CoChat
Sign Up
Cited by 4Writing and Handwriting EducationArtificial Intelligence in Healthcare and EducationMachine Learning in Materials Science

Ink and algorithm: Exploring temporal dynamics in generative AI ‐assisted writing

Open access
Yang, Kaixun, Cheng, Yixin, Zhao, Linxuan, Raković, Mladen, Swiecki, Zachari, Gašević, Dragan, Chen, Guanliang·2026·British Journal of Educational Technology
DOI 10.1111/bjet.70031
Open this paper in CoChat — free accountOpen PDF

Abstract

The advent of Generative AI (GAI) has transformed writing, marking a shift towards GAI‐assisted writing in education. However, the dynamics of human–AI interaction in the writing process are not well‐understood, and thus, it remains largely unknown how human learning can be effectively supported with such technologies. This study addresses this gap by investigating how humans employ GAI during writing and examining the interplay between patterns of GAI usage and writing behaviours. To capture these patterns, we applied Dynamic Time Warping time‐series clustering to identify temporal trajectories of GAI use and employed Epistemic Network Analysis to examine how these trajectories relate to cognitive processes such as knowledge telling, knowledge transformation and cognitive presence. Our analysis revealed four distinct temporal patterns of GAI usage (ie, AI‐critical writers, AI‐dependent writers, AI‐independent writers and AI‐balanced writers), each associated with different cognitive engagement strategies. The findings suggest that some writers tend to rely excessively on GAI, which may limit opportunities for meaningful learning by reinforcing surface‐level strategies such as knowledge telling. These insights highlight the need for researchers and educators to design GAI‐based writing assistants that are sensitive to students' temporal dynamics and can scaffold more productive cognitive engagement. Practitioner notes What is already known about this topic Researchers and educators can gain practical insights into achieving intelligence augmentation through critical engagement by studying effective user behaviours for enhanced human–AI partnership in writing. Generative AI‐assisted writing can be evaluated using an evidence‐centred assessment framework, which relates to writing cognitive processes. Current studies on Generative AI‐assisted writing usually classify writers according to their overall AI usage behaviours throughout the entire writing session. What this paper adds We propose using time‐series clustering to identify and analyse common temporal patterns in AI usage during generative AI‐assisted writing processes. We uncover the correlation between temporal patterns in AI usage and human writing behaviours, which reflect cognitive processes, through Epistemic Network Analysis. We identify four major distinct temporal patterns in AI utilization and highlight that each pattern is correlated with different cognitive processes. Implications for practice and/or policy Researchers and educators should be aware of the risks associated with students overly relying on GAI for writing tasks, as this dependency can hinder opportunities for meaningful learning. Researchers and educators should carefully consider the types of suggestions provided by GAI‐based writing tools when integrating them into educational contexts.

Integrity signals

Registry

Crossref · 10.1111/bjet.70031

Cites

81 of 81 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

12 of 47 cited sources checked · 8 supported · 1 flagged · accuracy not yet measured

  • Interpretationcitation

    Previous studies revealed limited insight into writers' writing behaviours and corresponding cognitive processes, offering limited implications for supporting human learning in GAI-assisted writing.

    “(ii) they revealed limited insight into writers' writing behaviours and the corresponding cognitive processes and thus offered limited implications for supporting human learning in this particular task (studies by Nguyen et al., 2024;Shibani et al., 2023;Wambsganss et al., 2023).”
    Quote in the paperSubstantial claim1 of 3 sources checked
    • cites Nguyen et al., 2024 source

      The Discussion identifies distinct writing behaviours and cognitive and metacognitive skills, then explicitly says these findings can inform targeted pedagogical interventions. The Conclusions and final remarks call the understanding preliminary and recommend deeper study of cognitive processes, but that qualification does not support the claim that the study offered limited insight into behaviours and limited implications for learning support. The indexed passages have no page numbers. (LLM, conf=0.86)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured
    • cites Shibani et al., 2023

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Wambsganss et al., 2023

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Interpretationcitation

    Previous studies on GAI-assisted writing often categorize writers based on overall AI usage throughout the entire writing session, overlooking temporal dynamics.

    “However, these existing studies often have two primary shortcomings: (i) They generally categorized writers based on their overall AI usage throughout the entire writing session, overlooking the temporal dynamics of these behaviours as the progression of a writing session (studies by Cheng et al., 2024;Shibani et al., 2023;Wambsganss et al., 2023;Yang et al., 2025);”
    Quote in the paperSubstantial claim2 of 3 sources checked
    • cites Cheng et al., 2024 source
    • cites Shibani et al., 2023

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Yang et al., 2025 source

      Section 3.2 (p. 6) classifies writers in this study by the overall frequency of AI-assisted behavior patterns, supporting that part of the claim. But Section 2.2 (p. 5) describes an earlier study that used time-series clustering, so this paper does not support treating the neglect of temporal dynamics as a general shortcoming of previous studies. (LLM, conf=0.83)

      Experimental — accuracy not yet measured
  • Findingown data

    Some writers often accept GAI-generated text with minimal modification before incorporating it into their work.

    “Our study shows that some writers often accept GAI-generated text with minimal modification before incorporating it into their work.”
    Quote in the paperSubstantial claim
  • Findingown data

    Cluster 3 was characterized by the weakest connection between compose and seekSugg among all clusters.

    “Cluster 3 was characterized by the weakest connection between compose and seek-Sugg among all clusters, suggesting that writers in this group were less inclined to frequently seek input from GAI during the writing process.”
    Quote in the paperNot a substantial claim
  • Findingown data

    Cluster 1 writers exhibited the highest rate of modification among all clusters, adjusting about 46% of accepted suggestions.

    “Instead, they exhibited the highest rate of modification among all clusters-adjusting nearly half of the suggestions they accepted (about 46%).”
    Quote in the paperNot a substantial claim
  • Classified as probably not a substantial claim. 40
  • May combine two claims. 18
  • The quote only partly states this claim. 3

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    A writer's sense of ownership may diminish as the proportion of system-generated content increases.

    “In particular, a writer's sense of ownership may diminish as the proportion of system-generated content increases (Lee et al., 2022).”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Lee et al., 2022 source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Yang et al. (2025) demonstrated that writers who actively engaged in higher order cognitive processes consistently produced higher-quality essays.

    “Similarly, Yang et al. (2025) demonstrated that writers who actively engaged in higher order cognitive processes consistently produced higher-quality essays.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Yang et al., 2025 source

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Cheng et al. (2024) revealed significant differences in writing processes across groups defined by authorship (User vs. GAI), writing type (Creative vs. Argumentative), and GPT temperature settings (High vs. Low).

    “Cheng et al. (2024) proposed a learning analytics-based methodology for evaluating GAI-assisted writing, revealing significant differences in writing processes across groups defined by authorship (User vs. GAI), writing type (Creative vs. Argumentative) and GPT temperature settings (High vs. Low).”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Cheng et al., 2024 source

      The Discussion and Conclusions reports an assessment method and statistically significant writing-process differences for user versus GAI ownership and creative versus argumentative prompts. The Regression Results (section 4.2; page not provided) says temperature was not statistically significant on either dimension, so the claimed significant High-versus-Low temperature difference is not supported. (LLM, conf=0.94)

      Experimental — accuracy not yet measured

  • Classified as probably not a substantial claim.

    Interpretationcitation

    How writers' relationships to systems evolve through extended use, especially in interactions with GAI tools, remains underexplored.

    “However, how writers' relationships to systems evolve through extended use (Lee et al., 2024), especially in the context of interactions with GAI tools, remains underexplored.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites Lee et al., 2024 source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Lee et al. (2024) emphasized that the effectiveness of a writing assistant often depends on how well it aligns with a user's specific needs.

    “Lee et al. (2024) emphasized that the effectiveness of a writing assistant often depends on how well it aligns with a user's specific needs.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Lee et al., 2024 source

  • Classified as probably not a substantial claim.

    Premisecitation

    Nguyen et al. (2024) found that doctoral students who engaged in iterative, highly interactive processes with GAI-powered tools generally performed better in the writing task.

    “Nguyen et al. (2024) proposed using Hidden Markov Models (HMM) and hierarchical sequence clustering to analyse human-AI interaction sequences in academic writing, finding that doctoral students who engaged in iterative, highly interactive processes with GAI-powered tools generally performed better in the writing task.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites Nguyen et al., 2024 source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Lee et al. (2024) categorized the outputs of writing assistants into three types: analysis, generation, and proposal.

    “Lee et al. (2024) categorized the outputs of writing assistants into three types: analysis, generation and proposal.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Lee et al., 2024 source

  • May combine two claims.

    Premisecitation

    ENA has demonstrated its effectiveness and validity as a method for assessing GAI-assisted writing in previous research.

    “ENA has demonstrated its effectiveness and validity as a method for assessing GAIassisted writing in previous research (Cheng et al., 2024).”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites Cheng et al., 2024 source

      The Discussion and Conclusions report statistically significant differences between writing conditions using an ENA-based assessment, supporting its usefulness as a proof of concept. But the Research Questions section says the dataset lacked writing-quality scores needed to test whether the method distinguishes high- from low-quality processes, and the Discussion calls the method plausible rather than validated. The claim overstates the evidence for validity. (LLM, conf=0.87)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Students expect AI to fulfil multiple roles, such as multi-tasking writing assistant, virtual tutor and digital peer, to support multifaceted writing processes and performance.

    “Some studies examined students' perceptions of using these tools during the writing process (Kim et al., 2025;Wang, 2024), finding that students expect AI to fulfil multiple roles, such as multi-tasking writing assistant, virtual tutor and digital peer, to support multifaceted writing processes and performance.”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites Kim et al., 2025 source
    • cites Wang, 2024

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Findingown data

    Cluster 2 had the lowest modification rate (about 17%) and the second-highest acceptance rate (about 66%).

    “At the same time, they had the lowest modification rate (about 17%) and the second-highest acceptance rate (about 66%), suggesting a pattern of frequent and unaltered use of GAIgenerated content.”
    Quote in the paperNot a substantial claim

  • May combine two claims. The quote only partly states this claim.

    Findingown data

    Cluster 4 produced the highest proportion of AI-generated characters (averaging 36%) and maintained a moderate modification rate (about 34%).

    “They also produced the highest proportion of AI-generated characters (averaging 36%) and maintained a moderate modification rate (about 34%), pointing to a willingness to both accept and adapt AI suggestions as needed.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Cluster 3 writers submitted the fewest GAI requests overall (fewer than 8 per draft on average) and their final outputs contained the lowest proportion of AI-generated content (approximately 18%).

    “Writers in this group submitted the fewest GAI requests overall (fewer than 8 per draft, on average), and their final outputs contained the lowest proportion of AIgenerated content (approximately 18%).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    AI-dependent writers consistently interacted with GAI across all stages of the writing process, whereas AI-balanced writers primarily engaged with GAI during the initial stages.

    “For instance, our findings reveal that AIdependent writers consistently interacted with GAI across all stages of the writing process, whereas AI-balanced writers primarily engaged with GAI during the initial stages.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    AI-dependent writers (Cluster 2) maintain consistent, predefined roles between human writers and GAI, aligning closely with the script-based writing approach.

    “This aligns closely with practices observed among AI-dependent writers (Cluster 2), who maintain consistent, predefined roles between human writers and GAI.”
    Quote in the paperSubstantial claim

  • May combine two claims. The quote only partly states this claim.

    Findingown data

    Cluster 4 displayed the second-highest average number of GAI requests and the highest acceptance rate of all clusters (approximately 77%).

    “In terms of their final outputs, this group displayed the second-highest average number of GAI requests and the highest acceptance rate of all clusters (approximately 77%).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S1 (asserted), S3 (load-bearing), S4 (based). The quote only partly states this claim.

    Findingown data

    Cluster 4 writers demonstrated a front-loaded pattern of engagement with GAI, using the tool most heavily during the early stages.

    “Writers in this cluster demonstrated a front-loaded pattern of engagement with GAI, using the tool most heavily during the early stages of their writing process.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingreasoning

    AI-balanced writers (Cluster 4) interchangeably utilize GAI either to propose content based on human-generated ideas or to modify AI-generated text to better meet their evolving requirements, mirroring the collective writing approach.

    “This mirrors the behaviour of AI-balanced writers (Cluster 4), who interchangeably utilize GAI either to propose content based on human-generated ideas or to modify AI-generated text to better meet their evolving requirements.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Cluster 1 exhibited notably stronger connections between compose and seekSugg, as well as between cursorFwd and lowModification, when compared to Cluster 3.

    “Cluster 1 exhibited notably stronger connections between compose and seekSugg, as well as between cursorFwd and lowModification, when compared to Cluster 3.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Time-series clustering methods are commonly divided into three classes: model-based, feature-based, and shape-based approaches.

    “Methods for clustering time-series data are commonly divided into three classes (Aghabozorgi et al., 2015): modelbased, feature-based and shape-based approaches.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Aghabozorgi et al., 2015 source

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Research exploring the use of shape-based time-series clustering in GAI-assisted writing tasks is notably absent.

    “Research exploring the use of shape-based time-series clustering in GAI-assisted writing tasks is notably absent.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

    Premisenone given

    These cognitive frameworks collectively suggest a set of claims for assessing the processes involved in co-writing with GAI: using GAI suggestions as is (Knowledge Telling); modifying GAI suggestions to fit the writer's own goals (Knowledge Transformation/integration); requesting suggestions from GAI (triggering); comparing different GAI suggestions (exploration); and arriving at a final product (

    “These cognitive frameworks collectively suggest a set of claims for assessing the processes involved in co-writing with GAI: using GAI suggestions as is (Knowledge Telling); modifying GAI suggestions to fit the writer's own goals (Knowledge Transformation/ integration); requesting suggestions from GAI (triggering); comparing different GAI suggestions (exploration); and arriving at a final product”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisereasoning

    In education, the majority of time-series clustering research has utilized feature-based and shape-based methods.

    “In the field of education, the majority of research has utilized feature-based (Hassan et al., 2022;Talebinamvar & Zarrabi, 2022;Zhang et al., 2019) and shape-based (He et al., 2023;Ma'ady et al., 2022;Shen & Chi, 2017) methods, primarily due to their ease of interpretation and lower computational complexity.”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites Hassan et al., 2022 source
    • cites He et al., 2023

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Shen & Chi, 2017

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Shibani et al. (2023) found that writers who used GPT-3 less frequently produced texts with higher lexical density and greater semantic overlap.

    “For example, Shibani et al. (2023) found that writers who used GPT-3 less frequently produced texts with higher lexical density and greater semantic overlap.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Shibani et al., 2023

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    ENA models the co-occurrence of coded elements in discourse to identify and compare patterns of thinking and learning.

    “In education, a cornerstone of QE is Epistemic Network Analysis (ENA), which models the co-occurrence of coded elements in discourse to identify and compare patterns of thinking and learning (Shaffer & Ruis, 2017).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Shaffer & Ruis, 2017

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    Existing research on GAI-assisted writing is relatively scarce.

    “Existing research on GAI-assisted writing is relatively scarce.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    ENA has been widely applied across various educational settings to investigate complex learning processes such as collaborative learning, self-regulated learning, and communities of practice.

    “ENA has been widely applied across various educational settings to investigate complex learning processes such as collaborative learning (Csanadi et al., 2017;Nachtigall & Sung, 2019), self-regulated learning (Lim et al., 2021;Uzir et al., 2020) and communities of practice (Hu et al., 2018;Peng et al., 2020).”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites Csanadi et al., 2017

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Lim et al., 2021 source
    • cites Hu et al., 2018

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning

    How temporal AI usage patterns interact with human writing behaviours and their subsequent impact on writing performance and human learning remain largely unexplored.

    “Additionally, how these temporal AI usage patterns interact with human writing behaviours and their subsequent impact on writing performance and human learning remain largely unexplored.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Epistemic Network Analysis (ENA) utilizes network models to depict epistemic actions and can evidence cognitive behaviours inherent to co-writing with GAI.

    “we conducted Epistemic Network Analysis (ENA) (Shaffer & Ruis, 2017), a popular learning analytics method that utilizes network models to depict epistemic actions, on each cluster to evidence the cognitive behaviours inherent to co-writing with GAI.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Shaffer & Ruis, 2017

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Premisecitation

    Learning is a process that unfolds over time, making temporal analyses essential to better capture its nature.

    “Additionally, several studies point out that learning is a process that unfolds over time, making it essential to conduct insightful temporal analyses to better capture its nature (Chen et al., 2018;Knight et al., 2017).”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Chen et al., 2018 source
    • cites Knight et al., 2017 source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Wambsganss et al. (2023) found no significant difference in bias between reviews written by students with and without GAI-generated suggestions, regardless of whether GPT-2 or GPT-3 was used.

    “Wambsganss et al. (2023) investigated gender bias in a GAI-assisted writing pipeline and found no significant difference in bias between reviews written by students with and without GAI-generated suggestions, regardless of whether GPT-2 or GPT-3 was used.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Wambsganss et al., 2023

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    Prioritizing proposal-type suggestions over generation-type suggestions leads students to maintain greater control over their writing processes.

    “By prioritizing proposal-type suggestions over generation-type suggestions, students maintain greater control over their writing processes.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    Because GAI generates suggestions based on human-created content, human-written content can influence GAI-written content, making separate writing less distinct in GAI-assisted writing tasks.

    “However, because GAI generates suggestions based on human-created content, the human-written content can influence the GAI-written content, making the separate writing less distinct in GAI-assisted writing tasks.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: GAI generates suggestions based on human-created content.

  • May combine two claims.

    Interpretationreasoning

    The convergence of behavioural and self-reported data suggests that Cluster 2 writers engaged with GAI frequently and incorporated its output extensively.

    “The convergence of behavioural and self-reported data suggests that these writers engaged with GAI frequently and incorporated its output extensively.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Zhang et al. (2021) found that text production and editing behaviours captured from keystrokes can be used to distinguish writers with different cognitive processes.

    “Similarly, Zhang et al. (2021) found that text production and editing behaviours captured from keystrokes can be used to distinguish writers with different cognitive processes.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Zhang et al., 2021

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Insights into writers' temporal behaviours in GenAI-assisted writing can inform the design of features for time-series models that capture writing performance over time.

    “Our insights into writers' temporal behaviours in GenAI-assisted writing can inform the design of features for time-series models that capture writing performance over time.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    Prioritizing proposal-type suggestions over generation-type suggestions supports student agency.

    “This approach supports student agency, encourages deeper engagement with writing tasks and ensures that final written products predominantly reflect students' authentic voices and ideas, rather than becoming a blend of human-AI contributions.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The analysis did not differentiate data according to essay genres, such as creative vs. argumentative writing.

    “Third, our analysis did not differentiate data according to essay genres, such as creative vs. argumentative writing.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    The dataset is well-suited to address the research questions and can provide meaningful insights into real-world educational scenarios.

    “Therefore, this dataset is well-suited to address our research questions and can provide meaningful insights into real-world educational scenarios.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Cluster 2 writers predominantly adopted the Knowledge Telling strategy and engaged in Cognitive Presence-Triggering.

    “As a result, these writers predominantly adopted the Knowledge Telling strategy and engaged in Cognitive Presence-Triggering.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Latif (2021) proposed a writing process model composed of seven components (eg, Linguistic rehearsing, Reviewing and Ideational planning) and offered explanations of the strategies used in each component.

    “Latif (2021) proposed a writing process model composed of seven components (eg, Linguistic rehearsing, Reviewing and Ideational planning) and offered explanations of the strategies used in each component.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Latif, 2021 source

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    Prioritizing proposal-type suggestions over generation-type suggestions ensures that final written products predominantly reflect students' authentic voices and ideas rather than becoming a blend of human-AI contributions.

    “This approach supports student agency, encourages deeper engagement with writing tasks and ensures that final written products predominantly reflect students' authentic voices and ideas, rather than becoming a blend of human-AI contributions.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    In the separate writing approach, writing tasks are divided into distinct parts among group members, rather than collaboratively producing a single document.

    “Additionally, in the separate writing approach, writing tasks are divided into distinct parts among group members, rather than collaboratively producing a single document.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Heinonen et al. (2020)

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Time-series models informed by writers' temporal behaviours can support the development of writing scaffolds and adaptive support systems, for example by using counterfactual modelling to identify key cognitive activities.

    “These models can, in turn, support the development of writing scaffolds and adaptive support systems (ie, by using counterfactual modelling to identify key cognitive activities that writers should engage in).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Writing encompasses a set of recursive and intertwined cognitive processes, for example planning, translating, reviewing and monitoring.

    “Writing encompasses a set of recursive and intertwined cognitive processes, for example planning, translating, reviewing and monitoring (Koppenhaver & Williams, 2010).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Koppenhaver & Williams, 2010 source

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Students overly relying on GAI for writing tasks may hinder opportunities for meaningful learning, for example by consistently adopting the knowledge-telling strategy.

    “Researchers and educators should be aware of the risks associated with students overly relying on GAI for writing tasks, as this dependency may hinder opportunities for meaningful learning (eg, consistently adopting the knowledge-telling strategy).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Baaijen et al. (2012) developed methods and measures for analysing keystroke logging with the goal of enhancing the correlation between keystroke data and cognitive processes.

    “Baaijen et al. (2012) developed methods and measures for analysing keystroke logging with the goal of enhancing the correlation between keystroke data and cognitive processes.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Baaijen et al., 2012 source

  • Classified as probably not a substantial claim.

    Premisecitation

    In the script-based writing approach, the script provides clear structure and explicitly outlines the responsibilities and tasks for each group member at specific points in the process.

    “In the script-based writing approach, the script provides clear structure and explicitly outlines the responsibilities and tasks for each group member at specific points in the process.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Heinonen et al. (2020)

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    Distinct motivational and contextual factors between Mechanical Turk workers and students may shape cognitive processes, temporal writing patterns and responses to GenAI assistance in fundamentally different ways.

    “These distinct motivational and contextual factors may shape cognitive processes, temporal writing patterns and responses to GenAI assistance in fundamentally different ways.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Mechanical Turk workers' motivations are likely tied to task completion and compensation.
    • Rests on an assumption: Student-authored educational writing often entails different goals, such as meeting academic expectations, demonstrating learning or engaging with disciplinary conventions.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Keystroke logging is a prevalent method for examining the writing process in educational settings, involving recording and timestamping keystroke actions to reconstruct the writing process.

    “Keystroke logging is a prevalent method for examining the writing process in educational settings, involving the recording and timestamping of keystroke actions to reconstruct the writing process (Leijten & Van Waes, 2013).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Leijten & Van Waes, 2013 source

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    DTW aligns well with the tasks of GAI-assisted writing.

    “This aligns well with the tasks of GAI-assisted writing.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    Prioritizing proposal-type suggestions over generation-type suggestions encourages deeper engagement with writing tasks.

    “This approach supports student agency, encourages deeper engagement with writing tasks and ensures that final written products predominantly reflect students' authentic voices and ideas, rather than becoming a blend of human-AI contributions.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Conijn et al. (2019) proposed the method of mapping features from the keystroke logs to higher level cognitive processes, such as planning and revising.

    “Conijn et al. (2019) proposed the method of mapping features from the keystroke logs to higher level cognitive processes, such as planning and revising.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Conijn et al., 2019 source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Conijn et al.'s (2019) findings suggest that keystroke features such as word count and number of revisions are associated with cognitive load.

    “For example, Conijn et al.'s (2019) findings suggest that keystroke features such as word count and number of revisions are associated with cognitive load.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Conijn et al., 2019 source
53 of 53

  • The paper relies on this without support.

    The CoAuthor dataset's log trace events provide a valid record of writers' AI usage and writing behaviours.

    “It offers a detailed log of keystroke-level interactions during these sessions, capturing 13 different event types (eg, text-insertion, text-deletion and suggestionget).”

  • The paper relies on this without support.

    Temporal patterns in AI usage can be meaningfully identified by clustering writing sessions with Dynamic Time Warping.

    “For this reason, to answer RQ1, we opted for Dynamic Time Warping (DTW) (Müller, 2007), a method capable of assessing the similarity between two action sequences irrespective of their length disparities.”

  • The paper relies on this without support.

    Humans vary in the speed of essay writing but exhibit comparable patterns in AI-usage behaviours.

    “Humans might vary in the speed of essay writing, yet exhibit comparable patterns in AI-usage behaviours.”

  • The paper relies on this without support.

    GAI generates suggestions based on human-created content.

    “because GAI generates suggestions based on human-created content”

  • The paper relies on this without support.

    Most existing approaches to writing scaffolds and adaptive support systems have not been specifically designed for the context of GAI-assisted writing.

    “most existing approaches have not been specifically designed for the context of GAI-assisted writing.”

  • The paper relies on this without support.

    Mechanical Turk workers' motivations are likely tied to task completion and compensation.

    “whose motivations are likely tied to task completion and compensation.”

  • The paper relies on this without support.

    Student-authored educational writing often entails different goals, such as meeting academic expectations, demonstrating learning or engaging with disciplinary conventions.

    “student-authored educational writing often entails different goals, such as meeting academic expectations, demonstrating learning or engaging with disciplinary conventions.”

  • The paper relies on this without support.

    Writing patterns can differ among genres.

    “Given that writing patterns can differ among genres”
8 of 8

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Cluster 2 is described as making no modifications despite a nonzero modification rate

    Minor

    The results interpret Cluster 2 as accepting suggestions without any modifications, but Table 3 reports that its writers modified an average of 17% of accepted suggestions. The absence of a network connection does not make that measured rate zero.

  • Q9 is described as measuring improvement rather than confidence

    Minor

    The Cluster 1 interpretation treats Q9 as an item about whether GAI improved the writing. Table 5 identifies Q9 as confidence in one's ability to write with the system's help, which is a different question.

2 of 2

Cites

Flagged first, then by citations

Research works printed but not linked: 13

  1. Allen L. K., 2016, {ENTER}ing the time series {SPACE}: Uncovering the writing process through keystroke analyses
  2. Bholowalia P., 2014, Ebk‐means: A clustering technique based on elbow method and k‐means in wsn, International Journal of Computer Applications, 105, 17
  3. Cobo G., 2011, EDM, 253
  4. Csanadi A., 2017, Collaborative and individual scientific reasoning of pre‐service teachers: New insights through epistemic network analysis (ENA)
  5. Dj M. Z., 2015, Using hypnoteaching strategy to improve students' writing ability, Dinamika Ilmu, 15, 185
  6. Liu Y., 2022, Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems (CHI ’22), New Orleans, LA, USA, 1
  7. Ma'ady M. N. P., 2022, 2022 Seventh ICIC, 1
  8. Senin P., 2008, Dynamic time warping algorithm review, 40
  9. Shaffer D. W., 2017, Quantitative ethnography
  10. Shen S., 2017, Clustering student sequential trajectories using dynamic time warping
  11. Shibani A., 2023, Visual representation of co‐authorship with GPT‐3: Studying human‐machine interaction for effective writing
  12. Tavenard R., 2020, Tslearn, a machine learning toolkit for time series data, Journal of Machine Learning Research, 21, 1
  13. Zhang M., 2021, Using keystroke analytics to understand cognitive processes during writing

Cited by

Sorted by citations

Authors
Yang, K., Cheng, Y., Zhao, L., Raković, M., Swiecki, Z., Gašević, D., & Chen, G.
Affiliations
Monash University (AU)
Corresponding author
Chen, Guanliang
ORCID
Venue
British Journal of Educational Technology
Publisher
Wiley
Published
2025-11-26
Volume
57
Issue
2
Pages
508-534
ISSN
  • 0007-1013 (ISSN-L)
  • 1467-8535
DOI
10.1111/bjet.70031
OpenAlex
W4416722686
Cited by
4
References
54
Citations by year
2026: 4
FWCI
16.84
Type
journal article
Language
en
Open access
Licence
Locations
British Journal of Educational Technology (publishedVersion)
Funders
  • Australian Research Council (DP220101209; DP240100069)
  • Jacobs Foundation (CELLA2CERES)
Topics
  • Social Sciences › Education › Writing and Handwriting Education
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
  • Physical Sciences › Materials Science › Materials Chemistry › Machine Learning in Materials Science
Keywords
human-AI interaction, knowledge transformation, cognitive presence, epistemic network analysis, dynamic time warping, time series clustering, AI dependency, intelligence enhancement

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-11.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-11.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review