AI Leaders Agree on Slowdown, But Hurdles Remain

Leading voices in the artificial intelligence industry have found rare common ground, agreeing that the rapid advancement of AI needs to be slowed. This consensus, articulated by figures such as Anthropic's Dario Amodei and OpenAI's Sam Altman, emerged following a stark warning from an Anthropic researcher about the potential threats AI poses to humanity.

Despite this agreement, translating the call for a slowdown into practice faces considerable challenges. Domestic and international competition, the drive for profit, and differing political stances, including past pushback from former President Donald Trump, all complicate efforts to pace AI development. Experts suggest that achieving this slowdown may require a fundamental reframing of what constitutes success in the field, moving away from speed and power towards reliability and safety, analogous to the aviation industry's focus on preventing accidents.

In theory, AI companies could unilaterally decelerate their model development. However, intense competition to build more powerful systems, often encouraged by a White House eager to maintain U.S. technological dominance over China, incentivizes rapid progress. Furthermore, major AI companies like Anthropic and OpenAI are pursuing potentially lucrative public offerings, adding financial pressure to the development race.

Proposed Solutions and Industry Commitments

Anthropic's Amodei has outlined a multi-faceted plan to address these concerns. One key proposal involves "frontier labs"—companies at the forefront of AI development—committing to provide "ongoing, employee-like access" to independent outside evaluators. These evaluators would be embedded within the companies, equipped with offices, access badges, and company laptops to monitor safety practices.

Anthropic has announced it is unilaterally adopting this measure, and OpenAI's Altman has expressed support, stating his company will follow suit. Altman also welcomed the idea of a federal framework for AI safety standards, emphasizing that companies do not need to wait for legislation to address safety concerns.

Amodei's plan also calls for government regulation and international coordination. He suggests that the U.S. government and other democratic nations could help frontier AI companies establish common safety standards and limits on unchecked progress. The proposal extends to coordinating with authoritarian governments, though Amodei acknowledges the significant difficulty in securing cooperation, particularly from China.

Amodei detailed several potential levels of international agreement. The most achievable, he noted, would be a ban on specific dangerous AI applications, such as the creation of biological weapons. A more challenging step would be a global agreement for pre-release testing of AI models for risks in cybersecurity or biological threats, potentially through a standards body. The most ambitious and difficult proposal is establishing a "speed limit" on AI models capable of recursive self-improvement—that is, models that can develop improved versions of themselves.

Amodei likened this to Cold War-era treaties that capped missile numbers to limit destruction while preserving deterrence. He argued that slowing the pace of AI development from "extremely fast" to "only somewhat fast" would yield significant safety improvements with minimal strategic disadvantage.

Obstacles and Skepticism

While the agreement on the need for a slowdown is noteworthy, experts express skepticism about its practical implementation. Sandra Wachter, a professor at the Oxford Internet Institute, described the required substantial coordination between companies and governments across nations as "highly unrealistic" given the current global political climate.

Wachter suggested that governments could increase accountability for AI companies regarding the risks their technologies pose and pursue regulations that limit access to essential resources like electricity and water needed for data centers.

The proposal for embedded external evaluators has received mixed reactions. While generally welcomed by the industry, some critics question the independence of these assessors and the process for establishing evaluation standards. Aidan Gomez, CEO of Canadian AI lab Cohere, voiced concerns that this could lead to further concentration of power among a few dominant labs. He also pushed back against the idea of antitrust waivers for frontier labs, arguing that the rules for such a consequential technology should not be written by a small group of commercially aligned companies.

Elham Tabassi of the Brookings Institution highlighted that without solidified and public measures, independent auditors would operate under a "voluntarily-provided, company-controlled" system. She stressed the importance of companies developing scientifically valid methods for testing and measuring AI models to ensure genuine safety improvements, which would bolster the effectiveness of any regulation.

The Trump administration's historical preference for a light-touch approach to the AI industry, aimed at fostering innovation and outpacing China, contrasts with the calls for a slowdown. The ongoing AI competition between the U.S. and China is often described as an arms race, a dynamic that experts believe is not conducive to collaboration.

Despite these challenges, some industry leaders remain cautiously optimistic. Zahra Timsah, co-founder of i-GENTIC AI, noted the rapid convergence of opinions among CEOs and the bipartisan attention from senators on oversight as positive signs. However, she cautioned that "attention is not the same as implementation" and the effectiveness of these developments remains uncertain.