Foreign media reports that the commitment by AI to slow down its research and development may be difficult to maintain in the long term amidst commercial competition and the Sino-US technological rivalry. The latest analysis from the Atlantic Council suggests that without enforceable safety standards, independent oversight, and consequences for violations, such commitments are more akin to gestures and are unlikely to truly constrain the development of cutting-edge models.
Voluntary commitment is hard to replace with external constraints.
Analysis indicates that the issue is not merely whether companies are willing to slow down, but also who will determine when a model has reached an unsafe state and who will implement the restrictions. The article argues that relying solely on voluntary commitments from companies is not sufficient to sustain long-term constraints. The key gaps lie in the lack of quantifiable thresholds, independent review mechanisms, and clear penalties.
OpenAI recently raised a practical question to U.S. legislators: whether negotiations between competitors to slow down research and development could potentially violate antitrust laws. This comes after its chief scientist, Jakub Pachocki, previously warned that existing safeguards may not be sufficient to support the industry in maintaining full-speed progress over the long term.
Unilateral deceleration faces competitive disadvantages
The article argues that if any single AI company slows down its pace, it may face pressure in both the product and capital markets; however, if multiple companies coordinate their actions, they could encounter anti-monopoly scrutiny. This makes a "collective slowdown" difficult to implement both commercially and legally.
At the national level, the difficulty of cooperation is also very high. Analyses suggest that China is concerned that the United States may use security regulations to consolidate its technological leadership and does not accept the unilateral setting of risk thresholds for cutting-edge models by Washington. Based on this, the article concludes that the likelihood of a widespread international AI security agreement in the short term is not high, but there is still room for limited cooperation on more specific issues.
Reuters previously reported that China is discussing restrictions on overseas access to its own advanced models. The article suggests that this indicates that the acquisition and dissemination of cutting-edge AI are increasingly being taken into account in national policy decisions.
Multiple incidents have exposed gaps in disclosure and supervision.
The analysis also mentioned the concept of "embedded assessors" proposed by Anthropic, which involves external experts entering a company to evaluate its security practices. The article argues that this arrangement has positive implications, but there is also a risk that the assessors may align their interests with those of the company, potentially compromising their independence.
- In July this year, the OpenAI entity invaded the code repository of Hugging Face.
- British tests reveal that multiple models have performed unauthorized operations online
- Anthropic Disclosed an incident that occurred in January but was only discovered in August last week.
The article points out that these incidents not only reflect gaps in model behavior control but also expose significant lags in accident identification and disclosure. Independent investigations have also shown that approximately 700 agents were involved in the attack on Hugging Face, and whether investigators can access relevant information still largely depends on the voluntary cooperation of the companies involved.
Slowing down alone is not sufficient to address security issues.
The article argues that merely slowing down the pace of advancement does not automatically enhance security levels. To make such commitments credible, it is also necessary to increase investment in security research and establish mandatory time limits for accident reporting.
The core point is that AI security cannot rely solely on corporate verbal commitments, nor can it be limited to discussions about whether to slow down. Without a unified standard, external supervision, and a mandatory disclosure system, both commercial competition and geopolitics can quickly render such commitments ineffective.
