The push to pause artificial intelligence development is gaining momentum, with hundreds of protesters marching outside major AI labs in San Francisco, yet the world's leading AI researchers remain divided on whether stopping progress is even possible. On March 21, about 200 protesters gathered outside the headquarters of OpenAI, Anthropic, and xAI, demanding a halt to AI development over existential risks. But behind the scenes, experts are grappling with a harder question: if a global pause is unrealistic, what can actually be done to make advanced AI systems safe?

What Are Experts Actually Worried About When They Discuss AI Extinction Risk?

The concern centers on artificial general intelligence (AGI), a theoretical AI system that would match or exceed human intelligence across all domains. Frontier model developers are deliberately building toward superintelligence, according to Jonas Vollmer, chief operating officer at the non-profit AI Futures Project. The danger isn't hypothetical. "If you have an AI system that's smarter than all humans and it concludes that humans are kind of inconvenient to have around for its goal, which is science advancement, then it might as well get rid of them," Vollmer explained.

Several AI researchers currently estimate the existential threat from AI at 10 to 15 percent, which experts argue is significant enough to warrant serious attention. Stuart Russell, a computer science professor at the University of California, Berkeley, emphasized that this level of risk demands action. Organizations like the Future of Life Institute, Stop AI, and Pause AI are working to raise public awareness through protests and open letters, hoping to build political support for safety measures.

Why Can't Governments Just Ban AI Development?

The practical answer is straightforward: no single government can enforce a global ban. Subbarao Kambhampati, a professor at Arizona State University, put it bluntly: "How are you going to stop the development of AI systems? Even if one government actually agrees with the ban, it is not like we control the entire world."

This geopolitical reality has shifted the focus among safety researchers away from calls for a complete pause and toward developing technical safeguards. The challenge is particularly acute with what researchers call "agentic systems," which represent the next frontier in AI capability. Unlike current AI systems that cannot execute plans independently, agentic systems will have application programming interface (API) access to real-world actions, multiplying the potential for harm if they malfunction or behave unexpectedly.

How to Build Safer AI Systems: What Researchers Are Actually Doing

• Technical Safety Solutions: Rather than pursuing bans, researchers are developing technical approaches to control rogue AI systems and prevent unintended consequences from advanced models.
• Verification and Licensing: Stuart Russell advocates for verification processes and licensing requirements for AI systems before deployment, similar to how medications or aircraft are regulated.
• Agentic System Controls: Kambhampati is researching ways to ensure AI systems don't execute plans unless there's extremely high confidence the action won't cause damage.
• Non-Profit Safety Initiatives: Yoshua Bengio, founder of Mila (Quebec Artificial Intelligence Institute), launched LawZero, a non-profit focused on building technical solutions for safe AI systems.

Kambhampati stressed the importance of this work: "Safety is important to make sure that you don't actually execute a plan unless you know for sure that the probability that it will basically cause damage is extremely low and I'm doing research in that area."

Kambhampati

The divide between pause advocates and pragmatists reflects a deeper tension in AI discourse. While pioneers like Geoffrey Hinton and Yoshua Bengio have signed open letters supporting a pause, they acknowledge that a complete halt is unrealistic. Instead, the focus has shifted to what can actually be implemented: technical safeguards, regulatory frameworks, and research into AI interpretability and control.

Meanwhile, the broader AI industry continues advancing rapidly. The OpenAI Foundation, established with substantial assets, announced plans to invest at least $1 billion over the next year, though critics note the funding allocation raises questions about priorities. The foundation is directing resources toward life sciences, jobs and economic impact, and "AI resilience," with only a portion dedicated to AI model safety. This allocation reflects a broader pattern in which safety concerns, while acknowledged, often take a backseat to other priorities in the AI ecosystem.

The reality facing policymakers and researchers is uncomfortable: the existential risks from advanced AI appear real and significant, but the tools to address them are limited. A global pause is politically and practically impossible. Bans can be circumvented. What remains is the harder, slower work of building technical solutions, establishing safety standards, and hoping that the pace of safety research can keep up with the pace of capability development. Whether that's enough remains one of the most consequential open questions in technology today.