AI Liability Is No Longer a Future Problem for Risk Managers
Artificial intelligence has moved from operational experiment to embedded infrastructure across insurance and corporate risk functions, but governance, liability frameworks, and coverage structures have not kept pace, according to a new report from WTW’s Willis Research Network.
The report documents a surge in AI-related incidents, up roughly 50% year over year from 2022 to 2024, with 2025 incidents having already exceeded 2024’s total before that year concluded. The report warns that organizations are accumulating exposure across general liability, professional indemnity, cyber, EPLI, and D&O, often without policy language that explicitly addresses AI risks.
Silent AI Risk Spreading Across Lines
The report frames AI less as a standalone peril and more as a risk amplifier across existing coverage lines, a dynamic it compares explicitly to silent cyber exposure before 2019, when ambiguity in policy language narrowed quickly once claims and litigation materialized. WTW describes this as the “silent AI” problem: AI-related losses can sit embedded across multiple lines and many insureds simultaneously, invisible until a claim forces interpretation.
AI-related losses are already reaching insurers through multiple pathways, the report said. When AI causes bodily injury or property damage, in autonomous vehicles, industrial systems, or medical devices, courts typically analyze traditional negligence and product liability frameworks, focusing less on whether AI was involved than on whether reasonable care was exercised in deployment and supervision.
When AI causes financial loss, attribution disputes arise across model developers, data providers, system integrators, and end users, with losses landing across technology E&O, professional indemnity, and D&O policies.
The report organizes AI risk into four overlapping categories: performance risk, misuse risk, governance risk, and systemic risk, each mapping to different liability theories and insurance triggers.
Systemic risk, driven by shared dependencies on a small number of models, vendors, or infrastructure providers, is characterized as the most challenging for insurers and reinsurers because it undermines diversification assumptions. A single AI failure could trigger losses across thousands of insureds and multiple lines simultaneously, the report said.
Trust Depends on More Than Accuracy
A separate section of the report, drawing on a joint study by the Willis Research Network and Rutgers Business School that assessed eight leading large language model assistants, found that high technical accuracy does not guarantee trustworthiness in enterprise settings. The evaluation measured AI systems across six dimensions: accuracy and reliability, consistency and robustness, privacy and data security, bias and fairness, transparency and explainability, and governance and accountability.
No system tested met the researchers’ “adequate” threshold for data protection, the report said. Privacy gaps were widespread, with consumer-tier offerings frequently lacking clear deletion guarantees, encryption assurances, or contractual commitments. Governance scores varied sharply: systems backed by formal policies and third-party audits scored higher, while open models that transfer responsibility to deployers were found to shift residual risk to the organizations using them.
The report warned that non-deterministic behavior, where identical AI prompts can produce different outputs across runs, complicates validation and creates challenges for regulated use cases. Bias testing revealed that holding clinical or financial facts constant while varying only demographic identifiers sometimes produced systematic differences in recommendations across repeated runs.
AI Adoption Leaders Are Separated by Culture and Data Infrastructure
The report, drawing on analysis conducted with the Wharton School of the University of Pennsylvania, found that organizations leading in AI adoption are differentiated less by access to advanced models than by the maturity of their data infrastructure, governance structures, and organizational culture. AI leaders treat adoption as a board-level transformation program with multi-year investment horizons; laggards acknowledge AI’s importance but leave strategy reactive, the report said.
Among the most consistent findings across aviation, insurance, and financial services: AI progress is ultimately constrained by data infrastructure maturity. Insurers, despite holding rich historical datasets, struggle with siloed architectures and poor documentation. The most persistent failure mode across all three sectors, the report found, is the inability to move from successful pilot to enterprise-wide deployment, driven by misaligned governance, insufficient standards, and lack of reusable infrastructure.
Obtain the full report here. &
