The framework’s central insight rests on a distinction that cognitive load theory has refined over decades: not all mental effort is equal, and not all of it deserves to be automated. Intrinsic load refers to the inherent complexity of a task — the number of elements a writer must hold in mind simultaneously to make sense of an argument. Extraneous load, by contrast, is the friction that contributes nothing to understanding: hunting for a citation format, wrestling with a template, untangling a garbled bibliography. Germane processing is the deliberate allocation of working memory resources toward building schemas, the mental structures that make expertise possible. The paper’s argument is that AI tools are remarkably well suited to absorbing extraneous load and strikingly ill suited to absorbing germane processing, because germane processing is not a task that can be handed off at all. When a tool performs the work, the processing does not transfer to the writer; it simply does not occur.

This claim gains force from recent empirical findings on what happens when writers lean heavily on generative assistance. Studies documenting reduced neural engagement, weaker recall of one’s own writing, and a diminished sense of authorship among heavy AI users suggest that the costs of offloading are not merely pedagogical abstractions. They are measurable changes in how deeply a writer encodes their own ideas. Writing, as Janet Emig argued in 1977, is a distinct mode of learning rather than a record of thinking already completed. The act of struggling to render a messy idea in precise language is itself the mechanism by which the idea becomes clear to the writer. Remove the struggle, and the clarity never arrives.

The evidence on writing-to-learn adds an important nuance. A review of 48 school-based writing-to-learn programs by Bangert-Drowns and colleagues found a consistent positive effect on achievement, and a subsequent synthesis by Graham and colleagues reported comparable effects across 56 experiments in science, social studies, and mathematics. But what predicted a stronger effect is instructive for AI design: writing that included metacognitive prompts helped more, while assigning longer pieces helped less. The benefit of writing comes from the requirement that writers think about what they are doing, not from sheer output volume. This is precisely the condition that AI can silently remove. A student who prompts a chatbot to expand a paragraph has produced more words while doing less thinking — the exact inversion of what the writing-to-learn literature identifies as beneficial.

The nine-phase framework responds to this problem by refusing to treat research writing as a single undifferentiated activity. Instead, it maps the process from initial curiosity and inquiry through references and submission, and for each phase specifies three things: the dominant cognitive demand, the appropriate AI role ranging from minimal involvement to full collaboration, and design guidelines for how tools and interactions should behave. The logic is that the right AI role depends on the cognitive demand of the phase, not on the tool’s capabilities. In early phases, where the writer must generate questions and confront ambiguity, AI involvement should be minimal, because the struggle to formulate a researchable question is germane processing in its purest form. In later phases, where formatting conventions and citation management dominate, AI can operate at full collaboration, because those demands are largely extraneous and offloading them frees working memory for the thinking that matters.

This phase-aware approach is an application of responsive pedagogy to tool design. Responsive pedagogy holds that instruction should respond to the learner’s current understanding through recursive feedback dialog rather than proceeding on a fixed script. Transposed to AI-assisted writing, the principle becomes: the tool should respond to the cognitive demands of the task at hand, adapting its level of involvement as the writer moves through phases. A system that offers the same degree of assistance at every stage — drafting, organizing, synthesizing, polishing — is pedagogically blind, because it cannot distinguish between the friction that should be removed and the friction that is doing the teaching.

The framework’s scope is unusually broad. It addresses research writing from secondary school students in grades 9 through 12 developing their first researchable questions, through doctoral candidates building dissertations, to postdoctoral and professional researchers preparing manuscripts for peer review, across disciplines. This range matters because the temptation to offload struggle is not confined to novices. The paper opens with an anecdote about a high school student who submitted three AI-generated papers, frustrated by what the student perceived as a double standard — teachers use AI all the time, so why is it not acceptable for students? The anecdote illustrates how pervasive the temptation has become, and how the absence of clear norms leaves students improvising their own ethical boundaries. But the same dynamics apply to a seasoned researcher facing a deadline, for whom the pull toward delegation is arguably stronger because the procedural fluency to prompt effectively is already in place.

The paper’s own composition models its recommendations. Each section includes a human-AI collaboration disclaimer documenting exactly which cognitive work the researcher performed and which was delegated. The researcher conceived the paper’s focus, proposed the conceptual framework, and implemented it. Grammarly was used for editing, with sentence revision suggestions accepted or dismissed by the researcher. When Claude Pro was asked to review the initial single research question, it suggested splitting the question into two — one addressing which cognitive demands constitute extraneous load versus germane processing, the other addressing how human-in-the-loop collaboration should be designed across phases — and the researcher, after review, agreed. In building a definitions table, the researcher set the structure, identified terms, and selected sources; the AI drafted definitions grounded in those researcher-selected sources and flagged terms the researcher had overlooked. The researcher then verified each entry against the cited sources and reorganized the categories. This is the framework applied to itself: human curation and judgment retained, AI operating within boundaries the human establishes.

The transparency of these disclaimers points toward a broader implication for how AI-assisted scholarship should be evaluated. If the value of a paper lies partly in the germane processing its author performed, then disclosure of the division of cognitive labor becomes a form of methodological honesty, analogous to reporting data collection procedures. Readers can then judge for themselves whether the claims rest on understanding the author actually built or on synthesis the tool performed. The framework implies that assessment practices — in classrooms, in doctoral programs, in peer review — may need to shift from evaluating final products toward evaluating the process by which they were produced, since the product alone can no longer reveal whether the thinking occurred.

Equity considerations complicate the picture. The framework’s implications for responsive digital education include tool selection, assessment, and support for writers with cognitive differences. For some learners, certain demands that are extraneous for others may function differently, and rigid phase prescriptions could either burden or liberate them inappropriately. A student with dyslexia, for example, may experience transcription demands as genuinely extraneous friction that assistive tools should remove, while still needing to own the conceptual struggle of argument development. The framework’s emphasis on human judgment as the operational mechanism — rather than fixed rules about which tools may be used when — is what allows this flexibility. The line between offloadable and non-offloadable work is drawn by a person who understands the learner and the task, not by a policy document or a software default.

What emerges is a design philosophy rather than a prohibition. The paper does not argue that AI should be kept out of research writing; it argues that AI’s role should be calibrated to preserve the cognitive work that makes research meaningful. The productive struggle literature, drawing on work by Hiebert and Grouws and by Warshauer, positions struggle as a means of engaging in and developing critical thinking — not as an obstacle to learning but as its engine. Anne Lamott’s familiar advice to her overwhelmed brother facing an overdue report on birds — take it bird by bird — captures the framework’s practical spirit: large cognitive tasks become manageable when broken into steps, but the steps themselves must be climbed by the writer. AI can hold the ladder, hand over the field guide, and check the citations. What it cannot do is climb, because the climbing is the learning. The framework’s contribution is to specify, phase by phase, where the ladder belongs and where the writer’s feet must stay on the ground.

Subject of Research: A nine-phase human-in-the-loop framework for AI-assisted research writing based on cognitive load theory

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

Article References: Chauncey, S. A. (2026). Designing a human-in-the-loop AI iterative process for research writing: a framework for preserving productive struggle. Frontiers in Computer Science, 8, Article 1947835. https://doi.org/10.3389/fcomp.2026.1947835

Image Credits: AI Generated

DOI: 10.3389/fcomp.2026.1947835

Keywords: artificial intelligence, research writing, cognitive load theory, productive struggle, human-in-the-loop, generative AI, education, academic writing, cognitive offloading, responsive pedagogy, NotebookLM, authorship

The study, published in the journal AI & Society by Yasmin A. Carruthers and Johannes M. Heim, compared podcasts generated by NotebookLM with naturally produced human podcasts on the same topics, examining two hallmarks of human conversation: the timing of turn-taking and the way speakers adopt contextually appropriate conversational roles. The researchers argue that naturalness in dialogue requires far more than getting the timings right; it demands a contextually appropriate, addressee-oriented approach that responds to how a conversation actually unfolds.

Human turn-taking is a marvel of predictive coordination. Decades of research, beginning with the landmark work of Sacks, Schegloff and Jefferson in 1974, have established that conversation is governed by an organized system in which only one person typically speaks at a time, transitions are smooth, and gaps or overlaps are minimized. Listeners exploit morphosyntactic, semantic and prosodic cues to anticipate when a speaker’s turn will end, preparing their responses while the previous speaker is still talking. Average turn transitions fall in the range of roughly 200 to 250 milliseconds, far shorter than the approximately 600 milliseconds humans need to plan and produce a spoken response. This means that human interlocutors must be predicting, not merely reacting, and traces of this ability appear remarkably early in development, with infants showing sensitivity to conversational turn structure from birth.

Yet the “no-gap, no-overlap” principle is not absolute. Subsequent research has shown that overlaps can account for a substantial share of speaker transitions, and that brief overlaps often signal engagement rather than disruption. Backchannels, the brief vocal acknowledgments such as “mm-hm” or “right” that listeners produce without claiming the floor, add a dynamic and expressive quality to human dialogue. In podcast settings, hosts and guests also follow genre-specific social scripts: hosts introduce topics and then step back into a novice role to give guests a platform, while guests predominantly adopt expert stances. These role expectations, rooted in the broadcasting heritage of podcasting, make conversational behavior highly predictable despite frequent shifts between expert and novice positions.

Against this backdrop, the researchers annotated approximately ten-minute excerpts from two human podcasts and two NotebookLM-generated podcasts built from their transcripts, using the acoustic analysis software Praat. They coded speaker contributions, role shifts, turn durations, overlaps and transitions, achieving very high inter-annotator agreement. The human podcasts came from a series on productivity, featuring a recurring host and invited guests, while the AI versions were generated by NotebookLM using Gemini-based script writing and Google’s own text-to-speech voices, with no additional training data supplied.

On the surface, the AI podcasts behaved similarly to their human counterparts: questions structured the conversation, topics were expanded through follow-ups, and both hosts and guests provided feedback through backchannels. A statistical model confirmed that guests, whether human or artificial, produced significantly more expert contributions than hosts, in line with the conventions of the podcast genre. But the deeper measurements told a different story. Human turns were dramatically longer, averaging 22.3 seconds for hosts and 49.4 seconds for guests, compared with just 11.3 and 14.8 seconds for the AI agents. Human distributions of turn length were skewed toward long, developed contributions, while AI turns were short and symmetrically distributed.

The most striking divergence appeared in overlapping speech. Human guests overlapped for an average of more than four seconds and human hosts for over two seconds, whereas AI overlaps lasted mere milliseconds. AI backchannels rarely overlapped at all; instead, the current speaker paused, waited for the backchannel to be uttered, and then resumed, producing a mechanical politeness that no human interlocutor would display. The researchers interpret human overlaps as signals of engagement and emotional alignment, made possible by the fact that human listeners continuously predict where turns will end. NotebookLM, by contrast, sidesteps the problem of endpoint prediction entirely by pre-planning the entire dialogue script before converting it to speech, strictly enforcing the avoidance of gaps and overlaps in a way that humans never do.

The study also tested a newer NotebookLM feature that allows a human to join the AI conversation live, and the results were revealing. Every human intervention disrupted the pre-planned script: transitions slowed dramatically, overlaps vanished, audio glitches caused sudden voice switches, and the agents could only acknowledge the human’s contribution after long, silence-based delays before returning to their prepared narrative. Because spontaneous intervention requires resource-intensive, real-time context integration that pre-planning was designed to avoid, the hybrid dialogues lost the fluidity that made the fully scripted versions sound convincing. The authors conclude that pre-determining role behavior and turn-taking only goes so far, and that NotebookLM clearly hits its limits when responding to unpredictable input.

Taken together, the findings suggest that AI-generated dialogue can mimic individual features of natural conversation in isolation but fails to synthesize them into the dynamic negotiation of information and social relation that defines human talk. The AI agents spoke shorter, overlapped far less, and behaved more like two equally informed participants than like a host giving a guest a platform, breaking with genre conventions that listeners intuitively expect. Human conversation, the authors argue, is not optimized for seamless information exchange but for building common ground, an emergent and collaborative process shaped by mutual attention, verbal feedback and shared, growing context. Variable turn-taking timings and adaptive conversational role behavior remain unmatched human skills, and achieving genuine naturalness will require AI systems that can interpret social contexts and respond dynamically to the fluid negotiation of interlocutor relations, not merely imitate their surface elements.

Subject of Research: Turn-taking timings and conversational role behavior in AI-generated podcast dialogues compared with natural human conversation

Article Title: Not your average podcaster: turn-taking timings and conversational role behavior in AI-generated dialogues

Article References: Carruthers, Y. A., & Heim, J. M. (2026). Not your average podcaster: turn-taking timings and conversational role behavior in AI-generated dialogues. AI & SOCIETY. https://doi.org/10.1007/s00146-026-03365-3

Image Credits: AI Generated

DOI: 10.1007/s00146-026-03365-3

Keywords: AI-generated podcasts, NotebookLM, turn-taking, conversational roles, dialogue generation, backchannels, overlaps, knowledge asymmetry, common ground, pragmatics, human-machine interaction, speech technology