Artificial Intelligence (AI) in Claims and Litigation: What You Need to Know

Artificial intelligence is rapidly changing how information is created, analyzed, and used. Its impact is increasingly being felt in insurance claims, litigation, and investigations. From AI-generated images and documents to automated analysis and decision-making, AI can create efficiencies while also raising an important question: Can the information be trusted?

Understanding these capabilities and limitations is increasingly important when AI-generated information becomes part of a claim, investigation, or legal matter.

AI is here to stay. From large investments in data centers, to government use, to computer-generated art and documents, it is becoming a mainstay of society. The equipment that hosts AI are servers and storage arrays, but the resources available are increased so it has more memory, storage capacity, and processing power. Sometimes, multiples of the same hardware are put together to form a cluster. Energy and cooling are vitally important to the infrastructure needed for AI. This equipment can and will fail, so it is important to understand the infrastructure to provide why it failed, the cost to repair or replace, and what AI functions are to the business.

An AI model is essentially a program or algorithm that is trained by data. The AI model then uses the data it ingests to recognize patterns, generate new content, and make predictions. There can be limitations created by rules and quality of data. Most AI models are inspired by human brains; the model forms a neural network, a computing network of hardware nodes that interconnect to form artificial neurons, which process data in a fashion that mimics the human brain. This gives rise to machine learning and all the advantages and disadvantages that come with it.

These concepts are important to understand as the data ingested is not always the data that is outputted. An AI model “learns” from the data that it is programmed with. The data ingested by the AI model needs to be accurate and unbiased; otherwise, the output may be inaccurate or biased. Therefore, it is important to have data that is free of bias or inaccuracies to effectively use an AI model. “Trust but verify” applies to AI models, and a level of understanding is needed to verify the output.

AI language processing, or Natural Language Processing (NLP), is a variant of AI that enables the network to understand and process natural, spoken language to the model, thereby enabling speech inputs that can be processed and understood by the model. It can also enable AI to generate human language and use machine learning to perform tasks such as translation, chatbots, and analysis of sentiments. You might recognize NLP from GPS systems, Amazon Alexa, Apple’s Siri, or Google Gemini. These are often known as assistants.

There are two main access models of AI: private and public. Private AI is used in-house by many businesses to assist with research and development, decision-making, or sharing information. Public AI is widely used by anyone with popular versions, such as ChatGPT, Claude, Grok, etc. Public AI is normally hosted by third-party suppliers. There are even versions that can be a hybrid of both, such as Microsoft Copilot, where one is free and the other is paid. The main difference is the usage and data ingested by the AI model. Most public AI is trained on broad data, while private AI is sensitive data in a controlled environment with stricter rules on usage. Both can be hosted in a cloud environment, which is hardware hosted by someone else. The main difference is public access via a cloud environment versus a private cloud. Public AI is widely shared, whereas private AI has restricted access along with more controls for data privacy and security. Public AI can be free for general use, while there are costs for premium features, but private AI has a higher initial cost, which can later have cost savings depending on how it is used.

One of the primary drawbacks of AI is accuracy. Accuracy is the ability of the model to generate output that is free of bias and flaws, which has been verified as correct based on the data provided/ingested to program the AI model. AI models can hallucinate output, which is where the AI generates wrong information that is not based on facts or the data it was trained on. AI models sometimes have additional parameters to adjust weight on some data to improve accuracy. This is why it is important to verify the data to confirm its accuracy. Utilizing incorrect output to form decisions has led to many mistakes in business and even law enforcement. This arises from accuracy issues in facial recognition or even fake evidence.

Generative AI is widespread as it can create audio, images, text and video, which allows malicious actors to create convincing digital media of events that never occurred. To the untrained eye, this “evidence” appears real and conventional methods of review become problematic if utilized. A forensic approach to authenticating the data is needed to verify the veracity of the digital media. This may include examining metadata, file characteristics, source information, and other digital artifacts to help assess the authenticity and integrity of the evidence. For example, the below was fabricated in around 2 minutes in Microsoft Copilot, a widely available AI model.

AI-generated example for illustrative purposes. Generative AI tools can create convincing digital content in minutes, reinforcing the importance of authenticating digital evidence when its origin or integrity is in question.

AI will continue to create new efficiencies and new challenges for claims professionals, attorneys, and investigators. As AI-generated content becomes increasingly sophisticated, determining whether digital evidence is authentic may no longer be possible through visual inspection alone. When the integrity of digital evidence could affect the outcome of a claim or case, a forensic examination can help establish what is authentic, what has been altered, and what the underlying data can support.