Thought Leaders
From Neural Networks to Negligence: Who’s Responsible When AI Fails?
This article discusses legal issues surrounding the harms caused by artificial intelligence, and where liability may lie if a neural network is not functioning correctly in the real world.
Artificial intelligence (AI) has slipped by research labs and into courtrooms, clinics, cars, and stock exchanges, with neural networks diagnosing disease, approving loans, and increasingly accomplishing tasks long considered the domain of humans. But when such systems fail in the real world, they do so with grave and sometimes fatal consequences.
As a plaintiff personal injury lawyer in Ontario, Canada, and a current Doctor of Business Administration candidate studying the intersection of business, law, and technology, I am increasingly asked a simple question: If AI harm occurs, who is to blame?
The answer is actually more complex. AI raises challenges to the doctrines of negligence, causation, and foreseeability, and prompts fundamental questions about how the law should respond to machine decisions.
The Expanding Role of AI in High-Risk Decision-Making
AI is no longer limited to automation of low-skilled tasks or to machine learning platforms. It now makes decisions on healthcare, finance, employment, transport, policing, and legal analytics.
Current AI models typically consist of deep neural networks that, while they can detect complex patterns within large datasets, remain substantially opaque to their developers. Academics have observed that AI systems introduce unpredictability and autonomy into the law, once strictly governed by foreseeability and intention. This creates a conflict between innovation and accountability.
While legal systems treat caused harm as being attributable to human beings, AI distributes responsibility to the dataset’s engineers, who create them, deployers, and end users. Responsibility becomes diffuse.
The Liability Gap: When Harm Cannot Be Easily Attributed
Legal scholars increasingly refer to an emerging “liability gap” in AI governance. Traditional tort law depends on identifying:
- A duty of care
- A breach of that duty
- Causation
- Damages
AI complicates each element. For example, developers may not know how a model will behave after deployment, and organizations may use third-party machine learning (ML) systems. From the end-user perspective, the mechanism for determining outputs is opaque.
One academic study noted how proving fault is more difficult if the AI systems act semi-autonomously or adapt autonomously through processes like machine learning. This fragmentation challenges doctrines that rely on human agency.
In personal injury litigation, courts traditionally examine whether a defendant acted reasonably in the circumstances. But how should courts assess reasonableness when decision-making is partially delegated to probabilistic models?
Neural Networks and the Problem of Explainability
Deep learning systems often operate as “black boxes.” Their internal decision-making processes are not easily interpretable, even by experts. This lack of explainability has serious legal implications.
If a medical AI system misdiagnosed cancer, who would be at fault? It could be:
- How the model was trained
- Whether the training data contained bias
- Whether validation processes were adequate
- Whether clinicians relied too heavily on the automated output
Legal literature recommends a distinction between causal responsibility, role responsibility, and liability responsibility for attributing responsibility for harms caused by AI.
However, in practice, responsibility could extend to a chain of actors:
- Data providers
- Software developers
- Model trainers
- Deployers
- Organizations that use AI outputs
- Professionals relying on AI recommendations
AI does not eliminate responsibility, but redistributes it.
Lessons from Autonomous Vehicles: A Case Study in AI Liability
Autonomous vehicle litigation provides an early glimpse into how courts may address AI-related harm. Courts have since applied the customary principles of negligence and product liability to new technologies causing injury. In recent cases, juries have begun to apportion liability between human drivers and the technology companies that created the automated driving systems.
Legal commentators have suggested that existing product liability doctrines, such as design defect, manufacturing defect, and failure to warn, are relevant to AI-enabled systems.
However, autonomous vehicles expose the limitations of current legal frameworks. Should liability attach to: The vehicle manufacturer? The software developer? The human operator? Or the data used to train the algorithm?
Some scholars have suggested that analogizing product liability principles can effectively delineate responsibility between upstream and downstream actors.
From the plaintiff’s point of view, the cases show that courts may still apply customary principles to new technologies, but only with truly advanced expert evidence and technical knowledge.
Strict Liability and Negligence: Competing Legal Theories
A central debate is whether customary negligence frameworks can be applied to AI systems. Some scholars advocate strict liability regimes, arguing that injured parties should not bear the burden of proving fault in highly complex technological environments.
Strict liability may be especially useful where harm is foreseeable but unavoidable, where an AI system is deployed at scale, and where the risks are socially distributed. Or technical causation is not easy to prove.
It is also argued that negligence law can adapt to technological changes. Comparative legal research has argued for building on existing doctrines, with their intrinsic stabilizing effects and incremental adaptation to new harms.
Strict liability and negligence illustrate policy preferences relating to innovation, fairness, and how risk and costs should be allocated.
Should innovators be responsible for the technological risk? Or should society broadly share the costs of progress?
The Business Perspective: Risk Allocation and Insurance
From a business perspective, AI liability is not just a legal issue but a matter of risk management. The organizations deploying AI systems have begun focusing on questions of contractual risk allocation: 2. professional liability insurance, 3. cybersecurity coverage, and 4. indemnification, and regulatory compliance frameworks.
Insurance markets may play an important role in establishing accountability for AI. Some research suggests that certain harms from AI may ultimately be best addressed through hybrid compensation schemes that combine insurance and tort.
Companies incorporating AI into operational decision-making should account for the risk of litigation in their digital transformation. Failure to do so may leave such organizations exposed to reputational damage, regulatory sanctions, and civil liability.
Ethical Responsibility vs Legal Responsibility
In many cases, legal liability is not equivalent to ethical liability. AI governance debates include the following principles:
- Fairness
- Transparency
- Accountability
- Explainability
Legal obligations, however, do not automatically follow from ethical considerations.
Recent work proposes conceptual frameworks for apportioning responsibility in multi-actor AI ecosystems and for constructing evidentiary rules that connect design decisions to legal outcomes. Thus, legal systems that seek to promote innovation must also account for those harmed by new technology. This balancing exercise will shape the future of AI governance.
The Future of Negligence in the Age of Artificial Intelligence
Artificial intelligence challenges the assumption that decision-making authority always rests with identifiable human actors. But legal responsibility ultimately remains human.
In the near future, it is unlikely that courts will recognize AI systems as legal persons. Instead, it is expected that individuals who develop, deploy, and profit from AI systems will be held liable.
As AI systems grow more independent, hybrid models may evolve, combining various forms of regulation:
- Negligence principles
- Product liability doctrines
- Regulatory oversight
- Insurance-based compensation schemes
Instead of replacing common law concepts, AI may simply force courts to clarify them. From a plaintiff’s standpoint, the question is what happened.
Who is the risk creator? Who can best prevent the risk from harming individuals?
Until legislatures adopt comprehensive technologies. In neural AI liability regimes, courts will apply existing legal doctrines to novel technologies. Neural networks may be new. Negligence is not.
