“You no longer need to study, AI does it for you”: PISA 2025 says otherwise

47% of Mexican 15-year-old students already use artificial intelligence chatbots at least once a week to help them learn.

So far, that might seem like just another data point on AI adoption. The problem starts when we look at what they're using it for: 40% use it for preliminary research, 34% to summarize a text they were supposed to read, and 35% to write assignments. Only 8% say they practically never use AI for the school activities analyzed.

These are figures published by the OECD in the results of PISA 2025 for Mexico.

And then comes a phrase I hear more and more, in different variations:

"Why study that if artificial intelligence already does it?"

PISA has just provided quite a bit of information to discuss that idea.

Submitting an assignment and learning are two different outcomes

A generative AI can summarize ten pages in seconds.

It can write an essay, explain a concept to you, solve certain problems, and quickly find information that used to take much longer to track down.

That's impressive. I use it practically every day, and it would be absurd to deny its usefulness.

The problem appears when we confuse productivity with learning.

Let's imagine a student receives a ten-page article they need to read and analyze.

They can read it, try to understand it, identify the main ideas, and then use an AI to compare their interpretation.

They can also paste the entire document into a chatbot, ask for a summary, and submit the conclusions they received.

In both cases, there's a finished document.

The learning behind those documents can be completely different.

What's interesting about PISA isn't that AI is "bad"

Here it's worth being careful with headlines.

PISA 2025 does not prove that using artificial intelligence causes worse academic results. We're mainly talking about associations observed between behavior and performance, not a direct demonstration of cause and effect.

But the pattern found by the OECD deserves attention.

The general PISA 2025 report notes that students who don't use AI chatbots for school assignments generally perform better than those who do.

The story changes when we analyze how much and what it's used for.

Among students who turn to AI specifically to “help me learn,” those who use it moderately show slightly better results than those who use it very little or intensively.

For specific activities, such as summarizing texts or doing preliminary research, moderate users also outperform those who turn to these tools constantly.

That changes the conversation quite a bit.

The way I read it is quite simple: AI helps more when it takes part in the thinking process than when it replaces it.

Mexico has another problem behind it

The Mexican data make the discussion even more important.

In mathematics, only 30% of the Mexican students assessed reached at least level 2 proficiency, considered by PISA as the basic level. The OECD average was 65%.

In reading, approximately 50% reached the minimum level, compared with 69% in the OECD.

In science, something similar happened: around 50% in Mexico versus 74% in the organization's average.

And now we are placing on top of that foundation a tool capable of producing complete answers with very little effort from the user.

That can turn out very well.

It can also turn out quite badly.

If a student has trouble understanding a text and uses AI to have it explained in other words, create examples, and answer their questions, they have at their disposal a kind of tutor available practically all the time.

If that same student avoids reading the text because the chatbot can summarize it, we are probably automating precisely the skill they needed to practice.

The calculator already went through something similar

The discussion reminds me quite a bit of what happened with calculators.

A calculator can solve a division much faster than we can. That did not make learning math unnecessary.

What changed was the level at which we can work.

We no longer need to invest as much time in certain mechanical operations, and we can focus on more complex problems.

The condition is important: first we have to understand what operation we are doing.

Someone who doesn't know math can also enter numbers into a calculator.

The problem comes when they get the result.

Is it reasonable?

Did they use the right operation?

Did they enter the right data?

Does the result make sense within the problem?

Something similar happens with artificial intelligence, but with an added difficulty: an AI can make a mistake using extraordinarily convincing language.

To detect that mistake, you need knowledge.

Knowing how to ask doesn't replace knowing

In recent years, another fairly common idea has emerged: it will no longer be necessary to know as much because the important thing will be learning how to ask the AI.

I only agree up to a certain point.

Knowing how to write good prompts helps a lot.

But a good question usually comes from someone who understands the problem well enough.

I see it constantly working with AI tools in software development.

You can ask a model to generate a complete solution. In seconds you'll have something that looks correct.

The real difference shows up afterward.

You have to check whether the architecture makes sense, whether there's a security issue, whether a decision will affect performance, whether the information used is correct, or whether the model simply made up a plausible answer.

Without prior knowledge, many of those things go unnoticed.

The exact same thing happens in education.

I would teach how to use AI, not try to hide it

Banning these tools completely seems like a fairly limited solution to me.

Students will end up using them outside the classroom.

I'd rather teach them to work with them.

For example:

• Try solving the problem first and then compare your answer with the AI.
• Ask it to explain why your answer is wrong.
• Request different examples until you understand the concept.
• Ask it for practice questions without showing the answers right away.
• Cross-check important information with other sources.
• Explain in your own words what you just learned.

There's an even simpler test.

After closing the chatbot, try explaining the topic without looking at it.

If you can do it, you probably learned something.

If you can't, maybe you just finished the assignment.

The new illiteracy can look quite competent

There's something that worries me more than a student using ChatGPT to do homework.

It's a student capable of producing perfect documents about topics they truly don't understand.

Technology can hide that difference for quite a while.

An essay can have good structure.

A presentation can look professional.

A technically correct answer can sound fine.

Until the moment comes to defend it, apply that knowledge to another problem, or detect that the information the AI itself provided was wrong.

PISA tries to measure precisely a part of that: not just recalling information, but using knowledge to solve problems, think critically, and face real-world situations.

That's where learning is still needed.

My takeaways

PISA 2025 doesn't leave me wanting to ban artificial intelligence in schools.

It leaves me wanting to use it better.

The important data point isn't that 47% of Mexican students already use AI weekly. That figure will probably keep growing.

What will matter is what happens during those minutes in front of the chatbot.

It can be used to avoid twenty minutes of reading.

Or to spend twenty minutes asking better questions about what we just read.

The same technology. A completely different educational outcome.

Conclusion

Artificial intelligence is eliminating a lot of mechanical work, and it will keep doing so.

That's a huge advantage.

But delegating a task and delegating learning are different decisions.

A few years from now, it will probably make even less sense to memorize certain data that a machine can look up instantly. Understanding concepts, connecting information, identifying errors, arguing, and making decisions will still be necessary.

In fact, when facing systems capable of producing increasingly convincing answers, those skills may become more important.

So when someone says "there's no need to study anymore, AI does it now," my answer would be quite simple:

AI can do the homework for you. Learning it is still up to you.

And that difference will probably be one of the most important skills we'll have to teach in the coming years.