The rise of artificial intelligence is the single most important trend of our time. The changes it will wreak on human life eclipse not only all other contemporary technological developments, but also any since the Industrial Revolution. The new pope recognizes its salience, choosing as his regnal name “Leo XIV,” because just as Leo XIII was confronted with the social change from the Industrial Revolution, he seeks to confront the even more dramatic change from AI.
Breakthroughs happen monthly now. OpenAI’s o3 model recently scored higher than 99.8 percent of competitive programmers—while the same lab’s Sora engine, launched in February and now being integrated into ChatGPT, can already generate minute-long, high-definition video from text. In July, frontier models from both Google DeepMind and OpenAI sat the International Mathematical Olympiad under the same four-and-a-half-hour rules given to the world’s brightest teens, solved five of the six problems, and earned gold medals. In April, a University at Buffalo team unveiled Semantic Clinical AI (SCAI), an architecture that grafts formal medical knowledge onto a large language model. SCAI scored as high as 95 percent on Step 3 of the US Medical Licensing Examination—better than most practicing physicians and ahead of every previous AI benchmark—showing that well-structured retrieval can turn AI into a skilled general diagnostician.
AI’s new methodological phase is the emergence of AI agents, systems that autonomously execute sequences of tasks in pursuit of a goal. This phase is called agentic AI, which builds on Large Language Models (LLMs) that have dominated AI in the last few years. Those LLMs are neural networks that understand the relation of words in text and can use that understanding to generate competent, even expert, answers on every subject.
Agentic AI marries the predictive eloquence of LLM to an institutional framework of memory, goal-seeking, and tool use. No longer confined to completing sentences, the system can now formulate a purpose, decompose it into ordered tasks, engage external software, monitor its own performance, and revise its course when it makes mistakes. In short, where the LLM offers fluent speech, the agentic overlay supplies the infrastructure necessary for transforming mere words into coordinated action. One way of measuring progress in agentic AI is the uninterrupted duration of its competent autonomy at human tasks.
A year ago: a few minutes spent sorting e-mail, drafting a paragraph of code, or a short speech.
Now: twenty to thirty minutes in which Deep Search on ChatGPT and other similar services produces memos on any subject. As Tyler Cowen observes, the answers “wipes [sic] the floor against any humans, pretty much across the board.”
By 2026–27: several hours, during which an agent can redesign a software module, or plan a fortnight of coordination for a business.
As the decade closes: days or even weeks, in which an agent may conceive a research project, run simulations, draft the article, and submit it for peer review.
Because of its current and future power, progress in AI has become a central national concern. For example, AI may offer a solution to the intractable fiscal situation in which the United States finds itself. The United States’ national debt stands at approximately 120 percent of gross domestic product. And it is rising since we are also running an annual deficit of about 6 percent a year. Neither party offers any solution to the problem. The Republican Party just renewed the tax cuts of Trump’s first term and added more without compensating cuts to the budget. When the Democrats were last in power, they added programs and government spending that also added to the budget deficit. Neither party provides plausible reforms to the middle-class entitlements that drive future deficits in an aging population. Both are committed not to raise taxes on any but the top two percent, which is completely insufficient to curb deficits and debt. The result is a sea of red ink as far as an economist can project.
Given the fiscal plight and political constraints, the only solution is sharply increased economic growth. AI provides the most plausible engine for that growth. As a general-purpose technology, like electricity, it can make almost every human enterprise more efficient. In business enterprises, AI can match output to demand in real time and trim the staffing once required to execute routine tasks. In medicine, it can help deliver correct diagnoses faster with fewer people. In my own field of law, it can draft contracts and briefs with human lawyers just doing the work of tweaking and revising.
In geopolitics as well, AI offers both opportunity and danger. If the West wins the AI race, it will be in a better economic and military position, particularly given that the battle will now be fought with AI technology—both to surveil the enemy and to launch attacks and create defensive shields. But if China gains mastery before the United States, it could replace it as the global superpower.
These three executive orders clear the regulatory thickets at home, wire friendly nations into an American-led AI ecosystem abroad, and insist on ideological neutrality in the models.
As a result of AI’s centrality, the Trump administration has created “Winning the AI Race: America’s AI Action Plan,” encompassing three companion executive orders issued on July 23. Together, they erect a federal legal architecture for artificial intelligence that seeks to secure US primacy by accelerating innovation, expanding infrastructure, and projecting American standards abroad. The plan sets out more than ninety actions, from specific regulatory sandboxes encouraging innovation at the FDA and SEC to a more general revision of federal rules that “unduly burden AI innovation,” all premised on the proposition that prosperity and power in the twenty-first century will flow to the nation that commands the AI frontier.
The plan recognizes that AI will succeed in rapidly increasing economic growth only if it is not stymied by bottlenecks. It has substantial demands for computational infrastructure and the energy needed to power that infrastructure. Indeed, one way to think of AI is that it converts electricity into intelligence without the biological constraints on such conversion in humans. But electricity generation faces regulatory constraints.
As such, the first executive order in Trump’s bundle, “Accelerating Federal Permitting of Data Center Infrastructure,” attacks the bottleneck of the computational power needed for AI. It aims to shave years off construction times for the facilities needed for frontier model training by creating exclusions to environmental regulations for AI infrastructure. Cheaper and faster computing is the indispensable fuel of the AI age. Thus, unclogging permits promises to liberate AI progress.
This executive order aids AI development more by deregulation than subsidization. The promise of AI is so powerful that the efficient capital market ecosystem is sufficient to supply funds. The roadblock is government regulation. Thus, the focus of the executive order is to ease the regulations that may block its smooth development.
The second executive order, “Promoting the Export of the American AI Technology Stack,” turns US economic strength outward. A new American AI Exports Program authorizes the Commerce and State Departments to finance and shepherd “full-stack” AI packages, including chips, models, software, and cyber safeguards, to trusted allies. Wiring friendly nations with American technology locks in governance norms, buoyant markets, deepens supply-chain reliance on US firms, and denies strategic rent to adversaries.
Thus, the second order is designed to deploy AI to strengthen our international alliances. For all the talk of Trump’s “America First” foreign policy, this executive order at least indicates the administration recognizes the importance of bringing allied nations within the ambit of America’s growing AI technology network. In the Cold War, the United States strengthened and integrated the West in two principal ways. First, it placed much of the West under a US security umbrella through NATO, additional multilateral pacts, and a web of bilateral alliances. Second, it was a principal mover in international trade agreements, like the GATT, which encompassed most Western nations and sought to tie them together economically. But today, the rise of AI makes our AI technological network the essential mechanism of integration and strength. And it is not Russia that is the principal challenger, but China.
As Dario Amodei (creator of the Claude Model) puts it in a memo that preceded the executive order:
a coalition [of liberal democracies integrated thought AI] would aim to gain the support of more and more of the world, isolating our worst adversaries and eventually putting them in a position where they are better off taking the same bargain as the rest of the world: give up competing with democracies in order to receive all the benefits and not fight a superior foe.
The third order safeguards viewpoint neutrality. “Preventing Woke AI” bars federal agencies from procuring language models that embed ideological filters, mandating “truth-seeking” and “ideological neutrality” as procurement conditions.
The aim is to make sure that AI does not compromise epistemic understanding by skewing answers to questions ideologically. There have been instances in which AI models seem to provide politically correct or biased answers. For instance, when asked to show a set of Founding Fathers, Gemini initially assembled a multiracial group of images in an obvious gesture to political correctness. Just as the Data Infrastructure executive order avoids subsidies to AI, this order avoids actual legal orders to the AI companies to be ideologically neutral—a matter fraught with First Amendment concerns. But the order does direct the government not to contract with companies that engage in distortions. The government has a substantial, indeed compelling interest, not to have its work distorted by ideological bias.
Taken together, the three executive orders clear the regulatory thickets at home, wire friendly nations into an American-led AI ecosystem abroad, and insist on ideological neutrality in the models the government pays for. If the bet pays off, long after the daily controversies of the Trump administration are forgotten, this set of actions may be remembered as the most important set of executive orders in the history of the American government’s relation with science.
