Via Negativa

The words come from the latin, the 'way of negation', and the principle can be used to understand or define something which is so complex or difficult to understand (such as God or AI), that it is beyond human comprehension. This approach enables decisions to be made whilst acknowledging the limitations of human understanding, language and complexity. Via negativa has rigorous practical applications in one’s personal life, decision making, religion and insurance underwriting.

The approach can be exemplified by comparing regular reasoning with via negativa:

• Regular reasoning:
  • God is good (vague and broad, unclear what this means)
  • God is love (vague and broad, unclear what this means)
• Via negativa:
  • God is not evil (can be exemplified in the real world)
  • God is not hateful (can be exemplified in the real world)

Via negativa is a powerful approach as it:

• challenges typical positive descriptions (via positiva) and provides real world examples of what something is not or what not to do
• recognises with humility, the incomprehensibility or complexity of the thing it is describing
• prevents attributing human-like qualities to the thing it is describing
• challenges the idea that we can have definitive or quantifiable knowledge of things and refers to descriptions based off actions rather than subjective intellectual assertions
• defines the factual, real world essence of something highly complex through inversion

Good decision making involves avoiding bad decisions, and this approach is key to Charlie Munger and Berkshire Hathaway’s success. Munger is famous for stating,

"All I want to know is where I’m going to die, so I’ll never go there”.

Applying via negativa to generative AI insurance

Generative AI risks are being speculated upon. The complexity of AI systems, hyper-scale and the pace of innovation make evaluating the risks associated with this technology difficult.

At Testudo, we provide insurance for losses arising from AI systems. We approach underwriting by removing known areas of accumulated AI risk, identified in our loss data. Given AI is a moving target, relying on historical data alone will not be sufficient, so we have developed a technology platform to help our clients actively track and mitigate identifiable risks. The result is standalone Gen AI liability insurance underwritten on Lloyd's of London paper, sized to the real exposure rather than speculation about hypothetical failure modes.

Our approach of removing known risk is more robust and practical than forecasting where losses may come from. Whilst we appreciate the entire AI ecosystem needs to consider systemic risk, theorising may lead to obvious risks being missed, by valuing complexity over simplicity.

Insurance is a way for a market to self-regulate as it puts a price (an insurance premium) on assurance, regulatory standards and frameworks. Financial claims allow us to quantifiably understand AI risks as they occur in the real world and not in simulated evaluation environments. Applying unsubstantiated claims of risk in areas that are not systemic or ruinous may cause more harm to the industry than good (Iatrogenics).

Next steps

• Read more of our research and insights articles to get the latest AI insurance insights
• Contact our team to find out more about insuring GenAI and how to monitor the litigation risks of the systems you use

How does Testudo underwrite generative AI risk?

Testudo underwrites generative AI risk by removing known areas of AI-related loss from its proprietary real-world litigation and incident database, rather than speculating about which risks might appear. This via negativa approach is more robust than forward forecasting because AI is a moving target. Testudo's AI Risk Engine then monitors each insured's exposure in real time. The result is standalone Gen AI liability insurance on A+ (Superior) rated Lloyd's of London paper, placed without an invasive technical audit of the insured's systems.

Why doesn't traditional insurance cover generative AI liability?

Traditional commercial insurance products (CGL, Cyber, Tech E&O) were not designed with generative AI liability in mind and are increasingly adding explicit exclusions for it. As of January 1, 2026, Verisk and ISO generative AI exclusions took effect on CGL policies across the United States. Standard insurers also lack the litigation data needed to price the risk confidently, which slows product development. Testudo addresses both problems: its underwriting is grounded in real loss data rather than speculation, and its policy is purpose-built to respond to third-party claims from AI outputs. Enterprises seeking cover before purpose-built policies existed explored alternative risk transfer mechanisms such as captives and parametric insurance, though none of these provide the same balance-sheet transfer as a standalone liability policy.