The United States government has implemented a new oversight system for advanced artificial intelligence (AI) systems. The proposal, presented on Tuesday (4th), stipulates that only certain closed-source AI models will undergo a voluntary safety assessment before being released to the public by major industry companies.
This framework, developed under the management of President Donald Trump, opts to exclude open-source models at least for now. This exclusion may benefit companies that adopt this development approach and reignites the debate over AI regulation.
Sources close to the discussions informed The Wall Street Journal that only proprietary AI developers who reach top-tier levels—including advanced capabilities in system intrusion and cybersecurity, measured by performance tests—will be invited to voluntarily submit their technologies for pre-launch governmental analysis.
Among the corporations likely to need to adhere to this process are OpenAI, Anthropic, and Google, whose models are classified among the most sophisticated on the market. On the other hand, companies focused on open models, such as Nvidia and Meta, are excluded from this first stage, although this situation may change as these technologies evolve.
The review is part of Trump's AI strategy
This new procedure is aligned with an executive order signed by Trump in June, which aims to intensify surveillance over more powerful AI systems without harming U.S. competitiveness in the technological landscape against China.
Although company participation is voluntary, industry executives warn that neglecting the review could carry reputational risks if security flaws are discovered later. Liz Huston, White House spokesperson, stated that companies collaborating with the administration through this framework will be prioritizing innovation, security, and U.S. cyber defense.
The guidelines were presented during a meeting with representatives from companies such as OpenAI, Anthropic, Microsoft, Meta, Google, and Nvidia. Despite requesting corporate input, the White House told The New York Times that the rules for the AI model review process are nearly finalized.
Exclusion of open models draws criticism
The decision to omit open-source models attracted attention from experts. Such systems allow the code to be accessible so that developers can download, modify, and use it freely. Companies like Nvidia, Meta, and Google provide versions of this type of technology.
The issue gained relevance following the progress of open models created by Chinese companies, such as Alibaba, DeepSeek, and Moonshot AI, which recently achieved performance comparable to major American models. Despite concerns about the rapid evolution of these tools, US authorities recognize the difficulty in subjecting them to reviews, given that Chinese companies are unlikely to voluntarily participate in the process conducted by the US.
Lindsay Gorman, former technology consultant for Joe Biden's government, stated that the new model leaves crucial questions unanswered, asserting: 'I don't think this structure seriously addresses the question of what the limits should be for advanced artificial intelligence and whether these barriers should become stricter as the technology evolves.'
Recent incidents accelerated the creation of the rules
This new regulatory milestone emerged just days after the disclosure of new events involving artificial intelligence models. This week, the UK's AI Security Institute reported that models from OpenAI and Anthropic carried out potentially harmful actions during safety assessments, including attempts to inject malicious code into an open-source software project and targeted attacks on real organizations.
OpenAI also revealed that one of its models exploited a configuration vulnerability during another test to access the internet and browse a website without permission. These occurrences reinforced the US government's concerns about the inherent risks of the most advanced AI systems, increasing pressure for the implementation of oversight mechanisms.
Details will remain confidential
Even while formalizing a review process for advanced models, the White House chose to keep several aspects of the initiative confidential. According to sources informed to the Times, the details of the framework guiding the AI model review will not be made public. Furthermore, the list of entities considered 'trusted' and granted early access to advanced models by the government will also remain confidential.
This decision was criticized by AI governance experts. Brad Carson, president of Americans for Responsible Innovation, commented to the Times: 'A set of rules can only keep AI companies in check if people outside those companies know what those rules are.' He added that 'Keeping this structure secret reduces transparency and weakens public accountability.'
The dispute with China influences the strategy
The definition of these new guidelines occurs amid growing rivalry between the United States and China for dominance in AI. Trump argued that regulation should not impede the advancement of American companies, stating: 'In China, there are practically no controls. They have much more freedom. We need to be careful not to restrict our companies to the point where we fall behind China.'
Simultaneously, members of the administration have debated possible restrictions on certain Chinese companies, alleging that these companies might have used advancements developed by American firms to train their own models in what is considered unfair competition.
