Inside the AI Safety Debate: Helen Toner on the Race Toward Superintelligence, Corporate Governance, and the Future of Regulation

As artificial intelligence continues to advance at a breakneck pace, the intersection of technological capability, corporate governance, and national security has become one of the most pressing policy challenges of the modern era. To explore these issues, Georgetown University’s Center for Security and Emerging Technology (CSET) Executive Director Helen Toner recently joined The Verge’s Decoder podcast for a wide-ranging discussion on the reality of AI safety, the lessons learned from her time on OpenAI’s board, and what effective regulation might actually look like.

Toner is widely recognized in the artificial intelligence safety and policy community, most notably for her historic tenure as one of the four OpenAI board members who voted to fire CEO Sam Altman. That turbulent period—often referred to by insiders as "the blip"—ultimately concluded with Altman’s reinstatement, the resignation of Toner and other dissenting board members, and a massive corporate restructuring. Toner notes that OpenAI’s original nonprofit board was designed specifically to prevent the company from abandoning its core mission of building safe, beneficial AI in favor of raw profit and aggressive commercial growth. When the board attempted to exercise that oversight mechanism, the system buckled, illustrating the immense difficulty of keeping powerful commercial incentives in check.

That friction between safety commitments and commercial pressures has only intensified. Toner points out that the fundamental vision shared by the core researchers and leaders at frontier labs like OpenAI and Anthropic was never to become ordinary consumer software companies. Instead, their long-standing goal has been to pioneer machines that match or exceed human intelligence, eventually utilizing techniques like recursive self-improvement to spark an intelligence explosion.

While the general public often views AI through the lens of consumer applications, voice assistants, or search engines, frontier labs are intensely focused on autonomous systems and long-horizon capabilities. This creates a massive divergence between everyday experiences with artificial intelligence and the high-stakes ambitions driving the industry’s top tier. Toner emphasizes that while terms like Artificial General Intelligence (AGI) have become fuzzy and nearly useless due to shifting definitions, the underlying push toward superintelligence is taken entirely seriously by those building the technology.

Addressing the regulatory landscape, Toner discusses the daunting gap between rapidly evolving technology and the slower, often bureaucratic pace of government institutions. Despite widespread cynicism regarding lawmakers’ ability to understand or keep up with Silicon Valley, Toner observes a genuine and growing bipartisan thirst for rigorous, actionable information among members of Congress and executive branch agencies. While acknowledging that political lobbying and massive campaign spending inject unprecedented amounts of money into Washington, she argues that money has fundamental limits. When policymakers are faced with complex national security and safety issues, independent research institutions like CSET often find a receptive audience among staffers and officials eager to understand the technical details.

However, the policy challenges remain formidable, particularly when trying to govern systems that present complex risks such as biological weapon proliferation or autonomous cyberattacks. Toner cautions against falling into the trap of assuming that all risks can be easily neutralized through basic technical alignment or rigid constraints. Concepts of alignment—whether defined as preventing a model from executing harmful requests, ensuring it adheres to a corporate constitution, or forcing it to behave in accordance with human values—are far from being solved. Furthermore, the debate is compounded by profound philosophical uncertainties, such as whether advanced models possess any form of consciousness or moral status, questions that remain entirely unsolved by modern science.

The conversation also touches upon the looming geopolitical dynamics of the global AI race, particularly regarding China. While critics frequently use the threat of foreign competition as an excuse to avoid implementing domestic safety standards, Toner argues that national security and safety are not mutually exclusive. Safeguarding algorithmic secrets, preventing the unauthorized theft or distillation of frontier model weights, and fostering a shared threat perception regarding autonomous systems are critical components of a comprehensive strategy. If the risks associated with loss of control are genuine, international stability requires recognizing those shared dangers rather than treating the race purely as a zero-sum game of unbridled acceleration.

As the industry prepares for major commercial milestones, including Anthropic’s upcoming initial public offering and ongoing corporate restructuring across the sector, the pressure from public market forces will introduce entirely new governance dynamics. Whether these external pressures and increasing regulatory scrutiny will successfully balance commercial ambitions with vital safety thresholds remains to be seen. As a new legislative cycle begins, Toner stresses that the ultimate test will not be whether institutions can halt innovation entirely, but whether they can establish realistic, enforceable standards to manage risk before frontier capabilities outpace society’s ability to adapt.