Mathematics: From Deductive Science to Actually an Experimental Science

Tiago Veríssimo, PhD student at Newcastle University

Disclosure: This text was written and envisioned by me, any tipos are due to me not being an AI.

I was someone who thought mathematics was purely a deductive science, even identified with the formalist philosophical point of view, and I have to say that until very recently this was what mathematicians used to do; after all we would prove theorems and give these theorems to the community.

With AI now, theorem proving is becoming to be ever more devalued as this process seems to be getting faster and better done by AI. The question is then what do mathematicians actually do in this case?

If you are not proving theorems, you must be doing something. I foresee that in the near future AI will get better at even explaining what these proofs are, even narrating exploratory stories about some theorem or even creating theories to explain unexplored mathematical phenomena. I dare to say, that finding “interesting” phenomena might be not so difficult to an AI, as it will gain “taste” after some training.

With such a technology, I believe the only route for humans to keep up and have a more interesting life than verifying mathematical proofs of AI is to do “mathematical experiments”, as in:

  1. Learn a lot of mathematics
  2. Explore “subjectively curious/interesing mathematical phenomena” via “mathematical experimentation”
  3. Feed the AI with that mathematical experimentation
  4. Take you formal conclusions from an AI workflow

The concept “mathematical experimentation” already exists, however it used to be a second category activity, with the main effort and social value actually in proving the results. This, however, might actually prove to be the golden key to evaluate researchers mathematical capacities. But what is mathematical experimentation?

Mathematical experimentation involves on the top of my head the following:

  1. Computational experiments
  2. Conceptual experiments

Computational experiments is what we have been doing for a while, the so-called “let’s make some examples”, either by hand or computer.

The conceptual experiments seems to me to be the new kid in the block and akin to “vibe coding” in programing where you have a given engineering objective that you want to acomplish, you put that in your team of 10 agents and is typically done with a good enough quality. In mathematics, conceptual experiments would be informed “mathematical vibes” that you get from your study of mathematics or speaking with other people or even AI interactions or even just random processes in your brain.

Look at this example:

You: “I feel that diffentiable functions are continous, they are smooth and surely something smooth is continous?”
AI: “Yes! In fact ….”

Evaluating researchers at the “I feel sentence” would be evaluating mathematical experiment capacity/taste of people, as in:

  1. Why did you proceed with this line of experiments?
  2. Why did you have that initial concept vibe?
  3. Why did you lined up those questions afterwards?

In a sense we would be measuring the subjective mathematical taste of the human researcher. I think this is would be a fair and pratical way to go forwards to evaluate us: the little human researchers.

This may even effect how we present materials in the sense that we should report our mathematical experiment in journals and talks, maybe hiring and grants should be based in the potential of the mathematical taste to do mathematical experiments? What do you think?

As an aside, the prior paragraphs develop my intuition on why the community is having a bad time with AI and Maths, people signed up to be in the deductive science business, not to do mathematical experiments. Furthermore, the actual career incentives were to proof theorems and so do deductive science.

Mathematics seems to have undergone, in pratice, an epistemic flip; this seems to be the reason behind all this crisis.

Yes, it still is deductive because the actual truth comes only after verifying the AI proof, but when it comes to real human work, this is will be all about mathematical experiments, therefore mathematics is becoming experimental in pratice.


Received 29 October 2026.

2 responses to “Mathematics: From Deductive Science to Actually an Experimental Science”

  1. M Avatar
    M

    I think evaluation will be based on what value is attached to someone’s work (in whatever form that work appears).

    An example Look at two people.

    • A researcher who used to publish 1-2 solid paper per year and spend the rest of the time giving talks, teach, peer-review, engage with the math community, and who after 2026 continues to put out 1-2 papers per year, which may with AI assistance now be much more comprehensive and deeper, while still going around and giving good talks and lectures about these papers, teaches, serves the community in some meaningful way.
    • A person who used to publish no paper, or very low level papers, and now suddenly has like somewhere between 1-100 solid or even spectacular results per yer. Certainly, that person would not be able to go around and give good talks, teach the theory, build and inspire people, serve the community in some noble way.

    Even if the results of the second person are maybe very valuable as mathematical statements, I could imagine that second person might not receive that much of applause and nice job offers (in academia, and maybe also not elsewhere depending on their skill). On the other hand, I can imagine that the first person may be a recognized part of a network of experts that have authority and are of relevance to society.

    Certainly, it’s not clear what a mathematician should do in the future to earn a good salary. Given the fact that everything in our universe has limitations, there is a chance that the research frontier may shift just drastically but still can benefit from human input, and that one needs to have a lot of hard earned skills to push the boundaries together with AI systems. And even if AI does literally everything better, the existence of a human society will give meaning for humans to understand the world around them, as a notable number of people might not be satisfied by answers from an oracle.

    I think you may be right. Doing math might be more like doing experimental science, which could include highly complex theoretical experiments to find a way to push AI system to break through some obstacles. And it may be that recognized top publications will often be those that really make an impact on the real world. The bridge from pure math to interdisciplinary applications is really complex, and even with current AI its often not just a button push, probably also not with the next version (maybe with the second next one, I don’t know).

  2. John Cosgrave Avatar
    John Cosgrave

    I’ve been compiling a number of questions that AI (or should that be ‘SI’?) cannot answer. I’ve played with free versions of ChatGPT and Claude (I won’t be spending part of my pension on anything else), but someone I know with access to top-of-the range Claude posed questions from me, questions that Claude couldn’t answer.

    Here, for now, I give but a single example. First recall Euler’s simple result that every prime factor q of M_p = 2^p-1 (p prime) satisfies q = 1 (mod p). I asked Claude if it could find a prime p for which such a q satisfied q = 1 (mod p^2), or else prove that was impossible.

    It couldn’t do either.

    What led me to pose this question?

    During COVID I wrote a substantial document called a ‘Mersenne-Wieferich Odyssey’ in which I stated as a conjecture (with appropriate motivation) that the above was impossible.

    (An equivalent conjecture is that the multiplicative group (let’s call it ‘Z’) of residues modulo M_p does not contain an element of multiplicative order p^2, despite the order order of Z is divisible by p^s, where s = the number of distinct prime factors of M_p.)

    Why did I feel that Claude couldn’t provide a counter-example? Well, I did what anyone else could also do: I checked every known completely factored M_p, together with all known factors of every partially factored M_p, p < 5000.

    Could Claude or OpenAI or … have come up with a counter-example? Of course they could (but I doubt it, as I believe in my conjecture): just throw ‘10,000’ agents and millions of dollars at it…

    But, even allowing that that could have happened, I very much doubt that it could (indeed ever will) answer this simple question: if the above conjecture fails, for how many primes {p} does it fail?

    Okay, I’ll keep it brief.

    Neither could it answer this question: can you find a prime p = 1 (mod 4) (other than p = 53, see a paper of mine with Karl Dilcher: paper #3 at https://johnbcosgrave.ie/publications.htm, see comment after Corollary 3) for which:

    ((p-1)/2)!^2 = -1 (mod p^2) … (i)

    But, again, even if it had managed to find one/more examples, it would not be able to answer this question: for how many primes {p} does (i) happen? (hats off if it ever will!)

Comments are moderated. Read our comment policy.

Add to the discussion

New posts by email.

Prefer a feed reader? Subscribe by RSS.

Also on Mathstodon.

Latest comments across the site.

  1. Minor correction to the part from the Gita (with some help from my ChatGPT):Arjuna already knew that Krishna was divine,...