This study experimentally analyzed the effects of large language model (LLM) usage on human brain activity and cognitive load during essay writing, aiming to understand the implications of AI dependence on learning processes.

The research employed an experimental design involving participants divided into three groups: LLM (e.g., ChatGPT users), Search Engine users, and Brain-only (no tools used). Participants performed identical essay writing tasks across three initial sessions. In a fourth session, some participants crossed over conditions: LLM users transitioned to the Brain-only condition (LLM-to-Brain), and Brain-only users transitioned to the LLM condition (Brain-to-LLM). A total of 54 participants completed sessions 1-3, with 18 completing all four sessions.

The core methodology involved objective and subjective measurements. Electroencephalography (EEG) was utilized to measure cognitive load and brain connectivity during essay composition. This provided data on the strength and breadth of brain networks engaged. Concurrently, the written essays underwent comprehensive evaluation:

1. Natural Language Processing (NLP) Analysis: This included Named Entity Recognition (NER), n-gram pattern analysis, and topic ontology analysis to assess linguistic characteristics and thematic coherence within and across groups.
2. Assessment by Evaluators: Essays were scored by both human teachers and AI evaluators.

Behavioral and linguistic observations were also collected, including self-reported ownership of the written essays and the accuracy with which participants could cite their own work. A follow-up tracking was conducted over four months to assess long-term performance.

Key findings of the study were substantial:

• EEG Analysis: Significant differences in brain activity were observed across groups. The Brain-only group exhibited the strongest and broadest brain connectivity, indicative of higher cognitive engagement. The Search Engine group showed intermediate levels of engagement, while the LLM group demonstrated the weakest connectivity. A general trend indicated that increased reliance on external tools correlated with decreased cognitive activity.
• Crossover Effects (Session 4): Participants who switched from LLM to Brain-only (LLM-to-Brain) showed a reduction in alpha and beta band connectivity, suggesting a state of cognitive hypoactivity or disengagement. Conversely, participants transitioning from Brain-only to LLM (Brain-to-LLM) exhibited enhanced memory recall capabilities and increased activation in the occipital-parietal and prefrontal brain regions, similar to search engine users.
• Behavioral and Linguistic Observations: Self-reported ownership of essays was lowest in the LLM group and highest in the Brain-only group. LLM users also struggled significantly with accurately citing their own written content.
• Long-term Impact: A four-month follow-up revealed sustained low performance across neural, linguistic, and behavioral levels for participants who consistently used LLMs.

The study concludes that while LLMs offer immediate convenience, their long-term use incurs significant cognitive costs, potentially leading to a "cognitive debt." It highlights a negative impact of AI dependence on learning and critical thinking abilities, advocating for a necessary re-evaluation of AI integration in educational contexts and a redesign of learning mechanisms in the AI era.