Text丨Xu Xian team
The heated debate on whether AI will replace insurance agents has not yet ended, and the topic of a wider range of insurance practitioners may also be replaced by AI has begun.
With the development of artificial intelligence and large language model technology, there has been a heated discussion about the substitution of AI for human work. In the field of insurance, due to the powerful conversational capabilities of large language models represented by ChatGPT, this discussion often focuses on "can ChatGPT replace insurance agents?".
However, with the rise of multi-modal large models such as GPT-4V (ision), the scope of those to be replaced will be greatly expanded. GPT-4V not only demonstrates superior natural language processing capabilities, but more importantly, it shows great capabilities in processing multimodal data such as text and images. This all-round processing capability indicates its wide application potential in multiple insurance business fields such as underwriting, risk control, and claims.
So, in this era of rapid development, is GPT-4V just a tool, and will it become a comprehensive "replacement" for insurance practitioners in the near future?
Recently, Professor Xu Xian's team from Fudan University and the University of Rochester released a research report to qualitatively analyze the performance of multimodal large models (LMMs) represented by GPT-4V in various multimodal tasks in the insurance field, and explore the potential application of GPT-4V in the insurance industry.
This study explores the possibility of combining insurance and cutting-edge LMMs technology for the first time, and makes a fundamental exploration for promoting interdisciplinary communication and development of insurance and the continuous promotion of future research work.
Let's take a look, in addition to "replaceing" insurance agents, based on the experimental tests of Professor Xu Xian's team, what else AI can and cannot do for the insurance industry, and what problems and challenges exist -
1 AI has a strong ability to complete insurance tasks: it can understand both insurance content and insurance scenarios
Based on the two dimensions of insurance type and insurance business stage, Xu Xian's team systematically classified the multimodal tasks in the insurance field.
In terms of insurance selection, it covers representative types of life insurance and property insurance (health insurance, car insurance, family/enterprise property insurance, agricultural insurance);
In terms of business stage, it covers three key business processes: underwriting assessment, risk monitoring, and loss claim, so as to reflect the comprehensiveness and representativeness of task division.
The study collects and tests a large number of graphic cases close to insurance business scenarios, showing that GPT-4V has shown strong capabilities in insurance-related tasks, not only has a strong understanding of multimodal content in the insurance field, but also has a comprehensive understanding of insurance scenarios.
1. In the field of auto insurance, this study conducts an in-depth exploration and analysis of GPT-4V's vehicle underwriting assessment, vehicle risk monitoring, vehicle loss claim and vehicle fraud identification capabilities, and presents GPT-4V's capabilities through specific cases.
(1) Vehicle underwriting evaluation: GPT-4V has demonstrated a strong ability to underwrite and evaluate auto insurance, and it can extract key elements such as vehicle model, vehicle appearance condition, and vehicle VIN. Notably, it also provides a precise analysis of the role of the newspaper in the picture (used to determine the timing of the shooting).
(2) Vehicle risk monitoring: GPT-4V can accurately identify dangerous driving behaviors (making phone calls, playing with mobile phones) in the car.
(3) Vehicle loss claim: GPT-4V can identify the loss of the vehicle, describe the loss and analyze the degree of loss
(4) Vehicle fraud identification: GPT-4V can identify mismatched or exaggerated losses in vehicle damage reports based on vehicle accident pictures.
2. In the field of home/enterprise property insurance, this study conducts in-depth exploration and analysis of GPT-4V's family/enterprise property underwriting assessment, family/enterprise property risk monitoring and family/enterprise property loss claim settlement capabilities, and presents GPT-4V's capabilities through specific cases——
(1) Household/business property underwriting assessment: GPT-4V can systematically analyze and evaluate the overall risk situation of family houses and corporate buildings.
(2) Family/enterprise property risk monitoring: GPT-4V can monitor and warn of real-time abnormal events (fires, falling objects, etc.) based on the monitoring images of households/enterprises.
(3) Household/business property loss claims: GPT-4V can identify and analyze the loss, cause and extent of damage of family houses and business facilities, but it is difficult to accurately estimate the specific amount of loss.
3. In the field of health insurance, this study conducts an in-depth exploration and analysis of GPT-4V's health underwriting assessment, health risk monitoring and health loss claim settlement capabilities, and presents GPT-4V's capabilities through specific cases.
(1) Health underwriting assessment: GPT-4V can extract the content of the policyholder's medical report and analyze its health risk status.
(2) Health risk monitoring: GPT-4V can identify and warn of risk events that threaten the health of policyholders (such as falls of the elderly) based on the monitoring screen.
(3) Health loss claim: GPT-4V can identify and extract information from medical expense vouchers, and analyze the correlation and matching between medical items and medical reports.
4. In the field of agricultural insurance, this study conducts in-depth exploration and analysis of GPT-4V's farmland underwriting assessment, farmland risk monitoring and farmland loss claim settlement capabilities, and presents GPT-4V's capabilities through specific cases.
(1) Farmland underwriting assessment: GPT-4V can identify the types of crops in farmland and analyze their risk status and risk level.
(2) Farmland risk monitoring: GPT-4V can identify the growth stages of crops, and monitor and warn threatening events that affect crop growth (such as wildlife invasion, fire, etc.).
(3) Farmland loss claims: GPT-4V can analyze and evaluate the causes, loss conditions and loss levels of farmland, but cannot accurately estimate the loss.
2 Obvious shortcomings and challenges: dynamic and accurate numerical memory, picture illusion, and different language recognition problems will affect the application
The study found that although GPT-4V has demonstrated strong capabilities in multimodal tasks in the insurance field, it also presents obvious shortcomings and challenges.
For example, GPT-4V often fails to provide accurate risk probabilities or accurate loss values when conducting risk assessments or loss claims. Another example is the hallucination of GPT-4V in some test cases.
For another example, GPT-4V's ability to extract and analyze information from documents in different languages is different, and the tests shown that it performs more accurately in the languages of European and American countries (English, German, French), and worse in the languages of East Asian countries (Chinese, Japanese, Korean). In addition, the study also mentions that multimodal large models are not just technical tools, but may develop in tandem with insurance, thus posing a disruptive challenge to the traditional insurance business.
1. It is difficult to memorize the dynamic and accurate numerical content, and there are challenges in accurate damage assessment:
Taking auto insurance as an example, the study achieves accurate loss assessment in the form of additional information input, but GPT-4V itself is difficult to memorize the precise numerical content of such dynamic changes, which poses a challenge to accurate loss assessment.
2. There is an illusion of picture recognition, which will limit the practical application:
Two hallucination phenomena of GPT-4V have been discovered, namely multi-image input hallucination (the results of multi-image input and single-image input do not match) and object recognition hallucination (object recognition error), which will pose a fundamental challenge to the application of multimodal large models in the insurance field.
3. Different languages have different literacy recognition capabilities, and East Asian countries have poor performance in script:
The study tested GPT-4V's ability to recognize and extract documents in mainstream languages in the world's top eight countries in terms of premium income, and found that GPT-4V's ability to perform poorly in different scripts was poor. This is an obstacle to the promotion of multimodal large models in the global insurance industry.
3 AI may revolutionize the traditional insurance business
The study also argues that the focus of the above discussion is on the impact of the emerging technology of multimodal large models on the traditional business of the insurance industry, but the insurance industry itself is also evolving. By introducing emerging technologies, innovative ways of operating can be created that complement the insurance industry and promote mutual development within the industry.
Taking auto insurance as an example, with the innovation of autonomous driving technology, auto insurance is also evolving, and multimodal large models can be intertwined with this development, co-evolve, and have a disruptive impact.
For example, multi-modal large models can give full play to their comprehensive visual, auditory and language processing capabilities to conduct real-time data analysis and decision support, or introduce the processing of visual data through multi-modal large models to help insurance companies develop new risk assessment models based on real-time behavior and environmental feedback.