Mathematician Terence Tao urges a verification revolution to manage the flood of AI theories

Terence Tao, often regarded as one of the most brilliant mathematicians of the modern era, has issued a profound warning and a roadmap for the future of intellectual inquiry in the age of artificial intelligence. His central thesis is that we are witnessing a fundamental economic shift in the production of knowledge: artificial intelligence has driven the marginal cost of generating new ideas down to nearly zero.[1][2][3] However, this abundance of hypothesis and theory has not yet translated into a corresponding abundance of discovery. Instead, the surplus of potential ideas has shifted the primary bottleneck of scientific and mathematical progress from the creative spark of generation to the rigorous labor of verification.[1][4][2] This transformation, Tao argues, requires a complete overhaul of the infrastructure of science, much like the introduction of the automobile required a transformation of urban planning.[5]

The current state of generative AI allows researchers to produce thousands of plausible-sounding theories for any given scientific problem in a matter of seconds.[3] Tao compares this capability to the internet’s impact on communication; just as the cost of sending a message plummeted to zero, the cost of proposing a conjecture has followed suit.[1][4][5][2] In the past, the bottleneck of science was the "scarcity of ideas." Human experts spent years, sometimes decades, developing a single robust hypothesis. Today, the problem is no longer a lack of ideas but an overwhelming "signal-to-noise" crisis. Scientific journals are increasingly flooded with AI-generated submissions that appear professional and rigorous but often contain deep-seated logical flaws or hallucinations.[3] This influx is effectively "clogging" the traditional peer-review system, which was designed for a world where humans produced research at a much slower, more deliberate pace.

To explain this friction, Tao employs a vivid analogy involving the evolution of cities.[5] He compares the impact of AI on mathematics to the impact of the automobile on urban development. Before the car, city streets were narrow and optimized for pedestrians and horses.[5][3] When cars were first introduced, they were undeniably faster and more powerful, but they initially created gridlock because they were forced onto roads built for a different era.[6][5] Simply making cars faster did not solve the problem; instead, it required the construction of entirely new infrastructure, such as highways, traffic lights, and parking systems, while also necessitating urban planning to ensure that cities remained livable.[6][5] Similarly, Tao argues that the current social infrastructure of mathematics—journals, conferences, and student-mentor relationships—is like those narrow horse-and-buggy roads.[5] Attempting to "retrofit" AI into these human-centric structures only creates congestion and "slop."[3]

The solution, according to Tao, is not just to build better AI models but to build "machine-friendly" infrastructure.[5] In the world of mathematics, this means moving toward formal verification systems like Lean, an interactive theorem prover.[7][8] Unlike traditional mathematical proofs, which are written in natural language for other humans to read and interpret, formal proofs are written in code that a computer can check with absolute certainty.[8] Tao notes that while formalization is currently a tedious and time-consuming process for humans, it is exactly where AI can be most transformative.[3] By shifting the focus from generating ideas to formalizing them, mathematicians can create a "closed-loop" system where AI proposes a solution and a verification engine instantly confirms its validity.[7][3] This would allow the field to handle an influx of thousands of theories by automatically filtering out the 99 percent that are incorrect, leaving only the "gold" for human experts to review.

The implications of this shift extend far beyond the ivory towers of pure mathematics.[3] In fields like coding, law, and drug discovery, the cost of generation is also collapsing. Large language models can write thousands of lines of code or draft hundreds of legal contracts, but the human expert is still required to painstakingly verify every line to ensure it is safe and accurate.[3] Tao points out that in certain mathematical experiments, AI systems have already been used to solve a "long tail" of ignored problems—specifically, dozens of problems from the famous Erdős list that had been overlooked for decades.[9] While the AI’s success rate per problem was only one or two percent, the sheer scale of its "breadth" allowed it to find solutions that humans had not bothered to seek. However, these wins are currently "cheap" because they lack the deep conceptual insight and "narrative path" that human mathematicians value.

This highlights a critical distinction in the new division of labor: AI excels at breadth, while humans remain the masters of depth.[2] Tao observes that human mathematical discovery often involves a messy, "embarrassing" process of trial and error, dead ends, and intuitive leaps that are rarely recorded in final publications.[10] Artificial intelligence, trained primarily on the "polished" results of human success, misses this essential subtext of failure.[10][11] Consequently, AI-assisted proofs can sometimes lead efficiently from a hypothesis to a result but lose the valuable "side effects" of human research, such as the development of new techniques or the mapping of conceptual terrain.[5] To preserve the "walkability" of the scientific landscape, Tao suggests the emergence of a new discipline he calls "AI planning," which would focus on managing the interaction between high-speed automated discovery and human-scale understanding.[5]

For the AI industry, Tao’s analysis suggests a pivot in value.[7][9][5][12][8][13][14][2] While the "hype cycle" has focused on the creative and generative powers of models, the true long-term value may lie in verification and "trust infrastructure." Companies that develop tools for automated labeling, formal verification, and AI-assisted peer review are building the "highways" that will allow the "cars" of generative AI to actually move science forward. Tao’s own experiments with "vibe coding"—using AI to translate informal mathematical ideas into formal code—show that even the most advanced researchers are shifting their workflows.[3] The mathematician is becoming less of a manual craftsman, painting each "wooden doll" by hand, and more of a strategic architect who directs a fleet of automated tools.

Ultimately, Tao views the AI revolution not as a replacement for human intelligence, but as a "cognitive Copernican revolution" that reorders our understanding of what is hard and what is easy. Tasks that once required years of human labor, such as searching through massive libraries or performing tedious calculations, are now trivial. Meanwhile, the task of establishing truth and building deep conceptual frameworks has become more valuable than ever. By acknowledging that idea generation is now a commodity, the scientific community can begin the hard work of building the new roads required to navigate an era of infinite hypotheses. The challenge of the coming decade will not be to think of more things, but to build the systems that can tell us which of those things are actually true.