A New Phase in the Safety Discussion
For years, the debate over advanced AI was a binary argument between acceleration and outright shutdown. One side argued for pushing forward at maximum speed to unlock economic and scientific benefits, while the other warned of existential risks, at times
calling for a complete pause on development. Now, a more nuanced and practical conversation is taking center stage. Industry leaders, including the CEOs of major labs like Anthropic and OpenAI, are moving the discussion from 'if' we should manage development to 'how'. The consensus is growing that an unchecked race to build the most powerful models is unsustainable. Instead, a deliberate, coordinated approach to pacing development—slowing down just enough for safety measures to keep up with new capabilities—is being championed as a viable path forward.
What 'Pacing' Actually Means
The concept of "pacing" is not about halting progress. Rather, it's about building guardrails into the development process itself. In a widely discussed proposal from September 2026, Anthropic CEO Dario Amodei outlined a framework for ensuring that companies devote adequate time to testing and safeguarding their most advanced models before deploying them. This means slowing the rate at which AI capabilities improve, giving researchers critical time to work on alignment and control without sacrificing commercial or strategic advantages. The idea is to prevent a scenario where a company, pressured by competition, releases a powerful but poorly understood system. Proposals include embedding independent, third-party evaluators within AI labs to monitor safety practices and verify claims before a new, more powerful model is trained or released. This is a significant shift from a race to the top in performance to a race to the top in safety.
The Push for Global Coordination
Effective pacing cannot happen in a vacuum. If one company slows down, another might speed up to gain an edge. This is why global coordination has become an essential part of the debate. The conversation now includes proposals for international bodies to oversee AI safety, similar to how the International Atomic Energy Agency monitors nuclear technology. Recent international meetings, like the AI Seoul Summit, have seen major tech companies and governments agree to collaborate on defining risk thresholds and improving transparency. OpenAI has called for the creation of global technical standards for measuring capabilities and reporting safety incidents. The goal is to create a unified framework that prevents any single actor from creating a dangerous imbalance by developing AI irresponsibly. Even geopolitical rivals are beginning to engage, with the US and China recently agreeing to establish a communication channel for AI-related incidents.
The Practical Hurdles Are Immense
Despite the growing consensus, the path to effective global coordination is fraught with challenges. The primary obstacle is geopolitical competition. Strategic rivalries, particularly between the US and China, risk politicizing what should be technical safety standards. Getting fiercely competitive companies to cooperate also raises legal issues, with some critics arguing that pacing agreements could be seen as anti-competitive cartels designed to manage investment costs under the guise of safety. Furthermore, there is no universal agreement on what constitutes 'safe' AI or at what capability level development should be slowed. Different nations have different priorities; some are focused on innovation and economic growth, while others prioritize ethical considerations and risk mitigation. Forging a truly global agreement will require navigating these conflicting interests, a diplomatic and technical challenge of unprecedented scale.
