Legal Frameworks Struggle to Address Unpredictable AI Behavior

The legal system in many jurisdictions is built on the assumption that machines behave deterministically. If you input the same data, you expect the same result. However, large language models do not work this way. They are nondeterministic, meaning they can produce a wide range of different outputs from the exact same prompt. This fundamental difference challenges existing tort laws that were designed for predictable machinery.

As these systems become more integrated into critical sectors, the lack of a clear liability framework is becoming a major concern. When an AI system causes harm, it is often unclear who is responsible because the behavior was not strictly predictable. Legal scholars are now arguing that courts must recognize this technical reality to properly assign blame and protect users from risky deployments.

Unpredictable outputs challenge traditional liability

According to recent legal analyses reported by GN technics/ai (en-US), the core issue is that developers deploy systems they know are unpredictable. By releasing tools that can generate unbounded ranges of responses, creators are introducing high-consequence risks. In safety-critical contexts, such as healthcare or finance, the cost of an error is severe. The argument is that by deploying such systems without sufficient guardrails, developers consciously accept a level of risk that should carry legal responsibility.

This approach aims to create a predictable liability regime for companies. If developers know they can be held liable for the inherent unpredictability of their models, they are incentivized to build safer guardrails. This protects users by ensuring that when things go wrong, there is a clear path to accountability rather than legal ambiguity.

Negligence as the primary legal tool

While some legal scholars have been hostile to using negligence laws for AI, others argue it is the most suitable framework. Negligence is broad and flexible, allowing it to cover various ways foundation models might cause harm. This includes internal corporate deployments, security failures regarding model weights, or even the release of open-source models that are later misused.

The flexibility of negligence law allows courts to adapt to the polymathic nature of modern AI. Unlike rigid statutory regulations, common law can evolve to address new types of harm, such as economic or emotional injury caused by AI behaving in ways that violate human norms. This adaptability is crucial as the technology continues to shift rapidly.

Regulatory gaps require judicial action

Currently, the development and release of foundation models are subject to minimal formal regulation. Governance is often haphazard and lacks clear ex ante standards. In this vacuum, the common law of torts serves as the primary mechanism for managing serious risks. Until comprehensive legislation is passed, courts will likely need to expand the scope of the duty of care to address the unique capabilities and risks of these systems.

This judicial development is necessary to reflect the reality of modern AI. By incrementally developing the law of negligence, courts can ensure that the legal system remains relevant and effective. The goal is to balance innovation with safety, ensuring that the benefits of AI are not overshadowed by unchecked risks and lack of accountability.