Image by Hotpot.ai
By Jill Maschio, PhD
One way to understand the benefits and consequences AI has for learning is by examining research on its application in essay writing. A 2025 research paper by Fan and colleagues, they compared the resources students used to write an essay. The 117 university students (participants who were English as a second language) were randomly assigned to one of four groups: used a large language model (ChaptGPT 4.0), a group with no support for completing the task, a group supported by a human expert, and a group that had the support of the writing analytics toolkit named Checklist Tools. All participants were trained on how to use learning tools and then began a 2-hour reading and writing task. A second training was provided where the participants watched a video about revising writing and were given a 1-hour revising task to improve their essays. Afterwards, they were asked to complete a post-test within one day.
The researchers measured the participants’ intrinsic motivation, data of their behaviors during the study (actions while on the Internet), metacognitive activities (orientation, planning, monitoring, and evaluation), and their performance (essay score, knowledge gained, and knowledge transfer). The researchers reported the following:
- There were no differences in motivation levels between the groups.
- There were significant differences in the frequency and sequences of the self-regulated learning (metacognition).
- While those who used ChatGPT outperformed in the essay score, the knowledge gained and transferred was no different.
In another study by Kosmyna et al. (2025), the researchers experimented on how different support systems—LLM (ChatGPT), Search Engines, or no tools—impact cognitive load during essay writing. The researchers compared how active people’s brains were when they wrote essays under three different conditions: writing completely on their own (Brain-only), using a search engine (like Google), and using an AI tool (ChatGPT). Using EEG, the authors track brain connectivity across Alpha, Beta, Theta, and Delta bands, they found a clear pattern in brain activity: people who wrote on their own showed the most brain activation, those who used a search engine showed moderate activation, and those who used ChatGPT showed the least.
This means that the more help people received from external tools, the less their brains had to work during the task. In other words, relying on technology like ChatGPT can reduce the mental effort we invest in thinking, planning, and writing—because the tool is doing a lot of that work for us. This lower mental effort was measured using EEG, which showed reduced connectivity in brain regions involved in attention, memory, and cognitive control.
The Cognitive Load Trade‑Off
The tools that people use may matter, according to this study. The researchers found a clear pattern in brain activity: people who wrote without any tools (Brain-only) showed the highest level of brain engagement, those who used a search engine had moderate brain engagement, and those who used a large language model (LLM) like ChatGPT showed the lowest brain engagement.
The Four‑Session Twist: When Tools Leave a Mark
The researchers didn’t stop there. In a fourth session, LLM users were forced to write without tools—and Brain‑only users got access to ChatGPT. The result? Those switching off LLMs after using them for three sessions showed lower Alpha/Beta connectivity than even the Brain‑only group.
This finding points to a phenomenon known as “skill atrophy” or cognitive debt. Just like muscles weaken when we stop using them, mental skills can fade when we stop exercising them. In the study, participants who had relied on ChatGPT for several sessions showed significantly reduced brain activation when they were later asked to write without any tools. This implies that the brain may become less efficient or less prepared to perform tasks it has been outsourcing.
In other words, if we regularly let AI do the hard thinking for us—like generating ideas, organizing thoughts, or constructing sentences—our brains may become less practiced and less ready to do those tasks on their own. When we then try to write or think independently again, it’s harder to “turn those mental gears back on,” because we haven’t been using them actively. This is what the authors refer to as cognitive debt—a kind of mental shortfall we accumulate when we substitute deep thinking with automated assistance too often.
This has important implications for learning: it suggests that repeated reliance on AI tools could gradually weaken the cognitive skills students and writers need most, such as critical thinking, memory, and sustained attention.
Language, Memory & Ownership
Beyond the EEGs, the LLM group’s essays were more homogeneous and formulaic. Participants struggled to quote their own writing—a sign of shallow memory encoding. They also reported feeling less ownership over their work than the Brain‑only group. In contrast, Brain‑only authors not only recalled their content better but also described a stronger sense of authorship.
Cognitive Debt: A New Form of Load
“Cognitive debt,” as the authors call it, accumulates when we lean heavily on LLMs. The neural data, linguistic patterns, and behavioral findings all converge to suggest this isn’t benign. It’s not just about short‑term convenience—it’s a potential long‑term hitch in how we think, remember, and write).
Why It Matters
Learning vs. Convenience: Using LLMs is undeniably efficient. But at what cost to deep learning and mental resilience?
- Educational Impact: If students habitually use AI for writing, could their critical thinking and retention suffer?
- Soft Skills: Skills like synthesis, memory, and creativity may weaken when the tool takes center stage.
- Tools: Utilizing AI enables fast task completion, but within the field of education, a more structured AI may be required for independent student learning. This is why:
If you don’t know what you don’t know, how can a user be most effective at using AI for learning? The user must know what questions to ask to get the most out of it, prompting. An expert in a specialized field would be able to prompt it for specific information and guide it quickly, for example, including when images are incorrect. The novice or beginner learner needs guidance and direction for structured learning to co-learn with AI. Oftentimes, the information AI generates pertaining to social sciences is brief and shallow, for example. Structured learning, such as having learning outcomes, can help. However, how would the student know what questions to ask to guide such structuring? AI models for education could be designed with this in mind – that the AI provides more structure and learning outcomes relevant to the topic.
References
Holstein, K., McLaren, B. M., & Aleven, V. (2020). The concept of shared adaptivity in human–AI hybrid learning systems. International Journal of Artificial Intelligence in Education, 30(2), 200–231. https://doi.org/10.1007/s40593-020-00191-1
Kosmyna, N., Hoogeveen, D., Chang, T., Patel, K., Zhao, R., Caliskan, A., & D’Mello, S. K. (2025). Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing task (arXiv:2506.08872). arXiv. https://arxiv.org/abs/2506.08872
thanks for info.