

3. Bias

Similar to what we mentioned above, biases in training data can cause unfair decisions, discriminating against certain groups of policyholders, leading to regulatory and reputational risks. 

4. Scalability challenges: 

Investing in maintenance, updates, and scalability is required. Failure to do so can lead to system inefficiencies and increased operational risks. Additionally, many security measures are required to use LLMs at scale to ensure no security or data breach. 

5. Transparency and explainability: 

LLMs can be complex, making understanding the reasoning behind their decisions challenging. This lack of transparency can be problematic when explaining decisions to policyholders or regulators.

The safety of LLMs in Insurance is a serious topic.

Despite the risks, have other insurance providers applied LLMs internally?

Leading carriers like AXA have already begun this journey. 

AXA, in partnership with Microsoft, has launched an in-house GPT platform for all employees. This platform provides AXA staff access to Generative AI and Large Language Models in a secure, data-privacy-compliant Cloud environment. It enables employees to generate, summarize, translate, and correct texts, images, and codes. Initially available to 1,000 AXA Group Operations employees, there are plans to expand access to all 140,000 employees worldwide soon.

A quick announcement: In collaboration with AXA Next and Microsoft, we’re excited to host a webinar for French speakers! Don’t worry if you don’t speak French – we’ll be producing a video with English subtitles. If you’re interested in receiving a link to the video in your inbox, drop your email below! Stay tuned!

Tactics to mitigate the risks of implementing LLMs

Despite the inherent risks, it has become evident that all industries, including insurance, will embrace Generative AI to optimize their operations. Below are some of the tactics and guidelines we use for our LLMs. 

These serve as our guidelines and tactics as we build a recommendation engine in the insurance domain. In context, our aim is to assist both insurance providers and policyholders in achieving their objectives by flagging and pushing suitable insurance coverage recommendations that fit the policyholder’s lifestyle and risk profiles. This means policyholders receive protection from the coverage that they were promised and paid for, while insurance providers can fulfill their commitments without negatively impacting their business. With this in mind, Zelros’ Recommendation Engine, leveraging Large Language Models (LLMs), delivers personalized and contextually relevant recommendations to insurance professionals. The engine generates customized insights that facilitate decision-making processes and enhance customer interactions by analyzing extensive customer data. This leads to improved efficiency (with up to 14% time saved) and accelerated training of new agents (reducing training time by threefold)

    1. Put in place Ethical or Responsible AI frameworks and processes

Zelros’ platform is intentionally designed with a strong emphasis on ethical AI practices. This includes the implementation of stringent data privacy measures, the establishment of Responsible AI protocols, the introduction of bias bounties, and the incorporation of explainability and transparency as the core components of our bias-detection algorithms. 

These measures collectively ensure that the content generated by our Recommendation Engine is accurate, impartial, and compliant with regulatory standards. 

For insurance carriers that already have an established data team, we also provide an ethical report to verify the bias level of the data. To learn more about fostering fairness and responsible AI in the insurance sector, you can refer to our blog titled 7 ways to foster a fair and responsible AI in the insurance service posted in 2020, and the most recent article we shared on Forbes this year.

    1. Make sure the AI recommendation is transparent and explainable

Yes, it’s crucial, so let’s emphasize it once more. To reduce the hallucination risks created by gen AI, the recommendation must clearly articulate why and how it generates these recommendations or solutions. This enables agents and other stakeholders to cross-check the reasoning and logic of the algorithm while allowing the recommendation engine to receive feedback from agents and other stakeholders. Ultimately, this operation aids ongoing training and the pursuit of more refined insights and less risky decisions. 

    1. Assign a human supervisor. 

We believe that the optimal approach in today’s landscape is to maintain a balance between AI and human expertise in complex insurance scenarios. This balance is achieved by regularly monitoring and auditing the performance of generative AI systems.

By utilizing Large Language Models (LLMs), insurance providers can receive timely notifications about potential sales and segment opportunities that might have otherwise been overlooked. These opportunities are identified based on changes in policyholders’ intentions and lifestyles.

In tandem, the algorithm collects feedback for continuous refinement through retraining. This collaborative approach strengthens the bond between insurance providers and policyholders, ultimately improving relationships within the industry.

It’s crucial to underscore the importance of a balanced partnership between AI and human expertise. This ensures consistently informed decisions, particularly in intricate and multifaceted insurance scenarios.

What’s next for LLMs in Insurance? 

A recent article from Andreessen Horowitz believed that there are 2 waves: Gen Ai and Synth AI,

Credit: Andreessen Horowitz. Graph 1: The Synth AI revolution

Wave 1 is now. Currently, AI-generated content is largely suitable for repetitive and low-stakes work, such as writing short ad copy or product descriptions. For instance, while AI can generate a generic sales email, it’s less reliable for accurate personalization. Wave 2 is about the shift from generating information to synthesizing it for improved decision-making. Andreessen Horowitz called this a  SynthAI, which is the capability to sift through large amounts of information, synthesize, analyze, and push out recommendations that will help humans make better decisions, faster. Instead of writing long-form responses based on a concise prompt, the idea is to reverse engineer from massive amounts of data the concise prompt that summarizes it. This represents an opportunity to rethink the user experience to convey large amounts of information as efficiently as possible.

To help insurers move toward wave 2 of AI, Zelros’ Recommendation Engine provides a Vertical SaaS that companies insurers with properly trained LLMs to optimize their operations. It seamlessly combines the power of LLMs with ethical considerations, empowering insurance professionals to tap into AI’s potential while proactively guarding against its potential pitfalls. We have prepared great and concrete use cases of the e-book on Gen AI, and we are pushing our new product release in a few weeks, input your email below if you are interested in receiving more materials directly in your inbox!