For decades, proving difficult mathematical theorems has been considered one of the most demanding intellectual activities. It requires creativity, logical precision, deep understanding of mathematical structures and the ability to connect ideas that may appear unrelated.
Traditionally, these tasks have been performed by human mathematicians, sometimes over many years. However, artificial intelligence is beginning to change what is possible.
Recent developments in AI research have demonstrated that advanced systems can do more than calculate answers or solve familiar textbook exercises. They can generate mathematical arguments, explore previously unanswered questions and produce formal proofs that computers can independently verify.
One particularly striking development is the use of AI to formalise a computer-checked proof of Fermat's Last Theorem. Other research has demonstrated that AI agents can solve previously open mathematical problems and contribute to ongoing research.
These developments raise an important question: Can AI now prove difficult mathematical theorems, or is it still relying on human mathematical knowledge?
The answer is more complicated than a simple yes or no. Nevertheless, recent progress suggests that AI could become an important research partner for mathematicians.
What Does It Mean for AI to Prove a Mathematical Theorem?
A mathematical theorem is a statement that has been established as true through rigorous logical reasoning.
For example, the Pythagorean theorem states that, in a right-angled triangle, the square of the hypotenuse equals the sum of the squares of the other two sides.
While simple theorems can be demonstrated using elementary mathematical techniques, advanced results may require hundreds of pages of reasoning, sophisticated mathematical concepts and connections between different areas of mathematics.
To prove a theorem, an AI system must do more than produce an answer that appears convincing.
It must:
Understand the mathematical statement and its assumptions.
Identify relevant definitions, lemmas and established results.
Develop a logically valid sequence of arguments.
Handle difficult intermediate steps.
Avoid unsupported assumptions and logical gaps.
Produce a proof that can withstand independent verification.
This distinction is important because AI systems sometimes generate mathematical explanations that look convincing but contain errors.
A correct numerical answer is not necessarily a mathematical proof. Similarly, a persuasive explanation is not automatically a rigorous demonstration.
Researchers are therefore increasingly combining AI reasoning with specialised mathematical verification software.
The Breakthrough: AI and the Formal Proof of Fermat's Last Theorem
One of the most remarkable developments in AI-assisted mathematics research came from Anthropic in September 2026.
The company reported that its Claude AI system had produced the first complete computer-checked formalisation of Fermat's Last Theorem in the Lean programming language.
Fermat's Last Theorem is one of the most famous results in mathematics. It states that there are no positive integers a, b and c satisfying:
aⁿ + bⁿ = cⁿ
for any integer n greater than 2.
The theorem was famously associated with the French mathematician Pierre de Fermat, who wrote the claim in the margin of a mathematical book in the seventeenth century.
For more than 350 years, mathematicians struggled to establish a complete proof.
The problem was eventually solved by British mathematician Andrew Wiles, whose proof was completed in the 1990s using sophisticated techniques from number theory and algebraic geometry.
What makes the 2026 AI development particularly interesting is that Claude reportedly worked largely autonomously for 11 days to translate the established mathematical proof into Lean, a language used to write formally verifiable mathematical arguments.
According to Anthropic, the system generated approximately 13 million lines of Lean code and proved around 29,500 intermediate theorems.
The resulting formalisation was described as the first end-to-end computer-checked proof of Fermat's Last Theorem.
However, an important distinction must be made: AI did not independently discover the original proof of Fermat's Last Theorem.
Andrew Wiles had already established the theorem decades earlier. The AI achievement was to formalise the existing mathematical result in a form that could be checked by a proof assistant.
This is still a significant accomplishment because translating advanced mathematics into machine-verifiable logic is itself an extremely difficult task.
It demonstrates that AI systems may be able to help mathematicians verify, formalise and extend mathematical knowledge at a scale that would otherwise require enormous human effort.
What Is Lean, and Why Is It Important?
A major technology behind modern AI theorem proving is Lean.
Lean is a programming language and interactive theorem prover designed to support rigorous mathematical reasoning.
Unlike a conventional calculator, Lean does not simply return a numerical result. It checks whether a proposed mathematical argument follows logically from its definitions, assumptions and previously established results.
Think of Lean as an extremely strict mathematical examiner.
If an AI system produces a proof containing an invalid logical step, Lean rejects it. If the proof satisfies the formal rules and uses accepted mathematical foundations, the system confirms that the proof is valid within that formal framework.
Lean also benefits from Mathlib, a large community-developed library of formalised mathematics.
This library contains definitions, theorems and mathematical results that researchers can reuse when developing new proofs.
The combination is powerful:
An AI model analyses a mathematical problem.
It proposes a possible proof strategy.
It translates the reasoning into Lean code.
Lean checks each logical step.
If an error occurs, the AI can revise the proof.
The process continues until a valid formal proof is obtained or the system fails to solve the problem.
This approach reduces the risk of accepting mathematical arguments simply because they sound convincing.
It also creates an opportunity for AI systems to work on mathematical problems that require more than ordinary language-based reasoning.
AI Is Beginning to Solve Previously Open Mathematical Problems
The Fermat's Last Theorem development is not the only evidence of progress.
In May 2026, researchers published a study titled Advancing Mathematics Research with AI-Driven Formal Proof Search.
The research investigated whether AI agents could solve open mathematical problems rather than merely reproduce solutions to familiar exercises.
The study examined 353 open problems associated with the work of mathematician Paul Erdős.
The researchers reported that their most capable AI agent autonomously resolved nine of these problems.
The system also proved 44 out of 492 conjectures drawn from the Online Encyclopedia of Integer Sequences (OEIS).
These results are important because open mathematical problems do not necessarily have readily available solutions in textbooks or training datasets.
They require systems to explore possible arguments, identify useful mathematical relationships and construct valid reasoning.
The researchers also demonstrated that AI-driven proof search could support work across several areas, including:
Optimisation.
Algebraic geometry.
This suggests that AI theorem proving is gradually moving beyond narrowly defined mathematical competitions.
Instead, researchers are investigating how AI can contribute to active mathematical research.
First Proof Challenge: AI Takes on Research-Level Mathematics
Another important development is the First Proof initiative, which evaluates whether AI systems can solve genuine research-level mathematical problems.
Unlike conventional mathematical benchmarks, these challenges involve problems drawn from unpublished mathematical work.
This makes the evaluation particularly interesting because AI systems cannot simply rely on reproducing familiar textbook solutions.
A report published by Harvard on October 19, 2026, discussed the third round of the First Proof evaluation.
According to the report, every problem in that round was solved by at least one participating AI system.
The strongest system, called ProofCouncil, solved nine out of ten problems. It used a multi-agent approach combining several advanced AI models.
Another system based on GPT-6 Astra solved seven problems.
The evaluation involved 30 human reviewers, who assessed the submitted solutions for correctness, clarity and proper attribution of existing mathematical work.
However, the results also revealed important weaknesses.
Although the systems demonstrated impressive problem-solving ability, reviewers identified shortcomings in how some solutions were presented. These included unnecessarily complicated arguments, poor organisation, excessive explanations and problems with mathematical notation and attribution.
This is a reminder that solving a mathematical problem is not the only requirement for successful mathematical research.
A useful proof must also be understandable, properly justified and connected to existing scholarship.
Why AI Theorem Proving Is Different from Ordinary AI Mathematics
Many people associate AI mathematics with solving equations, performing calculations or answering examination questions.
However, advanced theorem proving involves a different level of reasoning.
Consider three levels of mathematical capability.
Level 1: Mathematical calculation
An AI system calculates an integral, solves a simultaneous equation or evaluates a matrix.
These tasks may require mathematical knowledge, but they often have established computational procedures.
Level 2: Mathematical reasoning
The AI develops an argument, identifies patterns, selects an appropriate theorem and explains why a mathematical result holds.
This requires more sophisticated reasoning.
Level 3: Mathematical discovery and formal proof
The AI investigates an unresolved question, develops a potentially new mathematical argument and produces a rigorous proof that can be independently checked.
This is the most demanding level.
The recent research developments are particularly exciting because they suggest that AI is beginning to make progress towards the third level.
However, success on selected research problems does not mean that current AI systems can solve every difficult mathematical question.
How AI Systems Actually Prove Theorems
Modern AI theorem-proving systems often combine large language models with mathematical software and specialised tools.
Rather than relying on a single response, they may use an iterative process involving several stages.
1. Understanding the Problem
The AI first examines the mathematical statement.
It identifies the assumptions, definitions, variables and conditions that must be satisfied.
For example, a problem involving groups in abstract algebra requires the system to understand the relevant algebraic structures and their properties.
2. Developing a Proof Strategy
The system explores possible mathematical approaches.
It may consider induction, contradiction, inequalities, algebraic transformations or previously established theorems.
More advanced systems can generate multiple candidate strategies and evaluate which ones are most promising.
3. Searching Mathematical Libraries
The AI can search formal mathematical libraries for existing results that may help solve the problem.
This is important because many difficult theorems depend on smaller intermediate results.
Instead of rebuilding everything from the beginning, the system can reuse established mathematical knowledge.
4. Generating Formal Proofs
After identifying a possible strategy, the AI translates its reasoning into a formal language such as Lean.
The formal proof must follow strict logical and syntactic rules.
5. Independent Verification
The proof assistant checks the generated argument.
If a logical step is invalid, the system receives an error and can attempt to correct it.
This feedback loop is one of the most important differences between ordinary AI-generated mathematical explanations and machine-verified theorem proving.
Can AI Replace Mathematicians?
Despite these developments, it would be premature to conclude that AI is ready to replace professional mathematicians.
Mathematical research involves more than proving statements.
Researchers must identify important questions, understand why particular problems matter, develop useful definitions, interpret results and connect discoveries to broader mathematical theories.
AI systems can assist with some of these activities, but they still have significant limitations.
1. Difficulty with Novel Mathematical Ideas
AI systems may struggle when a problem requires a genuinely new mathematical framework rather than an application of familiar techniques.
2. Dependence on Existing Mathematical Knowledge
Many theorem-proving systems rely on established libraries and previously formalised results.
When the required definitions or supporting theorems are missing, the AI may struggle to construct the necessary mathematical infrastructure.
3. Weaknesses in Long Mathematical Arguments
A proof may involve numerous intermediate steps, each depending on earlier results.
Maintaining consistency throughout a long argument remains challenging.
4. Difficulty Explaining Mathematical Significance
A system may successfully produce a formal proof without explaining why the result is important or how it changes our understanding of a mathematical field.
5. Verification Does Not Guarantee Mathematical Importance
A computer-checked proof establishes that a statement follows from its formal assumptions.
It does not automatically establish that the theorem is original, important or a major scientific contribution.
For these reasons, human mathematicians remain essential.
The more realistic future is likely to involve collaboration between human researchers, AI systems and formal verification tools.
What Are the Implications for Mathematics Research?
The progress in AI theorem proving could have major implications for universities, research institutions and scientific discovery.
Faster Mathematical Verification
Checking complex mathematical proofs can take months or even years.
Formal verification tools could reduce the burden of checking logical steps and identifying errors.
New Research Possibilities
AI systems may help researchers explore mathematical questions that would otherwise require considerable time and computational resources.
Improved Mathematical Education
AI theorem provers could help students understand how proofs are constructed, identify errors in their reasoning and explore alternative approaches.
Stronger Scientific Computing
Mathematical proofs are important in computer science, cryptography, optimisation, statistics and physics.
More reliable automated verification could support research in these areas.
More Accessible Advanced Mathematics
Formal proof systems could make it easier for researchers from different disciplines to verify sophisticated mathematical results without having to manually inspect every step.
What Does This Mean for the Future of Mathematical Discovery?
The most important lesson from recent research is that AI is gradually moving from mathematical assistance towards mathematical participation.
Systems that once struggled with complex reasoning are now capable of generating long formal proofs, solving selected open problems and supporting researchers working on advanced mathematical questions.
However, the distinction between formalising an established theorem and independently discovering a new theorem remains crucial.
The reported formalisation of Fermat's Last Theorem demonstrates that AI can handle an extraordinarily complex mathematical formalisation task. The Erdős problem research shows that AI agents can make progress on selected open problems. The First Proof results suggest that advanced AI systems are becoming increasingly competitive in research-level mathematical problem solving.
Together, these developments raise an exciting possibility: mathematicians may soon work alongside AI systems that can explore mathematical ideas, generate proof strategies, identify errors and formally verify discoveries.
The future of mathematics may not be a competition between humans and machines.
Instead, it may be a new research environment in which human creativity and machine-assisted reasoning work together to push the boundaries of mathematical knowledge.
Frequently Asked Questions (FAQs)
1. Can AI prove mathematical theorems?
Yes. Modern AI systems can generate proofs for mathematical statements, including some advanced and research-level problems. Systems connected to formal proof assistants such as Lean can also produce proofs that are checked by computers. However, their abilities remain limited, and they cannot reliably prove every difficult theorem.
2. Which AI has proved Fermat's Last Theorem?
Anthropic reported in September 2026 that its Claude AI system produced the first complete computer-checked formalisation of Fermat's Last Theorem using Lean. The achievement formalised the established proof developed by Andrew Wiles rather than independently discovering the original mathematical proof.
3. What is Lean in artificial intelligence?
Lean is a programming language and interactive theorem prover used to express and verify mathematical proofs. It checks whether every step in a formal argument follows logically from the accepted assumptions and definitions.
4. Can ChatGPT solve advanced mathematical problems?
Advanced AI models can solve many mathematical problems, including some difficult research-level questions. Their performance depends on the problem, model and tools available. Connecting an AI model to formal verification software can improve the reliability of its mathematical work.
5. What is an AI theorem prover?
An AI theorem prover is a system that uses artificial intelligence, mathematical reasoning and sometimes formal verification tools to construct proofs of mathematical statements. Some systems generate possible arguments, while others translate those arguments into formal languages that computers can check.
6. What are open mathematical problems?
Open mathematical problems are questions that have not yet been solved or conclusively answered by the mathematical community. They may require new theories, mathematical techniques or extensive reasoning.
7. What is the Erdős problem collection?
The Erdős problem collection contains mathematical questions associated with the work of Paul Erdős, one of the most influential mathematicians of the twentieth century. Many of these problems concern number theory, combinatorics and graph theory. Researchers increasingly use them to evaluate AI mathematical reasoning.
8. What is the First Proof Challenge?
First Proof is an initiative that evaluates AI systems on research-level mathematical problems. It is designed to test whether AI can produce meaningful mathematical solutions rather than simply answer familiar competition questions.
9. Can AI discover new mathematical theorems?
AI systems have demonstrated the ability to solve selected previously open problems and generate new mathematical results. However, establishing that a result is genuinely original and mathematically significant still requires careful verification and assessment by researchers.
10. Can AI replace human mathematicians?
Not at present. AI can assist with calculations, proof generation, formalisation and mathematical exploration. Human mathematicians remain important for identifying research questions, developing new concepts, evaluating significance and interpreting discoveries.
11. Is an AI-generated mathematical proof always correct?
No. An ordinary AI-generated explanation can contain logical errors or unsupported assumptions. A formal proof checked by a trusted proof assistant offers a much stronger guarantee that the argument is valid within the specified formal system.
12. What is the difference between a formal proof and an informal proof?
An informal proof uses conventional mathematical language and explanations. A formal proof expresses the argument using a precise logical language that a computer can check step by step.
13. Can AI prove the Riemann Hypothesis?
There is no accepted, independently verified AI proof of the Riemann Hypothesis. It remains one of the most famous unsolved problems in mathematics. Claims that an AI has solved it should be treated cautiously until the mathematical community has thoroughly examined the proposed proof.
14. How could AI theorem proving help scientific research?
AI theorem proving could help verify mathematical foundations in physics, computer science, cryptography, statistics and engineering. It may also help researchers develop new mathematical arguments and identify errors in complicated derivations.
15. Will AI eventually become better than human mathematicians?
It is too early to make that prediction. AI systems are improving rapidly in selected mathematical tasks, but mathematical creativity, research judgement and the development of entirely new concepts remain difficult challenges.
Final Thoughts
AI theorem proving is becoming one of the most fascinating developments in artificial intelligence.
The ability to formalise a theorem as complex as Fermat's Last Theorem, solve selected open mathematical problems and compete in research-level mathematics demonstrates how far the technology has progressed.
Nevertheless, these achievements should not be confused with the complete automation of mathematical discovery.
AI has not solved every major mathematical problem, and human expertise remains essential for evaluating its contributions.
What is becoming increasingly clear is that the relationship between mathematics and artificial intelligence is changing. The next major mathematical discovery may involve a human researcher working alongside an AI system capable of exploring ideas and verifying proofs at unprecedented speed.
The exciting question is no longer simply whether AI can do mathematics. It is how much further mathematics can advance when humans and AI work together.
Post a Comment