In 1997, IBM’s Deep Blue supercomputer stunned the world by defeating chess champion Garry Kasparov, sparking widespread debate about whether machines could truly think like humans. While the answer at that time leaned toward “no,” the rapid advancements in artificial intelligence (AI), especially generative AI models in 2026, have reignited that question. But instead of chess, researchers are now turning to a different arena to test AI’s intellectual limits: mathematics.
Unlike the typical math problems many encounter in school, professional mathematicians focus on far more complex and abstract questions. These are not simply tasks with straightforward numerical answers, but rather inquiries into the truth or falsehood of intricate mathematical statements that describe objects and patterns often beyond human visualization. For instance, while most people might understand simple shapes like triangles or squares, mathematicians work with exotic shapes in multiple dimensions and strange curvatures, which require advanced proofs to understand their properties.
This distinction is critical when considering AI’s performance on math tasks. AI systems like Google’s Gemini Deep Think have recently achieved impressive feats, such as earning gold-level scores on the International Mathematical Olympiad and solving multiple problems posed by the legendary mathematician Paul Erdős. However, these accomplishments mostly involve problems resembling homework or competition puzzles rather than the original, frontier research questions mathematicians tackle. These research problems involve proving new theorems or lemmas—smaller, foundational statements within larger proofs—which contribute to expanding the boundaries of mathematical knowledge.
To better assess AI’s capabilities in genuine mathematical research, a group of eleven prominent mathematicians launched the First Proof challenge. Their goal was to push AI models to the limits by presenting them with real, unsolved research problems drawn from their own upcoming academic papers. Crucially, these problems had never been published online, ensuring the AI could not have encountered their solutions during training. Each problem focused on proving specific lemmas, key building blocks in more extensive proofs.
The challenge invited AI developers and mathematicians alike to submit solutions purely generated by AI, without human intervention, to see if machines could independently advance mathematics. Publicly available AI chatbots were initially tested and managed to solve only two out of the ten problems, highlighting the difficulty of the task. Meanwhile, major AI companies such as OpenAI and Google Gemini separately claimed to have solved approximately half of the problems (five or six out of ten), though some of these claims required verification and revision upon closer expert examination.
An intriguing aspect of the challenge was the vibrant response from a global community of mathematicians and math enthusiasts, many of whom are not professional researchers. These individuals eagerly used AI tools to attempt proofs, sharing their results on social media and specialized forums. While many AI-generated proofs were flawed or nonsensical, some showed genuine promise, revealing both the potential and current limitations of AI in mathematical reasoning.
One notable insight from the First Proof challenge was the difference in quality and methodology between AI-generated proofs and traditional human proofs. The AI’s proofs often resembled “19th-century-style math,” a phrase used by researchers to describe proofs that rely heavily on existing mathematical tools and brute-force techniques rather than introducing novel concepts or elegant reasoning. Human mathematicians cherish proofs that are not only correct but also “beautiful”—proofs that illuminate why a result must be true, often by creating new mathematical objects or frameworks that simplify complex problems.
Currently, AI tends to assemble known methods in complex, roundabout ways that reach correct conclusions but lack the deeper insight and creativity prized by mathematicians. For example, while a mathematician might invent a new concept to distill and understand a problem better, AI appears more comfortable recombining existing tools than inventing fundamentally new ideas. This limitation raises questions about whether AI can ever truly replicate the creative process central to mathematical discovery or if it will remain a powerful but ultimately utilitarian tool.
However, the situation is not entirely one-sided. Some AI-generated proofs have impressed experts for their correctness and elegance, suggesting that as AI models continue to improve, they may begin to approach or even surpass human creativity in mathematics. The ongoing rounds of the First Proof challenge, with tighter controls and clearer documentation of AI-human collaboration, promise to shed more light on AI’s evolving capabilities.
The debate among mathematicians about AI’s role in advancing mathematics is lively and varied. Some believe AI will soon revolutionize the field by solving major unsolved problems, accelerating research at an unprecedented pace. Others caution that AI will never replace human curiosity, intuition