Donald Trump promised to repeal the Biden-Harris administration’s executive order on artificial intelligence, with the argument that it stifles innovation in capabilities research. But Executive Order 14110 is already so weak in its provisions that Trump’s promise to repeal the order is mostly meaningless.
Vice President-elect JD Vance, a former venture capitalist with deep Silicon Valley connections, also expressed concerns about overregulating AI. But, he supports open-sourcing model weights, claiming this will expose left-wing biases in large language models. This approach is dangerously shortsighted, because it allows foreign adversaries to access highly-refined, proprietary model weights.
Admittedly, it’s quite disheartening to see AI safety treated with such triviality, particularly by Vance.
While partisan biases in AI products are problematic, they do not constitute an existential risk, at least according to the Bostrom Scope-Severity Matrix. Enforcing open-source model weights in the name of preventing left-wing biases will threaten national security and undermine the United States’ AI lead. At worst, reconfiguring model weights for use by bad actors may pose a catastrophic risk.
What if, as a consequence of Vance’s enforcement of open-sourcing model weights, the People’s Liberation Army had open access to the weights of frontier American models? The implications might be detrimental, and here’s why.
The AI triad — data, algorithms and compute — are the raw materials that collectively determine the development of machine learning. Chinese AI developers enjoy relaxed data privacy laws and aggressive data gathering via state-sponsored surveillance activities — these provide a significant advantage in data collection for model training. However, American laboratories and universities are clear leaders in producing advanced algorithms for capabilities. China often recycles these algorithms for their own research.
Also, the U.S. has access to superior semiconductor design, photolithography equipment and cloud infrastructure to support model training. It’s safe to say that China is losing the AI arms race, fortunately.
However, if model weights are made open-source, the U.S. largely forfeits its advantage in algorithms, as Chinese laboratories could easily reconfigure and fine-tune frontier models while bypassing the extensive American investments into research and development. Simply reconfiguring pre-existing models allows China to avoid the computationally expensive training stage, freeing up its limited compute resources despite disadvantages in this area.
With more access to data than U.S. labs, China could produce further capable models. Ultimately, this chain reaction jeopardizes American leadership in capabilities research and, by extension, undermines the American ability to produce AI that are aligned with liberal principles.
This would also initiate a decisive Chinese acceleration toward superintelligent agents that are misaligned with liberal principles — or worse, altogether misaligned with humanity. Thus, American dominance in the AI race is an existential imperative.
I’m highly confident a Trump-Vance administration won’t push laboratories to openly publish model weights, and I’m highly confident that they won’t haphazardly deregulate this industry — it would be plainly imprudent. I hope the Trump-Vance administration will include advisors who are educated on the risks of advanced AI and receptive to sensible safety measures, even though this is somewhat antithetical to a Republican Party that is often inclined toward deregulation.
Given the supposedly deterministic nature of such technologies, I think the success of effective AI governance will require very close cooperation with leading labs without curtailing American innovation. On the whole, I think the Trump-Vance Administration has a great opportunity to use wise AI governance to guarantee national security and uphold humanity’s interests.
