The challenge of insuring artificial intelligence
The core question in this article by Josephine Wolff is whether conventional insurance can cover the novel risks that AI systems create for companies.
Key points:
- AI risks are hard to quantify. Insurers need historical loss data to price premiums, but AI-related incidents are too new and too varied for actuarial tables to exist. This makes underwriting extremely difficult.
- Existing policies have gaps. Standard commercial general liability and cyber insurance policies were not designed with AI in mind. Insurers are beginning to add explicit AI exclusions, leaving companies exposed to coverage gaps for harms caused by AI outputs, such as biased decisions, hallucinations, or autonomous system failures.
- The cyber insurance parallel is instructive but imperfect. Cyber insurance took decades to develop into a functional market. AI insurance faces the same bootstrapping problem, but AI risks are potentially broader and more complex than cyber risks, involving physical harm, reputational damage, IP liability, and discrimination claims simultaneously.
- Moral hazard is a concern. If companies can offload AI risk to insurers, they may underinvest in safety. Good insurance design would need to condition coverage on safety practices, essentially using premiums to incentivize responsible AI deployment.
- Market development is nascent. A few specialty insurers are beginning to offer standalone AI liability products, but coverage is limited, expensive, and not standardized. The article calls for clearer regulatory frameworks to help the market mature.
The overall takeaway is that insurance could become a valuable governance and risk-transfer tool for AI, but significant market, regulatory, and technical infrastructure still needs to be built first.
Can Companies Insure Against AI’s Growing Risks?
By: Josephine Wolff – Tufts University
AI’s growing risks are actually insurable
Dedicated AI insurance is necessary and feasible, argue Lukasz Szpruch, Agni Orfanoudaki, Carsten Maple, Matthew Wicker, Yoshua Bengio, Kwok-Yan Lam, and Marcin Detyniecki, but requires a fundamental paradigm shift namely from historical actuarial data to a priori performance-based underwriting.
Four core contributions:
- Underwriting challenges unique to AI: The paper identifies five structural obstacles: a) absence of historical loss data; dynamic model behavior (models change post-deployment through retraining or even prompt updates); b) emergent, unpredictable behaviors (hallucinations, sycophancy); c) opacity and attribution difficulties; and d) correlated/systemic risk, where a single flawed foundation model can cascade losses across many firms simultaneously.
- Novel pricing and risk transfer mechanisms: The authors propose three mechanisms tailored to AI: a) parametric insurance (automatic payouts when objective performance metrics cross a threshold, avoiding lengthy causation disputes); b) usage-based pricing (premiums tied to queries, transactions, or volume); and c) bonus–malus schemes (dynamic premium adjustments rewarding demonstrated safety and penalizing repeated failures).
- Insurance as a governance lever: This is the paper’s most distinctive contribution. Because underwriters directly bear the cost of claims, they have strong financial incentives to demand rigorous testing, monitoring, and validation from AI developers. Without insurance, AI assurance audits risk becoming “box-ticking exercises.” With it, insurers become de facto guardians of AI governance, setting standards that markets adopt faster than regulation can.
- Market structure and technical enablers: The paper anticipates an oligopolistic AI insurance market (mirroring cyber insurance), with large incumbents capturing first-mover advantages. It also details the technical toolkit insurers will need: LLM-based automated evaluation, adversarial red-teaming, causal inference for liability attribution, and conformal prediction for uncertainty quantification in pricing.
- Key distinction from existing coverage: AI insurance is explicitly not cyber insurance (which covers malicious external attacks) nor professional liability (which assumes inspectable standards of practice). AI failures arise from normal operation diverging from intended behavior, a fundamentally different risk category requiring its own coverage architecture.
Insuring AI: Incentivising Safe and Secure Deployment of AI Workflows
Authors: Lukasz Szpruch, Agni Orfanoudaki, Carsten Maple, Matthew Wicker, Yoshua Bengio, Kwok-Yan Lam, Marcin Detyniecki
From: University of Edinburgh, University of Oxford, University of Warwick, Imperial College London, University of Montreal, Nanyang Technological University, AXA Group
