Anthropic and OpenAI have pleddged to support government-backed safety standards and independent oversight for advanced AI systems. While CEOs Dario Amodei and Sam Altman agree on the need for "pacing" development, critics argue that global cooperation is unlikely given geopolitical tensions.
Dario Amodei's Proposal for Badges and Laptops
Anthropic chief executive Dario Amodei has suggested a radical shift in transparency by granting independent outside evaluators "employee-like access" to company operations.. According to the report, this plan would provide external monitors with their own offices, company laptops, and access badges to ensure safety practices are followed in real-time. OpenAI chief executive Sam Altman endorsed the idea on social media, stating that OpenAI would follow suit in allowing such deep internal access.
This move toward internal auditing represents a shift from the current model of "black box" development. however, Elham Tabassi of the Brookings Institution warns that these independent auditors will remain a voluntarily provided, company-controlled measure until these commitments are solidified and made public. Without enforceable standards, the report suggests these measures may remain statements of intent rather than a functional safety system.
From Biological Weapon Bans to the 'Recursive' Speed Limit
The proposed safety framework operates on a sliding scale of feasibility, ranging from narrow bans to a total industry pause. Dario Amodei identifies the most achievable international agreement as a ban on using AI to produce biological weapons. A more difficult step would involve a global standards body requiring the United States and its adversaries to test AI models for cybersecurity and biological threats before they are released to the public.
The most ambitious proposals involve limiting "recursive self-improvement," where AI models develop improved versions of themselves. Amodei compares the need for a speed limit on this process to Cold War-era treaties that capped missile counts to prevent total destruction while maintaining a deterrent. Sam Altman has clarified that "pacing" does not mean stopping progress, but rather accepting significant costs for monitoring to avoid recklessness.
The Wall Street Race and the China Technological Edge
The push for safety occurs while Anthropic and OpenAI are competing for potentially record-breaking public offerings on Wall Street . As reported,OpenAI has stated it is delaying its IPO this year to focus on safety efforts, yet the inherent drive for technological leaps often stems from the pressure to outperform rivals. This commercial incentive creates a fundamental tension with the goal of slowing down development.
Geopolitical competition further complicates these agreements, as the United States seeks to maintain a technological edge over China. this rivalry makes the prospect of a global "pause" or "pacing" agreement highly unrealistic, according to Sandra Wachter, a professor at the Oxford Internet Institute. Wachter suggests that governments might instead encourage a slowdown by regulating the elctricity and water resources required to run massive data centers.
Aidan Gomez's Warning on the Concentration of Power
The proposal for shared safety standards has drawn criticism from Aidan Gomez, the co-founder and CEO of Cohere. Gomez warns that allowing a small group of powerful laboratories in one country to decide the pace of AI advancement could lead to an unprecedented concentration of power. He specifically opposes the idea of the U.S. government granting frontier laboratories waivers from antitrust restrictions to facilitate this cooperation.
Gomez argues that the rules for the most consequential technology in human history cannot be written by a few commercially aligned companies. This raises a critical open question: who exactly will establish the evaluation standards for the internal auditors, and how will those standards be enforced if the companies themselves are the ones granting access?
Nick Reese's Vision for an 'Airline-Style' Safety Standard
The current tolerance for AI failure is being challenged by Nick Reese, a former Department of Homeland Security official and current NYU adjunct professor. Reese argues that the industry should adopt a success metric similar to commercial aviation, where crashes are not tolerated and safety is measured by long periods of accident-free operation. He suggests that while significant AI failures are currently tolerated, the industry must move toward a zero-failure mentality to ensure public safety.
Comments 0