Designing a human-in-the-loop AI iterative process for research writing: a framework for preserving productive struggle

1 Introduction and problem statement

1.1 Human-AI collaboration disclaimer

A human-AI collaboration disclaimer statement is included in each section of this paper. The researcher conceived the paper’s focus and problem. The researcher proposed and implemented the conceptual framework. In this section, the researcher used Grammarly to support editing. Sentence revision suggestions were either accepted or dismissed to improve correctness, clarity, and readability.

1.2 Focus

Recently, I met with a student who had submitted three “plagiarized” papers for their English class. The student used AI for each. They explained that “the work was hard,” and that they “didn’t understand the assignments,” and that “they didn’t know what else to do.” To be clear, the student had access to support from the teacher, teaching assistants, and other staff who were available to support them during and after school, but they shared that they did not want the English class to take time away from their other classes. Visibly frustrated, they argued that there was “a double standard” regarding AI use, specifically, “teachers use AI all of the time, why is that OK?” AI introduces a real and pervasive temptation: asking the tool to take on the struggle so we do not have to.

We are engaged in a challenge that requires thoughtful attention. In my exchange with this student, we recognize this challenge as it pertains to critical thinking when responding to readings in a high school English class. AI tools designed to support writing risk undermining the cognitive work that makes written responses personal and meaningful. Recent research highlights the importance of productive struggle as a means of engaging in and developing critical thinking (Hiebert and Grouws, 2007; Warshauer, 2015). Writing offers an opportunity to engage in productive struggle, to bring messy ideas into focus, and to make the complex accessible to others. Emig (1977) argued that writing is a distinct mode of learning rather than a record of thinking already completed. The research supports this with a condition attached. Bangert-Drowns et al. (2004), reviewing 48 school-based writing-to-learn programs, found a consistent positive effect on achievement, and Graham et al. (2020) report a comparable effect across 56 experiments in science, social studies, and mathematics. What predicts a stronger effect is instructive. Writing that includes metacognitive prompts helps more. Assigning longer pieces helps less. Writing supports learning when it requires the writer to think about what they are doing, and AI can remove that condition. Anne Lamott (1995) shares a story from her youth about her brother being overwhelmed by a large, overdue school report on birds. Their father advised him to simply take it “Bird by bird, buddy,” (Lamott, 1995, p. 19), breaking down a huge task into small, manageable steps.

This paper focuses on how AI can serve as a thought partner in the research process. From a theoretical framework perspective, the researcher, in collaboration with AI, considers cognitive load theory and productive struggle. From a conceptual framework perspective, the paper proposes a nine-phase interaction design framework mapped to the sections of a research paper. The framework applies to research writing at every level at which it occurs: a secondary student developing a first researchable question, a doctoral candidate building a dissertation, and a postdoctoral or professional researcher preparing a manuscript for peer review.

The paper’s contribution belongs to responsive pedagogy, a framework in which instruction responds to the learner’s current understanding through recursive feedback dialog (Smith et al., 2016; Bicsák, 2024). The framework applies that logic to AI-assisted research writing: the right AI role depends on the cognitive demand of the research phase.

1.3 Research questions

1.3.1 Human-AI collaboration disclaimer

The researcher asked Claude Pro (Anthropic) to review the initial single research question, RQ 1: The critical design research question is not whether AI should reduce cognitive effort, but which effort, a distinction that requires human judgment, not smarter tools. In response, Claude Pro (Anthropic) suggested splitting the question into two research questions to address the theoretical and conceptual frameworks. After reviewing, the researcher agreed to split the question into RQ1 and RQ2 to specifically address the when and how of using AI as a collaborator.

1.3.2 Two research questions

This research asks whether human-AI collaboration can keep researchers doing the hard thinking that produces real insight, while letting AI handle the procedural work that does not. The two research questions ask which kinds of cognitive work belong on each side of that line and how to build the line into the collaboration itself.

The two research questions for this study are:

H4a: Social networks are positively associated with leadership self-efficacy.

RQ1: Which cognitive demands in the research writing process constitute extraneous load that AI could reduce, and which require germane processing that must remain the writer’s own?

RQ2: How can human-in-the-loop AI collaboration be designed across the phases of research writing so that this distinction is operationalized through human judgment?

1.4 Definitions and terms

1.4.1 Human-AI collaboration disclaimer

The researcher set up the table structure, identified the terms, and selected relevant sources from the paper’s reference list to support the definitions. Claude Pro (Anthropic) drafted definitions grounded in those sources, proposed breaking the table into three categories, and pointed out terms the researcher had not initially included, but that were important as evidenced by their inclusion in this paper. The researcher read each entry against the cited sources, revised wording to ensure the sources were accurately referenced, and reorganized the categories. The collaboration follows the framework recommendations: human curation and judgment; AI operating within the researcher-selected sources.

1.4.2 Researcher and research writing

In this paper, a researcher is anyone who identifies a problem or topic worthy of study, develops research question(s), finds and evaluates sources, chooses an approach to interpreting the topic, and builds an evidence-based argument. This includes secondary students, college and graduate students, doctoral candidates, and professional scholars. Their experience levels differ, so the risks of using AI also differ. Sommers (1980) found that student writers treat revision as rewording at the level of individual words, while experienced writers treat it as finding the shape of an argument and imagining a reader. Bereiter and Scardamalia (1987) describe the same gap in cognitive terms: novices tell what they already know, while experienced writers work between content and rhetoric in ways that change what they understand. The distinction matters for AI use because word-level substitution is what AI editing tools do best, and it is the operation a developing writer is least equipped to evaluate. Novice writers may rely on AI too heavily and miss the struggle that builds skill. More experienced researchers may be better able to detect problems such as false information or arguments that seem stronger than the evidence supports. Research writing means writing that follows the main parts of a research paper, such as the introduction, literature review, framework, methods, findings, discussion, and conclusion. Swales (2004) established that the research article is a genre with recognizable rhetorical moves, and subsequent corpus work has found that a small set of those moves appears in nearly every article while the rest vary considerably by discipline. A secondary school paper and a journal article differ in depth and expectations, Whether they move through the same stages is an assumption of this framework rather than a finding, since the genre literature examines research articles and theses rather than school writing.

1.4.3 Human–AI collaboration design

Terms in Table 1 describe the design space in which humans and systems share intellectual work. They establish the structural vocabulary that the framework uses to specify where human judgment is preserved and where AI assistance operates.

Table 1. Human-AI collaboration design.

1.4.4 Foundational constructs

Constructs in Table 2 come from cognitive science, learning research, and writing theory. Together, they identify the kinds of cognitive work that produce understanding through writing, providing the theoretical basis for distinguishing intellectual effort that should be preserved from procedural effort that AI may legitimately reduce.

Table 2. Foundational constructs.

1.4.5 AI-specific failure modes

Terms in Table 3 name the characteristic ways AI integration into writing can go wrong. Each describes a pattern observed or argued for in recent literature and motivates the design of the proposed framework.

Table 3. AI-specific failure modes.

Artificial intelligence is evolving at an extraordinary pace, and the impact on teaching and learning is significant. Middle and secondary schools and institutions of higher education are navigating the challenge of integrating AI, or not, while preserving human thinking, judgment, and meaningful engagement, and reconsidering how learning is assessed. Shirky (2026) describes AI as a “wicked problem” for which there is no single, definitive solution. He argues that institutions must prepare students to use AI in a world that increasingly demands it while ensuring that its use does not come at the expense of intellectual effort and critical thinking.

The framework shared in this paper is intended as a starting point for structuring thinking about human-AI collaboration across the phases of the research process, while keeping human thinking, judgment, intellectual ownership, and meaningful engagement at the center. The goal is a framework that can be tested, refined, and adapted as AI technology evolves and as other researchers contribute new perspectives, evidence, and critical feedback.

Wagner (2025a, 2025b), a professor of digital media at Drexel University, shares his AI-infused writing process and personal disclosure. Wagner’s work, along with that of others, could inform future iterations of the framework or suggest the need for a new one.

Looking ahead, a refined framework could become an interactive guide for writers, suggesting AI tools based on the writing in progress and the phase it is in. For younger students, this could be designed as a research quest in which tools are introduced at key points in the process based on student contributions.

2 Literature review

2.1 Human–AI collaboration disclaimer

The references began with articles the author read and curated over the past 3 years. Articles were organized by category and by relevance to this research, which was conceived during the article collection process. Several tools supported the literature review: Claude Pro, Elicit Pro, Perplexity Pro, and Google Scholar were used to identify literature, a process that continued during the drafting of this paper. Section 2.3 (AI as a research assistant) was suggested as a subtopic during Claude Pro’s review of the draft literature review. Grammarly was used throughout to ensure clarity and grammatical accuracy. All references were consolidated into a single file, each DOI was resolved against the publisher record, and every in-text citation was verified against that file. 2. Because sources were identified across several tools over 3 years, entries moved between drafts without a final check against a single authoritative list, which produced citation errors identified in peer review. For this revision, the author built one reference file, resolved every DOI against the publisher record, and verified each in-text citation against that file before submission.

2.2 A note on sources

This paper draws on three kinds of source, and each does different work. Peer-reviewed studies establish what happens. The paper’s claims about effects on learning rest on experimental work (Bastani et al., 2025; Fan et al., 2025; Kestin et al., 2025; Wu et al., 2026; Zeitlhofer and Zumbach, 2026). Preprints report findings that have not been reviewed. Three studies central to this paper are preprints (Georgiou, 2025; Kosmyna et al., 2025; Liu et al., 2026); they are identified as such here and treated as provisional throughout. Practitioner writing names patterns. Six of the paper’s constructs come from practitioner sources: the Butler, Thinking Partner, and Sparring Partner modes from Kentz (2026); premature coherence and the cognitive offloading reframe from Underwood (2025a, 2025b); cognitive debt from Sarathy and Ganesan (2025); synthetic knowledge from Stricker (2025); the three-document system from Karelin (2025); and the Ethical Wheel of Prompting from Law (2026). Where the paper uses one of these names to claim an effect, the evidence for the effect is cited separately: premature coherence names a pattern documented by Kosmyna et al. (2025), Bastani et al. (2025), and Fan et al. (2025); cognitive debt names what Liu et al. (2026) found when AI assistance was withdrawn. The name comes from practice; the evidence that the pattern is real comes from controlled or peer-reviewed work. The audience for this paper crosses boundaries: researchers, K–12 and higher education practitioners, instructional designers, and school and district leaders are all making AI decisions in real time, and the literature each group reads differs. The literature on AI in education is also emerging faster than peer-reviewed publication cycles can accommodate. Researchers and educators are publishing on platforms such as Substack, arXiv (a moderated, not peer-reviewed, open-access repository for preprints), institutional blogs, and preprint servers to test ideas, gather feedback, and map the landscape before submitting peer-reviewed articles or longer works. A literature review restricted to peer-reviewed work would describe an AI landscape that no longer exists by the time it appeared, and the practitioners this paper addresses would find it unusable. The honest position is to use all three kinds of source and be explicit about which does what. Tables 1–3 label each citation by source type so the division of labor described here is checkable rather than asserted.

2.3 AI as a research assistant: existing collaboration frameworks

The question of how we should engage with AI, particularly for scholarly writing, has given rise to frameworks and models that offer guidance (Anthuvan et al., 2025; Hutson, 2025; Sutherland-Harris, 2024; Law, 2026), and we anticipate that others will emerge. The struggles teachers and students share about AI use in grades 6–12 and higher education settings clearly reveal that the evolution and use of AI are outpacing educational systems’ readiness to develop sound practices and policies (Nathan, 2026; McClain et al., 2026; Eisikovits and Burley, 2026; Nurenberg, 2026; Requarth, 2026).

The frameworks shared in this section describe AI as a tool for participation and prescribe structures for ethical or productive participation. What these frameworks do not address is the question and interaction design this paper tackles, specifically, which cognitive demands of research writing should be preserved as human work, and why. While these frameworks tell researchers how to manage AI participation, they reveal a gap in identifying where human effort serves a function that AI participation would displace. This paper argues that preserving cognitive work is key. AI integration should be designed around the cognitive load each phase of the work demands, rather than around AI participation protocols that vary without regard to the task at hand.

Anthuvan et al. (2025) propose the Scholarly HI–AI Loop, a stage-based model that maps human and AI contributions across ideation, drafting, and revision in academic writing. The framework names writers as “epistemic anchors” and treats AI as a participant whose outputs the writer evaluates and integrates (Anthuvan et al., 2025, p. 1). The Scholarly HI–AI Loop names stages but does not specify which stages require human cognitive effort and which can be redistributed.

Hutson (2025) offers a multidimensional account that organizes human–AI collaboration along axes of creative authorship, intellectual ownership, and procedural assistance. These axes identify the touchpoints relevant to designing a supportive framework, but Hutson stops short of building one. The proposed framework in this paper takes that next step.

Sutherland-Harris (2024) presents the SHARE framework, which gives scholars practical heuristics for selecting, handling, attributing, reviewing, and evaluating AI contributions to co-authored writing. SHARE is grounded in a sociocultural account that treats agency as a sociopolitical condition: who gets to claim authorship, under what norms, and with what accountability. The framework developed for this paper treats agency as a cognitive condition shaped by what the writer does or does not do during each stage of the writing process. The two framings are complementary rather than competing: SHARE investigates the social contract around AI use, while the framework in this paper considers individuals’ engagement in cognitive work.

Alyasin and Shah (2026) locate student agency in the social and institutional structures around tool use. This explains why agency varies across contexts but not why it matters. We fill that gap by considering the potential cognitive loss when structures do not address it, and what the writer loses cognitively when those structures fail. The framework in this paper supplies that missing layer.

Anthropic (2026) positions AI as a collaborator bounded by human judgment but does not specify where those bounds should fall or on what basis. Treating cognitive load as the criterion for defining boundaries for AI, human, and human-AI opportunities is key to collaborative work with AI, in which AI assists when load is extraneous; the human carries the work when load is germane to learning.

Parker and Becker (2026) offer a capability map that defines AI literacy across functional, critical, and rhetorical dimensions and maps these capabilities across stages of the research lifecycle, positioning “AI literacy as a form of scholarly practice rather than a set of technical skills.” (Parker and Becker, 2026, p. 1) Their framework specifies what literate AI use looks like at each research stage. The authors consider the question of which cognitive demands should constitute the writer’s irreducible intellectual work as a matter of researcher judgment rather than design specification, which this paper takes up.

2.4 Human–AI interaction design models

The models shared in this section provide vocabulary for specifying AI’s role but do not yet indicate which tasks are best suited to which forms of human engagement. The framework in this paper draws on the modes of Raisch and Fomina (2025) for the AI Role column’s structure, and on the practitioner vocabulary of Kentz (2026) for its terminology, to describeAI’s role at each phase, grounding the assignment in cognitive load theory rather than in design preference. The goal is to preserve the friction that produces productive cognitive struggle, the kind that supports learning, critical thinking, and decision-making, which is central to the framework shared in this paper.

Considering human–AI collaboration from an interaction design perspective, Raisch and Fomina (2025), writing in the Academy of Management Review describe three modes of collaboration: AI as a tool, AI as a teammate, and AI as a partner in interactive search. Through the lens of writing, each mode implies a different relationship between human and machine. The tool mode preserves full human authorship; the teammate mode shares cognitive load; the partner mode positions AI in dialog with the human’s developing thought. Kentz (2026) refines this with three interaction modes specific to AI as a thinking aid: (1) Butler (AI completes a defined task), (2) Thinking Partner (AI helps the human develop ideas), and (3) Sparring Partner (AI challenges the human’s thinking). The framework adopts Kentz’s terms because teachers and researchers find them more usable when deciding how to configure a tool, while the underlying distinction remains Raisch and Fomina’s. The Sparring Partner mode is particularly relevant to research writing because it positions AI as a source of friction rather than fluency. Section 4.4 operationalizes this distinction phase by phase.

Riedl and Weidmann (2025) study what determines whether human–AI collaboration produces synergy or substitution. They find that the human’s Theory of Mind, that is, the capacity to model what the AI does and does not understand, predicts collaborative success. This finding implies that effective AI use requires the human to maintain an accurate working model of the system’s limitations, not just its capabilities. Osmani (2025, 2026) extends the human-in-the-loop principle from software engineering into broader knowledge work, arguing that as AI outputs become more plausible, teams need stronger critical practices to avoid accepting outputs that sound right. Osmani’s argument matches the design problem this paper addresses: fluency masks the question of where human judgment was applied.

2.5 Writing-specific evidence on AI-assisted composition

Empirical work on how writers actually use AI is starting to accumulate. Nguyen et al. (2024) document distinct patterns of AI use in academic writing and find that those patterns correlate with different writing outcomes. Writers who use AI as a sounding board produce work that differs from writers who use it as a drafter. Jelson et al. (2026) studied how students use ChatGPT for essays and found similar patterns of variation, with use ranging from full delegation to bounded scaffolding. Guo et al. (2025) examine creative writers’ integration of AI and observe that experienced writers tend to use AI for friction (testing, challenging, and generating alternatives) more than for drafting. Handley (2025) makes a related practitioner argument about how skilled writers selectively incorporate AI.

Potkalitsky (2025), writing in a Harvard Business Impact (2026) collection on AI in undergraduate teaching, argues that effective AI use requires students to develop what he calls possibility literacy, which “teaches students how to recognize when AI enhances versus constrains their thinking—and how to maintain conscious control over how these tools influence their learning process” (p. 3). The distinction specifies a cognitive mechanism that the literacy scholars (Wagner, 2025c; Nurenberg, 2026) and assessment scholars (Corbin et al., 2026; Perkins et al., 2025) describe in more general terms: the work of recognizing, in the moment, which AI uses preserve the thinking the writing was meant to develop and which uses displace it.

Possibility literacy, as Potkalitsky describes it, assumes a writer with the cognitive and metacognitive maturity to make the enhance-versus-constrain determination in the moment of use. The assumption fits the undergraduate writing classroom he describes, but it raises a developmental question that is not addressed: where and when do students develop the capacity Potkalitsky’s literacy requires? If the determination is first asked of college writers, those who arrive without prior practice are not positioned to meet it (see Ma et al., 2025; Underwood, 2025a). The framework proposed in this paper is designed to support that developmental arc, specifying, for each of the nine phases, what cognitive work AI can support and what work it must not displace, in a form applicable from high school through graduate and post-graduate research writing.

The timing of AI use within the writing process matters. Liu and Barshay (2025) show that AI feedback at the revision stage can be more productive than at the drafting stage, and that early use of AI on drafting tasks is associated with reduced engagement with the writer’s own developing ideas. Underwood (2025a) makes a related argument from a cognitive perspective, identifying drafting as the phase where AI most directly displaces the writer’s generative processing. The convergent finding is that what AI does is less important than when in the process AI does it.

A separate strand examines how the quality of interaction with AI shapes writing outcomes. Shibani et al. (2024) and Tu et al. (2024) find that critical interaction (challenging the AI, asking for alternatives, evaluating outputs) produces stronger writing than passive acceptance. Grobe (2026) argues for treating writing as inquiry rather than production and frames AI’s appropriate role as a thought partner in inquiry, not a content generator. Brack (2026) makes the same point in a practitioner’s terms: skilled writers know when not to use AI. Nathan (2026) reports student testimony about the conditions under which AI use displaces learning, providing first-person evidence that students themselves recognize the difference between productive and unproductive AI use.

A final concern in this literature is epistemic. Stricker (2025) introduces the term synthetic knowledge to describe the phenomenon in which AI produces content that has the surface features of scholarship (citations, organization, confident assertion) but lacks grounding in actual sources Stricker’s concern is about what AI produces. A related question is why it produces it. For research writing specifically, this risk is acute: an AI-generated literature review can be read as a synthesis while containing fabricated or misattributed claims that a reader cannot detect from the prose alone. The literature converges on the conclusion that critical evaluation is non-negotiable in AI-assisted writing, but this requirement itself imposes cognitive load on the writer.

2.6 NotebookLM as a source-grounded research system

A small but growing literature examines Google’s NotebookLM as a distinctive design point in AI-assisted research. Yeo et al. (2025) describe the system’s central design feature: AI responses are grounded in human-curated sources rather than in the model’s training data. This architectural choice converts the open-ended generation of standard LLMs into bounded analysis within a researcher-defined evidence base. Caplan (2024) discusses the implications for teaching, framing source-grounded AI as a tool that preserves the researcher’s evidentiary boundary. Panke (2025) provides an autoethnographic account of NotebookLM use in scholarly work and observes that the system’s audio Deep Dive feature transforms curated sources into a format that supports a different mode of engagement than reading. Shor et al. (2026) evaluate NotebookLM specifically as a tool for conducting literature reviews and find both strengths (bounded analysis, retrievable evidence) and limits (still requires human evaluation of source quality, and synthesis must still be researcher-led).

This literature treats NotebookLM as a productive case for examining how interaction design can encode the boundary between human and AI work. The system’s source-grounding does not, by itself, preserve germane cognitive load (a writer can still ask NotebookLM to summarize and accept the summary without engaging), but it constrains the AI’s operational boundary in a way that makes it inspectable. Section 4 of this paper develops NotebookLM as an extended exemplar at the literature review phase.

2.7 AI literacy, equity, and assessment

A parallel literature examines AI in education through the lenses of literacy, equity, and assessment. Wagner (2025c) argues that AI literacy is best understood as a form of critical thinking and that surface-level prompting instruction is insufficient. Potkalitsky (2025), Potkalitsky (2026a) develops a similar position with greater specificity for K–12 contexts. Nurenberg (2026) proposes a middle path between prohibition and full adoption, arguing for AI integration that protects the cognitive work English class is designed to develop. Ma et al. (2025) and Underwood (2025a) document a developmental threshold: novice writers and expert writers benefit from AI in different ways and are vulnerable to different harms. Novices risk displacing the formative struggle that builds writing capacity. Experts can use AI more productively but face epistemic risks (fabrication, premature coherence) that require expertise to detect.

Assessment in AI contexts is its own concern. Corbin et al. (2026) and Perkins et al. (2025) examine how AI changes what assessment can measure and propose redesigns that surface process rather than only product. Kentz (2026) connects this to interaction design: assessment in AI contexts must evaluate the writer’s collaborative practices, not only the resulting text. Eisikovits and Burley (2026) frame the broader risk in stronger terms, arguing that the central threat of AI in higher education is not cheating but the erosion of learning itself. The shared implication of this paper is that any framework for AI-assisted research writing must address not only what good practice looks like but also how to develop and evaluate it.

Potkalitsky (2026b) draws on Corbin et al. (2025) to distinguish discursive from structural responses to AI in education, where discursive responses “operate through what might be called linguistic commands” (Corbin et al., 2025, p. 1091) while structural changes “create assessment environments where the desired behavior emerges naturally from the assessment design” (Corbin et al., 2025, p. 1093).

2.8 AI as cognitive scaffold

Law (2026) introduces a complication that the AI as a cognitive scaffold framework must address. Her autoethnographic analysis of 1,560 ChatGPT conversations argues that for neurodivergent writers, AI can function as a cognitive scaffold that sustains intellectual work rather than displacing it. Her data show that AI use is clustered around brainstorming and editing, the cognitively dense phases, and that her interactions became shorter and more iterative over time, suggesting growing facility rather than dependence. Law’s Rhetorical Prompting Method (keeping planning, drafting, and revising visible within a single thread) and her Ethical Wheel of Prompting (evaluating output for usefulness, relevance, accuracy, and harmlessness) supply practitioner-designed structures that work across multiple writing phases. Panke (2025) reaches a parallel conclusion from a different starting point, describing how AI can support researchers whose cognitive resources are constrained, including attention, executive function, and language.

This work does not invalidate the phase-by-phase analysis proposed by the AI as a cognitive scaffold framework. The distinction between extraneous and germane load must be applied at both the writer and phase levels. For a neurodivergent researcher, the executive-function demands of organizing a discussion section may constitute an extraneous load that AI can legitimately reduce, freeing cognitive resources for the interpretive work that constitutes the germane load for that writer. The design question remains the same (which effort?), but the answer can differ across writers. This paper treats Law’s contribution not as a refinement. The logic of the AI as a cognitive scaffold framework is a structure for asking the question, not a fixed prescription with the answer.

3 Theoretical foundations

3.1 Human–AI collaboration disclaimer

The researcher selected theoretical lenses; all references collected over time and selected for this paper are tied to definitions and the phased conceptual framework, which draws on the theoretical foundations.

3.2 Cognitive load theory as design framework

Cognitive load theory (CLT) supplies the central theoretical lens of this paper. Sweller et al. (2019), in a review of two decades of empirical work, distinguish intrinsic load from extraneous load. Intrinsic load reflects the inherent complexity of a task in relation to the learner’s prior knowledge. Extraneous load is imposed by the way the task is presented and includes any demands that do not contribute to learning. Earlier formulations of CLT also proposed germane load as a third, independent source, defined as the effortful processing through which the learner builds, refines, and integrates mental schemas, a redundancy Kalyuga (2011) argued against. Sweller (2023) reports that no instructional consequences of germane cognitive load were identified and that it is no longer treated as an independent source. Germane processes are instead the working memory resources used to handle intrinsic load rather than extraneous load and this paper uses that current formulation. The instructional design implication is well established: effective design minimizes extraneous load, manages intrinsic load relative to expertise, and directs germane processing toward the intrinsic load that matters.

Flower and Hayes explain why writing is demanding. Bereiter and Scardamalia (1987) explain why some writing changes the writer, and some does not. In knowledge-telling, the writer reports what they already knew. In knowledge-transforming, the writer works between what to say and how to say it for this reader, and moving between the two changes what the writer understands. Kellogg (2008) adds a condition: this happens only when the writer has enough working memory available to hold the content, the text, and the reader’s likely reading at the same time. Two consequences follow for this framework. The discussion phase demands the most because it requires all three at once. And what counts as germane processing shifts as a writer develops, which is why the framework applies from secondary school through professional research without prescribing the same answer at each level. This also sharpens premature coherence. Fluent AI output has the surface features of knowledge-transforming writing with no transforming behind it.

Kalyuga (2007) develops the expertise reversal effect, which adds a critical refinement. Instructional strategies that reduce load for novices can suppress germane processing for experts, who need the effortful processing to extend their schemas. The implication for AI use in writing is direct. A tool that lowers extraneous load may help a novice writer engage more productively, while the same tool used by an expert writer may bypass the schema construction that produces new understanding. The CLT framework does not yield a single answer about whether to use AI. It requires that the question be asked anew for each writer, task, and stage.

Flower and Hayes (1981) supply the writing-specific foundation. Their cognitive process theory describes writing as a recursive activity in which planning, translating, and reviewing draw on limited working memory simultaneously. The implication of combining the cognitive process theory of Flower and Hayes with CLT is that writing is unusually demanding on working memory because it requires idea generation, organization, language production, and audience modeling to occur simultaneously. Any tool that reduces the demand on working memory may help or harm, depending on which component of load is reduced. This recursive activity is highlighted in the AI frameworks shared in Chauncey and McKenna (2023a), Chauncey and McKenna (2023b), Chauncey and McKenna (2024a), Chauncey and McKenna (2024b) and Chauncey (2026). These frameworks address demands on working memory while supporting critical thinking, evaluation, and integrity.

Underwood (2025b) explicitly reframes cognitive offloading through CLT. His argument is that offloading is not categorically helpful or harmful. The evaluative question is which load is being offloaded. Offloading extraneous load is benign and often beneficial. Germane processing cannot be offloaded in the same way: when a tool performs the work, the processing does not transfer to the writer, it does not occur, and the understanding that processing would have built is what is lost.

This reframing converts the open-ended question of whether to use AI into a targeted question: which kind of load does this AI use reduce? The framework shared in this paper operationalizes that question across the phases of research writing.

3.3 Productive struggle and generative processing

Hiebert and Grouws (2007) introduce the construct of productive struggle in mathematics education research, defining it as the effortful, generative engagement that occurs when a learner works with content just beyond current mastery but within reach through sustained effort, in which cognitive difficulty is not an obstacle to learning but a precondition for it. Warshauer (2015) examines the specific moves teachers make to maintain productive struggle without letting it tip into unproductive frustration. When all difficulty is removed from a learning task, understanding is compromised.

Zeitlhofer and Zumbach (2026) adopt the same formulation and the same term defined in Section 3.2. They found that a two-day delayed posttest favored the more effortful learning conditions. Constantly high task complexity and interleaved practice each produced faster completion and fewer trials than their less demanding counterparts, and the least effortful condition performed worst on both measures. Drawing on cost–benefit accounts of mental effort (Shenhav et al., 2017; Skulmowski and Xu, 2022), they argue that learners invest germane resources only when they cannot complete the task without them. That argument carries a direct implication for AI-assisted writing. If AI renders a phase solvable without effortful processing, the condition that prompts the writer to invest germane resources is gone. The writer need not choose to disengage. The task no longer requires engagement.

Underwood (2025a) identifies cognitive effort through which a writer discovers what they think by working to articulate it. Underwood warns that AI drafting tools can produce premature coherence, organized prose that satisfies the surface demands of a writing task while bypassing the meaning construction the writer would otherwise perform.

3.4 Cognitive offloading and failure modes

Cognitive offloading research is converging on a consistent finding. Kosmyna et al. (2025) found reductions in neural engagement, recall, and reported authorship among LLM-assisted writers. Georgiou (2025) found reduced cognitive engagement among ChatGPT users. Liu et al. (2026) found that AI assistance reduces persistence and impairs independent performance once the assistance is removed. Baldeo (2026) documented behavioral evidence of executive function attenuation in high-use adults, which suggests the effects extend beyond the task at hand.

What matters for design is the last of these. The harm appears after the assistance stops, which means it cannot be detected while the writing is going well. Hardman (2025) and Liu and Barshay (2025) draw the instructional conclusion: the design question is when within a task AI should be available, not whether.

3.5 The challenge of evaluating AI output

At every phase of the framework, the researcher remains the evaluator of what AI produces. This evaluation is cognitive work the writer must carry out regardless of how the framework structures AI’s role, and it draws on the same finite working memory resources that germane processing requires. Applying the criterion in Section 4.2, the demand is extraneous. Verifying that a citation exists or that a summary reflects its source has low element interactivity, and it originates in the tool rather than in the research task. Both estimates point the same way, and the removal test settles it: take the verification away and the writer’s understanding of the research question is unchanged. Where source evaluation is itself a learning goal of the phase, removal would change what the writer knows at the end of it, so the demand is intrinsic and the writer’s engagement with it is germane processing. Where AI takes the lead, the evaluation demand intensifies rather than diminishes, because the writer must judge output produced. Mollick (2025) and Potkalitsky (2025) argue that the jagged frontier requires sustained critical evaluation of every output rather than periodic spot checks. The skill the writer must exercise is not whether to trust AI in general but whether to trust this particular output, on this particular task, at this particular moment.

Synthetic knowledge (Stricker, 2025; Section 2.5, Table 3) is to detect than the jagged frontier suggests, because fluent output does not signal its own unreliability. A fabricated citation reads exactly like a real one. A misattributed claim reads exactly like an accurate one. The evaluation burden therefore extends beyond noticing when AI struggles to noticing when AI produces confident output that should not be trusted. Kalai et al. (2025) explain why this happens. The way these systems are trained and scored rewards a confident answer over an admission of uncertainty, because a guess sometimes earns credit and “I don’t know” never does. Confidence is therefore built in, and it tells the reader nothing about whether the answer is right. This is why spot-checking does not work. A writer cannot sample a few claims, find them sound, and infer the rest are sound, because the tool sounds equally sure either way. This evaluation burden is extraneous load. Checking whether a citation exists or whether a summary reflects its source has low element interactivity: each check can be completed without reference to the others. The demand originates in the tool rather than in the research task, and removing it would leave the writer’s understanding of the research question unchanged. The burden is a cost AI introduces rather than work that builds the writer’s schemas. The exception is the phase where source evaluation is itself a learning goal. At Phase 3, judging whether a source is sound and whether its claims bear on the argument is intrinsic to research writing, and the writer’s engagement with that judgment is germane processing. Verifying that the tool reported the source accurately is a separate demand and remains extraneous.

The classification carries a consequence the framework must account for. AI adds extraneous load at every phase where it operates, including phases where it also reduces extraneous load. The net effect is not uniformly positive, and it varies with the writer’s prior knowledge. Section 5.3 develops this point.

3.6 What remains irreducibly human

Thomson and Barr (2025) argue that the drive to inquire, the recognition of a question worth asking, emerges from intrinsic necessity and inner complexity that AI cannot replicate. Spencer (2025) makes a related argument about creativity, locating it in the friction between perspectives and the iteration that produces unexpected synthesis. Suleyman (2025) approaches the question from a design perspective, framing AI’s appropriate goals as complementing rather than replacing human judgment. In summary, certain kinds of intellectual work are not optional features of research writing but constitutive of it, and AI integration that displaces this work changes what research writing produces.

4 Conceptual framework formulation

4.1 Human-AI collaboration disclaimer

The author developed the first draft of the framework. Claude Pro was used to review and expand certain sections. After completing the literature review, the author revisited the framework to identify gaps, update cognitive considerations, integrate relevant references, and develop the text for the AI Role, Design, and Tools Guidelines columns. The author’s and AI’s work was iterative, and the final table is a product of that collaboration.

4.2 The framework

The nine phases described in the proposed framework are not tool-focused; rather, they specify which cognitive demands must be preserved at which phases as humans work, and why. The framework draws on cognitive load theory to distinguish extraneous load from intrinsic load and the germane processing directed to it; and on productive struggle and generative processing to account for what is lost when a tool performs the work instead, as outlined in cognitive offloading research.

The criterion for classifying each phase is element interactivity. Following Chen et al. (2023), element interactivity determines task complexity, and Sweller (2023) notes that element interactivity cannot be measured exactly but can be estimated. Each phase is classified by three questions: (1) which elements must be held in working memory at the same time; (2) whether the demand comes from the task’s structure or from the tools and formats through which the work is carried out; and (3) whether removing the demand would change what the writer knows at the end of the phase. The first two questions estimate the demand; the third decides it. A demand whose removal leaves the writer’s understanding unchanged is extraneous. A demand whose removal changes it is intrinsic, and the resources the writer allocates to it are germane processing. Where the first two questions and the removal test point in different directions, the removal test governs, because extraneous load is defined by its failure to contribute to learning rather than by where it originates. Table 4 applies these three questions to each of the nine phases and names the interacting elements on which each classification rests.

Table 4. 9-Phase conceptual framework interaction design model.

Recent research reveals that the question posed and addressed in this paper’s framework is being raised across the field and in varied settings. Austin et al. (2026) reach it from disciplinary literacy and metacognition, “metacognitive checkpoints,” (Austin et al., 2026, p. 1)asking when a student should let AI into a task and when they should do the work themselves. Their research and this paper seek to protect the cognitive work that builds skill and being thoughtful about when to assign or share the work with AI. What this paper’s framework adds is a phase-level approach to questioning AI use in research writing, in that it specifies, across nine phases, which loads of AI should reduce and which it should leave with the writer.

4.3 Practitioner connection

The Adirondack Paradox is one of three practitioner cases drawn from Chauncey (2026) in which NotebookLM was the primary AI tool usedin the human-in-the-loop workflow to synthesize regional demographics, financial data, population data, and the role of schools in communities to support rural school funding advocacy. This is where curiosity, stating a purpose, developing a question, and a preliminary researchable claim (Phase 1) require germane processing, and where AI cannot tackle this essential, human-identified work (Thomson and Barr, 2025).

Phase 3 of the proposed framework used AI to support divergent thinking, while the researcher retained convergent judgment in decision-making and boundary-setting. The current paper considers researchers’ efforts in terms of cognitive demand, setting the boundary between human and AI work by what the writer’s understanding requires rather than by what the tool can perform. The same project shows both ends of the framework. This is one researcher’s recognition of one question. It shows the germane inquiry work preceded the tool; it does not prove that AI could never assist inquiry. The Adirondack Paradox is an example and is not intended to be a controlled study; it covers only one phase, not all nine. Its purpose is to show what the literature review row of Table 4 (Phase 3) looks like in a real workflow: where a source-grounded AI system reduces the extraneous load of working with a large body of sources, and where it leaves the writer’s germane processing in place or risks displacing it.

The research question for the Adirondack Paradox was established before selecting the AI tool; the purpose and question came first, before NotebookLM and the sources. The researcher decided what documents would be considered as sources. Responses were generated from the researcher-curated corpus rather than from model training data (Yeo et al., 2025). NotebookLM supported the generation of multimodal output summaries, briefings, and video explainers (Panke, 2025).

NotebookLM operates on the boundary the Phase 3 row of Table 4 now names. Locating relevant passages across a large body of sources is extraneous: each retrieval stands alone, the demand comes from how the material is stored rather than from the research question, and removing it produces the same passages found faster. The tool did that work. Judging which sources enter the corpus, and what their claims establish about the question, is intrinsic: accepting one source changes what the others contribute to the argument. That work stayed with the researcher.

Choosing and loading sources is deliberately effortful, and the effort is the point. The researcher identified and selected every source, read each one before adding it, and did not use AI to generate candidate sources for this case. The retrieval tool that followed operated within a boundary the researcher had drawn.

Source discovery features shift where that boundary falls without removing it. A researcher who reviews AI-proposed candidates still decides what enters the corpus, but searching, encountering, and setting aside sources are part of judging what counts as relevant evidence at this phase. Delegating the search narrows germane processing to approving a list someone else assembled. The framework’s Phase 3 guidance therefore concerns how the corpus is built, not only who approves it.

Multimodal output resets on the researcher actively selecting, organizing, and integrating the sources deemed relevant(Fiorella and Mayer, 2016). The video and audio explainers convert those curated sources into formats that invite viewing and listening rather than reading. Because the researcher read the sources before adding them, the outputs were checked against existing source knowledge rather than accepted on their fluency. This reduces the risk of premature coherence (Underwood, 2025a) without removing it, since a summary can still misrepresent a source the reader recalls only in outline.

Source-grounding constrains where the tool operates. It does not determine how the researcher uses what the tool returns. A researcher can ask NotebookLM to summarize a corpus and accept the summary without working through what it means for the question. In that case, the synthesis does not transfer to the tool. It does not happen. The framework’s guidance for Phase 3 therefore concerns the researcher’s own engagement as much as tool selection.

4.4 A 9-phase conceptual framework for human-AI collaboration in research writing: an interaction design model

The paper’s core contribution is an iterative framework for human-AI collaboration in research writing. While this framework focuses on research writing, it can be adapted for other writing genres. Specifically, this paper proposes a model for human–AI collaboration and the roles involved. The framework’s steps are iterative, and the principles apply at each phase regardless of the order. The framework aligns with the research writing process, making it immediately recognizable to any researcher, and seeks to answer the question: where and how should AI be involved in each section of my paper?

Figure 1 presents the framework at a glance. Each bar corresponds to one phase, mapped to its section of a research paper; bar length indicates the relative AI involvement specified by the framework, with the lowest where intrinsic load is highest and the greatest where extraneous load dominates. Table 4, which follows, provides the detail behind the figure. For each phase, the table specifies what happens in the phase, the dominant cognitive demand, the appropriate AI role, and design and tool guidelines.

Table 4 provides the detail behind the figure. For each phase, the table specifies what happens in the phase, the dominant cognitive demand, the appropriate AI role, and design and tool guidelines. Each phase in the writing process is considered as follows: (1) what happens in the phase, (2) cognitive demand, (3) AI role, and (4) design and tool guidelines. Each phase is defined by its primary cognitive demand, whether to be preserved, distributed, or offloaded. This framework maps directly to the sections of a high-quality research paper as follows: Phase 1: Curiosity and inquiry; Phase 2: Introduction and problem statement; Phase 3: Literature review; Phase 4: Theoretical or conceptual framework; Phase 5: Methodology; Phase 6: Findings or results; Phase 7: Discussion; Phase 8: Conclusion and Implications; Phase 9: References, formatting, and submission.

5 Implications for responsive digital education design

5.1 Human-AI collaboration disclaimer

In this section, the ideas, connections to points made throughout this paper, summaries, and the identification of challenges in using the framework are the researcher’s. AI was used to review the author’s writing, to ask questions for clarification, and to offer suggestions for wording and grammatical correctness.

5.2 What the evidence supports

The implications in this section translate the framework into responsive pedagogical practice: decisions about tools, assessment, equity, and cognitive support that respond to the phase a writer is in rather than to the availability of AI.

The president of the University of Chicago asked his staff to be “skeptical, ethical, and ambitious” about AI (Alivisatos, 2026, para 3). This framework aims at the same three things. We are early in this work. In some schools and workplaces, AI is a thought partner. In others it runs on its own as an agent. Practice has run ahead of the evidence Some claims about AI and learning have stronger support than others. For example: “AI makes students worse thinkers over time.” The worry is reasonable, but no long-term study confirms it (Gerlich, 2025; Kosmyna et al., 2025). Wu et al. (2026), pooling 35 experiments with 4,193 participants, found the opposite in the short run: students using ChatGPT outperformed those who did not (g = 0.670).

A narrower claim has better support: “AI can do the practice for you, so you stop building the skill, and you feel fine until a test.” Bastani et al. (2025) gave high school students access to GPT-4 during math practice. They did better with it and worse than the control group once it was taken away. Fan et al. (2025) found students using ChatGPT produced better work on the assisted task without learning more. Liu et al. (2026) found AI assistance reduced persistence and hurt independent performance afterward.

What separates these results is not whether AI was present. It is what the AI was set up to do. Bastani et al. tested two versions. Unrestricted GPT-4 caused the drop. A tutor version with guardrails, which prompted students rather than answering for them, did not. Kestin et al. (2025) built a tutor on the same principle for a college physics course and found students learned more than in an active-learning class, and in less time. Wu et al. (2026) found the same pattern in their moderator analysis: how the tool was used predicted the result.

This is what the framework builds on. AI that does the thinking displaces the skill. AI that prompts the thinking supports it. The question at every phase is which one the tool is doing.

5.3 Aligning AI tools to the writing phases

When you select an AI tool, ask which load the phase requires. A tool can be considered a fit when it reduces extraneous load and leaves the germane work with the researcher. That work is the productive struggle writing depends on. When AI use removes that struggle, the writer and the reader can both end up with a false sense of security. The writer sees finished prose and believes the thinking is theirs, and it is complete. The reader sees the referenced prose and believes it is a complete, accurate, and relevant response to the research purpose and the research questions posed. Neither the researcher nor the reader recognizes there are gaps. A single tool can fit one phase and harm the next. The same tool used well at Phase 9 can erode the writer’s work at Phase 1. So the fit depends on the phase and the researcher’s judgment for a specific research project.

Here the approach Alivisatos (2026) describes becomes practical. Phase-aware selection is “ambitious” where AI reduces extraneous load. It “stays skeptical” when the load falls on the researcher. The ambitious part is harder than it sounds. Deciding which demand is extraneous and which is intrinsic is an estimate rather than a measurement. That is true throughout cognitive load theory; Sweller (2023) notes that element interactivity cannot be measured exactly but can be estimated. The estimate is documented rather than asserted: the three questions in Section 4.2 record how it was reached, which lets another researcher dispute a classification by naming which question was answered wrongly. It is also a moving target. The effort a writer spends in one phase builds knowledge that changes what counts as intrinsic in the next. Work that was intrinsic on a first pass, the effortful struggle that builds understanding, can become routine once that understanding exists. The writer who returns to a phase is not the writer who first entered it. So the line between extraneous and intrinsic moves as the work proceeds, and the framework asks the writer to redraw it rather than apply it once. The line shifts with the task, the phase, and the writer’s expertise, and two researchers working from the same framework may draw it differently. That variation is a real limitation at this stage. The framework provides a structure for asking which effort matters;, and Section 4.2 sets out the questions for answering it, but the answers have not been tested against writing outcomes. The framework phases are not a checklist, and phase order is iterative.

Identifying the line between supportive and overcompensating is not trivial and requires an understanding of AI LLMs and the ability to evaluate every step of the way. Novice and expert writers benefit from AI differently and face different harms (Ma et al., 2025; Underwood, 2025a). Novices risk losing the formative struggle that builds writing capacity. Experts face fabrication and premature coherence they have the experience to catch. Access to tools is not a scarce resource. The scarce resource is instruction, that is, learning to evaluate AI output and to use it phase by phase. Without that instruction, more tool access can widen the gap rather than close it, because unguided offloading does the most harm to the writers who most need the struggle it removes. The verification burden AI creates is extraneous, as Section 3.5 sets out, and it falls hardest on novices, who have the least prior knowledge against which to judge whether an output is accurate. The framework’s implication for novices therefore concerns tool architecture rather than reduced AI involvement. Stated as a testable design hypothesis: holding phase and level of AI involvement constant, novice writers using source-grounded tools, in which outputs are constrained to a human-curated corpus and traceable to it, will carry less verification burden and sustain more germane processing than novice writers using open-ended generative tools.

5.4 Applying the framework across writers

For a student with executive-function challenges, the work of organizing a discussion section may be an extraneous load that AI can reduce, freeing effort for the interpretive work that is germane for that writer (Law, 2026; Panke, 2025). The framework’s question stays the same. Which effort matters here? The answer can differ from one writer to the next. The important question is, how can the framework be applied to support and serve a writer with varied challenges and cognitive load issues? In this case, the learner may be unable to engage with the framework’s phases without the support of an experienced human guide.

The same question applies to writers composing in a second language. Silva (1993), synthesizing 72 studies, found second-language composing more constrained and more difficult than first-language composing across planning, transcribing, and reviewing. Work that is automatic for a first-language writer, retrieving a word, building a grammatical sentence, consumes working memory that would otherwise go to meaning. AI language support at Phase 9 may therefore reduce a real load unrelated to research quality, or it may do composing work in a phase the framework treats as safe to delegate. Which one occurs depends on the individual writer.

6 Limitations and future directions

6.1 Human-AI collaboration disclaimer

This section did not rely on AI for content. The author revisited the points made throughout the paper and read emerging research. AI was used to review the author’s writing, ask clarifying questions, and offer suggestions on wording and grammatical correctness.

6.2 The proposed framework

As noted earlier, this Hypothesis and Theory paper addresses a design problem at the center of AI-assisted research writing. The framework maps how AI could distribute cognitive load across the nine phases, but it has not been tested against learner use and outcomes. The next step is an action research project to see whether it gives researchers enough direction in practice, and whether discretionary use preserves productive struggle and the germane processing that supports human curiosity, creativity, and insight. In short, whether it preserves the cognition we value. The framework addresses composition through submission. Responding to peer review is a distinct cognitive task, since someone else has done the evaluation and the writer’s work is to judge whether it is correct. Whether the framework extends to it is a question for a later version.

6.3 NotebookLM exemplar

The NotebookLM example in Section 4.3 demonstrates the framework at a single phase rather than across the full sequence. Each of the remaining phases needs its own exemplar (action research) before the framework can be evaluated for its efficacy in guiding research through the nine phases.

6.4 Neurodivergent scaffolding

The argument that AI scaffolding can reduce extraneous load for neurodivergent writers is relevant to the framework’s proposed reasoning. No direct evidence with this population has been tested. This claim requires future study with neurodivergent participants that considers how extraneous and germane processing would be assessed for learners with varied challenges.

6.5 Tool architecture

Section 5.3 states a design hypothesis: holding phase and level of AI involvement constant, novice writers using source-grounded tools will carry less verification burden and sustain more germane processing than novice writers using open-ended generative tools. This hypothesis is untested. No study in the literature reviewed here compares source-grounded and open-ended architectures against writing outcomes for novice writers, and the NotebookLM exemplar in Section 4.3 describes an architecture without measuring its effects. The hypothesis is a candidate for the action research proposed in Section 6.2. It is separable from the framework’s other claims: the phase mapping could hold while the architecture prediction fails, or the reverse.

6.6 What would falsify the framework

The framework makes predictions that could fail. Three are stated here with the observations that would refute them.

First, the framework predicts a rank order. Delegating Phase 7 to AI should degrade delayed recall, retention, and reported authorship more than delegating Phase 9. The prediction fails if delegation at the two phases produces equivalent outcomes, or if Phase 9 delegation produces the larger decrement.

Second, the framework predicts that phase-matched AI use produces better outcomes than uniform AI use at equivalent total involvement. The prediction fails if writers who distribute the same quantity of AI assistance evenly across phases perform as well as, or better than, writers who concentrate it where the framework directs.

Third, the framework predicts that independent researchers applying the derivation procedure in Section 4.2 to the same phase will converge above chance. The prediction fails if interrater agreement on phase classification does not exceed what chance would produce.

The intrinsic and extraneous categories earned their standing through four decades of randomized trials in which their instructional consequences were tested (Sweller, 2023). This framework has no such record. That is its central limitation, and the action research described in Section 6.2 is designed to begin addressing it.

6.7 Reference verification

This paper’s own reference errors, identified in peer review and corrected as described in the Section 2.1 disclaimer, are the evidence for the Phase 9 change in Table 4: AI completing reference work without a verification step is insufficient even when the underlying research judgment is sound.

6.8 Emerging research

This framework (Section 4.4) considers research writing and does not specify how cognitive load, germane processing, or productive struggle, as defined for research, would translate into classroom instruction across disciplines. Recent conceptual work suggests that this differentiation and understanding is important. Austin et al. (2026) propose discipline-specific metacognitive checkpoints for AI-mediated learning, arguing that what counts as productive friction differs between, for example, a history seminar and a chemistry lab. This raises the question of whether the paper’s research framework might be revisited and redeveloped to better serve varied disciplinary contexts, offering an opportunity for future research. Both this paper and the Austin et al. (2026) paper share the same gap: they have not been tested to determine whether the cognition they aim to preserve actually produces the predicted outcomes.

6.9 Multilingual writers

The framework was developed with first-language writers in mind. The Phase 9 caveat in Table 4 names the problem without resolving it. Determining where AI language support reduces load and where it displaces composing would require study with multilingual writers, and it is the highest priority extension of the action research proposed in Section 6.2. The writers most likely to be told to use AI editing tools are the ones for whom the framework currently offers the least reliable guidance.

7 Conclusion

7.1 Human-AI collaboration disclaimer

This section did not rely on AI for content. The author revisited the points made throughout the paper and read emerging research. AI was used to review the author’s writing, ask clarifying questions, and offer suggestions on wording and grammatical correctness.

7.2 Questioning in field

As noted in the proposed framework, limitations, and future directions, this paper enters a field of study where questions regarding AI-human calibration are still in their infancy. Concerns regarding the impact of AI on human cognition, and the touchpoints where AI can be supportive or detrimental, are being raised across disciplines, but the field has not yet settled how to answer them. What this framework contributes is a structured, phase-level account of where AI should be used to reduce load and where it should not be used to preserve the human cognition that builds skill.

Work from cognitive load, disciplinary literacy, metacognition, and ethics is converging on the same problem from different directions (for example, Austin et al., 2026). One common point is that using AI in a way that preserves productive struggle helps. When AI removes germane cognitive load and struggle, it may thwart the work humans do best. As noted earlier, when AI is used to answer questions, it can provide answers but may give the impression of understanding, leaving the user unable to explain them when the AI is removed (Bastani et al., 2025; Fan et al., 2025). In short, tools that preserve productive struggle help; tools that remove it do not. The proposed framework seeks to provide the user with guidance across nine phases of the research process, with the goal of preserving germane processing and productive struggle. The framework needs testing with learners and in practice before it can be offered as a practical tool for researchers to use at various levels and across disciplines.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript. The author takes full responsibility for the use of generative AI in the preparation of this manuscript. Each section of the paper outlines how AI was used. The purpose of the paper is to consider how AI can be used in the nine steps of the research process as outlined in the framework and described in the paper.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Glenda H. Gay, The University of the West Indies, Cave Hill, Barbados

Reviewed by: Rohib Adrianto Sangia, University of Aberdeen, United Kingdom

Harald Eriksen, Oslo Metropolitan University, Norway