If Mathematicians Oppose AI, What Will Physicists Do?—The Danger of “Skipping the Steps”

The other day, I saw a news story about a group of mathematicians who had jointly expressed concerns about the use of AI in mathematical research.

The issue was not simply that “AI makes mistakes.” With AI, researchers can skip the calculations, proofs, and logical reasoning that once took a long time and move closer to an answer. Their concern was that, as a result, young researchers might also lose the opportunity to develop the ability to think for themselves and discover the underlying logic.

Reading this made me wonder about something else.

So, what would physicists think?

I majored in physics at university and went on to graduate school, although I did not become a researcher. Looking back on those days, however, I suspect physicists would react to this issue a little differently.

Physicists Will Use Anything That Works

Physics also uses a great deal of mathematics.

But while proof itself matters in mathematics, in physics, mathematics is only a tool for understanding nature.

So if a computer can solve something, we use a computer. If it cannot be solved analytically, we approximate it. If a simulation can provide the answer, we run a simulation.

Adding AI to that toolkit would not be particularly surprising.

What is nature actually doing? If a tool can help us find out, we use it. Physics is a thoroughly pragmatic world.

Reality Always Wins in the End

In physics, no matter how impressive a theory may be, reality eventually punches it in the face.

A mathematical formula may be beautiful, and a theory may come from an eminent scholar, but if it does not agree with experiments or observations, something is wrong somewhere.

The story of Einstein and the cosmological constant is a symbolic example. When general relativity was applied to the universe as a whole, it did not fit well with the static universe assumed at the time. Einstein therefore introduced the cosmological constant into his equations, but observations indicating that the universe was expanding later emerged.

In the end, Einstein withdrew the cosmological constant.

In physics, reality is more powerful than theory.

Seen in that light, even if AI discovers a method of calculation or a law that no human had conceived, physicists are unlikely to reject it simply because “AI came up with it.”

If its predictions agree with experiments and can be reproduced, AI becomes a perfectly legitimate tool.

However, Getting an Answer and Understanding It Are Different Things

This is where the mathematicians’ concerns begin to matter.

At university, physics students are made to perform calculations endlessly. You differentiate something incorrectly. You apply the wrong boundary conditions. The numbers do not work out, and you start again from the beginning.

At the time, it felt like nothing but a nuisance.

Yet that process also gave me an intuition for things like:

“With an equation like this, the result should have roughly this form.”

“The dimensions of this answer are wrong.”

“It is strange for this to diverge in this limit.”

AI can produce an answer in seconds. But I do not know whether simply looking at the finished answer can develop the same intuition.

The Same Question Applies Directly to Programmers

This issue also connects to the work I do today.

I am now an IT engineer, and I use AI almost every day: for research, code generation, testing, reviews, and as a sounding board for design. I no longer want to return to a development environment without AI.

So if someone said, “For the sake of training junior developers, let’s go back to doing everything by hand,” I would not agree. There is no need to do all our work with longhand arithmetic when calculators exist.

The problem lies elsewhere.

We still cannot distinguish between the skills that AI has made unnecessary and the skills that have become even more necessary precisely because we use AI.

The ability to write code one character at a time may become less important than it once was. But the ability to notice that AI-generated code is wrong will not become unnecessary.

The same is probably true in physics.

The ability to spend hours calculating integrals may no longer be necessary. But it would be dangerous if there were no humans left who could look at an answer and say, “There is no way a physical phenomenon would behave like that.”

Was Everything AI Eliminated Really a Waste?

Ultimately, the question is not whether to use AI.

The struggle involved in reaching an answer used to contain a mixture of mere busywork and the parts that genuinely developed our ability to think. AI sweeps both away at once.

Students in the AI era do not need to spend the same 100 hours acquiring a skill that students in the past did.

But if those 100 hours contained 10 hours that would later help them detect an AI mistake, we must not eliminate those 10 hours as well.

AI can easily remove the processes that humans have found tedious.

The problem is that we may only realize, after removing them, that they were never a waste.