The metaphor recently employed by legal scholars at Harvard University is as vivid as it is unsettling: Is a bot a cuddly 'Pomeranian' that merely yaps, or a 'wolf' in disguise capable of inflicting irreparable harm? As we move through 2026, the debate over the ethical and legal liability of Artificial Intelligence (AI) has migrated from academic halls to the courtroom. The central issue is no longer whether AI can make mistakes—we know it can—but who pays the price when those mistakes have real-world consequences on human lives.

The Fallacy of the 'Tool' and the Reality of Autonomy

For decades, the law treated software as a simple tool, akin to a hammer or a car. If the tool failed, liability fell either on the manufacturer (if there was a defect) or the user (if there was misuse). However, Generative AI disrupts this balance. When a Large Language Model (LLM) produces a defamatory statement or provides incorrect medical advice leading to injury, the traditional concept of 'fault' becomes blurred. Developers argue they cannot predict every output of the model, while users often blindly rely on the machine's perceived authority.

Professor Ruth Okediji of Harvard points out that our legal infrastructure is ill-equipped for entities exhibiting 'semi-autonomous' behavior. If we treat AI as a 'Pomeranian,' we underestimate the risk of systemic bias and misinformation. If we treat it as a 'wolf'—an inherently dangerous object—we enter the realm of strict liability. In this scenario, the owner or creator is responsible for damages regardless of whether negligence was proven, much like the laws governing the ownership of wild animals.

The 'Black Box' as a Legal Fortress

One of the greatest obstacles to achieving justice is the so-called 'black box problem.' The complexity of neural networks makes it nearly impossible for a plaintiff to prove exactly how and why an AI arrived at a harmful decision. Tech corporations often use this opacity as a shield, claiming the algorithm's behavior was 'unforeseeable.' However, ethically and legally, ignorance cannot serve as an alibi for creating a public risk.

  • Civil Liability: Who is the 'proximate' cause? The provider of the training data, the model's creator, or the end-user?
  • Consumer Protection: Existing product liability laws must be adapted to cover intangible services that 'think' and evolve.
  • Transparency: The need for 'Explainable AI' (XAI) is becoming a legal necessity rather than just a technical preference.

The European Approach vs. Global Divergence

While the US debate often focuses on litigation and common law precedents, the European Union, through the AI Act, attempts to prevent harm via stringent regulation. Europe appears to be leaning toward a risk-based categorization, where 'high-risk' systems are subject to much stricter accountability measures. In Greece and other civil law jurisdictions, discussions are underway to adapt Civil Codes to the challenges of the digital age, seeking models that protect citizens without stifling innovation.

"Liability is not just a legal obligation; it is the glue of social trust. If we don't know who to sue, we won't know who to trust," remarked a participant at the Harvard symposium.

In conclusion, the transition from AI-as-a-tool to AI-as-an-agent requires a radical overhaul of our social contract. Whether it is a 'Pomeranian' or a 'wolf,' society must define the limits of the 'leash' and, more importantly, the hand that holds it. Justice in the 21st century will be judged by our ability to assign a face and responsibility to the anonymity of code.