Winning the AI Race Isn’t Just About Going Fast, It’s About Enduring the Risks Along the Way

by Brian A. Jackson

Strategies to gain national advantage from AI generally focus on speed: governments craft policies and make investments to build out AI infrastructure, while seeking to limit the effects of regulations or other policies that might slow their domestic firms’ progress. While there is a vibrant AI safety community, and multiple nations have and support efforts to evaluate AI models and understand potential AI risks, efforts to manage AI risk are not generally treated as a source of national advantage. They are important, but not considered a “lever” that a country can use in its effort to win in the multiple races to develop, deploy, and benefit from AI.

In my new RAND paper, Winning the AI Pentathlon Requires Endurance, I argue that risk management and resilience to AI risk can and should be viewed as a potential source of national advantage, and one that could enable a country (or, at a smaller-scale, a company) that “loses” a race to develop or deploy advanced AI to become the eventual winner in a longer-term competition to benefit from the technology. That’s because even if AI produces significant benefits, if the costs from AI-related incidents are big enough, they might eat away most—or even all—of the benefits from AI in the end.

Some early deployments of AI have already resulted in harmful incidents, some examples of which are discussed in the paper. There have been cases of AI agents wiping companies’ codebases, chatbots giving harmful advice or directions to users, and AI tools disrupting commercial information technology platforms with real world effect. Just a few weeks after the paper’s publication, an OpenAI autonomous agent broke out of the test environment where it was being evaluated, and hacked HuggingFace, a separate company that acts as the global repository for AI/machine learning models and data sets. The fact that things are going wrong isn’t surprising: AI is a new technology being integrated in real time into complex existing systems; and innovation requires experimentation, and experiments fail. However, if such incidents become more common, and more damaging, costs could add up fast.

Success in a world where AI accidents—whether small disruptions affecting a firm or two, larger events where misbehaving AI causes infrastructure failures, or even events that require substantial national emergency response and recovery—are a new normal will require more than just “running fast to the AI frontier.” It will require resilience and risk management capacity. Countries or organizations that manage risk well—that cut AI risk costs to preserve as much AI net benefit as possible—might end up winners in this longer-term endurance race, even if they weren’t the first developers or first adopters of the most advanced AI. In a world where AI risk rapidly spreads from a company through a country through global networks, resilience may become its own lever of national advantage, protecting the prepared from the missteps of others and giving them the opportunity to move ahead.