The world must decide how to govern AI, but the decision is still ours

Seventy years old, two documents say more about artificial intelligence than any other model it has ever created. The first was a proposal to convene a Dartmouth College conference in the summer of 1956 that gave the field its name. The second one was recently published on July 13th and is only four sentences long. “We must act now” is a statement sponsored by Stanford University’s Digital Economy Institute and signed by more than 200 economists and AI researchers, including 16 Nobel Prize winners. The first document exuded confidence. The second one creates a sense of urgency. This technology has arrived at a destination that its successors admit is incomprehensible, along a path not foreseen by its creators.

It’s worth rereading the original proposal to gauge distance. John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon requested funding to bring together 10 researchers for two months on the assumption that “every aspect of learning and other features of intelligence can in principle be described so accurately that we can simulate it in machines.” They believed that great progress could be made in one summer. The concerns they raised were all technical. How to make machines use languages, form abstractions, and improve themselves. The question of what will happen to human labor, income distribution, and economic power if the project is successful is nowhere to be asked. Not out of negligence. The political economy of the machine was not part of the problem they believed they were solving.

The subsequent trajectory systematically punished optimism. The field experienced its first winter in the 1970s, when promise far exceeded ability to deliver and funding dried up. Reincarnated by expert systems, refrozen in the late 1980s, reshaped by machine learning, accelerated by deep neural networks since 2012, and exploded in popularity by transformer architectures, enabling the launch of OpenAI’s ChatGPT in 2022. In every cycle so far, the discrepancy has been the same. Technology promised more than it could deliver. The current cycle has reversed sign. For the first time, installed features are being executed before any understanding of their consequences.

It is precisely this reversal that the Stanford letter records in its telegram text. The paper states that AI could become fundamentally more powerful over the next decade, driving economic transformation “bigger than the industrial revolution”, but that it could be compressed into a significantly shorter period of time. Acknowledging both the risk of large-scale job losses and the potential for rising living standards, the report concludes that governments and industry must start now to create incentives, guardrails, and institutions that can guide technology to complement, rather than replace, humans. In a separate statement, Joshua Bengio, a Turing Award winner and one of the signatories, added that the choice must be collective and democratic, rather than “leaving it to market forces and risk leaving large segments of the population behind.”

This rush can be explained, given the circumstances surrounding this publication. Amazon announced in October 2025 that it would cut about 14,000 jobs, months after its CEO acknowledged that some tasks would be reassigned to generative AI and agents. Recent graduates in the U.S. are facing the toughest job market in years, with a study by the same Stanford University research group finding a relative 16% decline in employment among workers aged 22 to 25 in the jobs most exposed to generative AI, while employment among experienced professionals remained stable. The changes announced in this letter are not research hypotheses. It’s showing up in the employment statistics of countries that are already leading the race.

But there are three omissions in this document that deserve as much scrutiny as their presence. The first is the person in charge. Signatories include executives and co-founders of companies pushing the very frontier, including Anthropic, OpenAI, and Google. It’s an ambiguous gesture. It can be read as the clarity of people who know the power of what they build, but it can also be read as outsourcing problems to “governments and institutions” while commercial competition continues at full speed. The letter calls for guardrails to be put in place without saying who will pay for it, for systems to be put in place without saying who is in control, and for immediate action by everyone without imposing specific obligations on anyone. Diagnosis is collective. Divide responsibility.

The second absence is due to geography. The text speaks of “our economy” in the singular, as if the transformation would be distributed evenly across the globe. That won’t happen. Computing infrastructure, frontier models, and new economy revenues are concentrated in a small number of companies in two countries. By contrast, job losses will be global and will hit particularly hard economies that export precisely the services that AI agents are learning to perform. The risks to the Global South go beyond the technological unemployment described in this letter. Providing data, energy, minerals and consumer markets, they participate in the transformation only on the cost side, while the promised productivity gains pile up on the balance sheets of those who manage the infrastructure. If the institution this letter calls for were to emerge solely from Washington and its allied forums, it would organize the transition of club membership and leave everyone else to manage the outcome.

Nothing illustrates this imbalance better than the calendar that led to his third absence. Stanford’s letter is primarily signed by members of institutions in the United States and Europe. While Western experts were publishing this paper, China was finalizing preparations to hold a new edition of the World Artificial Intelligence Conference in Shanghai on July 17th. This year’s conference will feature opening remarks from Chinese President Xi Jinping himself and a high-level meeting on global AI governance with representatives from dozens of countries. The event will showcase over 1,000 exhibitors, 3,000 cutting-edge technologies, and 300 world-first products, all part of an industry expected to exceed 1 trillion yuan in 2025 and grow more than 30 percent in 2026. At the beginning of the conference, China and 28 other countries created the World AI Cooperation Organization (WAICO) to address issues related to global governance of AI with a Global South vision and multilateral approach.

At the same time, the United States is building AI governance the same way it builds most things these days: without treaties, charters, or votes, but through executive powers and corporate blueprints. Over a five-week period in June and July 2026, the chief executives of Japan’s three largest AI companies announced what amounted to a governance structure in several parts, with their proposals clustered together like the floor of a single building. Anthropic’s Dario Amodei designed a hard domestic layer modeled after the FAA. It would require third-party testing of all models that exceed computing thresholds across four risks: cybersecurity, bioweapons, loss of control, and automated research and development, with legal powers to prevent release and revenue penalties. Google DeepMind’s Demis Hassabis modeled the operations layer after FINRA. FINRA is an industry-funded body that receives Frontier models up to 30 days before release, and submissions are initially optional and mandatory once proven robust. OpenAI’s Sam Altman modeled the exterior facade after the IAEA. The IAEA is a Washington-led international forum that sets global standards and makes compliance with them a condition of access to technology. The experts’ letter signaled the need for such efforts by calling out the institutions without naming them. The three manifestos have advanced to fill this gap with pre-designed regimes by regulated parties.

Meanwhile, Washington was building faster than its manifesto could account for. The June executive order that lifted Frontier models from global markets with a single stroke of a pen has already created Gold Eagle, a vulnerability clearinghouse run by the Treasury Department along with CISA and the Department of Defense, that leverages Frontier models, including Anthropic’s Mythos, to protect critical infrastructure. The order envisages a review of the Frontier model before its release to trusted partners, and a process to translate the manifesto’s core provisions into national policy in August. The result is what one analyst called an involuntary pre-licensing regime, a de facto governance architecture fused with security mechanisms in which regulators are invited to become shareholders in the regulated entities, but which no Congress, not even the U.S. Congress, has voted on or approved.

The juxtaposition of the two scenes sums up the moment. On the one hand, there is a four-sentence manifesto calling for institutions without building them, signed by those who continue to push the technological frontier while acknowledging that they cannot control the outcome. On the other side are states that treat AI as public policy, show tangible results, and provide the world with the very governance platform that the letter lacks in the same week. Read in this context, Stanford’s letter comes across as more of a confession in a vacuum than a plan of action. And the geopolitical vacuum will not remain empty for long.

The participants at the 1956 Dartmouth Conference had the time wrong, but the direction was right. Machines exist that simulate aspects of intelligence. What they could not have foreseen, because it was not a matter of time, is that the crucial challenge would not be to build it, but to decide who would appropriate the value it would create and who would bear the costs of what it would destroy. The 2026 letter has the benefit of putting this issue at the center of the agenda with the authority of 16 Nobel laureates. There is a limitation in that it is not possible to clarify who is responsible for the requested action. Seventy years after the summer of 1956, everyone’s bill for that two-month experiment was due. It is up to each country to negotiate payment terms or accept a ready-made payment issued in dollars and written in English.