Every couple of months, a new frontier AI model breaks benchmarks once thought impossible. These tools are increasingly acting as autonomous agents, rather than simple chatbots, helping in scientific research, writing, software development, and much more. The economic incentives of these booms cannot be overlooked. AI has surpassed cybersecurity as the top technology spending priority for businesses, and Big Tech is raking in billions in investment.
With this backdrop, the AI Futures Project recently released AI 2040: Plan A. At first glance, the title seems like another long-range prediction of where AI may be in 15 years, like in the previous report AI 2027. In reality, the report is asking a more fundamental question: if AI capabilities continue to advance at today’s pace, what decisions should be made to ensure those systems remain under human control? The report argues that society’s ability to develop increasingly sophisticated AI is outpacing our ability to govern it responsibly.
To understand why the report reaches this conclusion, it helps to know what researchers mean by “frontier AI”. Frontier models are not just better versions of chatbots; instead, they demonstrate sophisticated reasoning, problem-solving, and capabilities. So, while today’s systems require significant human oversight, experts believe that future models will require minimal human intervention to perform increasingly complex work sequences, blurring the distinction between an AI model and an autonomous worker.
Over the past several years, researchers have observed that increasing computing power and training data creates predictable improvements in AI performance. While breakthroughs in algorithms remain important, much of modern AI progress has come from simply building larger models with more compute. This dynamic has fueled a global race among Big Tech companies to keep building ever-larger data centers and international actors who don’t want to fall behind. This race is also one of the central concerns that AI 2040 raises. If one company slows development to invest more heavily in safety, another may choose to keep pushing forward. If one country imposes stricter safeguards, another may continue acceleration.
This is where the report introduces one of its most discussed concepts: compute governance. Compute refers to the specialized hardware that powers modern AI development. Unlike software, which can be easily copied and distributed, frontier-scale compute is far scarcer and more expensive. Access to compute ultimately determines who can build the most advanced AI systems, which, this report argues, will become one of the most efficient points for AI governance. So, instead of attempting to regulate every AI application individually, policymakers could oversee and slow development via the compute that enables frontier AI in the first place.
However, slowing development is only valuable if that additional time is used productively. Much of AI 2040 centers on alignment, which is the practice of ensuring that increasingly autonomous AI systems reliably pursue human intentions, even in novel situations that developers could not explicitly anticipate. Solving the alignment issue is one of the defining scientific challenges of advanced AI, especially as these systems become more capable.
Yet even if one disagrees with the recommendations in AI 2040, the report’s greatest contribution is that it shifts the conversation away from the question of whether AI should continue to advance. AI has huge potential to transform society if it progresses responsibly. AI alone could contribute between 2.6 trillion and 4.4 trillion USD annually to the global economy. The World Trade Organization projects AI could boost trade by 40 percent by 2040. However, these opportunities will not negate the risks of deploying without robust safeguards.
The question is not whether America should lead in artificial intelligence. It absolutely should. The more important question is whether our public institutions can keep pace with the technology itself. AI 2040 highlights how, as AI capabilities advance, the governance choices will be as consequential as the technology itself.