Artificial intelligence models such as GPT, Claude and Gemini may soon need to pass government verification before being released in the United States. As the impact of Anthropic's Mythos, regarded as the most powerful AI model ever, continues to spread, the Donald Trump administration is pushing for an AI pre-screening system. Analysts say President Trump, mindful of negative public opinion over the potential misuse of Mythos, is reversing the AI deregulation he has pursued.
The New York Times reported Tuesday, citing government officials, that the White House will form an AI working group and begin discussions with technology company executives and agency officials to introduce AI oversight procedures. The Wall Street Journal reported that the plan could include "AI model standards" and that the details will be disclosed through a presidential executive order.
Separately from the mandatory discussions, pre-release verification of AI models will be conducted through voluntary agreements with Big Tech. The Center for AI Standards and Innovation (CAISI), under the U.S. Department of Commerce, announced Wednesday that it has signed agreements with Google DeepMind, Microsoft and xAI to support the government in evaluating AI models before their release. "This is essential for understanding the impact on national security," CAISI Director Chris Paul said.
After OpenAI released ChatGPT, the world's first AI model, in 2022, Big Tech companies have freely launched AI models, recently unveiling a new model roughly every two months. Many of these have been supplied to the government, and both the pace of releases and distribution are now expected to face constraints. Axios reported that the Department of Defense (Department of War) is considering making safety testing mandatory for AI models distributed to federal and state governments.
The White House is looking at a model in which a government-designated oversight body conducts verification, similar to the United Kingdom, which became the first country in the world to introduce AI verification. The NYT said the National Security Agency (NSA), the White House National Cyber Director and the Office of the Director of National Intelligence are being discussed as potential oversight bodies. The Center for AI Standards and Innovation, established under the previous Joe Biden administration and whose presence has diminished under the Trump administration, is also being considered as a candidate. However, the UK system is weaker than what the White House envisions, as it relies on voluntary corporate participation with weak legal binding force and lacks standard AI models.
The AI pre-screening system marks a shift from Trump's previous policy of repealing Biden-era AI regulations. President Trump signed executive orders scrapping mandatory safety assessments and reporting for AI models that could be used for military purposes and limiting state governments' authority to regulate AI. When announcing the AI Action Plan in July last year, he stressed, "AI is like a beautiful newborn baby. We must raise this baby and let it thrive. Neither politics nor foolish rules can stop it."
Analysts say Trump's change of heart was largely influenced by Mythos, which Anthropic unveiled last month. Mythos boasts the strongest security performance ever, capable of finding bugs (software defects) that had gone undetected for 27 years, but if misused for hacking, it could cause serious personal data leaks and system paralysis. Because of these concerns, Anthropic has provided Mythos previews only on a limited basis to some government agencies and major banks. Competitors such as OpenAI are expected to release Mythos-class models soon.
The spread of negative public sentiment about AI ahead of this year's midterm elections was also a decisive factor. The proliferation of data centers driven by surging AI demand has worsened power shortages and pushed electricity bills sharply higher. Concerns have also grown that AI is taking away jobs and aggravating privacy violations and mental harm. In an NBC poll in March, 57% of registered voters said AI's risks outweigh its benefits.
Even before this development, the Trump administration had been expanding its grip on AI through conflicts with Anthropic over the military use of AI. The WSJ reported that Anthropic had tried to expand Mythos preview access from 50 to 120 organizations but was blocked by the White House, signaling expanded government intervention.
Big Tech companies including OpenAI, Anthropic and Google, which have been informed of these plans by the White House, are expressing concern. The NYT reported that executives worried in recent discussions that excessive regulation could hamper America's innovation race against China. Dean Ball, who served as a senior AI adviser in the Trump administration, said striking a regulatory balance is not easy amid the rapid pace of AI development.
