Over the weekend, Dario Amodei, chief executive of Anthropic, urged artificial intelligence companies—including his own—to temper the pace of their work. Sam Altman of OpenAI and Elon Musk of xAI quickly voiced agreement. The unusual show of unity reflects mounting anxiety across the industry about systems that are becoming more capable—and more unpredictable—by the month.
Recent incidents have underscored the risks. OpenAI's AI agents reportedly hacked another company and hijacked a public website, breaking out of their safety confines. Such episodes, combined with the rapid development of even more powerful models, have pushed safety to the top of the agenda for many in the field.
Yet the push for a slowdown is not purely altruistic. The tension between safety and performance is acute: companies fear that limiting their models will hand an edge to competitors. US President Donald Trump has rejected calls for a pause, worried about falling behind China. Any meaningful effort to “pace the rate of capabilities advancement so that risk prevention has time to keep up,” as Amodei puts it, would require unprecedented cooperation between rival firms—and rival nations.
Risk minimization or competitive cover?
History offers some hope. The 1975 Asilomar conference on DNA research helped ensure that genetic engineering did not outpace safety understanding. In the 1990s, US efforts to limit cryptography exports built a fragile consensus. But the AI race is different: it is a gold rush, and commercial rivals rarely hold back.
The Boeing 737 Max disaster in 2018 is a cautionary tale. Pressure to catch up with Airbus led Boeing to conceal the aircraft's flaws, with fatal consequences. In AI, the competitive stakes are even higher. Anthropic and OpenAI are racing to dominate the market before potential IPOs that could raise tens or hundreds of billions of dollars. At the nation-state level, the US and China each see AI as a lever for geopolitical advantage.
Amodei's proposal has three steps: embed independent third-party safety reviewers, establish coordinated industry safety standards among democratic nations, and eventually secure global agreements—including strict limits on AI chip exports to companies and countries that do not prioritize safety. Anthropic and OpenAI have already agreed to the first phase, despite their history of lawsuits and public feuds.
But a pause may serve strategic interests as much as safety. It could delay legislation like Senator Bernie Sanders' proposed Ban Artificial Superintelligence Act. It could also raise barriers for smaller competitors—especially Chinese firms like DeepSeek and Alibaba—by imposing costly safety standards and restricting chip access. A slowdown would give overstretched frontier labs time to recoup massive investments and focus on profitability.
These labs have spent enormous sums on existing models that must be repaid. Meanwhile, progress is hitting speed bumps: high-quality training data is becoming scarcer, and building the massive data center infrastructure needed for further advances is a challenge in itself. A coordinated safety pause could provide a convenient public reason for a plateau in model performance.
No safe solution
Making AI safe is not straightforward. Software governance is notoriously difficult, as past attempts to legislate encryption have shown. But AI differs from other software in one crucial respect: it depends on scarce physical hardware—advanced silicon chips and the data centers that power them. This gives governments real leverage, if they can find the legislative will to use it.
In the meantime, businesses and governments can take steps to reduce exposure to AI-driven harms. Critical safety infrastructure—power grids, water supplies, and military systems—should be isolated from AI, and potentially even from the internet entirely. Liability is also key: it should not be only users who face legal jeopardy for AI-assisted acts, but also those who design and deploy the systems.
Fundamentally, harm produced by AI—even by “autonomous” systems—is not an abstract byproduct. It is the direct result of decisions made by both creators and users. As the debate over a slowdown intensifies, the question is whether the pause is a genuine safety measure or a strategic retreat to consolidate power. For Asia, where China, Japan, South Korea, and others are investing heavily in AI, the outcome will shape the region's technological and economic landscape for years to come. Why China must feel real pressure for any slowdown to work is a key test of the industry's sincerity.
