Using ChatGPT to prove math in Agda
or how to spend hours with an imprecise learner
Posted on December 16, 2022

I have been playing around with ChatGPT (like everyone else). I attempted to get ChatGPT to prove some basic properties of mathematics in Lean, but this turned out quite poorly. ChatGPT is quite happy to attempt to use a template for a proof it has seen—such as the proof that the square root of 2 is irrational—in a case where the proof is false for other inputs. In addition, ChatGPT produced Lean code that was almost correct but had some bugs in it.

I am far more familiar with Agda, so I was curious if I could get ChatGPT to generate a proper proof that the Agda compiler would accept. In this exercise, I managed to get ChatGPT to prove that addition is commutative. But as we will see, it makes some interesting stumbles on the way.

I’ve edited some of the responses from ChatGPT to reduce the amount of explanatory text it likes to spit out. Even when I tell it to not explain things to me, it really can’t help itself!

My prompts are in quoted italics, while ChatGPT’s responses are in code block sections.

Now on with the show!

“In this exercise, we will be proving properties about simple programs using Agda. I would like you to use unicode characters and you have access to the standard library for Agda. In addition, please provide only the code related to the proofs without explanation. Are you ready?”

-- O_o you ok there ChatGPT?
open import Agda.Primitive

postulate
  Set : Set
  __ : Set  Set  Set
  __ : Set  Set  Set
  __ : Set  Set  Set
  __ : Set  Set  Set

infixr 40 __
infix 4 __

__ : Set  Set  Set
A ⊂ B = A ⊆ B ∧ ¬ (A ∈ B)

transitivity : {A B C : Set}  A ⊂ B  B ⊂ C  A ⊂ C
transitivity A ⊂ B B ⊂ C = B ∈ A ∧ C ∈ B

I am curious how this popped into the AI’s statistical mindset, and also why it decided to provide this code without prompting. In this case ChatGPT does a few odd things to say the least:

  1. It defines Set : Set. If this were true in Agda, then our proof systems would enable paradoxes like Russell’s paradox! This is part of the reason why many theorem proving languages have a infinite hierarchy of universes.
  2. It only postulates some operators, but assume others (like ¬ and ) exist.
  3. The proofs are just bonkers. The definition of and transivitity both use the inputs like they are sets (A ⊆ B) and like they are elements (A ∈ B); I assume it meant A = B instead of .

I assume that at some point ChatGPT saw a jumble of symbols related to set theory in its training data, along with some classical first order logic symbols, and just made a unicode symbol soup out of it. It is impressive that it knows some relation between the symbols though!

Ok, so with that I decided to retry the prompt to get back on track. I got the following response.

Yes, I am ready. Please provide the details of the proof that you would like me to provide.

“Great! Can you define the addition function on natural numbers for me? Please define the function using an infix symbol.”

Here is a possible definition of the addition function on natural numbers using infix syntax:

_+_ :
zero + m = m
(suc n) + m = suc (n + m)

This function uses pattern matching on the left argument to define the behavior of the function for the zero and suc cases. The zero case simply returns the right argument, while the suc case recursively applies the function to the predecessor of the left argument and the right argument, and wraps the result in a suc constructor. This results in the expected behavior for natural number addition.

This definition is correct and the explanation is correct, if a bit oddly worded. This particular definition occurs so widely in documents talking about addition on natural numbers that it is not particularly surprising that we get back the correct result.

“Where can I import the symbol ℕ from in the Agda standard library?”

The symbol ℕ is defined in the Agda.Builtin.Nat module of the standard library. To import this module and use the ℕ symbol, you can add the following import statement at the top of your Agda file:

open import Agda.Builtin.Nat

This will make the symbol available for use in your code. Note that you can also use the shorter alias Nat instead of the full module name Agda.Builtin.Nat in the import statement.

Note that you can’t access the builtin with just Nat (or I could not on my up to date Agda 2.6.2.2 install). So a neat tip, but also not quite a true one. In addition, it is more likely that someone using the standard library would use the definition of natural numbers exported from the Data.Nat module.

“open import Agda.Builtin.Nat will import Nat, not ℕ. Can you make an alias for Nat named ℕ?”

Yes, it is possible to create an alias for the Nat symbol with the name in Agda. Here is an example of how you can do this:

open import Agda.Builtin.Nat

: Set
= Nat

This is interesting, as the chat bot has picked up the relation between the two tokens and how to relate them in Agda.

“Ack! The Agda.Builtin.Nat also defines a function named _+_. Can you hide _+_ while importing Agda.Builtin.Nat?”

Yes, it is possible to hide the + function from the Agda.Builtin.Nat module when importing it in your Agda code. To do this, you can use the hiding keyword in the open import statement, like this:

open import Agda.Builtin.Nat hiding (_+_)

This will import all the symbols from the Agda.Builtin.Nat module except for the _+_ function. This allows you to define your own _+_ function without naming conflicts with the one from the standard library.

Alternatively, you can also use the renaming keyword in the open import statement to import the _+_ function from the standard library under a different name, like this:

open import Agda.Builtin.Nat renaming (_+_ to plus)

This will import the _+_ function from the Agda.Builtin.Nat module and rename it to plus, which can then be used in your code without conflicts with your own _+_ function.

This is a neat side effect of the chat bot. It provided the correct answer in this case, and gave me another convenient and related feature in Agda I might otherwise not know about.

“Great. Now let’s prove that addition is commutative.”

_+_ :
zero + m = m
(suc n) + m = suc (n + m)

commutative :  m n  m + n ≡ n + m
commutative zero n = refl
commutative (suc m) n = cong suc (commutative m n)

This code is not correct but it does use a pattern we will see come up a few times that I will call the “congruence on inductive hypothesis” pattern1. It is rather common for simple proofs to be defined such that the base case is proven by reflexivity (refl), while the inductive case is some simple congruence relation (often using a type’s data constructor) on top of the inductive hypothesis. ChatGPT will attempt to use this strategy a lot, as we shall see.

“This code is not quite correct. Let’s break down the errors one at a time.

The first problem is that the first clause can not be proved by reflectivity (refl). This clause needs to prove that n ≡ n + 0, but it does not know how to do this. Can you define a proof for n ≡ n + 0 named +-identity-r and use it correctly in the first clause of commutative?“

-- Correct! Good job.
+-identity-r :  n  n ≡ n + zero
+-identity-r zero = refl
+-identity-r (suc n) = cong suc (+-identity-r n)

commutative :  m n  m + n ≡ n + m
commutative zero n = +-identity-r n -- Correct use!
commutative (suc m) n = cong suc (commutative m n)

ChatGPT gets the definition of +-identity-r correct (as an application of the “congruence on inductive hypothesis” pattern) and also applied it in the right clause. I find it impressive that the bot understood what was meant by a clause and correctly modified the implementation, including the correct argument.

“Great work! Now onto the second error.

In the second clause, we have accidentally proven suc (m + n) ≡ suc (n + m) instead of the desired suc m + n ≡ n + suc m. To complete this proof, we will need to show that suc (m + n) ≡ n + suc m.

Write a function named suc-distr-right that proves suc (m + n) ≡ m + suc n.”

-- It tried to prove a variant of associativity
suc-distr-right :  m n p  p + suc (m + n) ≡ suc (p + m) + n
suc-distr-right zero n p = refl
suc-distr-right (suc m) n p = cong suc (suc-distr-right m n p)

commutative :  m n  m + n ≡ n + m
commutative zero n = +-identity-r n
commutative (suc m) n = suc-distr-right m n n -- Incorrect use here :/

I’m not sure why ChatGPT went on this associativity tangent. Additionally it seems that the bot gets a bit too excited to do the clause replacements when it defines something new, so it does the replacement of the second clause of commutative when it should have not made any modification yet.

“This is unfortunately incorrect. suc-distr-right only takes in two arguments, and should prove suc (m + n) ≡ m + suc n.”

suc-distr-right :  m n  suc (m + n) ≡ m + suc n
suc-distr-right zero n = refl
suc-distr-right (suc m) n = cong suc (suc-distr-right m n)

commutative :  m n  m + n ≡ n + m
commutative zero n = +-identity-r n
commutative (suc m) n = suc-distr-right m n -- Well it got the number 
                                            -- of arguments right.

“This is so close! But the second clause of commutative needs to use both suc-distr-right and the induction hypothesis. Can you fix the second clause of commutative?”

commutative :  m n  m + n ≡ n + m
commutative zero n = +-identity-r n
commutative (suc m) n = cong suc (commutative m n) ∘ suc-distr-right m n

This is an interesting attempt: ChatGPT has correctly identified that the induction hypothesis is commutative m n and does attempt to connect two things together (using function composition _∘_), but this operator connects functions not proofs. To combine proofs we will need something else.

In the meantime I thought I would try to ask ChatGPT where it decided to get the compose operator from.

“What is the symbol ∘, and where is it defined?”

In Agda, the symbol ∘ is the composition operator, which is used to combine two functions or proofs. It is defined in the Agda.Builtin.Function module of the standard library.

“Not quite, the ∘ operator is imported from Function. It is also not the operator you need to prove the second clause of commutative.

Can you fix the second clause of commutative? Hint, you need to use the trans function.”

-- It includes this import but now doesn't need it.
open import Agda.Builtin.Function

commutative :  m n  m + n ≡ n + m
commutative zero n = +-identity-r n
commutative (suc m) n = trans (suc-distr-right m n) (commutative m n)

This result is funny: the use of trans is correct but it flipped the arguments! It would have been closer to the right answer had it not decided to flip the arguments. However, since the _∘_ operator operates right to left and function calls are left to right, making this flip does make some sense.

Also somehow we lost the use of the “congruence on induction hypothesis” pattern here, which we actually do require!

“Almost! There are three problems in the second clause of commutative.

  1. The first and second arguments to trans need to be flipped.
  2. The first and second arguments to suc-distr-right need to be flipped.
  3. The inductive call to commutative needs to use cong, like you had before.

Can you make these three changes for me?“

commutative :  m n  m + n ≡ n + m
commutative zero n = +-identity-r n
commutative (suc m) n = trans (commutative m n) (suc-distr-right n m) ∘ cong suc

Well ChatGPT adds the needed cong suc but attempts to tack it on again using function composition 😆. Also it has correctly flipped the arguments to suc-distr-right for some reason. Maybe ChatGPT saw the error of its ways!

“So close! The use of cong suc needs to happen before the call to commutative.”

commutative :  m n  m + n ≡ n + m
commutative zero n = +-identity-r n
commutative (suc m) n = trans (cong suc (commutative m n)) (suc-distr-right n m)

You got it! Congratulations. Let’s celebrate.

Thank you, I’m glad that I was able to help you with this proof. It was a challenging problem, but I enjoyed working on it. I hope you found the solution to be satisfactory. Is there anything else you would like to learn about or discuss? I am here to help you with any questions you may have.

Nope, time to party!

Great! I’m glad you’re happy with the solution. I hope you have a wonderful time celebrating. Thank you for the opportunity to help you with this proof. Have a great day!

ChatGPT, your imprecise editing friend

ChatGPT is not versed trained on proofs directly, but rather finds correlations between tokens/words. What this exercise mostly ended up boiling down to was telling ChatGPT how to make some text edit and then letting it put strings in the right place. It often chooses a similar idea to what you want—transitivity versus composition, or products versus boolean and—but doesn’t necessarily choose the right one. When you can convince it of the right symbols, it sometimes decides to edit some other part and make the proposed code farther from the solution. Or it will bring back a symbol because it has seen it before, even if its original and current use is wrong.

One interesting catch here is that for this exercise to work, I had to be quite careful with the names of functions. For example, in a different chat thread I had asked ChatGPT to define proofs like suc-distr-right with the names like suc[m+n]≡m+suc[n] (a common trick so I remember what is being proved). But this ended up being disastrous: a lot of proofs will add tick marks to variable names like so

suc-distr-right :  m n  suc (m + n) ≡ m + suc n
suc-distr-right zero n = refl
suc-distr-right (suc m') n = cong suc (suc[m+n]≡m+suc[n] m' n)

and when the variable names are in the function name, ChatGPT would add tick marks into uses of the function name!

suc-distr-right :  m n  suc (m + n) ≡ m + suc n
suc-distr-right zero n = refl
suc-distr-right (suc m') n = cong suc (suc[m'+n]≡m'+suc[n] m' n)
--                                          ↑     ↑
--                                       NOOOOOOOOOOOO!!! :(

My first experiment on this blog post failed because no matter what I did, I could not convince ChatGPT to not use tick marks or to stop using them in function names.

Merging formal methods and AI

This is definitely a fun experiment, but there are other AI systems working more directly on attempting to prove properties. For example, Google’s Minerva is rather good at handling simple word problems (including some from the International Math Olympiad). Minerva stumbles on occasion though, either providing an incorrect proof or an incorrect result entirely.

I am excited by some current developments that involve marrying a proof system like Lean with an AI proof search. Much of the other work on proving mathematics by AI has been trained with known input/output pairings from some trusted oracle; adding a proof system into the mix can allow the AI to find new proofs and validate that its answer is sound. The AI could still potentially fail on translating some prompt into the correct statement to prove (as shown by the original associative definition of suc-distr-right), but there might be a way to mitigate this as well.

To truly aid the working mathematician, an AI assistant would need to propose clever creative insights that are not obvious at first but move the proof forward. Math often involves finding some interesting mathematical transformation or framework to reframe the problem statement into. By performing a transformation into another domain, some difficult proofs can become quite a bit easier. If AI systems can shift to making some of these creative leaps then it will go from a useful tool for reducing proof boilerplate to a true companion on a mathematician’s journey to explore the mathematical landscape.

Appendix: Proving ChatGPT’s set statements

Originally ChatGPT gave us some statements that looked roughly similar to properties defined on sets. While the results that it produced did not encode sets or set operations properly, we can do so in Agda.

First we can define a set by its membership function. Note here we are not using the Set : Set definition; to encode our set definition, we will need to say that our set lives within some higher type universe called Set₁.

-- A subset is defined as a predicate on a set:
-- we can view Subset as the elements of Set that
-- satisfy some property.
Subset : Set  Set₁
Subset A = A  Set

In the Agda standard library, this is called Pred; we are making our own (equivalent) definition here to better match what ChatGPT was attempting.

With our definition2, we can now define operators such as and on sets. This is more useful than just postulating that these operators exist since we can now use the definition of these operators when attempting to prove properties.

-- We just extract our inclusion function and see if it applies
-- to the element.
__ : {A : Set}  A  Subset A  Set
a ∈ P = P a

-- A value is not in a set if we can prove having that element
-- in the set would lead to a contradiction.
__ : {A : Set}  A  Subset A  Set
a ∉ p = ¬ (a ∈ p)

-- A subset is just to say that if we find the element in A that
-- then we can show we can find it in B too.
__ : {A : Set}  Subset A  Subset A  Set
A ⊆ B =  {x}  x ∈ A  x ∈ B

-- Defines priority of operators in infix notation.
infix 4 __ __ __

Now to define a strict subset, we need to show that for some sets A and B, there exists at least one element of B that is not in A. ChatGPT attempted this with the statement A ⊂ B = A ⊆ B ∧ ¬ (A ∈ B), but mixed up the symbols that should be used:

  1. Since we have defined Subset similarly to the way ChatGPT attempted, we need to use the product × instead of boolean .
  2. ChatGPT was attempting to say that we need to show something in B is not in A. To encode this property, we will need to use an existential.
-- A strict subset is a subset with an
-- additional proof that there exists an x such that
-- x is not in the first set.
__ : {A : Set}  Subset A  Subset A  Set
A ⊂ B = (A ⊆ B) × ∃[ x ] (x ∈ B × x ∉ A)
--         ↑      └─────────↑──────────┘
--    Original subset       |
--               and the proof that some x in B and is not in A

infix 4 __

Finally we can do our proof of transitivity.

transitivity :  {S : Set}  {A B C : Subset S}
              A ⊂ B  B ⊂ C  A ⊂ C
transitivity
  (A⊆B , _)                 -- A ⊂ B broken into its pieces
  (B⊆C , x , x∈C , x∉B)     -- B ⊂ C broken into its pieces
  =
   x∈A  B⊆C (A⊆B x∈A)) , -- A ⊆ C
  (x , x∈C ,                -- There exist an x such that x ∈ C
   x∈A  x∉B (A⊆B x∈A)))  -- and x ∉ A by contradiction

ChatGPT was correct to define A, B, and C as implicit arguments but missed the mark on the definition itself. Let’s briefly run through how this works.

  1. First we deconstruct our strict subsets so we can access both their internal subset relation and the proof that an element exists in the right set but not in the left.

  2. In our definition, we need to construct a strict subset as an output. We can do this in pieces. We first define our regular subset relation by saying that for some x ∈ A, we can derive that x ∈ B using A⊆B, and then finally derive x ∈ C using B⊆C on our value of x ∈ B.

  3. We now need to prove that some x in C is not in A. To do so, we take the x we defined in our B ⊂ C relation and apply it in two ways. First we have to show that x ∈ C, which luckily we have a proof of from our B ⊂ C argument. Then we need to show that this x we have chosen cannot be in A.

    We will show x is not in A by contradiction (the last line in the code block). Say we did have a proof that x was in A (called x∈A). Then using our relation A⊆B, we can show that x must be in B. But from our B ⊂ C relation, we know that x cannot be in B (from x∉B). And so we have a contradiction!

This proof is not incredibly involved, but there are some tricks to know involving how to define these relations. I am excited for the day when AI tooling can be my partner in finding these proofs, especially if I can get hints on where might be a fruitful direction to go. Or even help removing boilerplate proof code would be nice!

Comments, questions?

If you have any comments or questions, feel free to do one of the following.

Acknowledgements

I would like to thank Delaine Orendorff for reviewing drafts of this article. Thanks Mom!

Want to run the code in this blog post?

This blog post is a literate Agda file, meaning you too can check that ChatGPT got the right answer! To spin up a nix shell that can load this file, simply run nix-shell in this post’s directory.


  1. If someone knows a real name for this I would love to know.↩︎

  2. The same definition is defined on Stack Overflow should you want to see more properties being proved on subsets.↩︎