ChatGPT acts as a “cognitive crutch” that weakens memory, new research suggests

A recent experiment provides evidence that relying on artificial intelligence to help study new material tends to reduce how much information students remember weeks later. The findings suggest that while these tools can speed up initial learning, they might actually weaken the deep mental processing required to store knowledge over the long term. The study was published in the journal Social Sciences & Humanities Open.

Generative artificial intelligence refers to computer programs capable of creating text, images, or other media in response to user prompts. These systems can answer complex questions, synthesize vast amounts of information, and write essays in seconds. Because of this convenience, millions of university students use these programs to assist with their coursework.

André Barcaui, a professor of the Business Management course at the Federal University of Rio de Janeiro, conducted the study to see how this technology affects memory. He wanted to test whether the ease of getting answers from a chatbot harms long-term learning.

When a student delegates mental tasks to an external tool, psychologists call it cognitive offloading. In the past, this meant using a calculator for math or a notebook to remember a grocery list. Generative AI platforms take this a step further by performing the actual thinking, analyzing, and problem-solving.

“What motivated me most were my own undergraduate students,” Barcaui told PsyPost. “I feel there has been a growing loss of cognition and interest among students over the years. As I work with and teach AI, I understand the full potential of the tool and consider myself an enthusiast for the beneficial and responsible use of AI.”

“However, I began to see risks for teenagers, as their interest in reading decreases with each passing year, and with that, their creativity, critical thinking, text interpretation, and even writing and articulation skills also diminish. I have spoken with fellow professors in other countries and I don’t believe this is a phenomenon unique to Brazil. That’s why I decided to research it and see if it corroborated other equivalent research pointing in the same direction.”

Barcaui grounded his research in the concept of desirable difficulties. This psychological principle suggests that learning actually improves when a student faces a certain amount of productive struggle. Activities like actively trying to remember a fact or puzzling through a tough concept force the brain to build stronger memory pathways.

By providing immediate and polished answers, an automated chatbot might remove these necessary hurdles. Without that mental friction, students might experience weaker memory consolidation. Memory consolidation is the biological process where fragile new memories are stabilized and stored securely in the brain over time.

To investigate this dynamic, Barcaui recruited 120 undergraduate business administration students. The participants included 68 males and 52 females between the ages of 18 and 24. None of the students had formal training in computer science or machine learning.

The students were randomly divided into two equal groups of 60. Both groups received the same assignment, which required them to research basic artificial intelligence concepts. Topics included ethics, societal impacts, and technical foundations.

The task required the students to spend up to two weeks researching their assigned topic and prepare a ten-minute presentation. The first group was instructed to use a specific AI chatbot, known as ChatGPT, as a study aid. They interacted freely with the program to explain concepts, generate examples, and structure their presentations.

The second group used traditional study methods. These students were barred from using automated chatbots. Instead, they relied on course notes, academic databases, and standard internet search engines.

After the initial research period, the participants delivered their presentations to small groups of their peers. The students then went about their normal academic lives for a month and a half. This delay allowed the researcher to test natural memory decay over a realistic academic timeframe.

Forty-five days after the initial study phase, the researcher surprised the participants with a retention test. Eighty-five students returned to take this twenty-question multiple-choice exam. The test was specifically designed to measure how well the students understood the concepts, rather than just their ability to memorize definitions.

The students who used traditional study methods performed better. On average, the traditional learners answered 68.5 percent of the questions correctly. In contrast, the students who studied with the chatbot answered only 57.5 percent of the questions correctly.

The negative impact of the chatbot was most pronounced when students were learning highly technical topics. While the software also impaired memory for less technical topics, such as ethics and society, the gap between the two groups was not as wide. This pattern suggests that productive struggle is especially important when mastering complex or structurally difficult material.

Barcaui also tracked how much time the students spent preparing for their presentations. The automated chatbot allowed students to finish their work much faster. The technology-assisted group spent an average of 3.2 hours on the assignment, while the traditional group spent 5.8 hours.

Even when adjusting the data to account for this difference in study time, the traditional learners still scored higher on the final test. This detail indicates that the learning deficit was not just a result of spending less time on the material. Instead, the quality of the study time appeared to be lower.

Traditional learners likely spent their extra time re-reading materials, testing themselves, and manually connecting different ideas. These traditional study habits naturally force the brain to practice retrieving information. The chatbot users likely spent their time writing prompts and reading generated responses, which feels productive but fails to exercise the brain’s retrieval networks.

The researcher also found that a student’s prior experience with chatbots did not change the outcome. Frequent users of the technology performed just as poorly on the retention test as those who were relatively new to it. The evidence suggests that students might experience an illusion of competence, where the software makes them feel like they understand the material better than they actually do.

Barcaui highlighted three main takeaways:

“Productivity does not replace Competence: There is an abysmal difference between delivering a piece of work and understanding the process of its creation. The indiscriminate use of AI can create an ‘illusion of competence,’ where the individual obtains results without developing the synapses necessary to replicate that reasoning independently.”

“The Atrophy of the Critical ‘Muscle’: Just as the constant use of calculators reduced mental calculation skills, delegating writing and text interpretation to AI can atrophy the capacity for synthesis and critical thinking. Without the mental ‘friction’ of reading and writing, we lose the ability to articulate complex ideas and question information.”

“AI as Co-pilot, not Autopilot: The main lesson is that AI should be used to expand human capabilities (increase reach), not to replace them (eliminate effort). Human value will increasingly shift from the ability to execute to the ability to ask the right questions and critically curate the generated data.”

As with all research, the study has some limitations. Nearly thirty percent of the original 120 participants did not return for the surprise test forty-five days later. While the dropout rate was similar across both groups, this loss of data could slightly skew the final numbers.

The sample was also limited to business students at one university in Brazil. Students in different academic disciplines, such as engineering or the humanities, might interact with the software differently.

Future research should explore how different guidelines for technology use might change student outcomes. Scientists could test scenarios where students attempt to solve a problem on their own before consulting a chatbot for feedback. Researchers could also use digital tracking software to get a precise measurement of exactly how students divide their time when studying.

The study, “ChatGPT as a cognitive crutch: Evidence from a randomized controlled trial on knowledge retention,” was authored by André Barcaui.