
Proof Assistant
A proof assistant is a program that doesn't just write down mathematical proofs, but checks them step by step by machine. This allows errors to be found in proofs and in safety-critical software that humans would overlook.
A proof assistant is a program that checks mathematical proofs. You formulate a proof in a formal language — a language in which every step is precise enough for a machine to read it unambiguously. The proof assistant then checks each individual inference: Does step three really follow from step two? Does the logic hold all the way to the end? A human can read through a hundred pages of proof and still overlook a tiny mistake. The program doesn’t. It only accepts a proof if it is completely correct — not a single step may be missing or invalid.
Why mathematical correctness is so hard to guarantee
Mathematical proofs can be very long and intricate. The proof of the four color theorem — which states that any map can be colored with four colors so that no two neighboring countries share the same color — comprises thousands of cases. When it appeared in 1976, the field doubted for decades whether it was really watertight. A proof assistant resolves this problem fundamentally: either the proof is accepted in full, or it is rejected. There are no gray areas.
The same applies to software. At its core, a program is a sequence of instructions, and one can mathematically prove that a program does exactly what it’s supposed to do — and nothing else. For software running in aircraft control systems, pacemakers, or encryption systems, this is no academic exercise. An overlooked bug there can cause real harm.
How one works with a proof assistant
First, you give the proof assistant a statement you want to prove — for example: “The sum of two even numbers is always even.” Then you build the proof step by step, in the system’s formal language. After each step, the proof assistant shows what still remains to be proven. This open target is called a proof goal. You work on it until nothing is left.
The process is interactive: the program is not an automaton that invents proofs on its own. It is more like a strict grader who checks off each step and immediately flags anything that’s wrong. Some proof assistants now have built-in tactics that automatically handle small sub-steps. But the actual proof is still devised by the human.
The best-known systems are called Lean, Coq, and Isabelle. They differ mainly in their formal language and in which kinds of proofs they support particularly well. Lean has received a lot of attention in recent years because a growing community of mathematicians is formalizing current research results in it.
Proof assistants in current AI debates
In AI research, proof assistants are currently seen as one possible way to make large language models more reliable. The problem: a language model like a chatbot can write convincing-sounding solutions to math problems that are nevertheless wrong. If you couple the model with a proof assistant, the assistant can immediately check each proposed solution by machine. Incorrect answers are automatically filtered out.
OpenAI, Google DeepMind, and other labs are working on systems that do exactly this. DeepMind’s model AlphaProof solved several International Mathematical Olympiad problems in 2024 by generating solution ideas and checking them directly in Lean. This is no longer a classic proof assistant, but it would not have been possible without one.
For the broader tech industry, this is relevant because verified software and verified AI outputs are becoming a growing topic in regulation. Whoever can prove that their system works correctly has a concrete advantage over regulators compared to someone who can only say: “We tested it thoroughly.”