Should AI companies be allowed to regulate themselves?

During a recent conference, Huang described AI simply as a complex computing system, rather than an unpredictable new entity. He argues that because AI is built by humans using software and hardware, it should be treated like any other technology product. He believes that if a company releases a dangerous or broken AI, the market will punish them, giving businesses a strong financial incentive to police themselves. His position effectively pushes back against the growing calls for new government safety standards specifically designed for AI.

At its core, AI is essentially a massive, sophisticated calculator. During a process called training, the system analyzes billions of patterns in data to learn how to predict the next piece of information, such as the next word in a sentence. When a company releases an AI product, they are offering an interface that lets users tap into those patterns. Safety in this context means trying to ensure the system does not produce harmful advice, leak private data, or assist in illegal activities. However, because these systems are statistical rather than logic-based, they are often unpredictable. An AI might work perfectly ninety-nine percent of the time, yet still generate unexpected and harmful responses in specific situations. Companies use various testing methods to reduce these risks, but achieving absolute safety is notoriously difficult, as the software is so complex that even its creators cannot always predict every output.

Huang’s approach suggests that we should trust the same companies that profit from these tools to determine their own safety limits. This is a high-stakes disagreement. On one side, proponents argue that government rules move too slowly and could stifle the pace of innovation, potentially preventing us from seeing the benefits of these tools sooner. On the other side, skeptics point to historical examples where companies have shipped products with serious flaws or harmful social consequences, arguing that the public should not have to wait for a disaster to occur before safety standards are enforced. The ultimate question is whether market pressure is actually strong enough to protect people, or if AI's potential to cause widespread harm requires a different set of rules than the ones we use for standard software.