High school students who solved math exercises with GPT-4 by their side did 48% better on those exercises. When the tool was taken away for the test, they did 17% worse than classmates who had never used it.
When AI does the exercise for the student
That result comes from a study published in PNAS (Bastani et al.), which tracked around 1,000 high school students using GPT-4 on math exercises: those are the +48% and -17% from the opening.
The same study tested a second version of the tool, a tutor designed to guide the student’s reasoning instead of handing over the finished answer. That tutor avoided most of the performance loss once the AI was taken away. The technology was the same in both cases; the outcome changed with what the AI was allowed to do in place of the person learning.
When AI explains the concept
A study by Anthropic itself, published in January 2026 (How AI Assistance Affects Coding Skills), had engineers, mostly juniors, learn a new library (Trio) with and without AI support. On the follow-up quiz, those who coded by hand scored 67% against 50% for those who used AI, with no significant time savings between the two groups. The study comes from a company that sells an AI assistant, which, if anything, makes it less likely to overstate the harm.
Among those who used AI, the outcome depended on how they used it: those who asked the AI to explain the concept kept most of the learning, those who simply handed the code over to it ended up among the lowest scores in the group. The authors themselves point to cognitive effort, including the effort of being painfully stuck, as something that likely matters for mastery.
Why effort is what makes it stick
A preprint from the MIT Media Lab, still without peer review (Your Brain on ChatGPT), used EEG to measure the brain activity of people who wrote essays with ChatGPT versus without help. The first group had more trouble quoting their own text from memory minutes later, with lower neural connectivity recorded during the task.
The groundwork for understanding why this happens is older: the concept of deliberate practice, from Ericsson and colleagues (1993), came from a study of violinists, in which the most advanced had accumulated more than 10,000 hours of practice by age 20. That is the finding of that specific study, and it does not hold as a general rule of learning. A replication by Macnamara and Maitra (2019) showed that practice does not explain everything: the effect is smaller than the original study suggested, though the effect survived the replication. All three point to the same place: what makes learning stick is effort, and mere exposure to content cannot replace it. That effort is exactly what AI removes when it hands over the finished result.
Skipping the learning has already been sold as a promise
Coding bootcamps grew about 8-fold between 2013 and 2018, according to Course Report, on the promise of trading years of a degree for months of an intensive course.
Part of that promise fell apart. The US Consumer Financial Protection Bureau (CFPB) sued Lambda School/BloomTech for advertising placement rates of up to 86%, when the real number was closer to 50%, and near 30% in some programs. Bootcamps shut down, and even 2U left the segment. The 2024 tech layoffs, more than 150,000 positions cut, mark the moment that promise collided head-on with the market it claimed to fill.
Nothing guarantees that history will repeat itself exactly with AI, but “you don’t need to learn this, the tool does it for you” has been a sales pitch before, and its results have already been measured.
What awaits new graduates now
Hiring of new graduates fell 65% at big tech companies and 76% at seed-stage startups compared to 2019, according to the SignalFire State of Tech Talent 2026 report; hiring of engineers overall, at big tech companies, fell only 11% in the same period, meaning the drop is concentrated among those entering the field.
The Indeed Hiring Lab recorded, in the first quarter of 2026, entry-level positions in just 4.5% of developer job postings, against 69% for senior roles. And the Stanford Digital Economy Lab found employment for 22-to-25-year-olds about 19% below expectations in the occupations most exposed to AI, a descriptive result, with no proof of causation, according to the authors themselves.
All of these numbers are from the United States, and nothing here licenses extrapolating them to the Brazilian market. It is also worth noting the counterpoint the New York Fed brings, to avoid overstating the picture: computer science graduates have high unemployment (7.0%), but relatively low underemployment (19% against a 39% overall average), which suggests a more selective market, but not necessarily a closed one.
Knowing how to judge the result is the part that cannot be outsourced
METR measured experienced developers working with AI: they were 19% slower, but believed they were 20% faster. According to Veracode’s 2025 report, 45% of AI-generated code samples failed security checks.
Google Cloud’s 2025 DORA report sums up the pattern in one line: AI amplifies what is already there. Someone who has learned enough to recognize a bad result can correct course; someone who skipped that step has no way of knowing what they are accepting. This holds for code, but not only for it: a structural calculation, a technical report, or an engineering design delivered by AI carry the same risk, and demand the same judgment from whoever signs off on them.
What to do with this, whether you are studying or teaching
The set of sources used in this piece started from an article by Fabio Akita on AI Lair (in Portuguese), written for people studying programming. Here the scope goes beyond a tech career and covers learning in general. One of the ideas he brings is his own estimate, about 1,400 hours of practice before a first junior job. The number is his personal yardstick, with no research behind it, but it helps give scale to what “practicing more” means.
For students, what these studies suggest is this: do by hand what you still want to learn, ask the AI to explain instead of solve, save the ready-made answer for tasks that fall outside what you are trying to learn, review everything it delivers before signing off on it, and count the real hours of practice.
For teachers (I have taught engineering since 2018 at CEFET-MG, a federal institution of technological education in Minas Gerais, Brazil), the most useful example is the tutor in the PNAS study: the AI stayed available, only the way it responded to the student changed, and that was enough to avoid most of the learning loss. The work of teaching now is to design that use, more than to police it.
Before asking for the next ready-made answer, it is worth asking: is that the exercise you want to learn to do on your own, or is it a task that just needs to get out of the way?