What AI really changes in teaching data analysis — and what should not change

Every student now carries a nearly free tutor, programmer and adviser in their pocket — one that almost never blushes when it says something ridiculous. That is a rather unusual teaching situation.

The first reaction is understandable: ban it. The second is to add twenty AI tools to the syllabus and feel like a citizen of the future. I am not especially fond of either.

I do not know the final recipe for “education after AI.” I suspect nobody does; some people have simply already made a very polished slide deck about it. But several changes in my own teaching are now hard to miss.

1. AI became an X-ray for our assignments

I think we have become a little too obsessed with the question, “How do I catch the student who used ChatGPT?” Sometimes a more useful question is: what exactly did my assignment measure before ChatGPT existed?

“Fit a logistic regression to this table.” Not long ago this looked substantial: open the data, select features, write the code, obtain coefficients. Today AI can do a large part of that faster than the student. And, somewhat annoyingly, faster than the lecturer.

That does not make logistic regression obsolete. It simply means that an excavator has arrived at a shovel-digging contest. Continuing to grade only the speed of the shovel work would be strange.

I am much more interested in questions such as:

• what is actually sensible to predict here;
• did the target leak into the features;
• why might 99% accuracy be almost useless;
• which error is genuinely more expensive;
• will this model still work next year, in another hospital, or in another country?

AI did not destroy a good assignment. It exposed a weak one very quickly. For that, I am almost grateful.

2. Answers became cheap. That is good news.

For years we lived in a world where answers were scarce. You had to wait for the lecturer, find a book, open Stack Overflow, or persuade a programmer friend to look at your traceback.

Now there are so many answers we could probably use them for insulation.

A student can ask: “explain it more simply,” “show me with five numbers,” “compare these two models,” “find my mistake,” “quiz me.” For teaching data analysis this is a gift: change the threshold, damage the train/test split, add an outlier, remove a feature — and immediately inspect what happened.

A lecturer does not have to compete with a machine in the speed of producing explanations. I increasingly see my job differently: create a situation in which a student has to make a meaningful choice, observe the consequences, and explain them.

3. Misunderstanding learned how to dress well

This is where things become less pleasant.

In the past, misunderstanding often looked honest: a blank page, a traceback, or a graph resembling the ECG of someone who has just seen the exam results.

Now misunderstanding can arrive in a suit: headings, a table, Python code, five bullet points, and a confident “Therefore…”.

AI has become very good at wearing the costume of understanding. Occasionally, there is nobody inside the costume.

You can explain a p-value beautifully and still not understand what would have been surprising under the null hypothesis. You can run a random forest and fail to notice the target among the predictors. You can achieve 99% accuracy on a rare event and feel like the champion of machine learning.

That is why I increasingly doubt that the key AI skill is “prompting.” In data analysis, something else matters more: designing a check that can destroy your own beautiful answer.

4. What I am changing in my own classes

I do not want a lesson to become an exhibition of neural-network achievements. If an idea is best explained with a marker and a board, the marker remains an outstanding technology: no account required, no subscription increase, and very few hallucinations.

In my course “Data Analysis in the Age of AI,” I increasingly build learning around a small loop:

• look at the data and propose an explanation;
• test it;
• get a result that slightly ruins your mood;
• ask AI what you may have missed;
• then test the AI.

It creates a small educational food chain: the student checks the data, AI checks the student, the student checks AI, and the lecturer walks around with a serious face pretending this was exactly the plan from the beginning.

I use Orange, my own interactive tools and web apps, visualizations, games, and AI assistants. But I am less interested in AI as a “button that does things.” I find AI much more useful as a skeptic: ask it not for the solution, but for an awkward question, a competing hypothesis, or an attack on your choice of metric.

5. What I am not ready to outsource to AI

Draft code? Of course. Syntax? Please. Another example? Excellent. Visualization ideas? Give me two.

But there are still things I want to see in the human:

The last one matters especially. In the age of AI, the ability to say “I don’t know” is becoming almost an elite skill. A machine often answers precisely when it would have been wiser to stare quietly out of the window. Humans may want to preserve that small competitive advantage.

6. And now, a place for a fight — an intellectual one

There is one claim I am especially interested in hearing objections to:

Perhaps the main educational crisis created by AI is not that students learned to cheat. Perhaps AI simply revealed how many assignments had been testing ritual rather than thinking for years.

I fully expect to disagree with parts of this article myself in two years. That is fine. I am less interested in guessing the pedagogical constitution of 2030 than in collecting real observations now.

So I am especially interested in hearing from people who disagree with me:

1. Which assignment have you already stopped giving because of AI — or deliberately kept?
2. What can a student safely delegate to AI without destroying the learning?
3. Which skill do you still want to see demonstrated by a human without AI?

And if your answer is, “Igor, you are burying old skills far too quickly,” even better. I would much rather have that argument with concrete examples than read another list of “37 AI tools that will change education forever by next Tuesday.”