Late last year (2025) three large insurers in the United States — AIG, WR Berkley, and Great American asked regulators for permission to exclude AI-related losses from their policies [1]. This does not mean that all Artificial Intelligence losses will become uninsurable, with for example, AIG later saying it had no immediate plans to implement the exclusions. However, the request marked an inflection point with insurers starting to move towards explicit carve-outs for AI-related liability

A fair objection is that AI is not
new in our lives. Artificial neural
networks — the same software
artifact that powers the large
language models — already
pervade many aspects of
interactions with business.

For example, neural networks sit inside the emergency braking systems of millions of cars [2], within the fraud-detection models that monitor transactions for hundreds of banks [3] and in the voice-synthesis services that are used in airports and within apps all around the world [4]. None of these systems led insurers to respond in the way they have towards generative AI.

Why now?

Firstly, each of the systems mentioned did just one job, with only a narrow output. A braking system decides whether to brake. A fraud model produces a risk score. A voice synthesizer turns text into sound. When they fail, the failure is contained inside a clearly defined function. In contrast, the Generative AI models that are now starting to be used have no defined function or boundary. The same model will write customer-service replies, draft contracts, generate code, summarize legal documents and approve expense claims. The output is open-ended language. There is no limit to what it can do, and therefore no limit to how it can fail.

Secondly, the same model is now embedded across many functions inside the same company, and across many businesses inside an industry. When a braking system misbehaves, the manufacturer recalls those cars. However, when a single generative model produces a bad output, the consequences can impact all the firms using it, across every function, all at the same time. This is what insurers mean when they warn of a thousand simultaneous losses from a single upstream failure [1].

Thirdly, those earlier systems were never really one single system. The braking model in Car Brand A is not the braking model in Car Brand B or Car Brand C; the fraud engine at Bank A is not the engine at Bank B. Each was built largely in-house and trained on that company’s own data, which differs widely from one company to the next. They may share a technique — the neural network — but they are distinct artifacts, and they fail in distinct ways. Generative AI breaks this pattern. The same Claude, GPT or Gemini model sits in thousands of firms at the same time; what each company adds is only a thin layer of customization — its own documents, its own prompts, its own context — on top of a foundation that everyone shares.

This changes where the risk concentrates. What counts as “out of sample” — the unfamiliar input that pushes a model past what it was trained to handle — is specific to each car brand, because each brand trained on its own data. The case that defeats Car Brand A’s system may be routine for Car Brand B’s. But what is out of sample for Claude, GPT or Gemini is, to a first approximation, out of sample for every client running it. The failure boundary is shared because the model is shared, and a single bad input stops being one firm’s problem and becomes, potentially, everyone’s at the same moment.

Finally, the data needed to price Gen AI risk does not yet exist and may not accumulate in any usable form. Insurance rests on a quiet assumption — that the past is a guide to the future, that enough years of claims describes a distribution stable enough to price against. Fire, theft, mortality, and motor accidents all satisfy this assumption. Gen AI does not. There are too few large losses to estimate a frequency, no agreed definition of what counts as “an AI event,” and the systems themselves do not stand still long enough to be measured. Nowadays, a foundation model is retrained, re-aligned, and redeployed on a cadence of months.

A truly systemic event — a defect
in a widely used foundation model,
cloud layer or authentication system
that triggers simultaneous losses
across many firms — could one day
look more like terrorism or natural
catastrophe and justify a public
response.

What happens next?

In the past, when insurers retreated from risks they considered too correlated or too catastrophic, governments have sometimes stepped in. After 9/11, the United States created the Terrorism Risk Insurance Act, a public-private backstop for insured terrorism losses [5]. Similar debates have appeared in cyber insurance, where policymakers and insurers have discussed whether the government should act as a reinsurer of last resort for catastrophic cyber events [6]. Natural disasters such as floods or hurricanes are another example. However, while a terrorist attack, a catastrophic cyberattack, or a natural disaster may justify some form of public backstop because the loss is external, systemic, or tied to national resilience, routine AI failures are different. Many arise from ordinary business decisions such as adopting an AI agent too quickly, failing to supervise outputs, giving a model authority it should not have, or relying on a vendor without adequate contractual guarantees.

Public backstops can preserve coverage, but if premiums do not reflect the underlying risk, they change the incentive to reduce exposure. If governments stand behind AI losses, firms will have weaker reasons to ask the questions they should: what is the model allowed to do, who supervises it, what evidence is logged, what authority has it been given, and who bears the loss when it acts outside its intended role?

So far, this is not where AI governance seems to be heading. For example, the National Institute of Standards and Technology’s AI Risk Management Framework treats AI risk as a problem of governance, mapping, measurement and management across the lifecycle, not exp-post compensation [7]. The OECD’s work on AI incidents pushes toward better accountability, reporting and evidence [8]; and the EU AI Act emphasizes risk classification, documentation, monitoring, and the responsibility of providers and deployers [9]. None of this points to a public AI-insurance backstop discussion.

That may change. A truly systemic event — a defect in a widely used foundation model, cloud layer or authentication system that triggers simultaneous losses across many firms — could one day look more like terrorism or natural catastrophe and justify a public response. For now, the right move is not to socialize the losses but make them harder to create. The exclusions being filed are an early warning that the market no longer treats “AI” as a harmless add-on to ordinary business risk. Firms should read that warning as a governance signal.

References 
[1] Financial Times, 23 November 2025. https://www.ft.com/content/abfe9741-f438-4ed6-a673-075ec177dc62 
[2] S. Grigorescu, B. Trasnea, T. Cocias and G. Macesanu, “A survey of deep learning techniques for autonomous driving,” Journal of Field Robotics, vol. 37, no. 3, pp. 362–386, 2020. https://doi.org/10.1002/rob.21918 
[3] J. West and M. Bhattacharya, “Intelligent financial fraud detection: A comprehensive review,” Computers & Security, vol. 57, pp. 47–66, 2016. https://doi.org/10.1016/j.cose.2015.09.005 
[4] A. van den Oord, S. Dieleman, H. Zen, K. Simonyan, O. Vinyals, A. Graves, N. Kalchbrenner, A. Senior and K. Kavukcuoglu, “WaveNet: A Generative Model for Raw Audio,” arXiv:1609.03499, 2016. https://arxiv. org/abs/1609.03499 
[5] Terrorism Risk Insurance Act of 2002, Pub. L. No. 107-297, 116 Stat. 2322. Overview: Congressional Research Service, “Terrorism Risk Insurance: Overview and Issue Analysis,” Report R47042. https://www. congress.gov/crs-product/R47042 
[6] U.S. Department of the Treasury, Federal Insurance Office, “Potential Federal Insurance Response to Catastrophic Cyber Incidents,” Request for Comment, 87 Fed. Reg. 59161, 29 September 2022; see also U.S. Government Accountability Office, “Cyber Insurance: Action Needed to Assess Potential Federal Response to Catastrophic Attacks,” GAO-22-104256, 2022. https://www.federalregister.gov/documents/2022/09/29/2022-21133/potential-federal-insurance-response-to-catastrophic-cyber-incidents 
[7] National Institute of Standards and Technology, “Artificial Intelligence Risk Management Framework (AI RMF 1.0),” NIST AI 100-1, January 2023. https://doi.org/10.6028/NIST.AI.100-1 
[8] OECD, “Defining AI Incidents and Related Terms,” OECD Artificial Intelligence Papers No. 16, May 2024; see also the OECD AI Incidents Monitor, https://oecd.ai/en/incidents. https://www.oecd.org/en/publications/defining-ai-incidents-and-related-terms_d1a8d965-en.html 
[9] Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act), OJ L, 2024/1689, 12.7.2024. http://data.europa.eu/eli/reg/2024/1689/oj