AI as the Hardworking and Perfect PhD Student

 

Some people seem very defensive, and sometimes openly hostile, about the use of AI in scientific work. Others go even further and become judgmental, almost as if they were sitting in judgment, practicing criticism far more readily than approval, as though using AI were a form of cheating, intellectual dishonesty, or something one ought to feel ashamed o

Recently, someone on YouTube asked me whether I had written my response using AI. Well, yes. Nowadays I often write whatever I want very quickly and then pass it to the LLM I prefer, without exhausting myself over language, grammar, spelling, or style. It may be an email, an answer to a question, a short note, in English or in Arabic, and then I let the machine rewrite it properly.

Why should I spend my time on linguistic and stylistic tedium if a machine can handle that efficiently?

Is that impersonal? Perhaps sometimes. But even if it is, why should that be a problem? I am not a literary writer. What matters to me is the substance: the idea, the argument, the physics, the mathematics. If an LLM can take care of the linguistic packaging, I am perfectly happy to let it do so.

And personally, I have nothing to prove on this point. I had already written and published several books before the advent of modern LLMs. I was writing in physics and philosophy, in both English and Arabic, at least a decade before these tools existed. Whether my work is popular or not is beside the point. I believe that what I write has substance and depth, and I know perfectly well that the ideas are mine.

So I am not going to be shamed, or pushed into feeling guilty, for using a machine to help me express those ideas more efficiently.

I see AI in the same broad category as other powerful scientific tools. Researchers already rely on Mathematica for symbolic calculations, MATLAB and numerical libraries for computation, Lean for formal proof verification, Monte Carlo methods, parallel computing, and many other tools that enormously extend what one person can do. If others use these tools effectively and you refuse to use them, you may simply become unable to compete scientifically. I say this from first-hand experience.

AI and large language models are, in my view, another tool of this kind — perhaps better described as a super-tool because of their breadth — but still a tool. As with any scientific tool, the important question is not whether it was used, but how it was used. Its role should be disclosed transparently, its output should be checked and calibrated, and the researcher must be able to judge the result independently.

But there is an equally important point on the other side. Using these tools without being able to control, direct, and evaluate them can become dangerous scientifically. The hierarchy must remain clear: you should be the supervisor, and the AI should be the hardworking PhD student, not the other way around.

You define the real question. You choose the direction. You decide which calculations are meaningful and which are nonsense. You detect the mistakes, judge whether a result is scientifically sound, and take responsibility for the final conclusion. The machine can calculate, suggest, derive, search, rewrite, and explore possibilities at extraordinary speed, but it should not become the authority that you simply follow.

If you start relying on the model in areas where you are unable to evaluate, correct, or defend what it produces, then the relationship has been inverted: the PhD student has become the supervisor. In serious scientific work, that will eventually expose you.

That is where the real distinction lies. A scientifically mature researcher can exploit AI as part of a broader methodology, together with symbolic calculation, numerics, formal verification, simulation, and their own expertise. Someone who does not yet possess sufficient command of the subject may instead become dependent on the model precisely because they cannot properly evaluate what it produces. Those are very different situations.

Personally, I use AI extensively, and I disclose that use very openly. I do not see anything embarrassing about it. The relevant question is whether the scientific ideas, judgment, verification, and responsibility remain with the researcher.

In fact, since I started using AI seriously, I have felt considerably more liberated as a researcher. It has given me something close to the collaborator I had often been missing. By incorporating these tools into my work, I sometimes feel that I have effectively built a one-person research group around myself: I can explore calculations, test ideas, check derivations, examine alternative approaches, and move between analytical, numerical, and editorial tasks with a speed that would otherwise require several collaborators.

And this may have consequences beyond the individual researcher. AI could potentially reduce, at least to some extent, the enormous structural gap between researchers working in developing-world universities and those working in the best-funded institutes of the developed world. It obviously cannot replace laboratories, funding, strong departments, or a genuine scientific community. But it can provide access to a kind of intellectual and computational support that was previously available mainly to researchers surrounded by large groups, strong collaborators, and substantial institutional resources.

For that reason, I see AI not simply as another convenience, but as a potentially important equalizing tool in scientific research — provided, of course, that it is used by people who already possess the scientific maturity to judge, verify, and take responsibility for what it produces.

Scientific maturity is not demonstrated by refusing powerful tools. It is demonstrated by knowing how to use them without surrendering your judgment to them.