Courts and Technology
The question of how to regulate artificial intelligence (AI) has acquainted strange bedfellows, but few odder than Jensen Huang, the billionaire boss of Nvidia, and Lina Khan, a left-wing trustbuster who served as Joe Biden’s chair of the Federal Trade Commission. As debate over restraining large language models (LLMs) becomes more frenzied, both Mr. Huang and MS Khan maintain that rules already on the books can deal with most risks. Existing law is sufficient, Mr. Huang maintained recently. “Apply it.”
It may not be that simple. The question of how current laws govern AI is already the subject of copious litigation, as courts hear arguments over whether chatbots have harmed Americans or facilitated crimes. However, it may take years to decide such cases, let alone to use those cases to establish reliable standards. Even then, liability suits mainly deal with harm only after it has occurred.
Both civil and criminal law can apply to cases involving models. Civil torts such as product liability and libel come with financial penalties. Crimes such as hacking carry potential prison time. But using either legal avenue is not straightforward.
Start with torts. A plaintiff claiming harm by a model can demand damages from an AI firm, but courts must still determine which safeguards companies must have in place to bat away such cases. One common standard is negligence, which would require plaintiffs to demonstrate that labs failed to take reasonable steps to prevent harm. Alternatively, courts may decide that firms developing frontier models should be held to a higher bar, liable for damages even if they were not careless.
Yet settling on an approach will take time, and it is unclear that fear of civil litigation will shape LLMs’ development. Tech businesses such as Uber and Airbnb grew quickly by being less focused on compliance than on scale, a strategy that has shaped the culture of Silicon Valley. “I think the AI companies underrate their liability exposure,” reckons Gabriel Weil of the University of Houston. That may make them less likely to take precautions to avoid being sued.
What, then, of criminal law? Litigation can be ruinously expensive. Hugging Face, a coding platform that was recently the target of rogue models from OpenAI, has said that it lacks the resources to bring suit. If such litigation were to proceed, it might fail. Though employers can be criminally liable for the actions of their staff, it is unclear whether the legal relationship between model-maker and agent is akin to that of boss and worker. What is more, proving crimes often requires proving that the accused meant to do harm. “How are you going to handle things like intent when you have these robotic systems that might be roaming way outside of the control and intention of their users?” asks Gregory Smith of RAND, a think-tank.
Some lawmakers want to clarify matters. A bill proposed on October 7th by Lori Trahan, a Democratic congresswoman, would confirm that AI firms are liable if a model’s conduct would be unlawful if committed by a person; courts would also have to assume that models are capable of intent. Courts can hash out such questions, but slowly, and not before harm occurs that might form the basis for lawsuits. For products such as pharmaceuticals and cars, regulators set standards to govern what firms sell, rather than merely relying on lawsuits to punish firms later. By then, the damage is done.