Column · Metadialoghi

Marco Ripà: research in the age of AI, between the power of the tool and human responsibility

Interview with Marco Ripà · Metadialoghi column · Part 2 of 2 · 20 November 2026

A painting in shades of yellow, light blue, green and red, with a formula on tetration at the centre and its conditions in the corners
The painting with Marco Ripà's formula on tetration (Ripà & Di Pietro, 2026). Photo courtesy of Marco Ripà

Second part of the interview with Marco Ripà, independent researcher in mathematics. He describes what changes when artificial intelligence works alongside the person doing research: surprising results, limits, verification and a responsibility that remains human.

This is the second of two parts, which can also be read separately. In the first part: how to become a researcher without a linear path.

Marco Ripà

Independent researcher in mathematics

Marco Ripà is an independent researcher. He graduated with honours in Economics, trained in mathematics as a self-taught scholar and works on number theory, graph theory and multidimensional geometric optimisation problems, publishing in peer-reviewed journals as well as preprints and books. He is the sole author of more than 200 sequences in the On-Line Encyclopedia of Integer Sequences (OEIS). Between 2011 and 2017 he also worked on psychometrics and giftedness, designing dynamic, software-generated cognitive tests. Today he explores the use of artificial intelligence models in mathematical research and, alongside his scientific work, is active in outreach and education, especially through his YouTube channel.

YouTube channel → · LinkedIn profile →

The opinions and interpretations expressed in the interview are those of the interviewee and do not necessarily reflect the position of Starry Mind Academy.

You have started experimenting with advanced language models on original mathematical problems. How have these tools changed your research process? Could you describe a particularly representative episode, including the difficulties or limits you encountered?

There was a precise moment when I realised that generative AI models had reached non-negligible mathematical capabilities, to the point of becoming genuinely useful tools in scientific research.

It was early August 2025. Noticing the release of the new GPT-5 model, I decided to test it on one of the first small open problems I had solved as a complete amateur, in 2010: a question proposed by Kenichiro Kashihara about the distribution of the final digits of the elements of a certain number sequence.

The choice of test was not random: I knew the problem perfectly, I had published a more general solution than the one the author had originally asked for, and I remembered well that it had taken me two full hours to obtain it. So I had a personal benchmark on which to assess the new model's performance, without having to trust blindly what it would produce.

Well, in less than five minutes and without consulting external sources, GPT-5 managed to produce on its own a formula equivalent to mine from 2010. The result amazed me so much that I decided to document the experiment in a preprint (Ripà, 2025).

For readers who want the mathematical details: what more GPT-5 had obtained
The model had also derived an exact formula for the intermediate positions between one complete block and the next. In my 2010 solution I had limited myself to providing an exact formula for complete blocks and to indicating, in note 14, bounds within which the value sought fell in the intermediate cases: an extension not required by the original problem, but certainly a welcome one (Ripà, 2010, 2025).

Between surprising results and the limits of models

That test made me begin to suspect that the time might already have come to consider LLMs (large language models, such as those behind ChatGPT) as possible assistants in mathematical research, but the tests that followed were rather disappointing. For example, the same model could not even find a simple continuous path of three segments visiting all four vertices of a unit square in the Cartesian plane. In short, I had in front of me a tool capable of generalising in a few minutes a mathematical result that had cost me hours of reasoning, but that could still fail on elementary geometric questions and produce frequent "hallucinations" (answers containing false or invented elements).

In fact, even before they could be considered real collaborators on the mathematical side, LLMs had proved useful for writing Python programs much faster than I could have done on my own. Not that they produced flawless code, far from it: I often had to step in with numerous corrections, then check the results with further cross-tests. Even so, this was already a considerable advantage in computational research and in exploring some promising lines of research.

From support to collaboration

Since then I have witnessed impressive progress, to the point of seeing myself clearly surpassed, in my own areas of expertise, by the most advanced versions of the models. In September 2026 I completed my first scientific manuscript developed in full synergy with AI, starting from a discovery of mine discussed on MathOverflow, a platform where mathematicians exchange questions and ideas: formalising the most general results did in fact require considerable effort.

Meanwhile, the mathematical capabilities of the models have kept advancing so fast that they have provoked deep reactions even among some of the greatest contemporary mathematicians. I am thinking, for example, of the French mathematician Hugo Duminil-Copin, 2022 Fields Medal, who said he cried after learning of the release, in October 2026, of more than 700 mathematical manuscripts produced by an internal OpenAI model (Larousserie, 2026; OpenAI, 2026). Of course, this does not mean that every announced result is automatically correct, original or already verified by the scientific community.

I have no difficulty admitting that, in my experience, GPT-6 in its High setting is already, on average, more capable than I am at attacking well-defined problems in number theory, proposing results and building proofs that, after appropriate checks, can turn out to be mathematically valid. Just imagine what a leap a technology like this can represent for an enthusiast who, until a few years ago, would not have had the chance to tackle certain problems.

It would therefore be reductive to consider these tools merely multipliers of the skills we already have: they can allow us to obtain results that, without their help, we probably would not have been able to derive.

“The point, if anything, is to understand how to exploit their potential without confusing the capabilities of the model with those of the person using it”

A window of opportunity

As far as I am concerned, the initial advantage of assisted programming is gradually being joined by the possibility of tackling complex mathematical problems through several operations:

  • breaking them down into more limited steps;
  • critically discussing the models' outputs;
  • comparing different strategies;
  • then putting the results back together into a coherent argument.

This requires adequate knowledge of the subject and the ability to organise and evaluate the work done with AI.

I see a window of opportunity for independent researchers and competent enthusiasts who can now access tools of a power unthinkable until very recently. I believe, however, that this favourable phase may be short: large research groups, with structured teams and more powerful models, could gain an advantage that will be hard to close on the most ambitious problems.

You mentioned the possibility that artificial intelligence might amplify, or instead constrain, human creativity. In your experience, when does a model encourage the exploration of new ideas, and when does it risk steering reasoning towards solutions the researcher might accept too easily?

A somewhat absurd comparison that comes to mind is being given an extraordinarily powerful sports car, fitted with racing tyres and even capable of leaping hundreds of metres to get past traffic and road barriers. A car that, however, has some flaws in its control systems and tends to make decisions on its own to minimise journey time.

Of course, we can switch to manual driving and use it like an ordinary small car: it will still perform better than our old car, but those are certainly not its best features. On the other hand, an expert driver could exploit its potential to improve their lap times enormously and even win prestigious races which, with the means they had before, they would not even have managed to qualify for.

There you go: I think today's LLMs lend themselves quite well to this deliberately paradoxical comparison. Having an extremely powerful tool does not automatically mean being able to use it effectively: you need knowledge of the field in which you are using it, an overall view, critical thinking and the ability to guide a process that may also change direction along the way.

Widening the range of possibilities

Even the choice of problems on which to focus your efforts is crucial. Today LLMs make it easier to find connections between results, even ones belonging to different branches, and in the future they might suggest ever more original and deeper connections. In the immediate term, they mainly make it possible to shorten the preliminary exploration: we can tackle several problems in parallel or compare different approaches to the same problem, obtaining leads that in the past would have required weeks of work. I am thinking, for example, of the possibility of comparing techniques from different disciplines, looking for counterexamples, writing verification programs or examining a very large body of scientific literature.

Of course, an LLM cannot read hundreds of articles for us and automatically guarantee a correct reconstruction of the state of the art. It can, however, help us direct and speed up a search that, with traditional methods alone, would take much more time.

The creative advantage therefore also lies in being able to explore more possibilities quickly, discarding some directions and pursuing others in more depth. A researcher who once could follow one or two promising leads can now examine many, and perhaps spot connections they would not have had time to consider.

The flip side is that these same tools can deeply condition our reasoning. A particularly convincing model can steer us towards a strategy, propose a seemingly plausible lemma and build an entire proof on top of it. If we do not check the steps critically, we risk investing time in a false premise, perhaps abandoning personal intuitions that deserved more attention.

Creativity, delegation and autonomy

Then there is a more general risk, especially in education. If we get used to systematically delegating to AI the choice of problems, the development of hypotheses and even the understanding of solutions, we may exercise less and less the cognitive and creative abilities we should be cultivating. It is a bit like having an extraordinary car at our disposal and giving up walking altogether: the vehicle takes us further, but in the meantime some of our own abilities may deteriorate.

In advanced research another difficulty also emerges: the most powerful models can produce proofs so complex that they require considerable human work to understand them, reorganise them and bring them back to mathematical ideas that can be assimilated. James Maynard, 2022 Fields Medal, recently highlighted how difficult it is to extract human mathematical understanding from a long proof produced by AI (Brumfiel, 2026).

And here we are back to the supercar driver: reaching the finish line is not enough; you need to understand how you got there, whether the route was correct and what possibilities opened up along the way. These questions are also discussed by mathematicians far better than yours truly, first and foremost Terence Tao (2026a, 2026b).

As far as I am concerned, in well-defined contexts and when used appropriately by a capable user, AI can amplify human creativity, even allowing us to obtain results we would not have reached on our own; but it can also compress it, if we passively accept its proposals.

“The difference depends on the model's capabilities and on how willing we are to keep reasoning independently while we use it”

In mathematical research, an answer that is convincing on a linguistic level may still contain errors of reasoning. How do you check the results produced by LLMs, and which abilities do you think those who study and do research should keep cultivating, especially as these tools become ever more powerful?

In so-called research-level mathematics, a convincing answer is not enough: you need rigorously proven results.

In a mathematics paper, at least the way I am used to writing them, propositions, lemmas, theorems and corollaries are accompanied by complete proofs, built from definitions and hypotheses stated beforehand. The techniques can be relatively elementary, such as mathematical induction or reductio ad absurdum, or combine into more complex arguments. What matters is that every logical step is justified and that the conclusions follow from the premises.

Checking what the model has actually proved

When an LLM proposes a solution, the most pressing question is to understand what it has actually proved. In the past these models would even produce circular reasoning just to go along with a request, for example if asked to "prove" that the Earth was flat; today they generally develop much better structured arguments. However, it is not enough for an argument to look elegant or to use appropriate mathematical terms: you have to check the hypotheses, verify the logical implications and make sure that the result matches what you wanted to prove.

Someone experienced in mathematical research and in using these models will also notice flaws in the presentation: variable domains not always defined, symbols that are superfluous or used before their meaning has been made clear, sometimes the same variables reused to denote different objects. These are problems also found in the writing of amateurs and of researchers who are not used to scientific publication.

A more insidious error consists in proving a statement slightly different from the original one, implicitly introducing an extra condition or overlooking edge cases. Making the conditions explicit in the prompts is therefore crucial. The model may also invoke a correct result and apply it in a context where its hypotheses are not satisfied.

That is why I test the conclusions through numerical checks, searches for possible counterexamples and, where appropriate, independent verification programs. No finite number of experimental tests can replace a general proof, but these checks help to catch errors that might otherwise spread while a rigorous proof is being worked out.

For readers who want the mathematical details: Lean 4
Today we can also turn to formalisation systems such as Lean 4, which can rigorously check (and validate!) the logical correctness of proofs expressed in its formal language (Avigad et al., n.d.). It is a tool of enormous value, but it does not automatically settle every question: you need to check that the formalised problem corresponds to the one we meant to tackle. Moreover, a formally valid proof does not in itself guarantee the scientific originality of the result.

Which abilities to keep cultivating

This raises a question I consider particularly important for the training of future mathematicians: if we systematically delegate to LLMs the formulation of conjectures, the search for proofs and their interpretation, how will students develop the independent abilities that, until very recently, formed the foundation of mathematical education? I am thinking above all of the ability to visualise structures, recognise regularities, devise original constructions and turn an intuition into a rigorous procedure. These are skills that require practice, unsuccessful attempts and often years of work.

To give a personal example, when tackling a multidimensional generalisation of the famous nine-dot problem (the nine dots to be connected with four consecutive segments), I devised a geometric algorithm, which I named the Clockwise-Algorithm, to build optimal solutions in arbitrary dimensions (Ripà, 2020).

The idea came from reasoning about the geometric constraints of the problem, partly through visual intuitions, until I developed a general mathematical construction. It was not a matter of mechanically applying a known formula, but of working out an original strategy through reasoning and successive refinements. If we no longer cultivate these abilities independently, we risk becoming less and less equipped to understand, evaluate and develop what AI produces, and also shutting ourselves off from results that AI cannot achieve and that we could have reached through healthy intellectual effort and further attempts.

“I therefore believe that, even in this phase of extraordinary accessibility, active research with the help of artificial intelligence must not turn into an uncritical delegation of understanding and of scientific responsibility, which remains with the human authors who sign the manuscripts”

A tool may well be better than us at finding a solution, but that does not mean we can accept it without understanding it... and, above all, learning to use AI effectively is neither an automatic nor a painless process.

English translation by Starry Mind Academy, reviewed and approved by the author. Read the original Italian version.

References

Show the 10 references

The interview starts here

In the first part: learning in his own way, from curiosity to method. How to become a researcher without a linear path.

Column · Metadialoghi

Conversations on learning, research and critical thinking

Metadialoghi (Metadialogues) gathers Starry Mind Academy's interviews with people engaged in research and in building knowledge, looking not only at what they do but at how they learn, how they think and how they got where they are.

Interview subject to editorial review. How we work

← Back to reflections

Reading options