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

Christian, a research fellow with the UC Berkeley Center for Human-Compatible AI, was part of a team behind a landmark study that found that spending just ten minutes relying on an AI tool significantly impairs our ability to focus and persist with difficult tasks. A preliminary draft of the work made waves earlier this year, and the team presented the updated, peer-reviewed paper this week at the Conference on Language Modeling. The collaboration included scholars from Carnegie Mellon University, the Massachusetts Institute of Technology, the University of Oxford and UCLA, and its central result is deceptively simple: a brief exposure to AI assistance measurably degrades independent performance the moment that assistance is withdrawn. “It’s a striking finding, but in a way it supplies ammunition for a story that a lot of us kind of feel in our gut,” Christian said. “I experience it too.”

To test how quickly people’s focus and persistence plummeted, the researchers designed a series of randomized controlled trials and recruited 1,222 participants online. The first experiment involved 354 participants. One group was tasked with solving 15 basic fraction problems without any help, while another group had ChatGPT open on the side to use for assistance, or even to ask directly for the answer. The group with AI access started off more accurate than the group working unaided, which is exactly what one would expect from a capable assistant. Then, after 12 problems, the researchers took the AI tool away. Almost immediately, the people in that group stopped solving questions accurately. The advantage they had enjoyed while the tool was available evaporated as soon as they had to rely on their own reasoning.

A similar pattern emerged when the team ran the experiment on a second, larger group of 667 participants. Whereas the AI users got answers wrong or gave up entirely when the assistance was withdrawn, the group that had never used AI stuck with the task and were more successful at the end. The data also revealed a telling behavioral shift: there was a significant drop-off in the solve rate for the group that had access to the AI tool after the tool was removed, and a sharp increase in the rate at which those participants skipped questions. In other words, the problem was not merely that people performed worse without the machine. It was that they became less willing to try at all, abandoning problems that the unassisted group continued to attack.

The researchers then ran a third experiment with 201 participants using an SAT reading comprehension prompt, a task far removed from arithmetic. The results echoed those from the fraction problems: when the AI tool was removed, both persistence and accuracy dropped. This consistency across two very different domains suggests that the effect is not about math anxiety or a lack of practice with fractions. It points to something more general about how quickly the human mind adapts to the presence of an always-available answer machine, and how poorly that adaptation serves us when the machine is taken away. “You almost couldn’t ask for a clearer story,” Christian said.

The findings back up what anecdotes and surveys have increasingly suggested in recent years: an overreliance on AI for answers erodes our cognitive fitness. The mechanisms the researchers describe are familiar to anyone who has studied how habits form. When AI produces answers in seconds, we get used to that speed, and anything that takes longer begins to feel inefficient. The friction that once accompanied difficult thinking, the slow and sometimes frustrating process of wrestling a problem into submission, starts to feel like a defect in the workflow rather than an essential part of learning. Relying on AI also removes what the team called “productive struggle,” the effortful engagement with difficulty that can lead people to discover their research strengths and passions in the first place.

There is a deep irony here that Christian is quick to acknowledge. These systems were built to be helpful, and by almost any benchmark they are. They answer accurately, quickly and on demand. But helpfulness in the short term and helpfulness in the long term can pull in opposite directions. “We’ve built these systems to be helpful, but are they being helpful to us in a way that’s actually helping us?” Christian said. “We showed, ironically, that they are often helping us in ways that are kind of unhelpful.” A tool that hands over the answer saves time on this problem and quietly costs capacity on the next one, leaving the user dependent on the very assistance that made them feel more capable.

The study attracted wide attention as its findings matured. Outlets from WIRED to Psychology Today covered the research when early results were released in the spring, and The New York Times published a story last week about the paper and similar work on how chatbots may affect thinking and brain development. Notably, frontier AI companies themselves have acknowledged how offloading tasks to chatbots can reduce a person’s baseline understanding and mastery of a subject. That admission from within the industry gives the academic findings added weight, because it suggests the builders of these systems recognize that the default interaction pattern, in which the model simply supplies the answer, may not be the one that best serves human learning over time.

Christian argues that the design of AI systems could change in response. Instead of defaulting to quick answers, AI responses could default to being more instructive, behaving more like a tutor that guides a learner toward understanding than like an oracle that dispenses solutions. The distinction matters because a tutor’s goal is to make itself unnecessary, transferring skill to the student, while an answer machine’s goal is to minimize the time to a correct response, which can leave the underlying skill untouched or even atrophied. “We need to work toward a future in which human abilities are, as much as possible, augmented rather than supplanted,” Christian said. The ten-minute finding suggests that the window in which that choice gets made is startlingly short, and that the direction of the effect, augmentation or replacement, is set almost as soon as the habit of asking takes hold.

The stakes, Christian believes, extend well beyond classrooms and test questions, and they are especially acute in research and academia. He worries about the way overreliance on AI tools in science could jeopardize what makes for a good scientist in the first place: staying close to data, being immersed in the material, and working as part of a team. Something significant could be lost, he suggests, when the team of ten developing the ideas that lead to a breakthrough is whittled down to two or three researchers and a chatbot. The informal exchange of half-formed hypotheses, the patient debugging of a stubborn analysis, the shared struggle that builds both expertise and judgment, all of these are the substrate of discovery, and none of them can be outsourced without cost. “This is not a story about fractions and SAT problems,” Christian said. “This is something that is happening to human knowledge and human expertise from top to bottom and from grade-school students all the way to the leading experts in the world.”

Subject of Research: The effect of brief AI assistance on human persistence and independent problem-solving performance

Article Title: Using AI for just 10 minutes erodes your ability to persist at hard things

Article References: Using AI for just 10 minutes erodes your ability to persist at hard things. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: artificial intelligence, cognitive persistence, ChatGPT, randomized controlled trial, productive struggle, human-compatible AI, education, cognitive offloading, Conference on Language Modeling, problem-solving, academic research, UC Berkeley

The research team approached the problem not by trying to pin down the causes of declining recruitment, but by examining how the roles of junior developers and the process of building professional expertise are actually changing on the ground. They chose South Korea as a case study in which these shifts appear in an early and highly visible form. According to data cited in the paper, AI is involved in 51.8 percent of work-related activities in Korea, nearly twice the comparable figure for the United States, while conventional entry-level hiring has contracted sharply. Entry-level job postings at the fifteen largest U.S. technology companies fell by 25 percent between 2023 and 2024, and IT job postings in Korea declined by 43 percent over the same period, with positions targeting entry-level candidates accounting for just 4.4 percent of all postings. The researchers caution that these declines cannot be attributed solely to generative AI, since post-pandemic workforce adjustments and the economic slowdown have also played a role, but the structural transformation of junior work was unmistakable in their data.

At the heart of the study are semi-structured interviews, each lasting approximately fifty minutes, with fourteen participants: six senior developers currently working in the industry and eight junior participants preparing to enter software development careers. The senior participants were required to have at least six years of professional experience and to have entered the workforce before the emergence of generative AI, while the junior participants were undergraduates whose university years spanned both the pre-AI and post-AI eras. The team analyzed the transcripts using reflexive thematic analysis, a qualitative method in which researchers iteratively examine their own interpretive role while identifying recurring patterns across interviews. Four major themes emerged, and together they sketch a system in which the loss of junior-level work is inseparable from the loss of junior-level learning.

The first theme was the shift of junior-level tasks toward what the researchers describe as “senior plus AI” workflows. Senior developers increasingly use AI to directly handle basic implementation, debugging, and documentation that would previously have been assigned to junior staff, and in doing so they absorb the very tasks through which novices once gained hands-on experience. One startup founder interviewed for the study said the company had substantially reduced its junior workforce without any measurable impact on productivity, asking pointedly, “What makes a junior developer better than a KRW 100,000-per-month AI subscription?” The economic logic of that question is difficult to dispute in the short term, and it explains why individual firms have strong incentives to make the substitution. The study’s point is that the costs of this rational choice do not disappear; they are displaced onto a different group of people at a different point in time.

That displacement is captured in the second theme: what is disappearing is not simply work, but the “opportunity to fail.” The researchers frame this through the concept of “productive struggle,” the well-documented principle in learning science that when learners make mistakes and correct them on their own, they develop not only knowledge but also the metacognitive ability to recognize what they do not know and to detect errors in the results they produce. Junior participants reported that they could now achieve good results with far less effort than before, yet many described a troubling sense of “not knowing what I don’t know.” One participant who took two courses in the same field received the same grade in both, but said that one course left them with actual knowledge while the other left them only with “the ability to use AI.” The single difference between the courses was whether the assignments could be completed using AI. The researchers found that current educational assessment methods do not adequately capture this distinction, meaning grades can mask profound differences in what students have actually learned.

Senior developers voiced the same concern from the other side of the career ladder. One participant explained, “The reason you become a senior is that experience teaches you what not to do. Now juniors can no longer learn from bad examples.” This observation connects directly to the study’s theoretical grounding in “situated cognition,” the idea that the problems a person can perceive depend on the position they occupy. Senior developers and the organizations with the power to address the problem may find it difficult to fully appreciate the circumstances facing juniors, while juniors who directly experience the problem lack the power to change educational and hiring structures. The asymmetry is captured starkly in the words of another senior participant: “We have 20 years of accumulated experience, so we can judge whether AI outputs are right or wrong. The next generation will not be able to reach that position. So we are fine.”

The third and fourth themes underscore that these dynamics are structural rather than matters of individual choice, and that they will not correct themselves. When every one of a student’s peers uses generative AI, grading on a curve effectively becomes a mechanism that compels AI use, since opting out means accepting a competitive disadvantage. In the courses taken by the interview participants, AI use was rarely restricted, so the loss of failure-based learning was built into university classrooms just as surely as into workplaces. Meanwhile, because seniors and juniors perceive the same situation so differently, no natural feedback loop exists to trigger correction. The researchers also found that the problem did not originate with generative AI alone: senior developers themselves largely built their expertise through hands-on work and trial and error rather than through formal training systems. In this sense, generative AI has not eliminated a newly created development pathway; it has accelerated the erosion of a pathway that was never institutionally protected in the first place.

The study argues that the solution lies in deliberate institutional protection rather than individual effort, and it draws on examples from other high-risk fields where automation has already forced this reckoning. In aviation, the Federal Aviation Administration’s SAFO 13002 advises pilots to maintain opportunities for manual flying so that automation does not erode their manual flying skills, and the nuclear industry requires regular simulator retraining to keep operators’ response capabilities sharp. These are examples of industries deliberately preserving the experiences necessary for developing and maintaining expertise even when machines can perform the underlying tasks. The research team proposes analogous reforms across three areas of software engineering. At universities, courses in which AI cannot achieve the learning objectives on behalf of students should be designated as required, and “reducing AI dependence” should serve as a criterion for evaluating educational quality. In hiring, employers should evaluate not only candidates’ ability to use AI to produce results quickly, but also their ability to recognize gaps in their own knowledge and detect errors while working alongside AI. Within companies, learning opportunities should be deliberately created for entry-level developers, for example by assigning them small modifications to actual products and letting them experience the full process from commits and code review through deployment.

Professor Moon, who supervised the research, framed the findings as a warning about a hidden ledger. “Some of the productivity we are gaining today may effectively be borrowed in advance from the expertise of the next generation,” he said. “Because the cost is borne by a different group and at a different point in time from those making the decisions, it is structurally difficult to recognize. This study is significant because it empirically reveals that blind spot.” He added that, as a department that develops AI technologies itself, SNU bears a particular responsibility to ask what those technologies are doing to the development of the next generation, and that the team plans follow-up research focused on university education and on incorporating the findings into the design of the undergraduate curriculum. The paper’s first author, Ph.D. student Sumin Yu of SNU’s M.IN.D Lab, conducts empirical research on how generative AI affects human learning, decision-making, and social institutions, and plans to continue studying the interaction between AI technologies and society across a range of fields.

The significance of the work extends beyond software engineering. AIES, jointly organized by the Association for the Advancement of Artificial Intelligence and the Association for Computing Machinery, is regarded alongside ACM FAccT as one of the two leading venues for research on the ethical and societal impacts of AI, and the ninth edition of the conference will be held in Malmö, Sweden, from October 12 to 14. The research was supported by the National Research Foundation of Korea, the Institute of Information & Communications Technology Planning & Evaluation, and the BK21 FOUR Education and Research Program at Seoul National University. What the study ultimately documents is a paradox at the center of the AI productivity boom: the judgment that makes senior engineers valuable is produced by exactly the kind of slow, error-prone, hands-on work that AI now performs so cheaply. An industry that optimizes away that work may find, a decade from now, that it has optimized away its own future experts, and the bill for today’s efficiency gains will come due in the form of a generation that never learned what it does not know.

Subject of Research: How generative AI erodes the professional development pathway of junior software developers

Article Title: SNU professor Taesup Moon’s team reveals how generative AI is eroding the development pathway for software developers

Article References: SNU professor Taesup Moon’s team reveals how generative AI is eroding the development pathway for software developers. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: generative AI, software engineering, junior developers, skill development, productive struggle, situated cognition, AIES 2026, Seoul National University, entry-level hiring, AI ethics, developer education, automation