PaperlayerView exact revision on arXiv

Source-linked AI summary

Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task

Nataliya Kosmyna, Eugene Hauptmann, Ye Tong Yuan, Jessica Situ, Xian-Hao Liao, Ashly Vivian Beresnitzky, Iris Braunstein, Pattie Maes

arXiv:2506.08872v2cs.AI

TL;DR

The cognitive implications of LLM use in educational writing remain a concern. This study compared essay writing with an LLM, a search engine, or no tools using EEG and essay analyses, finding stronger brain connectivity without assistance and reduced overall connectivity with AI assistance.

  • Problem

    Concerns about LLM use in education motivate comparing brain activation during essay writing with an LLM, a search engine, or no tools.

    Source: Introduction · pp. 10–11

  • Method

    The study compared essay writing with an LLM, a search engine, or no tools using EEG, NLP essay analysis, questionnaires, interviews, and a fourth session that switched some participants’ conditions.

    Source: Introduction · p. 11, Summary · p. 117

  • Results

    Unassisted writing increased brain network interactions across multiple frequency bands, while AI assistance reduced overall neural connectivity and shifted information flow.

    Source: Summary · p. 97

  • Takeaways & Limitations

    The findings support delaying AI integration until learners have engaged in sufficient self-driven cognitive effort.

    Source: Behavioral Correlates of Neural… · p. 151

    The study had few participants from a geographically concentrated area, prompting a call for larger, more diverse, and more gender-balanced samples.

    Source: Limitations and Future Work · p. 151

Abstract

from arXiv · show

This study explores the neural and behavioral consequences of LLM-assisted essay writing. Participants were divided into three groups: LLM, Search Engine, and Brain-only (no tools). Each completed three sessions under the same condition. In a fourth session, LLM users were reassigned to Brain-only group (LLM-to-Brain), and Brain-only users were reassigned to LLM condition (Brain-to-LLM). A total of 54 participants took part in Sessions 1-3, with 18 completing session 4. We used electroencephalography (EEG) to assess cognitive load during essay writing, and analyzed essays using NLP, as well as scoring essays with the help from human teachers and an AI judge. Across groups, NERs, n-gram patterns, and topic ontology showed within-group homogeneity. EEG revealed significant differences in brain connectivity: Brain-only participants exhibited the strongest, most distributed networks; Search Engine users showed moderate engagement; and LLM users displayed the weakest connectivity. Cognitive activity scaled down in relation to external tool use. In session 4, LLM-to-Brain participants showed reduced alpha and beta connectivity, indicating under-engagement. Brain-to-LLM users exhibited higher memory recall and activation of occipito-parietal and prefrontal areas, similar to Search Engine users. Self-reported ownership of essays was the lowest in the LLM group and the highest in the Brain-only group. LLM users also struggled to accurately quote their own work. While LLMs offer immediate convenience, our findings highlight potential cognitive costs. Over four months, LLM users consistently underperformed at neural, linguistic, and behavioral levels. These results raise concerns about the long-term educational implications of LLM reliance and underscore the need for deeper inquiry into AI's role in learning.

Summary of Results

Session 4 showed contrasting neural patterns: Brain-to-LLM participants had higher connectivity than the LLM group’s earlier sessions, while LLM-to-Brain participants showed less coordinated effort in most bands. The latter group’s essays scored highly but showed LLM-specific vocabulary bias and less distinctive NER and n-gram usage.

Session 4

  • Brain-to-LLM participants showed higher directed connectivity across alpha, beta, theta, and delta bands than the LLM group’s sessions 1–3.The authors suggest that rewriting with AI after prior AI-free writing engaged more extensive brain-network interactions.Source: Summary of Results · p. 3
  • LLM-to-Brain participants showed less coordinated neural effort in most bands and bias toward LLM-specific vocabulary.Their essays scored highly with both the AI judge and human teachers, but showed less distance in NER and n-gram usage compared with other groups’ sessions.Source: Summary of Results · p. 3
  • At the topic level, few topics deviated significantly and almost orthogonally between LLM and Brain-only groups, including HAPPINESS and PHILANTHROPY.Source: Summary of Results · p. 3

18 participants

This section situates the essay-writing study in prior research on learning and information seeking, then describes its participants, tool-use groups, EEG procedures, prompts, session reassignment, and post-task interviews. It reports patterns of essay homogeneity and ownership, as well as recall and connectivity among participants using different tools.

18 participants

  • Essay analyses described homogeneous ontology and n-grams, reduced ownership, and impaired ability to quote one’s own essay; some essays relied on copy-paste with minimal editing.Other reported patterns included frequent location and date NERs, translation use, person-name NER impact, and little deviation from the default ChatGPT response to the SAT prompt.Source: §18 · p. 3
  • Brain-to-LLM participants showed higher memory recall and directed connectivity across frequency bands than LLM-only sessions, while LLM-to-Brain participants showed lower interconnectivity and engagement.Brain-to-LLM participants also integrated content better than in their prior Brain sessions, used more information-seeking prompts, and scored mostly above average across groups; their connectivity remained below the Brain group, with split ownership.Source: §18 · p. 3

Introduction

  • The study compares essay writing with an LLM, an internet search engine, and no tools, using EEG, NLP, questionnaires, interviews, human teachers, and an AI judge.It treats essay writing as cognitively complex, requiring simultaneous organization of ideas and management of word choice, grammar, and syntax.Source: Introduction · pp. 10–11

LLMs and Learning

  • Prior work describes LLMs as offering personalized, adaptive responses, but warns that passive consumption may weaken deep engagement, critical thinking, and memory.LLMs can also produce inaccurate information or fabricated references, making fact-checking important.Source: LLMs and Learning · pp. 11–12

Web search and learning

Cognitive Load During Web Searches

  • During web searches, query formulation and relevance judgments place particular demands on memory, while search engines may shift users toward remembering where to find information rather than its content [36].Search interface complexity and individual working-memory capacity or expertise also affect cognitive load.Source: Cognitive Load During Web Searches · p. 14

Cognitive load during LLM use

  • Prior research reports that LLMs reduce cognitive load, but lower effort may come with shallower engagement and weaker reasoning than search-based learning.Users may shift from retrieving and curating information to verifying AI responses, while outcomes depend on competence and whether users actively synthesize material.Source: Cognitive load during LLM use · pp. 15–16

Engagement during web searches

  • Web-search engagement depends on sustained attention, task complexity, novelty, and feedback, whereas LLMs may boost immediate engagement without sustaining intrinsic motivation or deep reasoning.LLM personalization and responsive feedback can support interaction, but traditional methods may better foster critical analysis and integration of complex ideas.Source: Engagement during web searches · p. 16, Engagement during LLM use · p. 17

Physiological responses during web searches

  • Prior research linked active web searching to broader neural activation than passive reading and to reward-related information seeking.Search-related activity included decision-making, working memory, and executive function; the Google Effect describes remembering where to find information rather than the information itself.Source: Physiological responses during web… · pp. 17–18

Search engines vs LLMs

  • Search engines support broad exploration and source-specific fact-checking, whereas LLMs synthesize contextualized responses and streamline retrieval; essay writing was chosen to examine both approaches.The passage characterizes LLMs as useful for concise explanations, brainstorming, and interactive learning.Source: Search engines vs LLMs · pp. 18–19

Learning Task: Essay Writing

  • Prior work found ChatGPT improved essay scores without significant knowledge gain or transfer [55], while ChatGPT-4 often outperformed human essays linguistically and ChatGPT 3.5 reduced creative writing.The cited work raises concern that AI-generated drafts may bypass synthesis and retention processes involved in learning [55].Source: Learning Task: Essay Writing · pp. 19–20

Echo Chambers in Search and LLM

  • Prior work reports that conversational LLM search can reinforce users’ views and narrow exposure across turns, potentially weakening critical thinking in education.Echo chambers can exclude alternative perspectives, leaving users unaware of research gaps and encouraging overconfident conclusions.Source: Echo Chambers in Search and LLM · p. 21

Participants

  • The study reported data from 54 adults aged 18–39 recruited at five Boston-area universities; 18 participants optionally attended an extra fourth session.The sample included 32 women, 19 men, 2 non-binary participants, and 1 participant who preferred not to provide gender information.Source: Participants · pp. 22–23
  • Participants were randomly assigned to ChatGPT-only, Google/search-only, or brain-only essay writing, with the respective tools prohibited in the other groups.Completing three sessions defined study completion; session 4 was optional and depended on availability.Source: Participants · p. 23

Protocol; Stage 1: Welcome, Briefing and Background questionnaire; Stage 2: Setup of the Enobio headset; Stage 3: Calibration Test

Stage 4: Essay Writing Task

  • Each of three essay sessions offered a choice of three SAT-derived prompts, totaling nine unique prompts across the study.Participants selected one prompt for each session.Source: Stage 4: Essay Writing Task · p. 25

The session 3 prompts

  • The session 3 prompt asked whether a perfect society is possible or desirable, presenting arguments for and against striving to create one.Source: The session 3 prompts · p. 27

Assignment: Is a perfect society possible or even desirable?

Assignment: Can people have too much enthusiasm?

  • Participants had 20 minutes to write on a selected prompt, with LLM, Search Engine, and Brain-only groups restricted to their respective tool conditions.The LLM group could use only the provided ChatGPT account; Search Engine participants could use websites except LLMs; Brain-only participants could use no tools.Source: Assignment: Can people have too much… · p. 27

The session 4 prompts

  • Session 4 reassigned returning participants to another group and offered personalized prompts drawn from topics they had written about in sessions 1–3.Participants were not told in advance about the group or prompt changes.Source: The session 4 prompts · p. 28

Stage 5: Post-assessment interview

  • After writing, participants completed an audio-recorded, template-guided interview; the interview stage took about 5 minutes and the full study about 60 minutes.The interview included eight questions, with four additional questions in session 4.Source: Stage 5: Post-assessment interview · p. 28

Stage 6: Debriefing, Cleanup, Storing Data

  • At debriefing, participants could provide additional comments, were reminded about pending sessions, and received shampoo and a towel before their devices were returned.The experimenter saved EEG data, essays, ChatGPT and browser logs, and audio recordings; recorded EOG data was excluded from the manuscript.Source: Stage 6: Debriefing, Cleanup, Storing… · p. 29

Post-assessment interview analysis

7. LLM group:

This section examines LLM-group participants’ essay recall and quoting, ownership and reflections, and patterns in essay content, language, and assessment across sessions. It also describes EEG connectivity findings for LLM-assisted writing and the analysis of directed connectivity.

Adherence to Structure

Question 3. Ability to Quote

  • Figure 6 reports the number of participants in each group and their ability to quote anything from their essays during Session 1.Source: Figure 6 · p. 31

Question 4. Correct quoting

  • For Session 1, no LLM participants produced a correct quote, while 3/18 Search Engine and 2/18 Brain-only participants failed to do so.The LLM group performed significantly worse than both other groups (p < .001); Search Engine and Brain-only did not differ (p = 1.0).Source: Question 4. Correct quoting · p. 32, Figure 7 · p. 32

Perception of Ownership

  • LLM participants reported mixed essay ownership in Sessions 1–3, while Brain-only participants overwhelmingly reported full ownership.In Session 1, 9/18 LLM participants reported full ownership and 3/18 reported none; by Session 3, one LLM participant reported no ownership, while all 18 Brain-only participants reported full ownership. Search Engine participants reported no absence of ownership in these sessions.Source: Question 5. Essay ownership · pp. 32, 33, 37, 41, 42

Satisfaction

  • Search Engine participants reported full essay satisfaction in all three sessions, while LLM satisfaction was 17/18 and Brain-only satisfaction rose from 15/18 in Session 1 to 17/18 in Sessions 2–3.The remaining LLM participant in each session reported partial satisfaction; Brain-only participants who were not fully satisfied reported partial satisfaction, uncertainty, or dissatisfaction.Source: Question 6. Satisfaction with the essay · pp. 33, 38, 42

Reflections and Highlights

Quoting Ability and Correctness

Question 5. Quoting ability

  • In Session 4, 7/9 reassigned LLM participants failed to reproduce a quote, compared with 1/9 reassigned Brain-only participants.The passage reports lower quoting accuracy for the LLM group.Source: Question 5. Quoting ability · p. 46

Question 6. Correct quoting

  • Only 1/9 reassigned LLM participants quoted accurately, compared with 7/9 in the reassigned Brain-only group.The study reports that the reassigned LLM group performed significantly worse on correct quoting.Source: Question 6. Correct quoting · p. 47, Figure 11 · p. 48

Question 7. Ownership of the essay

Question 9. Preferred Essay

  • All participants preferred their Session 4 essay, while Brain-only showed stronger continuity with prior work and reassigned LLM participants focused on improving earlier essays but struggled with quoting.Reassigned LLM participants described adding and elaborating more of their own ideas in Session 4; one contrasted ChatGPT’s completion with the greater detail in the unassisted essay.Source: Question 9. Preferred Essay · p. 50, Summary for Session 4 · pp. 50–51

Latent space embeddings clusters

  • Essays clustered by original group with a distinct Session 4 subcluster; switched-to-Brain-only essays differed from earlier essays, and word-length variability was smaller for LLM and Search Engine than Brain-only.The embedding analysis reports significant distance between same-topic essays after switching from LLM or Search Engine to Brain-only.Source: Latent space embeddings clusters · pp. 52–53, Figure 12 · p. 52, Figure 13 · p. 53

Similarities and distances

  • Search Engine essays had the greatest cosine distance from an AI-generated essay, while LLM and Brain-only essays were similar; LLM essays showed greater within-group similarity and topic-specific KL divergence patterns.Brain-only diverged from other groups on CHOICES; Search Engine diverged most on LOYALTY and HAPPINESS, while Brain-only diverged from LLM on ART, CHOICES, COURAGE, FORESIGHT, and PHILANTHROPY.Source: Similarities and distances · pp. 56–58, Figure 18 · p. 56, Figure 19 · p. 57

Named Entities Recognition (NER)

  • NER-use association was strongest in LLM essays, moderate in Search Engine essays, and weak in Brain-only essays, with different entity-type distributions across groups.LLM essays most often used Person, Work of Art, Organization, Event, Title, GPE, and Nationalities; Search Engine essays used Person about half as often as LLM essays, while Brain-only NERs were more evenly distributed except for more frequent Instagram.Source: Named Entities Recognition (NER) · pp. 59–61, Figure 21 · p. 59

N-grams analysis

  • The n-gram “perfect societi” appeared across groups, while FORETHOUGHT and HAPPINESS showed different leading phrases by group.“Perfect societi” was most used by Search Engine, less by LLM, and least by Brain-only; in HAPPINESS, LLM led with “choos career,” Search Engine with “give us,” and Brain-only with “true happi.”Source: N-grams analysis · pp. 62–64, Figure 27 · p. 65

Ontology analysis

  • Prompt use varied by topic, with more back-and-forth on ART, PERFECT, HAPPINESS, and LOYALTY; the researchers also generated and manually checked ontology graphs for essays.LOYALTY had the most guidance prompts, while CHOICES and ENTHUSIASM showed less prompt variety and mostly information retrieval; the ontology agent was designed to produce parseable outputs.Source: Ontology analysis · pp. 67–68, Figure 33 · p. 68, Figure 34 · p. 68
  • The LLM group’s ontology edges dominated the distribution and overlapped with Search Engine essays, especially around justice and innovation, while Brain-only showed few intersections.The analysis matched differently phrased nodes using Levenshtein distance <= 10.Source: Figure 37 · p. 70, Ontology analysis · p. 70

AI judge vs Human teachers

  • The AI judge tended to score essays around 4 and identified almost half as LLM-assisted, while teachers generally scored lower and disagreed on content and organization.For content, teachers averaged between 3 and 4 and sometimes scored essays 1 or 2 when the AI judge gave 4; teachers consistently rated structure and organization around 3.5.Source: AI judge vs Human teachers · pp. 73–74

Scoring per topic

  • Assessments varied by topic: Session 4 was scored highly by both evaluator types, while ART uniqueness and CHOICES content showed contrasting group or evaluator patterns.In ART, teachers rated uniqueness highly while the AI judge was usually below the mean; in CHOICES, both rated Brain-only content low and LLM content high.Source: Scoring per topic · pp. 76–77

Interviews

  • Interviews surfaced Brain-only complaints about limited time, the importance of owning ideas, and ChatGPT’s popularity for generating introductions in Sessions 1–3 but not Session 4.The interview analysis used PaCMAP to identify 15 clusters.Source: Interviews · p. 85

EEG ANALYSIS

  • EEG analysis used a 32-electrode headset at 500 samples per second, filtered and cleaned signals with ICA, and min-max normalized values to 0–1.Reported bands included low/high alpha, low/high beta, low/high delta, and theta.Source: EEG ANALYSIS · p. 86

Dynamic Directed Transfer Function (dDTF)

  • The study used dDTF to fit dynamic multivariate autoregressive models and estimate directional, frequency-domain connectivity through pairwise electrode analysis.Unlike coherence, the resulting connectivity is asymmetric: A→B need not equal B→A.Source: Dynamic Directed Transfer Function… · p. 86
  • The analysis used 1-second EEG windows and AIC/BIC for pairwise MVAR order selection; both criteria generally favored order 5, though about 25% of pairs were order 6–8.The authors report that AIC-selected order 5 gave the best multi-band and subband dDTF results for their data and window size.Source: Dynamic Directed Transfer Function… · pp. 86–87

Alpha Band Connectivity

  • 79 versus 42 significant alpha connections favored Brain-only; its P7→T8 dDTF was 0.053 versus 0.009 for LLM (p=0.0002).The authors associate Brain-only’s higher alpha connectivity with internal attention and semantic processing during unassisted idea generation.Source: Alpha Band Connectivity · p. 88, Figure 57 · p. 89

Beta Band Connectivity

  • Low-beta total connectivity was 2.854 for Brain-only versus 2.653 for LLM; high-beta patterns were more balanced, with no beta connection stronger in LLM.The authors interpret Brain-only’s greater beta connectivity as reflecting sustained cognitive and motor engagement during composition.Source: Beta Band Connectivity · p. 90, Figure 58 · p. 91

Delta Band Connectivity

  • Brain-only showed 78 delta connections versus 31 in the opposite direction, with T7→AF3 dDTF of 0.022 versus 0.007 for LLM (p=0.0002).The authors link the pattern to broader slow-frequency integration and monitoring during unassisted writing.Source: Delta Band Connectivity · p. 92, Figure 59 · p. 93

Theta Band Connectivity

  • Brain-only had 65 theta connections versus 29 for LLM, and its Pz→T8 dDTF was 0.041 versus 0.009 (p=0.0012).Frontal theta connectivity was stronger in Brain-only, which the authors associate with greater working-memory load and executive control.Source: Theta Band Connectivity · p. 94, Figure 60 · p. 95

Summary

  • Brain-only writing showed stronger connectivity across bands, while LLM writing had lower overall connectivity and lighter working-memory and executive demands.The findings also suggest a shift in information flow: Brain-only writing showed more posterior-to-frontal flow, whereas LLM writing showed relatively greater frontal-to-posterior flow in some bands.Source: Summary · pp. 96–97

Alpha Band Connectivity

Across sessions 1–3, Brain-only participants showed stronger alpha, theta, and delta connectivity than Search Engine users, while beta results were mixed: Brain-only had slightly greater total strength, but Search Engine dominated more links.

Alpha Band Connectivity

  • Alpha dDTF was higher for Brain-only than Search Engine participants (0.423 vs. 0.288), with stronger incoming connectivity to F4 (~0.203 vs. 0.074).Brain-only also showed greater posterior-region outputs, while Search Engine’s few stronger links were fewer and weaker than the Brain-only frontal-hub pattern.Source: Alpha Band Connectivity · p. 98, Figure 61 · p. 99
  • Beta dDTF was slightly higher overall for Brain-only (0.417 vs. 0.355), but more connections favored Search Engine (11 vs. 7).Search Engine had greater beta input to Pz (0.151 vs. 0.052), whereas Brain-only had greater input to T8 (0.246 vs. 0.085).Source: Alpha Band Connectivity · p. 100, Figure 62 · p. 101
  • Theta dDTF totaled 0.644 for Brain-only versus 0.331 for Search Engine, and 22 connections favored Brain-only versus 4 favoring Search Engine.F4 received 11 significant theta inputs in Brain-only (0.336 vs. 0.127), making it a prominent hub.Source: Alpha Band Connectivity · p. 102, Figure 63 · p. 103

Delta Band Connectivity

  • Low-delta connectivity was 1.051 for Brain-only versus 0.537 for Search Engine; high-delta connectivity was 0.637 versus 0.261.Brain-only had 32 versus 6 low-delta connections and 26 versus 1 high-delta connections favoring its group, respectively.Source: Delta Band Connectivity · p. 104
  • Delta dDTF totaled 0.588 for Brain-only versus 0.264 for Search Engine, with 21 connections favoring Brain-only and 1 favoring Search Engine.T8 was a major Brain-only delta sink, receiving four links totaling 0.204 versus 0.044 for Search Engine.Source: Delta Band Connectivity · p. 104, Figure 64 · p. 106

Summary

The section summarizes EEG connectivity across writing conditions and repeated sessions, including changes associated with unaided writing after prior AI assistance and with switching from unaided writing to AI assistance.

Summary

  • Brain-only participants showed stronger, more extensive connectivity than Search Engine users, especially in delta/theta, while Search Engine users showed more focal beta connectivity around Pz.The Brain-only pattern was associated with internal memory and planning; Search Engine patterns were interpreted as handling external information.Source: Summary · p. 107, Figure 65 · p. 108

Alpha Band Connectivity

  • Alpha dDTF was comparable (Search Engine 0.901 vs. LLM 0.891), but Search Engine showed greater AF3-directed inflow while LLM showed stronger high-beta motor-region outflow.Search Engine also showed enhanced low-beta flow toward AF3, whereas LLM connectivity was interpreted as involving typing-related procedural fluency and feedback.Source: Alpha Band Connectivity · pp. 109, 111, Figure 66 · p. 109, Figure 68 · p. 111
  • Theta connectivity was higher overall for LLM than Search Engine (0.920 vs. 0.826), although Search Engine retained more posterior-to-frontal connections into AF3.Search Engine’s AF3-directed links included PO3 and C3, while LLM’s stronger total connectivity included parietal and central sources toward frontal targets.Source: Alpha Band Connectivity · p. 113, Figure 70 · p. 113

Delta Band Connectivity

Cognitive Adaptation

  • Across repeated Brain-only writing, delta, theta, and alpha connectivity rose sharply from Session 1 to 2 before plateauing or declining, while beta rose through Session 3.Session 2’s delta, theta, and alpha peaks were attributed to task familiarity; the Session 4 post-LLM attempt did not show the same revision-related surge.Source: Cognitive Adaptation · pp. 117–118, Cognitive processing · p. 124

Cognitive offloading to AI

  • After prior LLM use, Session 4 alpha and beta connectivity was below practiced-session peaks but above Session 1, consistent with an intermediate level of neural engagement.The authors suggest that prior AI assistance may have reduced later recruitment of planning, reasoning, and working-memory networks during unaided writing.Source: Cognitive Adaptation · p. 118, Cognitive offloading to AI · p. 118, Figure 74 · p. 122

Cognitive processing

  • Post-LLM Session 4 still showed theta connections matching or exceeding Session 3 and delta connectivity about 45% above Session 1, despite lower alpha/beta engagement.The authors interpret retained theta/delta activity as continued retrieval, planning, and attention rather than complete disengagement.Source: Cognitive processing · pp. 120, 122, Figure 73 · p. 120, Figure 74 · p. 122
  • LLM-to-Brain essays reused the LLM group’s “before speaking” n-gram, and teachers scored uniqueness low for the associated FORETHOUGHT and PERFECT topics.The authors connect these findings to possible limits in originality and independent critical work after AI-supported writing.Source: Cognitive processing · p. 122

Cognitive 'Deficiency'

  • Repeated unaided writing strengthened connectivity across multiple bands, whereas prior AI assistance was associated with reduced high-frequency connectivity during unaided writing.LLM-to-Brain participants nevertheless showed sustained theta and delta activity, suggesting continued cognitive effort that may have focused more on recall than complex reasoning; top-down control pathways may also be less engaged with LLM use.Source: Cognitive 'Deficiency' · p. 126
  • The authors recommend hybrid writing strategies that reserve idea generation, organization, and critical revision for users while allowing AI to handle routine composition tasks.They propose full neural engagement during early learning, with selective AI support in later practice to improve efficiency without undermining established writing networks.Source: Cognitive 'Deficiency' · p. 126
  • Session 4 connectivity sums exceeded the LLM group’s earlier-session sums across all four EEG frequency bands.Alpha: Session 4 was 0.823 versus Sessions 1, 2, and 3 at 0.547, 0.285, and 0.107. Beta: 1.924 versus 1.656, 0.585, and 0.275; theta: 1.087 versus 0.394, 0.260, and 0.132; delta: 1.948 versus 0.637, 0.408, and 0.188. The theta and delta patterns included frontoparietal working-memory links and visual-to-prefrontal influence, respectively.Source: Cognitive 'Deficiency' · pp. 127–128, Figure 77 · p. 129

Interpretation

The interpretation examines how external writing tools relate to neural engagement, essay language, and behavioral outcomes across unaided and AI-assisted writing. It also considers how prior unaided writing may shape later AI-assisted engagement, while noting limits in the Session 4 evidence and generalizability.

Band specific cognitive implications

  • Session 4 theta/alpha flows were interpreted as greater attention and memory demands, beta flows as active monitoring and revision, and parietal-to-central or occipital-to-frontal flows as integration and executive evaluation.The authors propose that AI may offload some cognitive processes while adding decision-making demands; the band-specific interpretations are presented as likely implications rather than direct measurements of those processes.Source: Band specific cognitive implications · p. 129, Inter-group differences: Cognitive… · p. 129

Neural Adaptation: from Endogenous to Hybrid Cognition in AI Assistance

  • After three AI-free essays, Brain-to-LLM participants showed stronger directed connectivity across all bands than initial LLM sessions, consistent with adapting to new AI suggestions by integrating prompts with existing plans.The authors contrast this network-wide Session 4 increase with declining connectivity across LLM sessions, which they interpret as possible neural adaptation or network specialization.Source: Interpretation · p. 129, Neural Adaptation: from Endogenous to… · p. 131

In-Depth NLP Topics Analysis Sessions 1, 2, 3 vs Session 4

LLM Group; Search Group; Brain-only Group

  • LLM users showed weak frontoparietal connectivity and goal-oriented phrasing, Search users showed delta connectivity interpreted as effortful integration of retrieved material, and Brain-only users showed stronger coupling with reflective, prosocial phrasing.The authors caution that the proposed link between phrase provenance and connectivity is speculative because it was analyzed for only one topic.Source: LLM Group · p. 136, Search Group · p. 138, Brain-only Group · pp. 140, 142

DISCUSSION

  • Across topics, Brain-only essays varied more, LLM essays were more homogeneous and used more named entities, and the groups differed in topic framing and vocabulary; the authors associate some patterns with tools, prompts, queries, and time pressure.They report Search and LLM essays converged in ontology more than Brain-only essays, while Brain-only essays more often addressed freedom/liberty and the tool-assisted groups justice.Source: DISCUSSION · pp. 143–145
  • Human teachers were more skeptical of essay uniqueness and quality than the AI judge, and identified recurring participant-specific styles that the AI judge failed to attribute.Teachers were not told participants’ assigned writing conditions.Source: DISCUSSION · p. 145

Neural Connectivity Patterns

  • Brain-only writers showed the strongest, widest-ranging connectivity, Search Engine writers intermediate engagement, and LLM writers the weakest, with up to 55% lower total dDTF magnitude than Brain-only writers in low-frequency networks.Search Engine connectivity was approximately 34–48% lower than Brain-only, depending on frequency band; Search Engine users also showed visual-cortex activity interpreted as screen-based scanning and evaluation.Source: Neural Connectivity Patterns · p. 146
  • Brain-only writers showed broad semantic, creative, and executive activation; Search Engine users combined visual management and regulatory control, while LLM use favored procedural integration over visual evaluation.The authors interpret these patterns as tool use restructuring the cognitive architecture of essay writing.Source: Neural Connectivity Patterns · p. 146

Quoting Ability and Memory Encoding

  • In Session 1, 83% of LLM users reported difficulty quoting their essays and none quoted correctly; by Session 3, 6/18 still failed, while Brain-only and Search Engine groups reached 100% reported ability.The authors link LLM users’ weaker quoting to reduced theta/alpha connectivity and shallow semantic encoding; correct quoting followed Brain-only > Search Engine > LLM.Source: Quoting Ability and Memory Encoding · p. 147, Correct Quoting · pp. 147–148

Essay Ownership and Cognitive Agency

  • Brain-only participants reported full essay ownership almost unanimously, while LLM users expressed fragmented authorship; the authors associate stronger connectivity and internal generation with memory, semantic accuracy, and ownership.Search Engine users showed more stable ownership than LLM users, but less certainty than Brain-only participants; the paper frames these differences as a trade-off between tool efficiency and internalization.Source: Essay Ownership and Cognitive Agency · p. 148, Cognitive Load, Learning Outcomes, and… · p. 148

Session 4

  • In Session 4, LLM-to-Brain participants exceeded Brain-only Session 1 connectivity but remained below Session 2 and lacked Session 3 consolidation networks, while Brain-to-LLM participants increased connectivity across EEG bands.The authors suggest introducing AI after initial unaided effort may support neural integration; the LLM-to-Brain group also lacked highly significant connections (p < 0.001).Source: Session 4 · p. 149

Behavioral Correlates of Neural Connectivity Patterns in Session 4

  • In Session 4, 78% of LLM-to-Brain participants failed to quote anything and 11% quoted correctly, versus 11% and 78% in the Brain-only group.LLM-to-Brain participants lacked robust fronto-parietal synchronization, while Brain-to-LLM participants showed stronger connectivity; repeated ideas in the LLM-to-Brain group were flagged as preliminary and need confirmation in a larger sample.Source: Behavioral Correlates of Neural… · pp. 150–151

Limitations and Future Work

  • The authors note limitations including a small, geographically concentrated sample, use of ChatGPT alone, an essay-only context, and no division of writing into subtasks.They call for larger and more diverse samples, comparisons across LLMs and modalities, and task-stage analysis.Source: Limitations and Future Work · pp. 151–152
  • The EEG analysis focused on connectivity, limiting assessment of spectral power and precise localization; the authors recommend fMRI and longitudinal research on memory, creativity, and writing fluency.They also propose collecting unaided writing samples to develop participant-specific style representations, noting that detection precision was limited by sample size.Source: Limitations and Future Work · p. 152

Energy Cost of Interaction

  • LLM queries were described as using around 10 times more energy than search queries; the authors also report reduced critical evaluation of LLM output and teacher detection of conventional, homogeneous essays.They call for continued research into LLMs’ effects on cognitive development and intellectual independence.Source: Energy Cost of Interaction · p. 152, Conclusions · p. 153

Conflict of Interest · Appendix

The authors report that the project was completed before Dr. Kosmyna’s Google affiliation, and that the research received no external funding.

Conflict of Interest

  • Dr. Kosmyna held a Visiting Researcher position at Google when the paper was published, but the project work was completed before that affiliation; the other authors declared no conflicts.The publication date was June 2025.Source: Conflict of Interest · p. 154

Funding

A: List of clusters for Figure 56. · B: Dynamic Direct Transfer Function (dDTF) for Alpha band for participants 36 and 43 · C: Aggregated dDTF for sessions 1, 2, 3 in LLM

Figure 56 groups interview insights across session 4 and sessions 1–3; accompanying dDTF panels describe alpha-band connectivity for two participants and aggregated connectivity in the LLM group.

A: List of clusters for Figure 56.

  • PaCMAP defined clusters of interview insights spanning session 4 and sessions 1–3.The listed insights correspond to positions on the map, read top to bottom and left to right.Source: A: List of clusters for Figure 56 · p. 167
  • Interviewees recalled essay topics and structures, but one could not reproduce a quote word for word.Recalled content included the benefits of choice and the importance of thinking before speaking.Source: A: List of clusters for Figure 56 · p. 167
  • Reported writing processes ranged from using ChatGPT for examples or an introduction to relying on personal ideas without outside sources.One respondent described revising ChatGPT output to fit their thoughts; another valued generating ideas independently.Source: A: List of clusters for Figure 56 · p. 167
  • Interviewees described choosing prompts for personal relevance or to gather information and challenge a text’s main tendency.One respondent also said 20 minutes might not allow a more analytical essay.Source: A: List of clusters for Figure 56 · p. 167

B: Dynamic Direct Transfer Function (dDTF) for Alpha band for participants 36 and 43

  • The alpha-band dDTF panels compare participants 36 and 43, who wrote on the same topic using no tools or an LLM.They display all 992 electrode pairs, mark significant pairs below p < 0.05, and show filtered, normalized significant dDTF values.Source: B: Dynamic Direct Transfer Function… · p. 168

C: Aggregated dDTF for sessions 1, 2, 3 in LLM

D: Aggregated dDTF for sessions 1, 2, 3 in Search · E: Aggregated dDTF for sessions 1, 2, 3 in Brain-only

The Search and Brain-only sections describe aggregated dDTF connectivity averaged across sessions 1–3, with area-level totals organized by connection direction.

D: Aggregated dDTF for sessions 1, 2, 3 in Search

E: Aggregated dDTF for sessions 1, 2, 3 in Brain-only

F: Alpha dDTF LLM vs Brain-only sessions 1, 2, 3 … Top 10 significant dDTF

The listed significant alpha dDTF connections show effects in both directions: some are stronger in Brain-only participants and others in LLM users.

Top 26 significant dDTF

  • Brain > LLM is reported for connections including P7→T8 (0.0114481049 vs 0.0647280365) and Fz→T8 (0.0149086686 vs 0.0680834651).The table labels both connections Brain > LLM.Source: Top 26 significant dDTF · p. 172
  • LLM > Brain is reported for connections including Fp2→Pz (0.0268734414 vs 0.0059849266) and FC1→Pz (0.0245060250 vs 0.0067248377).The table labels both connections LLM > Brain.Source: Top 26 significant dDTF · p. 172

Related papers

Loading 2506.08872v2…