21 November, 2026

The strange thing about AI proofs

Earlier, whenever there was a scientific breakthrough, media would be abuzz with several efforts to explain the key insights of this breakthrough in simpler language.

But these days, with AI-based scientific breakthroughs--be it the Navier-Stokes problem, or the Riemann Hypothesis--we don't see any explanations of what the proof was. Except that it was a breakthrough, and AI achieved it. And the buzz then shifts to AI.

So why are we not seeing any explanations of the proofs? Because we have not understood the proof yet. The proofs generated by AI are thousands of pages long and they are still being validated.

So how then do we know that AI has proved something?

It is like this. There is this formal theorem-proving language called Lean in which logical arguments can be strung through. The theorem to be proven is the final assertion. If the Lean program compiles successfully, then it means that the theorem entails from the premises. And what AI has generated is a proof written in Lean that has successfully compiled! So we know that there is a proof, just that we still do not know the logical flow.

But this kind of proofs also reveal something else--that these open questions had computable proofs! What does this mean?

Lean, or any other formal system of logic can generate an infinite number of logical entailments. But this infinity has a specific property--it is countable! That means that we can assign a unique natural number index to each entailment. We can, for example, say that the proof generated is the 309823674651244735273683467th proof of the Lean universe. And that would be a unique tag that represents this proof. 

In fact, all that we can compute using any formal system of logic is a countable universe. 

Note that, the LLM itself may be doing some uncomputable magic (which is debatable!). But the proof is in the Lean code--which is plain formal logic, that is in the computable universe, and no more expressive than computer programs.

One of the first mathematical theorems to be proven using computers was the Graph Colouring problem. This problem states that a planar graph can be coloured in a way that no two adjacent nodes have the same colour-- using just four colours. The proof, when it was published, was an anticlimax of sorts. The proof basically showed that there are a finite number of colour combinations that can occur in a graph that is planar (that is, where edges do not cross one another). And showed that four colours are sufficient for colouring all these cases. 

The main argument against such computer generated proofs was that it was primarily a brute-force search in a countable, formal logic space. A proof often requires deeper intuition that leads to fundamental insights about the nature of reality. It is not just the result itself that gets celebrated, but the proof technique itself becomes a tool that opens up new possibilities. 

Fundamental scientific breakthroughs made by humans, have often arisen from intuition and conjecture that can be termed uncomputable. For instance, why did Riemann believe that all zeros of the zeta function lay on a single line? Also, there are even older unsolved problems about prime numbers--older than the Riemann hypothesis. One example is the Goldbach's conjecture, which states that: Every even integer greater than 2 can be expressed as the sum of two prime numbers. This conjecture has been validated for very large numbers like 4x10^18 but no proof has arisen yet. But what made Goldbach come up with this conjecture? 

The fact that not all problems are computable, has its roots in a result from set theory whose intuition was got by Georg Cantor. He basically showed that there are an infinite number of infinite sets with "bigger" and "smaller" infinities! Georg Cantor himself spent his last days in a mental asylum, trying to figure out whether this infinite set of infinities was itself countable or uncountable! (An open problem called the Continuum Hypothesis). 

In order to prove this intuition that we have in our minds, we will need to invent clever techniques and new conceptualisations that convert this intuition into a logical proof. That technique that is invented is in fact, the primary breakthrough, which will then go on to become an important mathematical tool in itself.

For instance, Georg Cantor, in order to show that there are infinite sets of different sizes, developed an argumentation technique called the "diagonal-slash argument". This is based on assuming the contrary--that any two infinite sets does have the same size, and devising a clever method to derive a contradiction with this assumption. This novel proof technique was the oyster in this proof. It would go on to get used by several other mathematicians to prove other important results. Most notably, this technique was used by Alan Turing to show that there exist some uncomputable functions--or that the set of all problems we can compute are of the same size as countably infinite sets! 

Similarly, much later, a group of three researchers: Fischer, Lynch, and Patterson wanted to show that we can never guarantee consensus in a distributed system that is asynchronous. In order to do this, they developed a proof technique called Adversarial Theorem Proving. This is a generalisation of proof by refutation that is applied to systems. This assumes that for whatever system we build, there is an adversary that is trying to disrupt the functioning of the system. If we can show that it is impossible for such an adversary to exist (which is unlikely in most cases), or that such an adversary requires an impractical amount of computational power or memory to disrupt the system, we have essentially "proven" that our system is resilient. This technique is now pretty-much the standard for proving any system theoretic results in computer and information sciences. For instance, all our web security is based on cryptographic algorithms for which, an adversary would require an impractical amount of computing, in order to break the encryption. And this "impractical amount of computing" property is true only when we consider classical computers. Quantum computers for instance, can create adversaries that can break current-day encryption techniques! So we are now looking for techniques that can be resilient against quantum adversaries (a field of study called post-quantum cryptography). 

This is the characteristic of any scientific breakthrough. The result itself is one thing--but the clever technique that was needed to be developed in order to get the result, is the one that will go on to have a much greater impact on the field of study.

An AI proof on the other hand, shows that the proof of this open problem is computable and just required enormous computational muscle to prove--rather than a new, clever approach like the diagonal-slash or the adversarial theorem proving. Here, other than the fact that the open question is now no longer open, it does not create any new possibilities. 

My advisor once had told me about what makes a good thesis. A run-of-the-mill thesis is one that fills a research gap, while a good thesis is one that opens up new avenues of thought. A mediocre thesis reduces the amount of unknowns in the field, while a real breakthrough is one that opens new avenues of thought and basically ends up creating new open problems.

As researchers, we are excited when we find problems and possibilities--not when we find solutions. 

Scientific breakthroughs happen not when we discover that a path exists from our current knowledge to an open question, but when we discover that a short path exists, that can be traversed by the human mind! 

The strange thing about AI proofs

Earlier, whenever there was a scientific breakthrough, media would be abuzz with several efforts to explain the key insights of this breakth...