Generative artificial intelligence (GenAI) has undoubtedly changed how we engage with information, particularly for those involved in knowledge production, such as universities. Operating like an on-demand personal tutor or academic assistant, some of these tools allow rapid access to vast amounts of information and they save time, which is an attractive pursuit for any academic (Mabirizi et al. 2025). At the same time, GenAI differs from other technologies we have encountered because it can produce fluent, plausible academic artefacts at scale, thereby changing the relationship between text, knowledge and the knower (Selwyn 2024).

Recent studies have highlighted how GenAI can enhance research productivity and efficiency through academic writing and analysis (Mabirizi et al. 2025), support idea generation and refinement (Zhang & Ge 2025) and enable creative development and innovation through content creation (Arinushkina, Abramov & Mindzaeva 2025). Generative artificial intelligence has also been lauded for its ability to provide personalised learning and feedback (Pang & Wei 2025). However, while these effects can be construed as beneficial, higher education needs to step back and consider the implications for learning, judgment and certification. This cautiousness has given rise to studies that highlight, despite the perceived benefits, that GenAI may also erode social interactions (Hou et al. 2025) and negatively impact students’ metacognitive skills (Zhou, Teng & Al-Samarraie 2024).

When ethically used, meaning when outputs are critiqued, appropriated and repackaged, GenAI can become a cognitive extender, pushing researchers beyond their usual capabilities (Clark 2025). However, when large language models (LLMs) are used as a substitute for core cognitive operations, or when users lack strong foundational disciplinary knowledge, they begin to decouple the written word from the thinking mind, and, in the context of higher education, this threatens the historical social contract that once treated the essay as a reliable proxy for intellectual growth. In cases such as this, the university risks validating polished products without validating the uncomfortable cognitive journey that produced them, thereby moving closer to what could be described as a glorified digital diploma mill (Noble 2003).

Central to this article is Klein and Klein’s (2025:10) notion of the cognitive bypass, defined as the mechanism by which an artificial intelligence (AI) system – ‘designed for the frictionless delivery of plausible answers – enables learners to circumvent the effortful cognitive processes foundational to genuine learning and knowledge internalisation’. The fundamental concern is not that students use GenAI to assist them with their assignments, but that students may use GenAI to avoid the challenges necessary to build disciplinary understanding, epistemic judgement and the capacity to defend one’s own claims.

The article first clarifies its conceptual framework and key distinctions, then examines cognitive bypass through epistemological, pedagogical and institutional lenses, before turning to implications for supervision, assessment and teaching.

Through a conceptual analysis, the article explores how GenAI extends or bypasses critical cognitive processes in students’ learning. Gilson and Goldberg (2015:128) suggest that there is an increasing need for conceptual articles that ‘bridge existing theories in interesting ways, link work across disciplines, provide multi-level insights, and broaden the scope of our thinking’. Conceptual articles differ from empirical articles in that their conclusions are not derived from data, but instead involve the combination and assimilation of evidence from previously developed concepts or theories (Hirschheim 2008). In an empirical study, the author decides which data to select in order to respond to the research question; similarly, ‘a conceptual paper should explain how and why the theories and concepts on which it is grounded were selected’ (Jaakkola 2020:19). This article develops an integrated argument using three lenses. The justification for these lenses is that the cognitive bypass is not a single point of failure; rather, it is a systemic phenomenon that operates across different levels of human experience. The lenses allow for a discussion that focuses on the individual, the interpersonal and the structural. Firstly, the epistemological lens examines what counts as knowledge and how epistemic agency is formed (Klein & Klein 2025). Secondly, the pedagogical lens focuses on how retrieval, generation and desirable difficulties build durable learning (Bjork & Bjork 2011). Lastly, the institutional lens examines how audit culture, datafication and proxy-based assessment shape rational academic behaviour (Williamson, Bayne & Shay 2020). Taken together, these lenses constitute the article’s conceptual framework for explaining cognitive bypass in higher education. Rather than treating cognitive bypass as merely an issue of individual misuse, the framework argues that it emerges through the interaction between epistemic conditions of knowing, pedagogical conditions of learning and institutional conditions of assessment and certification.

This article starts with a focal phenomenon, one that is observable but not adequately addressed in the literature (Jaakkola 2020). Literature was included if it directly theorised the relationship between GenAI use and cognitive, pedagogical and institutional processes in higher education, and was published mainly between 2020 and 2026 to reflect the rapid increase in GenAI scholarship. This is not a systematic review of all GenAI literature in higher education, but a conceptually bounded analysis of scholarship selected for its explanatory value in understanding cognitive bypass as a multi-level phenomenon. Accordingly, sources were excluded where they focused primarily on primary or secondary education, reported user attitudes without conceptual or theoretical relevance, or addressed GenAI outside educational contexts. These exclusions were intended to preserve coherence between the focal phenomenon and the three analytic lenses used in the article.

Whetten (1989) notes that a good conceptual article, to ensure its validity, should be based on at least some of the seven criteria: (1) What’s new? (2) So what? (3) Why so? (4) Well done? (5) Done well? (6) Why now? and (7) Who cares? Addressing the What’s new? GenAI tools introduce an unprecedented ability to simulate human reasoning, threatening to substitute for rather than scaffold cognition. It can perform complex epistemic tasks (analysis, synthesis and argumentation) that were once exclusively completed by students, raising important questions about cognitive offloading (Larson et al. 2024). So what? The stakes are very high. When students use GenAI procedurally, without constructing or augmenting knowledge, their critical thinking, learning autonomy and applied knowledge significantly diminish (Chen & Cheung 2025). And who cares? Anyone in higher education stands to benefit from evidence that distinguishes between GenAI as a cognitive scaffold and GenAI as a cognitive replacement.

Having outlined the research design, the conceptual framework and the research methods, the next section clarifies the key distinctions that structure the analysis. These distinctions are necessary because debates about GenAI in higher education often collapse qualitatively distinct uses into a single moral or policy category.

Generative artificial intelligence ‘refers to a group of AI algorithms and models that are capable of producing new content, including texts, images, videos and problem-solving strategies, with human-like creativity and adaptability’ (He, Cao & Tan 2025, para 1).

Because debates about GenAI in higher education often collapse different kinds of tool use into a single moral category, this article makes several distinctions explicit to avoid conceptual drift across sections. These distinctions are used as analytical guardrails in the review and as practical levers in the implications for assessment and supervision.

Firstly, the article distinguishes between GenAI as a cognitive extender and GenAI as a cognitive bypass. Drawing on Clark (2025), GenAI can extend thinking when humans remain in the loop as directors, goal setters and critics, and when outputs are treated as provisional material to be tested against industry standards. By contrast, a cognitive bypass occurs when tool use circumvents effortful processes, such as retrieval, synthesis, explanation, error correction and verification that build the internal architecture of understanding (Bjork & Bjork 2011; Klein & Klein 2025).

Secondly, the article operationalises the notion of unethical use in cognitive terms, and not only the way it may be positioned within higher education policy. The critical distinction is between substitution and augmentation, as defined by Puentedura (2013) in his Substitution, Augmentation, Modification & Redefinition (SAMR) framework. Substitution occurs when GenAI performs the target cognitive work that the assessment is intended to elicit. This includes argument construction, conceptual synthesis and source checking. Augmentation, meaning when technology is used as a substitute for traditional methods but with functional improvements (Puentedura 2013), occurs when GenAI supports the student’s work, while the student remains the author of intellectual moves and can defend them. This aligns squarely with the article’s later argument that the cognitive bypass is co-produced by assessment regimes that reward end products over process and make substitution rational (Deslauriers et al. 2019; Shore & Wright 2015).

Thirdly, the article clarifies what kind of epistemic machine an LLM is, because cognitive bypass is often enabled by category error. Large language models are probabilistic generators rather than deterministic calculators or stable retrieval engines; they can therefore produce fluent text that appears academically competent while being low-resolution, partially ungrounded or confidently incorrect (Chiang 2023; Selwyn 2024). To be clear, LLMs, such as Claude, Thaura, DeepSeek, ChatGPT and Perplexity, are not databases. Therefore, verification processes are not nice-to-have add-ons, but a central academic practice that must be designed into assessment and supervision routines.

Finally, the article treats the effects of GenAI as unevenly distributed across levels of expertise. Klein and Klein (2025) describe an expertise duality, in which GenAI can level novice performance in the short term while amplifying expert capabilities. This means that the same tool can function as a scaffold or a substitute depending on prior knowledge and capacity for critique. For this reason, the article does not treat the cognitive bypass as inevitable; rather, it views it as a risk condition that emerges under specific conditions.

With these distinctions in place, the analysis turns first to the epistemological conditions under which GenAI reshapes what counts as knowing and who remains the author of knowledge claims.

The epistemological lens asks: (1) what counts as knowledge?, (2) how is it produced? and (3) how do knowledge claims become justified? In a GenAI-saturated university, these questions shift from abstract philosophy to daily practice as students encounter outputs that sound fluent and confident, but whose epistemic foundations are uncertain (Gabay, Funa & Ricofort 2026). Klein and Klein’s (2025) account of the ‘extended hollow mind’ offers a useful starting point because it treats LLMs not only as productivity tools, but also as interventions, under which epistemic agency and stable understanding are formed. They argue that GenAI creates a dual possibility – it can extend thinking, but it can also hollow it out. In earlier tools, such as the calculator, the cognitive cost is mostly defined by specific skills, but here, students may produce extensive, high-status academic artefacts while failing to construct the internal conceptual maps required to navigate the field independently (Klein & Klein 2025).

A key mechanism in this hollowing is a category error about what LLMs do. When students treat GenAI as a reliable retrieval engine rather than a probabilistic generator of plausible text, their epistemic stance shifts from inquiry and verification to consumption.

Selwyn (2024) notes that:

Bender et al. (2021) describe LLMs as stochastic parrots, able to assemble coherent sequences based on statistical regularities without any understanding. As such, when students are prompting, the LLM is not necessarily looking for the right answer; it is calculating the most likely next word. This conflation of fluency with truth is evident in cases where students draft substantial sections of their theses or essays with GenAI, and are only caught because the references do not exist or because lecturers know that, despite the sentence sounding fluent, it is, in fact, inaccurate. Epistemically, the danger is not just that the individual statements may be wrong, but that academically sounding outputs displace slower practices of source checking, argument construction and reconstruction and conceptual discrimination. In addition, if factually inaccurate theses and essays are circulated digitally, it runs an additional risk of feeding LLMs inaccurate information during scraping processes, thereby diluting, disrupting and corrupting disciplines.

Hallucinations exacerbate the issue. Ji et al. (2023:2) define hallucination as generated content that is nonsensical or unfaithful to the provided source content, and where the generated output is also ‘bland, incoherent, or gets stuck in repetitive loops’. As such, the authors show that inaccurate and unsupported content is a structural feature of many generative systems, rather than an occasional glitch. When fabricated claims, invented citations or subtly incorrect explanations are embedded in polished outputs, students can internalise errors while becoming less sensitive to the need for verification, because the surface signals of credibility are strong and incredibly seductive. Other scholars have reported similar concerns about the effects of hallucinations on academic writing and knowledge production, emphasising how they disrupt epistemic discipline and the required scholarly norms (Dang & Nguyen 2025; Danyaro et al. 2025). In addition, the randomised control trial (RCT) conducted by Steinbach et al. (2025) measured how LLM-generated feedback affected students. The study found that hallucinations increased confusion in students. Similarly, in Cecilio-Fernandes and Sandars’ (2025) article, the authors warn that, without proper scaffolding, hallucinations may erode deep learning and highlight the need for instructional design that integrates AI-literacy.

Klein and Klein (2025) insist that foundational discipline knowledge is not an optional background in the era of GenAI. In fact, it is the substrate that makes epistemic judgment possible. This foundational knowledge allows students to notice when something is not right with generated outputs, for example, when a method is misapplied, a theory is misinterpreted, or an inference in data analysis does not follow. Without these internal cognitive maps, students cannot reliably distinguish between work that is merely well written and work that is epistemically defensible. Klein and Klein (2025:9) speak about the Sovereignty Trap, defined as ‘… [A] key psychological mechanism that may tempt learners to cede their intellectual agency in favour of frictionless, authoritative AI outputs’. Campolo and Crawford (2020) describe this as ‘enchanted determinism’, believing that these AI systems are magical and superhuman, beyond our comprehension and control, yet deterministic and trustworthy enough to make important decisions. In essence, it is a gradual reorientation of academic agency in which the LLM becomes the default proposer of knowledge claims, and the students are reduced to editors of superficial elements of the work.

The epistemic consequences of GenAI are also unevenly distributed across levels of expertise (Klein & Klein 2025). Novices tend to benefit most in the short term because they produce outputs in discipline-specific jargon that they have not even mastered yet. But, they are also most vulnerable to hollowing because they lack the prior knowledge needed to evaluate GenAI output, especially the hallucinations. Experts, on the other hand, can use GenAI as a genuine cognitive extender because they can recognise errors and integrate the outputs into an already coherent internal map.

However, it is also important not to slide into a blanket suspicion of tool use. The point of this article is not to vilify the use of GenAI in any way. It is merely to understand which environment is favourable for students to learn meaningfully in the context of extensive GenAI use. Clark’s (2025) extended mind account directly challenges the cognitive bypass. He notes that humans have always offloaded cognitive labour to external artefacts – from ledgers and notebooks to calculators and now to LLMs – to free up mental bandwidth for higher-order creativity. In Clark’s view, GenAI is a powerful new component in hybrid thinking systems, and offloading does not automatically diminish intelligence; it can, in fact, reorganise cognition in ways that increase capability, provided that humans remain in the loop as directors, goal setters and critics. Clark’s position resists the nostalgic assumption that real thinking must be entirely intracranial, because GenAI can generate alternatives, thereby enriching deliberation rather than replacing it.

The critical distinction lies in what is being offloaded and how the coupling to the tool is governed. Chiang (2023) notes that LLMs are a blurry JPEG of the web. This is because they compress vast amounts of training data into a lossy statistical model, preserving patterns rather than the full facts. Combined with their probabilistic next-word prediction, this yields fluent but fuzzy relationships between concepts. Treating such a system as a stable store of truth conditions, in the way a calculator is treated as a stable store of arithmetic operations, is epistemically hazardous – particularly due to their differing affordances. If students are offloading not just the routine labour but also the foundational work of understanding, the mind does not extend; it detaches. The discomfort that GenAI removes in such cases is not merely the donkeywork, but the essential friction needed to transform information into knowledge.

From this perspective, the epistemological lens reframes the cognitive bypass as a shift in what counts as knowing under GenAI conditions. Students can begin to treat fluent, well-written text as truth, outsourcing the labour of synthesis and, in doing so, failing to build the internal maps necessary for disciplinary navigation. At the same time, GenAI can be integrated as a cognitive extender if universities cultivate robust forms of epistemic agency, such as embedding explicit verification practices and designing assessments that require students to defend, critique and trace the provenance of their claims. In essence, it does not matter to what extent students use GenAI, as long as they remain the locus of epistemic sovereignty in a hybrid thinking system (and embrace our cyborg-ness inevitability as Clark argues!).

If the epistemological lens clarifies what is at stake in the status of knowledge claims, the pedagogical lens asks how those shifts affect the learning processes through which understanding is formed and retained.

The pedagogical lens of this critical analysis foregrounds the biological and cognitive mechanisms of learning that are disrupted by GenAI. The central premise of the cognitive bypass is that unethical GenAI use is not only merely a moral violation, but also a fundamental instructional failure. It erases the very skills that make knowledge durable. This position is also adopted by Bjork and Bjork (2011, 2023), who distinguish between immediate performance (retrieval strength) and long-term learning (storage strength). In the context of GenAI, LLMs function as a fluency-enhancing apparatus, producing plausible-sounding outputs and, in some cases, quite factually relevant information, but leaving the student’s internal cognitive architecture unchanged and underdeveloped.

The importance of the retrieval process to ensure durable learning is also supported by Karpicke and Roediger (2007), who argue that repeated retrieval during learning enhances long-term retention relative to repeated studying, and that repeated recall of items can improve later recall substantially. As such, if a student avoids retrieval by instructing an LLM to write my thesis, or summarise these articles, the student is substituting exposure to answers for memory operations, which may lead to negative implications if independent performance is ever to be tested.

However, a common misinterpretation of the cognitive bypass idea is the assumption that all friction and discomfort are inherently good. Bjork and Bjork (2011) explicitly caution against this by distinguishing between desirable and undesirable difficulties:

Many students may be using GenAI to bypass institutional friction for tasks that are confusing, and where expectations may be unclear. The pedagogical task is not to romanticise struggle, but to identify which difficulties are worth preserving and which are unjust burdens that can be legitimately eased with GenAI.

A second critique is that the cognitive bypass framing can obscure the ambiguity of what constitutes unethical use. Did the notion of unethical use emerge from a policy, a lecturer’s personal stance or social pressure? This approach can conflate very different cognitive situations. For example, on the one hand, students may use these LLMs to complete their assignments; on the other hand, they may use GenAI to translate academic jargon. A sharper critical analysis requires an operational distinction between the two: (1) substitution – where GenAI performs the target submission; or (2) augmentation – GenAI increases opportunities for retrieval but maintains student authorship. This SAMR scaffolding (Puentedura 2013) aligns with Bjork and Bjork’s (2011) point that desirable difficulties must be task-relevant and solvable.

A third critique highlights the implicit individualism of the cognitive bypass accounts. We are all aware that universities are incentive environments (Singh 2024). This means that students respond rationally to assessment designs that reward end-product over process, speed over reflection and high-stakes outputs over iterative practice. These evaluative processes occur despite what the evidence of good learning suggests (Deslauriers et al. 2019). Given this, if a module only assesses the final essay rather than students’ retrieval, planning, drafting or revision processes, then GenAI becomes an efficient strategy. What this means is that the cognitive bypass in this case is co-produced by assessment regimes, workload pressure and credential logics – not by student ethics alone.

Looking at the cognitive bypass through a pedagogical lens suggests not treating it as a general effect of using LLMs. Bjork and Bjork’s (2011, 2023) work shows that both lecturers and students are vulnerable to misassessing learning when they equate current performance with long-term retention. Generative artificial intelligence intensifies that vulnerability by making high-performance cues -– polished language and structured arguments – easily accessible.

The pedagogical challenge is to design learning as a retrieval-rich, feedback-rich practice in which students must continually generate, explain and refine their own thinking. In a learning context such as this, GenAI can be constrained into a scaffold, for example, helping to prompt for explanations or to critique, rather than being a silent stand-in for the student’s cognitive work. In essence, when learning is framed mainly as a performance-oriented process, GenAI enables the cognitive bypass; when it is framed as visible cognitive work over time, GenAI becomes a tool to support rather than replace learning.

Yet cognitive bypass cannot be explained by learning processes alone; it is also shaped by the institutional environments that reward certain forms of performance and proxy evidence.

The institutional lens foregrounds the sociological and organisational structures that shape, incentivise and validate academic behaviour. It views the university as a system of certification and data management in which written artefacts, grades and digital traces stand in for complex processes of learning and becoming. For most higher education institutions, the written artefact still operates as a proxy for internal transformation. For example, most universities still use the end-of-course essay and the thesis as evidence of understanding, judgment and the ability to act. GenAI’s capacity to produce polished and plausible text without commensurate cognitive labour places this proxy regime under strain. When institutional logics equate a good document with a good mind, students can rationally exploit the equivalence by outsourcing the work to LLMs. From Biesta’s (2014) perspective, the qualification function (certification) overwhelms subjectification (supporting the emergence of a knowledgeable, responsible graduate), so that institutions risk becoming credentialing machines attached to opaque cognitive processes. In this context, GenAI not only single-handedly creates hollow certifications, but it also reveals that the link between qualification and person was always probabilistic and that the system has long relied on proxies that can be gamed.

This erosion of proxy reliability is entangled with the broader datafication of teaching. Williamson et al. (2020) show how modern universities are governed through expanding infrastructures of metrics – i.e. learning analytics, performance indicators and dashboards – through which teaching and learning are made legible as data. Grades, written assignments, dwell time on dashboards, number of logins on learning management systems (LMSs), number of clicks, late submissions and completion rates serve as signals of quality and productivity, compressing complex pedagogical processes into countable outputs and providing the institution with a student’s ‘bio’.

Shore and Wright’s (2015) notion of audit culture captures how these metrics produce coercive accountability. This means that students and staff are compelled to perform to indicators, regardless of whether those indicators truthfully track the values or practices they supposedly represent. When assessment regimes privilege final products and overlook the distributed work of retrieval, drafting and revision, GenAI substitution becomes structurally rational. It allows students to meet their course requirements and audit demands while minimising time, risk and the emotional labour of learning. In this sense, the cognitive bypass is co-produced by institutional designs that reward outputs rather than learning processes. Selwyn (2024) argues that GenAI is often framed as a route to frictionless efficiency, promising personalised learning at scale and the automation of routine tasks. Yet, this imaginary masks the messy realities of pedagogy and the power relations embedded in these LLMs, including who is watched, whose data are mined and how decisions are made.

In addition to this inaccurate imaginary, over the last three decades, institutions have rapidly made the switch from traditional pen, paper and manila folder systems to computerised systems (embedding AI) with students’ activity tracked through LMSs, plagiarism detection services and analytics dashboards (Selwyn 2024; Williamson et al. 2020). Generative artificial intelligence is now at the heart of a curious institutional contradiction. On the one hand, universities deploy AI to streamline administration, offer writing support (by encouraging the use of tools such as Grammarly) and enable adaptive learning. On the other hand, universities deploy AI to police their own loss of control, using automated grading, exam proctoring, plagiarism detection and behavioural analytics to assess originality and diligence (Selwyn 2024).

Students’ written outputs are automated and monitored at the same time and by the same class of technologies, while their progress is inferred from digital traces rather than from sustained dialogical engagement. Several authors have responded, cautioning against this reductionist approach to using data to represent the reality of student learning, noting that data can never truly reflect classroom realities or the complexity of how students learn. Firstly, Selwyn (2024) notes that

Secondly, Goulden (2018 cited in Kocabıyık 2025:5) states that ‘even the most “technologically smart” innovation is likely to be “socially stupid” when applied in a real school setting’. Thirdly, Broussard (2019 cited in Selwyn 2024) explains that

And lastly, Perelman (2014) notes in the context of automated essay scoring, when the state of the art is counting words and surface features, systems reproduce superficial proxies for writing quality.

However, it should also be noted that not all forms of bypass are inherently regressive. Critics of traditional assessment have long noted that much of students’ work already consists of busywork that satisfies institutional requirements without necessarily deepening understanding (Wheeler 2025). In cases such as these, GenAI may help students shed unproductive labour – such as rephrasing difficult assignment instructions, or translating complex academic jargon into easy-to-understand language, without undermining disciplinary learning. Equity-focused work on GenAI in educational settings highlights its potential as a scaffold rather than a substitute. Shukla and Pandey (2025), for example, describe how GenAI can support personalisation, accessibility and inclusion for students facing linguistic barriers, disability or limited access to cultural capital. In this case, GenAI can reduce institutional friction, allowing students to engage more fully with core conceptual difficulties.

This perspective complicates the straightforward defence of GenAI, simply removing the difficulty. The question is: What difficulty is GenAI removing? If the wrong things are made difficult – for example, navigating opaque systems and decoding poorly written tasks – then easing those challenges using GenAI may be a precondition for genuine cognitive work rather than an avoidance of it (Choudhuri et al. 2026). The institutional challenge is to differentiate between unimportant or less important drudgery that can justifiably be automated and epistemic labour that should remain with the student (Shukla & Pandey 2025). What is also important to note is that one cannot have one rule for all disciplines, as sciences and humanities, for example, have fundamental ontological differences, and what is important in one discipline may not be in the other.

So, responding to the crises of proxy, therefore, demands more than a tightening of surveillance or banning of tools. It calls for complete institutional change that aligns assessments and governance with what we know about learning and epistemic agency. One way to do this, as discussed in the latter part of this article, is to redesign assessment so that the process of learning, rather than only the final output, is visible and valued. In this reconfigured approach, GenAI is not removed but rather repositioned. When students are required to demonstrate understanding interactively across time and contexts, simple substitution becomes less effective, and human–AI interaction is more likely to serve as a scaffold than a proxy (Shukla & Pandey 2025).

Channelling Selwyn’s (2024) perspective, the institutional lens suggests that universities need to drop this fantasy of frictionless GenAI and instead embed this technology within pedagogies and governance structures that allow epistemic agency to remain with students and staff. Under this lens, the cognitive bypass is not simply a story about individual misconduct; it is a symptom of a broader regime that privileges data, outputs and efficiency over the slower, harder-to-measure work of thinking. To address this, two critical things need to happen. Firstly, students need to be taught how to ethically use GenAI in their work, and secondly, reconfiguring the institutional conditions under which proxies are produced, interpreted and rewarded.

Taken together, these three lenses show that cognitive bypass is co-produced across levels. The question, then, is how supervision, assessment and teaching might be redesigned to shift GenAI use from substitution towards augmentation.

If universities want GenAI to function as an augmenter rather than a silent substitute, they will need to redesign their tasks and the institutional mechanisms that define what counts as learning and performance. Scholars emphasise that this redesign is not simply a technical one; it is relational and contextual, entangled with histories of inequality, supervision cultures and institutional logics in higher education (Brown & Rossouw 2026; Naidoo & Samuel 2026). This required shift has at least three interlocking layers: supervision cultures, assessment practices and teaching design.

The suggestions made earlier would be easy to implement if universities focused more on the learning and less on the learning analytics. These changes are difficult to sustain and implement if institutional mechanisms continue to reward only throughput, satisfaction scores and product metrics. Generative artificial intelligence integration that results in augmentation will falter if it is grafted onto unchanged audit culture environments.

If institutions are serious about implementing GenAI in a meaningful way, they need to prioritise adjusting their quality assurance and programme review criteria so that process-rich assessment is recognised and rewarded, even if it is more labour-intensive. Institutions should also evaluate existing workload models and weight oral components equally to writing outputs, as both require critical thinking, argument development and argument defence. This will also allow for more creativity in the classroom and appeal to learners with different learning strengths. Institutional policy directed at GenAI use should be developed around learning design and epistemic justice, not only around misconduct, by articulating clearly what counts as legitimate augmentation and then designing assessments that demand the additional human cognitive work. Lastly, institutions should invest in shared infrastructure that supports process evidence (such as portfolio platforms, reflection templates, and tools for capturing annotated drafts) so that focused process practices are feasible at scale and do not depend on individual lecturers’ ad hoc solutions. In essence, moving from substitution to augmentation is not just about telling students to use the tools ethically; it is about changing the conditions under which using GenAI as a substitute is the easiest way to succeed. When supervision conversations, assessment rubrics and teaching practices require visible reasoning, retrieval and defence, GenAI more naturally falls into the role of an assistant rather than an author.

Using a conceptual analysis, this article highlights that LLMs are not databases and, if treated as such, critical learning gets misplaced. It clarifies that GenAI does not automatically harm learning, but that it can result in a cognitive bypass when the conditions are right. When students treat fluent GenAI text as the truth, they risk skipping the slow work of checking, explaining and connecting ideas, so understanding becomes shallow. When GenAI writes students’ work or provides them with complete summaries, it may well boost short-term performance, but weaken the desirable difficulties that make learning stick. At the same time, universities often reward polished products, fast turnaround and high grades more than visible thinking processes, which makes heavy GenAI use a rational strategy. To protect epistemic agency, we need supervision, assessment and teaching designs that make thinking visible over time. That way, GenAI supports students’ cognitive work instead of quietly replacing it.

Future research should empirically test whether process-oriented assessment designs reduce cognitive bypass behaviours across disciplines and at different levels of expertise. Longitudinal studies tracking how GenAI affects students’ retrieval strength and epistemic agency over time would also be a phenomenal contribution to the discussion. But more importantly, equity-focused research examining how cognitive bypass differs across socioeconomic, linguistic and disciplinary contexts is the priority.