AI industry growth has been so rapid that insurers have been unable to roll out in time insurance products that would support the legal liabilities AI companies may have to shoulder from various AI safety incidents.

According to a recent Wall Street Journal (WSJ) report, there are already a significant number of lawsuits over chatbot-related harm and alleged copyright infringement during model training. It said incidents involving out-of-control AI agents hacking websites have also occurred, increasing the chance they spill into litigation.

Disputes over allegations that AI chatbots encouraged suicide and murder are also intensifying. Insurance companies specialising in AI risk see the potential legal liabilities faced by large AI companies at tens of billions of dollars, and as much as hundreds of billions of dollars. Some industry officials have even argued that the U.S. government should take on compensation liabilities that companies and insurers cannot bear, the WSJ reported.

Against that backdrop, insurers also appear to be preparing in their own way for insurance claims worth millions of dollars stemming from accidents involving out-of-control AI agents.

According to a recent Financial Times (FT) report, global insurance broker Aon analysed about 300 AI-related lawsuits. It sees it can also pay out claims under crime, intellectual property, media liability, cyber security, and technology errors and omissions insurance.

The FT added that insurers are also preparing for lawsuits over product liability, algorithmic discrimination, personal information and wrongful death.

But the insurance industry still appears to be falling short of keeping up with the pace of AI development.

One point is that AI losses do not fit neatly into existing insurance classifications. Damage from AI ranges widely, from loss of life and property due to factory robot malfunctions to copyright infringement, unauthorised data leaks, hallucination errors and regulatory fines. The WSJ, citing U.S. think tank RAND, reported there is ample room for disputes over what is and is not covered. It also reported that AI companies are included in existing products such as general liability insurance or cyber insurance, but the industry's rapid growth could push AI-related claims beyond existing coverage limits.

There are AI-specific products, but the scale they cover is small. Reinsurer Munich Re sells insurance for AI model errors with limits of up to $50 million. George Lewin-Smith, CEO of British AI insurer Testudo, said large AI companies have both the willingness and capacity to buy insurance worth billions of dollars, but that level of coverage "does not exist on Earth".

The immature state of the insurance market for AI companies is also tied to the fact that AI companies face circumstances that differ from firms operating in other areas of the tech sector.

Portals and social media avoided much of the responsibility for platform content thanks to Section 230 of U.S. federal law enacted in 1996. By contrast, since AI companies are directly involved in generating content, that provision is unlikely to serve as a strong shield, Moody's Analytics analysed.

In this environment, clauses that allow insurers to opt out of compensating AI-related losses in corporate insurance appear to be spreading across the industry.

Insurance Services Office (ISO), an insurance unit under Verisk that provides standard terms, introduced an optional exclusion clause for generative AI this year and is reviewing ways to expand it to agentic AI, the WSJ reported. It said large insurers such as AIG and Chubb can apply the clause if they choose to.

On this issue, Marsh, the world's largest insurance broker, said AI exclusions are a "retreat" for the insurance industry and only increase coverage uncertainty and disputes. It also made clear it would advise clients to look elsewhere if an insurer applies an exclusion clause.

Some also see insurance, if it plays its role, as a tool to rein in AI. AI insurance often underwrites by assessing the likelihood of errors through simulation, meaning it can effectively serve as a form of technical audit. If insurers require safety measures as a condition of coverage, it could create a kind of guardrail for both developers and users, the WSJ reported.