Five years ago, using AI to study meant risking an automatic fail on your assignment. Today, a new study from the Federal University of Rio de Janeiro suggests that ChatGPT can actually help you learn faster. The catch? You might remember less than you think.
ChatGPT Study Finds Faster Learning With a Hidden Trade-off
Researchers led by AI expert André Barcaui at the Federal University of Rio de Janeiro ran a controlled experiment with 120 university students learning about artificial intelligence. Half were allowed to use ChatGPT as a study assistant. The other half stuck to traditional methods. The ChatGPT group spent an average of 3.2 hours on the assignment, compared with 5.8 hours for the non-AI group. They clearly moved through the material faster.
But when Barcaui sprang a surprise test on the participants 45 days later, the results shifted. Students who used ChatGPT scored an average of 5.75 out of 10. Those who took the traditional route scored 6.85 out of 10, roughly an 11 percent margin that could equate to a full grade level on a standard exam. The AI group learned faster in the moment. The traditional group remembered more over time.
This is not the first study to find a gap between short-term performance and long-term understanding when AI is involved. Researchers at the University of Illinois enrolled ChatGPT in a full undergraduate engineering course, where it earned a B overall. On structured math problems it scored an A, but on open-ended tasks requiring deeper reasoning it barely scraped by with a D. The process of generating answers, it turns out, can bypass the cognitive struggle that builds real comprehension. The Rio de Janeiro study puts a finer point on it: speed and depth are not the same thing.
Why Cognitive Struggle Actually Builds Memory
To understand why this matters, you need to understand how memory works. When you wrestle with a concept, your brain builds stronger neural pathways. That uncomfortable feeling of not knowing the answer, then working through it, is not a bug in your learning process. It is the feature. Psychologists call this desirable difficulty. The harder you work to retrieve or construct knowledge, the more durable it becomes.
ChatGPT removes much of that difficulty. You ask a question, you get a clean, structured answer. You ask for clarification, you get a simpler version. The friction disappears. And with it, apparently, does some of the memory encoding. As Barcaui writes in his published paper, unrestricted ChatGPT use 'impaired long-term retention, likely by reducing the cognitive effort that supports durable memory.'
Think of it like using a GPS every time you drive somewhere. You arrive faster than if you read a map. But if someone asked you to draw the route from memory a week later, you would struggle. Your brain never had to build the spatial map because the GPS did that work for you. ChatGPT functions as a cognitive GPS. It gets you to the answer, but it may not help you build the mental map you need for later.
The AI Cognitive Crutch Problem in Education
What the Rio de Janeiro study really reveals is the shape of a new problem in education technology. We have spent years asking whether AI tools are accurate enough for learning. That question misses the point. The more pressing issue is whether the process of using AI undermines the very thing it promises to deliver.
Barcaui's team effectively demonstrated that AI can become a cognitive crutch. A crutch helps you walk when you are injured. But if you use one long enough, your leg muscles atrophy. The same logic applies to thinking. When you offload reasoning, synthesis, and explanation to an AI, you are outsourcing the mental muscles that make learning stick.
A separate meta-analysis of 51 studies, published in Humanities and Social Sciences Communications, adds nuance here. ChatGPT showed a large positive effect on learning performance overall. But the benefits depended heavily on how the tool was used. Students gained the most in structured, problem-based learning environments, and over-reliance over longer periods sometimes showed signs of diminishing returns. The context matters enormously.
What This Means for How You Should Use AI to Study
So what does a responsible study session with ChatGPT actually look like? Based on what this research tells us, the key is timing. If you ask ChatGPT for the answer before you have tried to generate it yourself, you short-circuit the learning process. The better approach is to attempt the problem first, get stuck, and then use the AI to unstick you, not to replace your effort entirely.
Some educators have started recommending what they call the blank page test. Before you open ChatGPT, write down everything you know about the topic. Even if it is wrong. Even if it is incomplete. That act of retrieval primes your brain. Then use the AI to fill gaps, correct mistakes, and deepen understanding. The order matters. Struggle first, assist second.
There is also a difference between using ChatGPT to explain a concept you have already studied versus using it to generate an essay on a topic you have not engaged with at all. The former is closer to a tutor. The latter is closer to a ghostwriter. One builds understanding. The other builds a false sense of competence.
Where the Research Needs to Go Next
This study opens more questions than it answers, and that is a good thing. We need longitudinal research tracking students over months, not just 45 days, to see how AI-assisted learning affects knowledge retention in real academic settings. We need studies that compare different prompting strategies to figure out which ways of interacting with AI preserve learning benefits while reducing the crutch effect.
We also need age-specific data. The meta-analysis of 51 studies found that most research so far has focused on college students, while primary and early childhood education are significantly underrepresented. A college student using ChatGPT for economics might experience the trade-off differently than a younger student learning foundational skills. The cognitive mechanisms are likely different, and the stakes of weak retention compound over years of schooling.
The researchers at the Federal University of Rio de Janeiro have given us a useful starting point. They have shown that speed is not the same as depth, and that convenience can be the enemy of retention. But the practical playbook for using AI well in education is still being written.
The real takeaway here is not that ChatGPT is bad for learning. It is that unthinking use of any tool, no matter how powerful, produces shallow results. The question is not whether to use AI when you study. It is whether you are using it to think more or to think less. How do you make sure you are actually learning when an AI can do the heavy lifting for you?
Comments