What Dusty Plasma Can Teach the Insurance Industry About AI

Feb 23, 2026

What Dusty Plasma Can Teach the Insurance Industry About AI

By James W. Moore

Key Takeaways

  • Physicists at Emory University used a domain-trained AI to discover new laws of nature from a small, purpose-built dataset. The model ran on a desktop computer and corrected assumptions that had been accepted for decades.
  • The insurance industry is sitting on the largest cooperative data asset in American business (Verisk’s 34.5 billion statistical records, 1.8 billion claims) and hasn’t taken this step.
  • The organizations best positioned to build insurance-specific AI models (Verisk/ISO, NAIC) are focused on selling finished analytics products and regulating AI use, respectively. Neither is building domain-trained models for the industry.
  • Until insurers move beyond general-purpose AI tools to domain-specific models, they’re leaving their most valuable competitive advantage on the table.

Physicists Just Did Something Insurance Executives Should Be Watching

A team at Emory University recently published findings in the Proceedings of the National Academy of Sciences that should interest every insurance executive thinking about AI.

They built a custom AI trained on laboratory data from experiments with dusty plasma, an ionized gas containing charged particles found everywhere from Saturn’s rings to wildfire smoke. They didn’t use AI to process data faster or predict outcomes. They used it to discover new physical laws that human scientists couldn’t see on their own.

The AI corrected theoretical assumptions accepted in plasma physics for decades. It achieved over 99% accuracy. And it ran on a standard desktop computer.

What made this possible wasn’t brute computing power or massive datasets. It was domain-specific data combined with domain knowledge embedded directly into the AI’s architecture. The researchers constrained the model with known physics principles, then let it discover what those constraints alone couldn’t reveal.

That approach has direct implications for insurance.

This Is Different from What Insurers Are Doing with AI Today

Most AI in insurance today makes existing processes faster. Chatbots handle customer inquiries. Document extraction pulls data from submissions. Predictive models score fraud risk. These are valuable, but fundamentally conservative. They automate what we already know how to do.

The Emory team did something different. They used domain-trained AI to find patterns and relationships that human experts had missed entirely. The result wasn’t faster physics. It was new physics.

Apply that to insurance. How many pricing assumptions, risk correlations, and underwriting rules have been carried forward for decades because traditional actuarial analysis seemed to support them? How many are, like the plasma physics theories the Emory AI corrected, “not quite right”?

We don’t know. And we won’t know until someone builds an AI model designed for insurance discovery, not insurance automation.

The Industry Already Has the Raw Ingredients

Here’s the frustrating part. The insurance industry has something most industries would pay a fortune for: a 50-year tradition of cooperative data sharing.

Verisk (ISO’s parent company) maintains over 34.5 billion statistical records. Their ClaimSearch database contains more than 1.8 billion U.S.-based claims from over 1,850 contributors, representing roughly 95% of the U.S. P&C market. The entire cooperative is built on a “give to get” model where carriers contribute data and receive insights in return.

The data infrastructure exists. The cooperative sharing tradition exists. The domain expertise exists.

What doesn’t exist is anyone using this pooled data to train insurance-specific AI foundation models and making them available to the industry.

Verisk is building AI-powered tools (XactAI for claims, predictive scoring, Rating-as-a-Service). Those are useful. But they’re fundamentally different from building a domain-trained model that carriers could deploy for their own discovery.

The closest example is EXL’s Insurance LLM, launched with NVIDIA, which achieved a 30% improvement in accuracy on insurance tasks compared to general-purpose models like GPT-4, Claude, and Gemini. Promising, but even EXL’s model focuses on claims processing and underwriting document tasks, not on fundamental discovery.

Why the Natural Players Haven’t Stepped Up

Two organizations are best positioned to build this capability. Neither is doing it.

Verisk/ISO has the data, the relationships, and the domain expertise. But their business model is built around selling finished analytics products. Building an insurance-specific foundation model would be the logical evolution of their cooperative data model, updated for the AI era. It would also represent a massive new revenue stream. The gap may be one of identity: Verisk sees itself as a data analytics company, not a model provider.

The NAIC has taken an entirely regulatory posture toward AI. They’ve been surveying carriers, developing model bulletins for AI governance (now adopted by 24 states), and exploring a potential AI model law. That work matters. But nobody at the NAIC is thinking about enabling the creation of industry-wide AI resources. Regulation and development aren’t mutually exclusive, but in practice, the NAIC is focused almost entirely on the former.

What Would an Insurance-Specific AI Actually Look Like?

The Emory team’s approach offers a blueprint. They embedded known physics constraints into their AI’s architecture, then trained it on a small but rich experimental dataset. The model didn’t need billions of data points. It needed the right data points, combined with domain knowledge baked into the model’s structure.

The insurance equivalent: an AI model constrained by actuarial principles, regulatory requirements, and insurance fundamentals, then trained on actual pooled carrier results (loss data, claims outcomes, pricing histories, underwriting decisions). Not to automate what actuaries already do, but to discover what traditional analysis can’t reveal.

What might it find? Previously invisible correlations between risk factors and loss outcomes. Pricing relationships that conventional methods miss. Claims patterns that challenge long-held underwriting assumptions. We genuinely don’t know, and that’s exactly the point.

The obstacles are real. Data privacy and competitive sensitivity. Regulatory concerns about model explainability. Antitrust considerations around pooled data for pricing. But the cooperative model already navigates similar concerns for traditional data sharing. These are solvable problems, not fundamental barriers.

The Question Executives Should Be Asking

The physics community is using small, rich, domain-constrained datasets to discover new laws of nature. Insurance is sitting on the largest cooperative data asset in American business and applying general-purpose AI tools that weren’t built for the domain.

The question for insurance executives isn’t “are we using AI?” Most carriers are, or plan to. NAIC surveys show 88% of auto insurers and 70% of home insurers reported they use, plan to use, or plan to explore AI and machine learning.

The better question is: “Are we using AI that was built for our domain?”

For almost everyone, the answer is no.

Gartner projects that by 2027, more than 50% of the generative AI models enterprises use will be specific to an industry or business function, up from roughly 1% in 2023. The insurance industry has every advantage needed to lead that shift. The data exists. The cooperative tradition exists. The domain knowledge exists.

What’s missing is someone willing to assemble those ingredients into a purpose-built model. Executives should be asking their boards, their industry organizations, and their technology partners one question: Who is going to build it?

Sources

Related Posts

468