In a high-stakes White House meeting on Tuesday, the Trump administration discussed a voluntary framework for testing cybersecurity risks in frontier AI models with industry giants, including Google, OpenAI, and Anthropic. However, the lack of transparency surrounding this framework has drawn sharp criticism from Senate Democrats.
Decrying 'Unpredictable' Governance
Senator Kirsten Gillibrand, joined by four other Democratic senators, launched a scathing critique of the administration's AI policy, labeling it "unfocused" and "ad-hoc." In a letter released before the meeting, Gillibrand argued that the White House's constantly changing dictates jeopardize America's economic security and push global customers toward Chinese alternatives. The letter highlighted the Commerce Department’s June 12 directive that forced Anthropic to pull its Fable 5 and Mythos 5 models offline, as well as a request for OpenAI to limit the rollout of its GPT-5.6 system.
The Chinese AI Allure: Cost and Customization
While political uncertainty is a factor, experts suggest the primary driver for adopting Chinese AI is financial. Sam Bresnick, a research fellow at Georgetown’s CSET, explained that Chinese open-weight models are enticing due to their low cost and fine-tunability. "Cost and fine-tunability are the drivers of Chinese AI adoption," Bresnick noted, adding that downloading weights for specific applications is often more efficient for companies than paying a premium for closed, proprietary US models.
Companies like DoorDash are already splitting workloads, keeping sensitive data with US providers while off-loading routine analytics to dramatically cheaper Chinese models. However, this strategy carries risks of "economic coercion" from Beijing, which has a history of using market access as a foreign policy tool and may eventually block foreign access to its most advanced models.
The Path Forward
Gillibrand's letter pressed the administration for clear standards on judging national security risks and the legal authority behind export controls. Bresnick concluded that for the US to remain competitive globally, it must address the need for more open and cost-effective models, as the current proprietary approach of leading US labs may not serve the broader tech ecosystem's needs.
