Vi må bygge systemer som tåler at KI feiler

As artificial intelligence takes on a larger role in safety-critical systems like power grids, industrial plants, and autonomous ships, we cannot expect the technology to become completely error-free or fully predictable. Modern AI models are often too complex to explain internally, and they can deliver convincing yet incorrect answers or fail in unfamiliar situations. However, we do not need to understand every underlying calculation to use AI responsibly.

Drawing on established safety principles, we must build systems that assume AI will sometimes fail and include barriers to limit the consequences. For example, if an autonomous ship misinterprets its surroundings and picks a dangerous course, the surrounding system must detect the error and intervene through independent monitoring, hard limits, human oversight, or automatic safety mechanisms.

Because Norway and Europe already lag behind the United States and China in core AI development, waiting for perfect technology is not an option. Avoiding AI altogether carries its own financial risks for industries that need to remain competitive, but poor implementation can lead to serious accidents. Instead of choosing between speed and safety, proper risk management allows society to move forward quickly where it is safe while maintaining strict barriers.

Norway holds a strong position in maritime sectors, energy, and process industries, supported by deep domain knowledge and long experience with safety-critical systems. By focusing on what can go wrong during specific tasks and building robust safeguards, high-risk industries can safely adopt artificial intelligence without needing to remove all uncertainty from the technology itself.