A new study published in the prestigious journal Proceedings of the National Academy of Sciences challenges the prevailing notion that any government intervention in the field of artificial intelligence automatically enhances safety. The authors reach a paradoxical conclusion: poorly thought-out, "weak" control measures can make the final product more dangerous than if it had been developed without any oversight at all.
The Mathematics of Risk: Game Theory vs. Intuition
The study is based on a complex theoretical economic model built on the principles of game theory. Researchers modeled the behavior of key players in the AI market: giants creating foundational models (such as OpenAI, Google, Anthropic) and companies implementing these technologies in real-world sectors — from medical diagnostics to chatbots in online stores.
The modeling revealed a critical vulnerability in the current logic of regulation. If rules target exclusively downstream companies, ignoring the developers of foundational models, this creates a dangerous imbalance. Developers of base systems begin to cut corners on safety measures, subconsciously calculating that their partners will take responsibility for the safety of the final product.
The Prisoner's Dilemma in Code
This process is described by the classic Prisoner's Dilemma. In this problem, two rational players can either cooperate or betray each other. If both cooperate, the result is optimal for everyone. If one decides to cooperate while the other betrays, the first ends up in the worst position. Not knowing the partner's intentions, each often chooses betrayal — and in the end, both get a worse result than they could have.
This is exactly how weak regulation works in the AI sector. The free rider effect emerges: regulation becomes a tool used by a large provider to shift the burden of safety onto a specialized downstream partner. Without clear and equal obligations for all participants in the chain, each side chooses a self-defense strategy at the expense of overall safety.
Regulation as a Tool for Trust
The authors of the study insist that the way out of the deadlock lies in strict regulation of the entire production chain. According to them, "stronger and properly directed regulation can bring mutual benefit to all participants," improving both the safety of the final product and the economic performance of companies.
In this context, regulation is viewed not as a restriction on profit, but as a tool that eliminates uncertainty and creates conditions for trust between market players. People often perceive AI as a single entity, but in reality, it is a complex system with many participants. To regulate it meaningfully, one must consider the entire value creation chain, not just a single provider.
Political Context: Race with China vs. Safety
The study comes at a time of sharp disagreements in American politics regarding the future of AI. One side, supported by the current Trump administration, advocates for minimal restrictions, appealing to the need to outpace China in the technological race. Proponents of strict regulation, however, warn that in the pursuit of profit, the industry underestimates real risks: from the psychological impact of AI services on users to the environmental consequences of building data centers and a massive wave of unemployment.
