Science Coup: US AI is Taking Control of Mathematics | Prof. Babak Haghighat

Artificial intelligence is no longer a mere instrument for computation; it has become a generative force reshaping the practice and power of mathematical research. The recent emergence of models capable of proposing and proving research-level results marks a turning point: math is shifting from an open, collaborative public good into terrain where frontier capabilities are concentrated in the hands of a few private actors. That concentration poses political and scientific risks—inequalities of access, strategic asymmetries between states, and the potential for an accelerating feedback loop that amplifies one side’s lead. The conversation with Prof. Babak Haghighat crystallizes these stakes and urges a rethinking of how the global research community organizes access, governance, and partnerships around AI-driven discovery. ## AI transforms mathematical research The scale and speed of change are striking. As Prof. Babak Haghighat recounts, advanced models recently produced hundreds of mathematical manuscripts addressing hundreds of open problems—work that would previously have required decades of human effort and specialized expertise. What distinguishes these models from earlier computational aids is not raw processing power but emergent creativity: they can formulate approaches, propose conjectures, and assemble proofs that connect previously distant fields. Where symbolic algebra or numerical solvers executed human-provided recipes, these systems now explore and ideate. That shift alters the epistemic landscape. Mathematicians can use AI to traverse disciplinary boundaries, test vast swaths of reasoning, and rapidly iterate on complex ideas that would otherwise be prohibitively time-consuming. The promise is transformative: researchers gain a telescope into structures and patterns they could not have seen alone. Yet the novelty also raises methodological questions—about verification, reproducibility, and the social processes by which mathematical truth is validated. Early examples already show a mixed record: some AI-produced papers were withdrawn when errors were found, while others can be formally checked using proof assistants. The mix of dramatic possibility and necessary scrutiny sets the stage for the debate over who should steward these capabilities. ## The frontier gap and control of science The conversation highlights a stark geographic asymmetry: the most powerful frontier models—those operating at the level of original research in mathematics—are concentrated within U.S.-based private firms. Prof. Haghighat warns that this is not a trivial competitive edge. Even a short lead at the frontier can compound. Better mathematical tools produce better models, and better models, in turn, produce deeper mathematics. That positive feedback loop can accelerate innovation in ways that quickly outpace those excluded from the pipeline. Beyond competitiveness, the central question becomes control. When the instruments of discovery sit behind corporate walls, decisions about what is explored, who may access the outputs, and how results are shared follow commercial logics rather than scientific norms. Mathematics historically functions as a global public good—developed across cultures and eras through open exchange. Private ownership of the most capable creative engines breaks with that tradition and risks aligning foundational knowledge with profit motives, access restrictions, and geopolitical strategy. ## Researchers face an access dilemma A central paradox emerged in the discussion: researchers outside the privileged access circle face a Catch-22. To remain competitive they must use frontier models; but using them often requires surrendering data, proprietary prompts, or involvement in closed systems that strengthen the supplier’s advantage. Prof. Haghighat describes this as an acute dilemma for mathematicians based in China and many parts of the Global South, where official or corporate policies bar access to the most advanced models. The consequence is not merely inconvenience; it is structural exclusion from the contemporary means of discovery. This access asymmetry reshapes career incentives and the contours of scientific collaboration. Young scholars may migrate toward institutions and firms that grant model access, leaving some universities and states disconnected from cutting-edge work. Meanwhile, scholars who do use closed models risk contributing to the very concentration that disadvantages their peers—feeding data, experimental practices, and insights back into the proprietary ecosystem. The result is a fractured research commons where the quality and visibility of scholarship increasingly depend on privileged hookups rather than merit or open exchange. ## An accelerating race and security risks What begins as an academic lead quickly acquires strategic and security dimensions. Prof. Haghighat emphasizes that mathematics underpins physics and engineering: advances in mathematical understanding can translate into superior designs across domains, including those with direct military applications. In an environment where the leading models are controlled by a single state’s corporate sector, the risk of an unbalanced technological lead is manifest. The danger is twofold. First, the pace of improvement—driven by recursive improvements in models that can do better mathematics—means a temporal lead can morph into a long-term dominance. Second, the dual-use nature of advanced capabilities means commercial inventions can be redeployed in security contexts. Without counterbalances—national programs, open international collaboration, or shared public infrastructure—these dynamics would favor rapid militarization of any relative advantage. That prospect, argued in the conversation, justifies urgent policy attention beyond the academy. ## Chips, sanctions and open collaboration Part of the story of concentrated advantage is geopolitical supply chains. Hardware—the GPUs and chip fabrication capacities essential to training the largest models—remains a chokepoint. Prof. Haghighat points to chip sanctions and export controls as one reason the United States and allied firms have sustained an edge in training scale. Even where design know-how exists elsewhere, fabrication and the specialized manufacturing footprint are constrained, and that gap reinforces asymmetries in who can build the largest, most capable systems. Yet technology ecosystems are not monoliths. The conversation suggests remedies focused on rebuilding and reorienting research partnerships. Universities and public-sector labs have a crucial role: they can host cooperative training environments, establish mathematical model development as a publicly governed resource, and create verifiable pipelines for scientific work that dissociates discovery from corporate capture. Encouragingly, that approach need not be limited by borders; confined and domain-specific models—for pure mathematics and closely related fields—could be jointly developed as international public goods, opening access to scholars worldwide while keeping certain capabilities transparent and auditable. ## Energy and the cost of intelligence Beyond chips and policy lies the material cost of frontier AI: energy. The training regimes for the most advanced models demand prodigious compute, which in turn consumes large amounts of electricity and generates significant carbon footprints. This raises ethical and practical constraints on the scaling of research-focused AI. An exponential race that privileges sheer compute raises sustainability questions: who bears the environmental cost, and how can research agendas be aligned with responsible resource use? Prof. Haghighat’s remarks prompt a deeper consideration of efficiency as a strategic variable. Nations and institutions can pursue not only bigger models but smarter, more parsimonious methods—hybrid approaches that combine domain-specific symbolic systems, proof assistant

Science Coup: US AI is Taking Control of Mathematics | Prof. Babak Haghighat All is not well in the world of AI. While China has done a remarkable job at keeping up with AI in general, there is a gap at the very frontier of AI models, especially in mathematics. And the US is exploiting this with great skill. Tsinghua mathematician Babak Haghighat joins Pascal to discuss how artificial intelligence is changing mathematical research and the global balance of scientific power. He warns that private control of advanced models could restrict researchers, deepen international inequalities, and accelerate an AI arms race. Their conversation examines China’s research ecosystem, American chip restrictions, energy demands, and the promise of open collaboration. Haghighat argues for public access, stronger university partnerships, and keeping AI a tool for humanity rather than its scientific master. Babak's Links: Tsinghua University faculty profile: https://ymsc.tsinghua.edu.cn/en/info/1032/1214.htm Neutrality Studies Links: Neutrality Studies Substack: https://pascallottaz.substack.com Support Neutrality Studies: https://neutralitystudies.com /donate Neutrality Studies shop: https://neutralitystudies.com/shop Time Stamps 00:00 AI transforms mathematical research 05:06 The frontier gap and control of science 10:48 Researchers face an access dilemma 16:36 An accelerating race and security risks 23:11 Chips, sanctions and open collaboration 35:42 Energy and the cost of intelligence 40:31 Human creativity and AI as a tool 45:40 Avoiding an arms race; science as a public good #Pascal Welcome back, everybody, to Neutrality Studies. Today, for the first time, we're joined by Dr. Babak Harihat, a German-Iranian mathematician at Tsinghua University in China. Babak, welcome. Welcome. #Babak Haghighat Yeah, thank you, Pascal. Thanks a lot for the invitation. Let me begin, then. #Pascal Well, just a second, because I want to say that we were connected through a mutual friend of ours. And we were connected because you are actually quite worried about something that's going on in the field of mathematics connected with AI. And since you're in China, there's a particular problem there. Can you maybe explain to us a bit what worries you so much? #Babak Haghighat -- 1 of 16 -- Well, to set the stage, so to speak, let me begin with a date and a number. On October 6th of this month, there was an announcement by the company OpenAI in which they published 722 mathematical manuscripts covering 372 research results, or families of research results, tackling completely open problems in mathematics. So the system, which is an internal model, was given the task of looking at 4,000 open mathematical problems and conjectures, some of them open for decades and some even for centuries. On average, it needed three hours of compute time and produced these results. There are 372 distinct results with proofs for some of these open problems. And this is actually something extraordinary. I think we should all recognize that something very peculiar has happened. And this is, in some sense, an evolutionary moment, if you will. So, for a long time, mathematicians have been using computers, but computers have been just a tool. You were basically programming these computers. They were very powerful, but they were just executing, you know, like a recipe, whatever the mathematician was giving them. So they were executing a recipe and producing some results, but there was no creativity involved. And these programs were just tools that you could start and stop, and they were just computing things. They were doing algebraic manipulations, symbolic manipulations, or numerical calculations, but everything was predetermined, and they were just executing a script. But recently, what has happened is that they have kind of crossed the line into creativity. So now they can creatively search for solutions, come up with new ideas, and present proofs for open problems. And this is something completely new. This change has happened within the last two years. And so now they are proposing approaches, finding counterexamples, and connecting distant areas of mathematics, something which would have been unimaginable before. So this is the new development that culminated in this October 6th announcement. And, yeah, this is what I want to talk about. And if you want, I can explain it a bit more. So this announcement, or this October 6th moment, did not happen in a vacuum. It actually started a bit earlier, and this is the culmination of a process that started, say, two years ago. It started in 2024, when DeepMind's models were able, for the first time, to actually solve mathematics problems at the International Mathematical Olympiad and get silver medals. So they were basically given the same amount of time as a contestant, and they could win a silver medal. Then, in 2025, the newer models from DeepMind, the Gemini models, were able to even obtain gold medals. And now they are collaborating with us in research, and they are basically solving research- level problems. So this has been the acceleration I have been observing. And this is what I want to talk about, yeah. #Pascal Yeah, just to give people an idea of the problem that arises from this: we now have AIs that are creative and are solving problems, but do all of these AIs do that at the same speed? Because I think one of your observations is that the US models are at the frontier when it comes to -- 2 of 16 -- mathematics, right? DeepSeek and so on are really, really, really good at everything else. But when it comes to mathematics, it’s the US models that are at the frontier, right? And can you maybe just tell us briefly why you think that’s an issue? #Babak Haghighat So what you're talking about now, of course, are these DeepSeek models and these Chinese models, which are very, very capable. They have been participating in math contests, and they have been doing very well. But what I'm talking about here is a different level. So what this October 6th announcement implies is that some of these models, if you look at it, are actually Fields Medal- worthy. So if a mathematician had solved it, it would have been a straight Fields Medal. Of course, I should be careful here. We shouldn't say that all of these results that have now been put out are valid and so on. Mathematicians have to go over them and check the proofs, and there might be mistakes. I mean, actually, three papers were wrong and were withdrawn just the next day, out of these 722. So the official number is 719 now. Some of them can be automatically checked using software called Lean. So they can be checked with Lean, and then you're basically sure they are correct. But you have to check them. If they now withstand these checks, then if a mathematician had done that, this would be a straight Fields Medal. So this is where the gap starts. I mean, these Chinese models are not at this level. So this is the first thing I wanted to clarify. Now, by itself, you might say this might not be a problem and so on, but there are certain restrictions I want to discuss that actually do create a problem, in particular for the mathematics community, but I think more widely for all of science and for the whole world. Okay, let me not be so bleak. I mean, we can, of course, focus on the problems, but let me say at the outset that this is actually positive. #Pascal Hey, just a very quick note. The best way to support this channel is by signing up for my free Substack. You can also help with a paid subscription there, or you can get some of our new merch on neutralitystudies.com. Links below. See you there. #Babak Haghighat Put in the right framework, this is a very positive development, because now mathematicians have an incredibly powerful tool that allows them to connect different fields in ways that were basically not possible before for an individual mathematician. You would need a lifetime to learn about the other field. You can connect different fields, you can explore vast possibilities for solutions, and you -- 3 of 16 -- can basically carry out vast research tasks that were not possible before. So we are basically climbing up a ladder, or, if you want to use the analogy of climbing a hill or a mountain, you have a much better view. So in this sense, this is something very good for science, but the main issue is: who controls this? Who controls this tool? And this is basically where the problem starts. So by itself, this doesn't have to be a bad thing. But the issue is who is actually controlling it. And this is not only a problem for mathematics, but for all of science, because if you think about it, mathematics is the foundation of everything in the universe. If you want to speak in the language of Galileo, the book of nature is written in the language of mathematics. So those who can basically understand and master mathematics can then, as a subsequent step, master all of physics, or most parts of physics, and other sciences. And this is an extremely powerful tool in that sense. And even AI itself—its foundations are based on mathematics. So once you create better mathematics, you can actually use it to construct a better AI. So this is, in that sense, a very major development. But as I said, an instrument can serve many ends, but it basically depends on the hand that wields it. And this hand is now controlled by a small number of private companies. #Pascal Yeah, and the small number of private companies that have the newest models available, which are needed for this kind of frontier research, are all in the United States. And if I understood our previous exchange correctly, then one of the issues is that researchers in China, or anywhere in the world, actually, are now kind of confronted with a catch-22 situation. Either you don't use those models and you can no longer compete at the frontier of mathematics, or you use them and thereby automatically feed the Americans' data kraken. It's either-or, right? You can't have it both ways. Exactly. #Babak Haghighat No, exactly. This is exactly the main issue. And yeah, this basically brings me to the next point, because I myself am such a researcher. I am based in China, and I am using some of the AI tools that I can access. But in principle, these AI tools are now indispensable for daily research, and this is something that is worrisome because the models that produce these results are, first of all, unreleased. I'm talking about this announcement. Even mathematicians inside the U.S. don't have access to these models. You cannot check how the result was produced. You cannot test the model, and so on and so forth. Secondly, the companies' policies don't allow their models to be made public or open source. And thirdly, under U.S. law and company policies, no one in China or the Global South, or in many other countries, can actually access these models. People based in China, for example, are not allowed to use these models. And this, of course, creates a huge problem. This is hugely unfair to -- 4 of 16 -- mathematicians because, as I said, the mathematical foundations of AI, and even AI itself, were developed through worldwide collaboration between researchers. But now only a few companies have access to these frontier models, and they basically decide who will control them. So this is basically creating this paradox. #Pascal Yeah. And can you describe the scale of the issue? Because some people might say, well, you know, in six months, China or Chinese models will have caught up with that, right? And so how bad can it be? Can that tiny little difference between the best model and the second-best model make that much of a difference for mathematics and for the whole development of the technology? #Babak Haghighat Absolutely. I mean, there are several points to mention here. As I said, the foundation of AI is mathematics itself. If you have better control of mathematics, then you