Over the weekend, leaders of several key AI companies called for a slowing—or pacing—of frontier AI development. Why now?

Anton Leicht: The pace of AI development is breaking away from the pace of governmental oversight and even internal safety measures. Researchers increasingly find that their AI models are helping them develop newer, better models faster and faster. Some see the beginnings of “recursive self-improvement,” a process that could rapidly develop extremely powerful AI systems. But at the same time, safety measures keep failing, and internal incidents keep piling up. Most developers agree they’d rather catch up on safety before continuing to accelerate.

Scott Singer: AI capabilities are moving lightning fast, and most institutions—including governments—are moving slowly. Frontier AI companies recognize that it’s in their enlightened self-interest to slow down—they see the benefit of learning more about the technology and its potential risks before proceeding further. The question now is what the companies and governments will do during this period of slower development, and whether China gains ground or faces similar trade-offs around safety in the next few months.

What would need to happen—both in government and in the private sector—to implement pacing the frontier?

Anton Leicht: First, industry needs to agree on an actual standard for what constitutes “paced” research and development. Is it about models’ capabilities and their thresholds? The allocation of computational power? How internal resources are distributed between deploying current models and developing new systems? Anthropic CEO Dario Amodei has laid out his ideas, but we haven’t heard much substantive engagement from the rest of the industry yet.

Second, once industry leaders do agree, they need to determine how to evaluate compliance. Embedded evaluators, as suggested by Amodei, seem like the most obvious mechanism: These independent third parties would have access to internal models, logs, processes, and communications channels so that they can assess safety practices and investigate incidents. But government agencies or mutual investigations by industry players could also play a role.

Third, the government would need to agree this pacing is desirable. It would need to remove regulatory barriers to coordination, codify industry consensus to make sure there are no defectors, and eventually internationalize the approach. Right now, President Donald Trump has indicated he’s not interested in that, and Congress seems unlikely to act in time.

Scott Singer: A statement signed by nearly 1,400 employees of frontier AI companies calls for “tools needed to deliberately pace the frontier of automated AI development.” In a commercial and geopolitical environment where trust is scarce, these verification tools are essential: They allow companies to demonstrate they are following through on certain claims they are making. But the science of verification is nascent. It is not yet clear what is technically possible or what would be most useful for both domestic and international governance. Nonetheless, investing in these tools and figuring out how to use them will be critical to actually pacing the frontier.

In the immediate aftermath of U.S. company calls for pacing, Beijing has rejected the idea of slowing down. China’s Ministry of Foreign Affairs characterized calls for pacing by Amodei as “fearmongering,” reflecting in part fierce opposition to the prospect of the United States containing China’s economic development. But China’s reaction probably reflects geopolitical dynamics rather than a permanent stance. Its position may well change as its companies begin to see more evidence of potential catastrophic risk manifesting in its own models.

What does it mean that Trump and the major AI companies are at odds?

Anton Leicht: The administration could make implementing industry near-consensus much more difficult in two ways. First, not all of the frontier AI industry is on board, so even if one major developer defects from the consensus, we still face almost as much risk. Second, some argue there may be antitrust provisions the president can wield to stop industry coordination.

Scott Singer: There is public tension, yes, but also extensive government-company collaboration. For all the Anthropic-Pentagon tension earlier this year, U.S. AI companies have become central collaborators in promoting U.S. national security. Beneath the conflict, companies such as Anthropic are powering the U.S. government’s integration of advanced AI systems.

How might this debate play into the upcoming meeting between Trump and Chinese President Xi Jinping?

Scott Singer: Managing AI risks has increasingly emerged as an important discussion topic for Trump and Xi. Even before Amodei’s call for pacing the frontier, the leaders were likely to discuss AI safety at their summit. But as the conversation around pacing the frontier grows in the United States, the goal for the Trump-Xi meeting should not be an international treaty. Instead, the goal should be to identify areas where both sides have self-interest in getting a deal done. For example, both sides share concerns ranging from misuse by nonstate actors in financial sectors to child safety. And if AI crises cross borders, which they may well soon, both sides will want to show that they can manage those international crises together effectively.

Anton Leicht: Domestically, some minimal interest in cooperation will be important to sustain the idea of a domestic pacing of the frontier. Both the AI developers and the U.S. government find winning the AI race important, so if a pacing effort might appear to erode the current American lead, it will face much more pushback in the medium term.

What guardrails should AI leaders and policymakers consider going forward?

Anton Leicht: Embedded evaluators are an obvious first step. This entity-based approach is more promising than model-related guardrails right now, because most of the most important risks arise from internal deployments of AI models in testing and development, long before they are released. We’ve seen that in the most recent hacking incidents, including the OpenAI agent swarm that attacked Hugging Face’s internal infrastructure. These models were never available to any customer; they broke out of their testing environment long before deployment. It makes sense for the risk to be concentrated where the most powerful models are first deployed—before they reach the public—but that’s also when they don’t have all their safeguards in place.

Scott Singer: The OpenAI–Hugging Face incident underscored the importance of learning from critical safety incidents before they cause widespread harm, such as significant property damage or even death.

Building robust federal incident reporting will be critical to make sure governments genuinely understand the risks that are emerging from frontier AI. In addition, well-designed incident reporting systems will allow governments to disseminate key lessons and takeaways across industry, reducing collective risks. Building strong domestic channels in both the United States and China will be essential if the two sides hope one day to share more information with each other and manage cross-border crises constructively.