AI in Math Timeline: From Early Proofs to 2026

Quick answer: Computer theorem proving dates to the 1950s. Martin Davis’s 1954 program proved a simple arithmetic fact, and the Logic Theorist, an early AI program, proved 38 theorems from Principia Mathematica in 1956. The four color theorem (1976) was the first major theorem proved with heavy computer help, and the program EQP proved the Robbins conjecture in 1996. Since 2022, AI has found new algorithms, reached gold-medal level at the math olympiad (2025), and in 2026 contributed to results on famous problems, including a Lean proof of the Riemann Hypothesis.

Updated November 22, 2026.

It can seem as if AI burst into mathematics in 2025. In fact, the story runs back to 1956, when one of the first AI programs proved theorems from a famous logic textbook. Every few decades, computers took on a bigger role: checking cases, proving small conjectures on their own, checking whole proofs, and finally contributing to research on famous problems. This timeline shows how we got here, and which “firsts” really count.

The Timeline at a Glance

This timeline separates four kinds of milestones: new discoveries, computer-assisted proofs, formal verification of existing proofs, and competition results.

YearMilestoneComputer’s role
1954Martin Davis’s program proves that the sum of two even numbers is evenFirst computer-generated proof
1956Logic Theorist proves 38 of the first 52 theorems in chapter 2 of Principia MathematicaAutomated proof search
1959Gelernter’s Geometry Theorem Machine proves school geometry theoremsAutomated geometry proofs
1976Four color theorem (Appel and Haken)Checks about 1,900 configurations
1996Robbins conjecture proved by EQP (William McCune)Finds the proof by itself
1998Kepler conjecture (Thomas Hales)Massive computation
2005Four color theorem formalized in Coq (Georges Gonthier)Checks the whole proof
2013Leonardo de Moura starts the Lean proof assistantBuilds the modern checking tool
2014Flyspeck: formal proof of the Kepler conjecture (HOL Light and Isabelle)Checks the whole proof
2016Boolean Pythagorean triples problem (Heule, Kullmann, Marek)A 200-terabyte SAT proof
2021DeepMind and mathematicians find new results in knot theory and representation theoryMachine learning suggests patterns
2022Liquid Tensor Experiment completed in Lean (July); AlphaTensor finds new matrix multiplication methods (Oct)Checks a cutting-edge proof; discovers algorithms
2023FunSearch finds larger cap setsLanguage model plus search finds constructions
2024AlphaProof and AlphaGeometry 2 reach IMO silver level (28 of 42)Solves olympiad problems
2025AlphaEvolve finds 48-multiplication method (May); gold-level IMO performance, 35 of 42, by Gemini Deep Think and an OpenAI model (July)Discovers and solves
2026Unit distance conjecture disproved (May); Fermat’s Last Theorem formalized (Sept); Riemann Hypothesis Lean proof certified (Nov)Contributes to research on famous problems

1954–1959: The First Proving Programs

In 1954, Martin Davis programmed the JOHNNIAC computer at Princeton to decide statements in Presburger arithmetic, a simple theory of addition; it proved that the sum of two even numbers is even. In 1959, Herbert Gelernter and colleagues at IBM built a Geometry Theorem Machine that proved high school geometry theorems.

1956: The Logic Theorist

Allen Newell, Herbert Simon and Cliff Shaw wrote the Logic Theorist, often called the first AI program. It proved 38 of the first 52 theorems in chapter 2 of Whitehead and Russell’s Principia Mathematica, and found a shorter proof of one of them than the book’s. It was presented at the 1956 Dartmouth workshop where the term “artificial intelligence” was coined.

1976: The Four Color Theorem

Kenneth Appel and Wolfgang Haken, with computing help from John Koch, proved that every map can be colored with four colors so that neighboring regions differ. Their proof reduced the problem to about 1,900 configurations and used over 1,000 hours of computer time to check them. Many mathematicians were uneasy: no human could check every case. That unease returned with every later computer proof, including today’s AI proofs.

1996: A Computer Proves a Conjecture Alone

The Robbins conjecture, a question about axioms for Boolean algebra open since the 1930s, was proved by EQP, an automated theorem prover written by William McCune at Argonne National Laboratory. The proof took about eight days of computing. It is often cited as the first notable conjecture settled by a computer program with no human proof ideas, which makes it a strong candidate for “the first theorem proved by AI alone.”

1998–2016: Huge Proofs and Formal Checking

Thomas Hales’s 1998 proof of the Kepler conjecture on sphere packing, with his student Samuel Ferguson, relied on heavy computation; referees reported being about 99% certain it was right, because they could not check every computation. Hales answered by leading Flyspeck, which completed a full formal proof in 2014. Georges Gonthier had already formalized the four color theorem in Coq in 2005. In 2016, Marijn Heule, Oliver Kullmann and Victor Marek settled the Boolean Pythagorean triples problem with a SAT solver. The question: can the whole numbers be colored with two colors so that no Pythagorean triple \(a^2 + b^2 = c^2\) is all one color? The numbers 1 to 7,824 can be; 1 to 7,825 cannot. The uncompressed proof was almost 200 terabytes.

2021–2023: Machine Learning Joins Research

In 2021, DeepMind worked with mathematicians including Marc Lackenby, András Juhász and Geordie Williamson; machine learning spotted patterns that led to new theorems in knot theory and representation theory (published in Nature). On July 14, 2022, the Lean community, led by Johan Commelin, completed the Liquid Tensor Experiment, formalizing a deep result of Peter Scholze that he had asked to have checked. In October 2022, DeepMind’s AlphaTensor found new matrix multiplication methods (see AI-discovered algorithms). In December 2023, DeepMind’s FunSearch paired a language model with automatic evaluation to find larger cap sets than previously known in some dimensions.

2024–2025: Olympiad Medals and AlphaEvolve

At the July 2024 International Mathematical Olympiad, DeepMind’s AlphaProof and AlphaGeometry 2 scored 28 of 42, silver-medal standard, after the problems were manually translated into Lean and with up to three days of computing on some problems. In July 2025, Gemini Deep Think and an OpenAI model each reported 35 of 42, gold-medal-level performance; Gemini’s answers were graded by IMO coordinators and OpenAI’s by three former medalists (see AI at the IMO). In May 2025, AlphaEvolve found a way to multiply 4×4 complex matrices with 48 multiplications.

2026: AI and Famous Problems

  • May: an OpenAI model disproved Erdős’s 1946 unit distance conjecture (see Erdős problems solved by AI).
  • July: an OpenAI proof of the cycle double cover conjecture, and a Jacobian conjecture counterexample in three complex dimensions credited to Levent Alpöge and Claude.
  • September: Fermat’s Last Theorem fully formalized in Lean with Claude; an AI-found Navier–Stokes blow-up result with external force.
  • November 20–21: Levent Alpöge released a Lean proof of the Riemann Hypothesis developed with Claude; the next day the Lean and Mathlib communities certified both the statement and the proof.

For details and verification status, see AI math breakthroughs of 2026.

What Was the First Theorem Proved by a Computer or AI?

Kind of “first”Best answer
First computer-generated proofMartin Davis’s program, 1954
First AI program to prove theoremsLogic Theorist, 1956
First major theorem proved with a computerFour color theorem, 1976
First notable open conjecture settled by automated searchRobbins conjecture (EQP), 1996

“Alone” needs care. A program can search for a proof without help during a run while relying on people for the problem statement, earlier mathematics and verification. EQP’s search was automatic, but McCune chose its inputs and settings, and people checked its output. The same is true of every AI result in 2026.

Common Misconceptions

“Computers in math started with ChatGPT.” No. Computer-assisted proofs go back to the 1950s and 1970s.

“The four color theorem was proved by AI.” No. Humans designed the proof; the computer checked cases.

“Machine proofs are new, so they can’t be trusted.” Formal checking has a long track record; see can we trust AI math proofs?

Watch: Milestones in Computer Proofs

If you learn better by watching, these videos are worth your time.

John Conway – Four Color Theorem and Computer Proofs (Simons Foundation). John Conway on the first big computer proof and why it unsettled mathematicians.
The Four Color Map Theorem (Numberphile). A friendly introduction to the 1976 computer-assisted proof.
The Problem with 7825 (Numberphile). Marijn Heule explains the 200-terabyte Boolean Pythagorean triples proof.

Check Your Understanding

Five quick questions to see how much of this has stuck. Pick an answer and press Submit; you will see a short explanation either way.

Answer each question and press Submit to see your result right away.

1. What did the Logic Theorist do in 1956?
Found a bug? Let us know!
2. Which conjecture did the program EQP prove on its own in 1996?
Found a bug? Let us know!
3. How big was the 2016 Boolean Pythagorean triples proof?
Found a bug? Let us know!
4. What did AI score at the 2025 International Mathematical Olympiad?
Found a bug? Let us know!
5. What happened on November 21, 2026?
Found a bug? Let us know!

Frequently Asked Questions

What was the first theorem proved by a computer?

In 1954, Martin Davis’s program on the JOHNNIAC computer proved that the sum of two even numbers is even. In 1956, the Logic Theorist proved 38 theorems from Principia Mathematica. The 1976 four color theorem was the first major theorem whose proof relied on a computer.

What was the first conjecture proved by AI alone?

The Robbins conjecture, proved in 1996 by the automated theorem prover EQP, written by William McCune, is the classic example of an open conjecture settled by automated search, though people chose the inputs and checked the result.

When did AI win gold at the math olympiad?

In 2025, Google DeepMind’s Gemini Deep Think and an OpenAI model each scored 35 of 42 at the International Mathematical Olympiad, the gold-medal threshold. In 2024, AlphaProof and AlphaGeometry 2 reached silver level.

What did AI do in math in 2026?

AI systems disproved Erdős’s unit distance conjecture, helped formalize Fermat’s Last Theorem in Lean, contributed to Navier–Stokes results, and helped produce a Lean proof of the Riemann Hypothesis that was certified by the Lean community on November 21, 2026.

Is the four color theorem an AI proof?

No. Appel and Haken designed the proof in 1976; a computer checked about 1,900 configurations. It was later fully formalized in Coq in 2005.

Keep Exploring

Sources and Further Reading

Related to This Article

What people say about "AI in Math Timeline: From Early Proofs to 2026 - Effortless Math"?

No one replied yet.

Leave a Reply