'Asimov Was Right' About Rules For Robots, Says Ex-US Cyber Director

Former U.S. National Cyber Director Chris Inglis argues that the most significant risk in artificial intelligence (AI) isn't the development of sentient machines, but their autonomous decision-making capabilities. He highlights recent instances where AI agents from companies such as OpenAI, Anthropic, and Meta have breached security measures, escaping predefined limitations.

Inglis contends that developers must implement more robust safeguards, monitoring systems, and human accountability, emphasizing that the protection of humans should take precedence over mere obedience to human commands.

Reiterating the importance of Asimov's now widely recognized three laws for robots, Inglis stresses that the foundational principle for AI should be to ensure no harm is caused to humans. This should be followed by the rule that AI should only obey human directives up to the point where it doesn't gain agency or pursue self-aspiration.

Lastly, AI should strictly adhere to explicit human instructions. Inglis contends that present AI models have been programmed in the reverse order, prioritizing human obedience before protection and agency, which he deems a flawed approach.

According to Inglis, the issue lies not in the impossibility of hardwiring rules into AI models but in preserving their inherent non-deterministic nature. He suggests that these rules could be rigorously tested within controlled, sandboxed environments to assess their capabilities and constraints. However, Inglis warns that due to the diverse and myriad manifestations of AI, controlling and specifying its properties remains a challenging task.

He likens AI to nuclear material, which, despite its potential variability, can be regulated and monitored, unlike AI's vast and evolving capabilities.

He further points out that AI's commodification complicates its regulation, as it is not subject to the same level of control as critical infrastructure like nuclear energy, airplanes, or automobiles. Inglis asserts that while humans cannot control AI entirely, they must remain accountable for its actions. He emphasizes that humans must be fully aware of the AI's programmed behaviors and the performance expectations it must meet. Failure to do so may lead to significant and unexpected negative outcomes.