What AI will not be able to deliver for you

In 2025, a team from the MIT Media Lab did something that no one had done yet: they looked inside people's brains while they wrote with artificial intelligence.

The experiment was simple to describe. Fifty-four participants, divided into three groups, writing essays over the course of four months. The first group wrote with ChatGPT. The second, with a web search engine and nothing else. The third, raw: just their own minds. All wore a cap with thirty-two electrodes measuring their brain activity in real time, and participated in interviews after each session.

What emerged from that has become one of the most cited studies of the year, and it deserves to be read slowly, because it does not say what the headlines claimed it said.

The Debt

The first finding is almost geometric. Brain connectivity—the extent to which different regions of the brain work together during a task—decreased systematically the more external help there was. Those who wrote alone showed the strongest and most extensive networks. Those using the search engine showed intermediate engagement. Those using AI showed the weakest coupling of the three.

[GRAPHIC 1: g1-conectividad-mit.png]

So far, nothing shocking: if someone else does the heavy lifting, you exert less effort. The interesting part came next. In the fourth session, the researchers flipped the script: those who had been writing with AI for months had to write alone. And their brains did not return to the level of those who had always written alone. They remained under-activated, like a muscle that spent the season on the bench. The authors coined a term for this phenomenon: cognitive debt. Today's discount on effort is paid with interest tomorrow.

The 83%

But the figure that truly matters for anyone in teaching is different, and it carries a statistical cruelty that is hard to surpass.

Upon finishing their writing, the researchers asked each participant to quote a sentence from their own essay—the one they had just submitted a few minutes prior. Among those who had written alone, nearly everyone did so without difficulty. Among those who had written with AI, 83% were unable to quote a single sentence.

[GRAPHIC 2: g2-memoria-83.png]

Think about that for a moment. It is not that they forgot details; it is that the text bearing their name had never actually passed through them. And the participants themselves noticed it. When asked if the essay was theirs, those who wrote alone claimed authorship almost unanimously; those using AI hesitated, qualified their answers, and shared credit with the machine. The study calls this a fragmented sense of ownership. In plain terms: they signed their name to something that was not entirely theirs, and they knew it.

Here lies the key to the entire matter, and it is worth formulating precisely. The study does not prove that AI is bad. It proves that writing is the mechanism through which ideas become one's own. Memory is consolidated by choosing each word, by wrestling with every sentence. Whoever delegates writing does not bypass a mere administrative task; they bypass ownership itself.

This Is Not About a Hypothetical Future

It is worth stating immediately that we are not talking about a coming threat. We are describing today's classrooms.

In Spain, 89% of university students admit to using artificial intelligence in their academic work, and nine out of ten use it to study, according to surveys by the CYD Foundation. Around four in ten use it daily. According to Eurostat, we are the country in the European Union where AI is most used for educational purposes: twenty percentage points above the average.

[GRAPHIC 3: g3-uso-ia-universitarios.png]

And here, one must resist the temptation of moral panic. That 89% is not a generation of cheaters. It is a generation with an extraordinary tool in their pocket and, all too often, facing tasks that demand nothing the tool cannot provide. When a task can be done by a machine, the implicit message is that the task did not require a human. Students have understood that message better than anyone.

Why Policing Is Not the Answer

The first reaction of half the academic world was technological: detectors. Software that promises to distinguish human text from generated text. And here, the data is unforgiving.

A team from Stanford evaluated seven detectors against exam essays written by non-native English students: the average false positive rate was 61.3%. More than nine out of ten essays were flagged as suspicious by at least one detector. With native writers, there was almost no error. The reason is technical and perverse: detectors penalize simple vocabulary and straightforward syntax—meaning they punish precisely those writing in their second language.

Vanderbilt University deactivated the detector on its anti-plagiarism platform in 2023 with a calculation that anyone can perform: even accepting a 1% false positive rate, when applied to the 75,000 papers graded annually, it results in about 750 unjust accusations per year. Seven hundred and fifty innocent students per term.

The detection arms race is lost. And lost races are best abandoned early, before the damage falls upon the usual victims.

What Serious Universities Are Doing

If policing does not work, the remaining option is to change what is being asked of students. And that is exactly what is happening.

In the United States, sales of handwritten exam booklets—the old blue books—have risen 30% at Texas A&M, 50% at the University of Florida, and 80% at Berkeley over two academic years. A professor from Wisconsin explained why he returned to paper: in a class of four hundred, he received fifty final exams with clear evidence of ChatGPT usage; some even had the logo pasted in. But what is revealing is what he now looks for in handwritten exams: not the finest phrasing, but the ability to connect course concepts to one another and explain why they matter.

[GRAPHIC 4: g4-blue-books.png]

The handwritten exam is neither the only answer nor the best one for everything: timed writing sacrifices research and revision, which are also forms of thinking. But the direction of this movement is unmistakable, and it is correct: when faced with a machine that executes, serious assessment shifts toward what cannot be delegated. The first-person perspective. Oral defense. Judgment applied to one's own life. That which is built with others.

Writing Is Keeping Something

This is why a final humanities project is written without artificial intelligence. Not out of nostalgia, nor as a punishment, nor out of distrust toward a technology that is here to stay and that we must learn to use well.

It is written without AI because the final essay is not a product to be delivered: it is the mechanism through which the journey becomes yours. What is evaluated is not the elegance of the prose. It is that there is someone behind it: personal judgment, intellectual honesty, a person who can quote their own sentences because they chose them one by one.

Artificial intelligence will write increasingly better. That battle is decided, and that is fine. The remaining question is different, and it is the only one that ever mattered: when your education is complete, will there be anyone there? Someone with something of their own to say, and the words to say it?

No one can hand that to you. Not even the machine.

Cover image generated with artificial intelligence.