Recent alarm over an alleged AI-driven breach of Medicare’s computer system has been contrasted with earlier Telstra and Optus outages that left Australians unable to reach emergency services. In those incidents, criticism fell on the corporations that ran the networks rather than on the hardware itself. The article suggests that the same approach should apply to AI-related failures.
The term “artificial intelligence” emerged roughly seventy years ago, when primitive mainframes were greeted with both fascination and fear. Early programmers often deflected blame by claiming “the computer made a mistake,” a phrase common in the 1960s according to the author. Over time, analysts recognized that errors usually stemmed from faulty data or poorly written code rather than from the machines themselves.
When a user asks a system such as ChatGPT or Claude to retrieve sensitive information and the request results in a security breach, responsibility does not rest with the software alone. Either the individual who entered the prompt or the company that created the model made a mistake, and both should be liable for any damage. If fault cannot be clearly assigned, the article proposes joint and several liability, meaning each party can be required to pay the full loss.
Traditional debugging involved locating bugs in operating systems, applications, or input data, sometimes even a literal moth caught in a relay. Agentic models, however, contain billions of parameters that cannot be inspected individually, making root-cause analysis far more challenging. The author argues that external “guardrails” are ineffective because the purpose of an AI agent is to bypass obstacles, and imposing constraints may simply be treated as another hurdle.
Imposing financial liability on AI firms is expected to slow the current “hyperscaling” race to build ever more powerful agents. The shift could encourage developers to prioritize safety over novelty, which the article claims would benefit both the environment and the broader economy. At the same time, the piece acknowledges that benign AI applications,such as improved search, document summarisation, translation, and code generation,will continue to evolve without posing existential threats.