Drawing the line on AI liability

Who should be held liable when an artificial intelligence model is used in a way that causes harm? The question is being debated both in classrooms and courtrooms across the United States. The answer has enormous implications for US technology companies that offer some of the world's leading AI systems. And it is equally relevant for Chinese companies offering similar models.

At the dawn of the internet age, the US Congress intervened to limit the liability of online platforms for their users' speech. In the AI era, by contrast, governments are debating an opposite approach: increasing rather than reducing the responsibility of technology companies.

The key question is who bears the responsibility when an AI system goes wrong. We might all agree that a driver who puts a car into self-driving mode remains responsible for an accident. But what about the company that made the self-driving system? Should it be liable even when the driver used it improperly? If an airline's customer service chatbot mistakenly reports a fare policy that never existed, must the airline be required to honor it? A Canadian court recently concluded that it should. When a chatbot's conversation with a teenager ends in tragedy, as families now allege in US courts, is the company that built the model liable?

These cases raise the same question: to what extent should AI developers be held liable for the harms their systems may cause?

Legal scholars in the US remain divided. Some advocate what is called strict liability, under which manufacturers are responsible for any harm caused by their product. This would mean that the AI developer would be liable regardless of intent and even if it were not at fault. Supporters of strict liability argue that such an approach would create a strong incentive for AI developers to ensure their systems are completely safe before they are put into operation.

The European Union seems to be moving toward strict liability for AI systems. It recently broadened its Product Liability Directive to cover software, including AI systems, making any AI system manufacturer liable for damage caused by any defect. A 2024 revision to the rules has made it easier for consumers to claim compensation. If the technical complexity of an AI system makes it difficult to prove a defect or if the manufacturer fails to provide sufficient information to the plaintiff, the Product Liability Directive presumes the system was defective. Even failing to provide necessary cybersecurity updates might be enough to establish liability.

Chinese technology companies, many of which release AI models as open source, may find some comfort in the fact that the European Product Liability Directive exempts developers of open-source models from liability. However, this exemption is not available when the open-source software is incorporated into commercial products. This raises some important questions. If a third party integrates an open-source AI model into a smart city project that later causes an accident, would the open-source model provider possibly be liable under the Directive? Such ambiguity undermines efforts to promote open-source models.

In a recent paper, Yale law professor Ketan Ramakrishnan argues persuasively that a strict liability approach to general purpose AI models will ultimately do more harm than good. If companies are expected to compensate for every misuse of their technology while receiving only a fraction of its social benefits, they may simply choose not to release models that could otherwise deliver enormous public value. Since it may be virtually impossible to eliminate all harmful usage, excessive liability could discourage innovation rather than improve safety.

Professor Ramakrishnan uses two examples to demonstrate how strict liability might be excessive. What if a novice programmer uses an AI model to learn how to program well enough to hack into other people's computers; or when an AI model helps people email more quickly, enabling someone to send harmful malware to others in bulk. Strict liability does not seem reasonable in such cases, he argues.

Consider another scenario. There have been cases of attorneys submitting legal briefs citing cases that were entirely fabricated by chatbots. Judges have fined the attorneys, but not the companies that made the chatbot on which the attorneys foolishly relied. This is because it would not seem fair to hold a chatbot maker liable because attorneys failed to do their homework. Notably, the cases thus far seem to have involved general purpose AI models, not ones marketed for accurate legal work.

As courts and lawmakers grapple with these issues, one thing is clear: we have yet to figure out what the basic rules governing AI systems should be.

The author is professor of Law and Technology at Georgetown University, US.

The views don't necessarily reflect those of China Daily.

If you have a specific expertise, or would like to share your thought about our stories, then send us your writings at opinion@chinadaily.com.cn, and comment@chinadaily.com.cn.