AI can solve the problem. Can you still do the research?

AI tools can now do in an hour what used to take us a week: the maths, the code, the figures, the search for papers. So a fair question is on many researchers' minds. If the machine can solve the problem, what is left for us? A recent essay by Pierre Marion, a researcher at INRIA in France, gives a calm and useful answer. Here is the plain version, and what I think it means for you.

What AI is now good at

Marion's starting point is simple. Today's AI models are very good at solving problems that are clearly defined. Give them a well stated maths question or a coding task, and they often do better than most researchers, for the price of a monthly subscription.

That covers a large part of our daily work: calculations, writing code, making figures, finding related papers. These tasks used to eat most of our time. Now they can be done much faster.

What AI still does not do well

Solving a problem is only one part of research. Marion points to two other parts where experienced researchers are still ahead, and where he expects them to stay ahead for at least the next few years.

• Doing the science itself. Choosing which problem is worth working on. Turning a real world situation into a clear question. Reading results and deciding what they really mean.
• Sharing knowledge. Teaching, training students, explaining work to other people. This is human work by nature.

His explanation is worth remembering. These models learn best on tasks where a computer can check the answer automatically, like a maths proof or a piece of code that runs. They learn much less well on tasks that need judgment, like deciding whether a question is a good one. Nobody knows how long this gap will last. But for now, it is real.

Two easy mistakes

Mistake 1. Refusing AI completely

Some researchers have good reasons to be careful with these tools: cost, privacy, the environment, and more. Those concerns deserve respect. But ignoring a tool that can speed up science a lot is hard to defend when our job is to move knowledge forward.

Mistake 2. Using AI to publish more and more

This is the more tempting mistake. If a paper takes less time, why not write twice as many? Marion's answer is that this road leads nowhere good. Conferences and journals are already flooded with submissions, and reviewers cannot keep up. When papers become easy to produce, the number of papers you have stops meaning much. Hiring and promotion will start to look at the quality of your best work instead.

The real risk for PhD students

Here is the part that matters most if you are early in your career. Research skills grow through struggle. You try something, it fails, you understand why, you try again. That slow process is how you build intuition and a sense of which problems are worth your time.

If AI does all of that struggle for you, you get answers but you do not grow. You become a person who checks the machine's work, instead of a researcher who can lead a project. Marion's hope is that we can have both: faster work and growing skills. He admits this is not proven yet. It depends on how we choose to use the tools.

Six habits for using AI the right way

These come from Marion's essay, with a few notes from my own work as a researcher and reviewer.

1. Publish less, and publish what matters. Before you write a paper, ask: does this have a real chance of being useful to someone, or of becoming part of what my field knows? If not, it can wait.
2. Use AI to explore questions, not only to answer them. Ask it to help you look at different ways to frame your problem. Finding a good question is the hard part, and AI can help you search wider. If you need a method for this, see how to choose a research question.
3. Build with the AI, do not just check its work. Work on the solution together, step by step. If you only read finished answers, you stop thinking for yourself.
4. Practise without AI on a regular basis. Airline pilots fly by hand from time to time so their skills stay sharp, even though the autopilot is very good. Do the same. Solve some problems, write some code, and draft some sections on your own.
5. Check everything. In mathematics, tools such as Lean can check a proof automatically. In most fields you have to do it yourself: run the code again, check the numbers, and make sure every cited paper really exists and really says what you claim.
6. If you supervise, add AI help slowly. A PhD is training first. Let students use more AI as their skills grow, not all of it on day one.

Why this matters to me

My research is about AI in breast cancer screening, and one idea comes back all the time. A model that is wrong and unsure is not the biggest danger. A model that is wrong and very sure of itself is. A doctor will double check the first one. The second one slips through.

AI writing tools have the same problem. They sound confident even when they are wrong. So the most valuable skill for a researcher today is not speed. It is judgment: choosing the right question, checking the evidence, and saying clearly where your claim stops. Marion makes the same point in his own way. If machines make exploring cheap, then careful human judgment becomes worth more, not less.

That is good news. And above all, as Marion reminds us, keep enjoying the work.