Serqet — Smarter Systems for UK Businesses

Critics of artificial intelligence in education argue, drawing on a long lineage of pedagogical conservatism, that generative AI undermines genuine learning. I will argue that it does not — and that the deeper stake of this debate is not the integrity of conventional pedagogy but the question of access.

The Two Charges

Critics contend, first, that AI permits students to bypass the cognitive labour that assignments are designed to elicit, producing polished outputs without commensurate understanding (Cotton, Cotton and Shipway 2024, 230). Second, they aver that habitual reliance on these systems risks atrophying the very faculties — retrieval, analysis, sustained writing — that formal education exists to cultivate, a concern that has acquired particular potency in a generation whose attentional architecture has, on a growing body of evidence, been compromised by the dopaminergic incentive structures of short-form digital media (Gerlich 2025, 3-4; Haidt 2024, 117-19).

However, in this essay, I will argue that artificial intelligence does not, in itself, undermine genuine learning, for two principal reasons. Principally, the shortcut behaviour that critics invoke is more accurately understood as a symptom of pre-existing structural deficiencies in mainstream education, particularly its tendency to reward visible output over genuine comprehension and to teach to a median learner rather than the individual student. In addition, when properly governed and combined with human tutoring, artificial intelligence offers something that conventional classroom architectures have struggled to deliver at scale: bespoke, adaptive, and continuously available tuition that addresses precisely the mismatch from which shortcut behaviour arises in the first place. I will further argue, in the closing sections, that the deeper stake of this debate is the question of access — whether the bespoke instruction long available to the affluent can now be extended to every child, and in particular to the neurodivergent and dyslexic learners whom an educational regime calibrated to the statistical mean has systematically failed.

Defining the Terms

I define artificial intelligence in education as the deployment of generative and adaptive computational systems — principally large language models and intelligent tutoring systems — within formal and informal learning contexts to produce explanations, generate practice material, deliver feedback, or substitute for written work. This essay’s focus will not be on the use of AI for purely administrative ends, such as timetabling or institutional analytics. Rather, the concern is specifically with artificial intelligence as it interacts with the cognitive activity of the learner. It is also necessary to distinguish, at the outset, between two modes of student use: replacement, in which the system performs a cognitive task that the student was meant to perform; and augmentation, in which the system scaffolds, explains, or interrogates the student’s own thinking. Much of the conceptual confusion in the present debate stems from a failure to maintain this distinction.

Charge One — Substitution

Critics assert that artificial intelligence undermines genuine learning by enabling students to substitute machine-generated output for their own intellectual effort. It is claimed that an essay produced by a generative model, even if subsequently edited by the student, does not reflect the cognitive process the assessment was intended to measure (Cotton, Cotton and Shipway 2024, 232) and that no educator or institution has the justification to treat such output as evidence of learning.

To illustrate this contention, consider an extreme thought experiment. Two students, Alice and Bernard, are enrolled in the same module on Renaissance political thought. Both must submit a 2,000-word essay analysing Machiavelli’s conception of virtù. Alice, conscientiously, reads the primary text, consults secondary literature, drafts and redrafts her argument over several weeks, and submits a piece that bears the imperfections of authentic struggle. Bernard, by contrast, supplies a generative model with the prompt and a few stylistic preferences and submits the resulting output with minimal modification. Both students receive comparable marks. According to the critic, this scenario instantiates serious harm: Bernard has not undergone the cognitive process the assignment was designed to elicit, and the parity of grades suggests that the institutional measure of learning has been decoupled from learning itself, with corrosive effects on the credibility of the qualification (Selwyn 2019, 67-68).

Proponents might rebut by labelling Bernard’s behaviour exceptional. If Bernard had used the model not to produce his essay but to interrogate his own draft — asking the system to identify weak inferences, to articulate the strongest counterargument to his thesis, and to suggest unfamiliar primary sources — the outcome could plausibly be that Bernard learns more, not less, than Alice (Mollick 2024, 91-93).

However, this rebuttal does not, on its own, dispose of the integrity charge. Even if augmentative use is genuinely educative, it does not follow that the prevailing pattern of student use is augmentative rather than replacement-oriented. Critics would consequently posit that institutions ought, in the absence of robust countermeasures, to restrict or prohibit student use of these systems altogether.

This prohibitionist conclusion can be rebutted on two grounds. First, it is unrealistic to expect that a technology of such ubiquity and utility, embedded already in the workplaces students will enter, can be successfully cordoned out of education by institutional fiat. Second, prohibition mistakes a symptom for the disease. Bernard’s substitution of machine output for his own thinking is, in part, a rational response to an environment in which completion is rewarded more reliably than comprehension. Expecting prohibition to restore the integrity of learning, while leaving the underlying incentive structure intact, addresses the surface manifestation of the problem rather than its cause — and may represent a greater impediment to educational reform than a properly governed integration of the technology (Selwyn 2019, 71).

The boundary between authentic and assisted thought is not given by nature but constructed by pedagogical convention.

The critic’s proposition has its strongest grip only when learning is conceived as a purely solitary cognitive activity, generated through the unaided labour of the individual student. Yet learning is in fact derived through an amalgamation of internal effort and external scaffolding — textbooks, lectures, peer discussion, search engines, calculators, dictionaries. Since no cognitive achievement of consequence is wholly endogenous to the learner (Vygotsky 1978, 86), the use of artificial intelligence as one further external scaffold does not, in principle, violate the sanctity of learning. The wholesale rejection of AI as illegitimate assistance rests on a conception of learning that few of its critics would consistently apply to other tools.

Consider the mathematics student who uses a calculator to perform arithmetic in the course of solving a calculus problem. Earlier critics might have insisted that the calculator interferes with the integrity of the student’s mathematical reasoning. However, since the cognitive achievement at stake is not arithmetic but calculus, and since the curriculum has been deliberately reorganised to privilege the higher-order task, the imposition of the calculator on the lower-order task is not a violation of mathematical learning but a reciprocal redistribution of cognitive effort towards that which the discipline now considers central (Holmes, Bialik and Fadel 2019, 142-43). The same logic, suitably extended, applies to generative AI: the question is not whether a tool has been used, but whether the cognitive achievement around which the curriculum is organised has, in fact, been displaced.

Charge Two — Cognitive Dependency

The second main argument is that artificial intelligence in education undermines genuine learning by inducing cognitive dependency. Even where AI is used augmentatively rather than replacively, the habitual outsourcing of cognitive effort risks atrophying the very faculties that education exists to cultivate. This reflects the underlying pedagogical principle that the difficulty of the task is itself partly constitutive of the learning (Bjork and Bjork 2011, 58) and that frictionless assistance, however well intentioned, can undermine the productive struggle from which durable understanding emerges.

Consider a student who uses a generative model to receive an instant explanation of every passage of a difficult text she does not immediately understand. She derives, in the short term, a sense of comprehension and a smooth path through the material. However, it is not obvious that the smoothness is benign; the smoothness itself may be the harm, since the moments of confusion she has bypassed are precisely those in which her own interpretive faculties would have been exercised and strengthened (Carr 2010, 115-16). Ease of comprehension is not the same as depth of comprehension.

The dopaminergic generation

The dependency objection acquires particular force in the case of younger learners, whose executive function and capacity for sustained attention have, on a growing body of evidence, been compromised by the dopaminergic incentive structures of short-form video and infinite-scroll media (Haidt 2024, 117-19; Lembke 2021, 64-67). On this stronger version of the argument, generative AI placed in the hands of a cohort already conditioned, by hours of daily algorithmic reinforcement, to seek the shortest path to gratification cannot reliably be expected to be used augmentatively. The disciplinary preconditions for the productive use of any cognitive tool — patience, tolerance for confusion, the willingness to dwell in the unresolved — have, on this account, been systematically eroded outside the classroom.

This is not a casual objection. It is, in my view, the most serious form of the dependency critique, since it cannot be dismissed by appeal to the abstract possibility of well-governed use. The relevant question is whether well-governed use is psychologically available to the learners in question, and there is reason to suspect that, at present, it is not.

The Strongest Response — Human Tutor Plus Machine

One of the weaker rebuttals to the dependency argument is that AI tutoring is justified, even if it introduces some risk of dependency, on grounds of educational throughput — a single classroom of thirty students can now receive immediate clarification of any unfamiliar concept without the bottleneck of a single human teacher (Kasneci et al. 2023, 5). However, this rebuttal does not challenge the fundamental proposition that the depth and durability of learning may be compromised. Critics will rightly assert that learning is composed not of momentary comprehensions but of robust cognitive structures, and that these structures should not be sacrificed for the sake of mere institutional efficiency (Carr 2010, 122-24).

The strongest response to the dependency argument, particularly in its dopaminergic form, is not to deny the objection but to draw from it a different conclusion than the one critics typically extract. To the extent that younger learners cannot be relied upon to supply the discipline that augmentative use of AI requires, it follows that the discipline must be supplied externally. It does not follow that AI must therefore be withheld. The architecture that responds most directly to the objection is not prohibition but a combined intervention: the human tutor supplies the executive scaffolding, the boundary-setting, and the relational accountability that the dopaminergically compromised learner cannot reliably supply herself; the AI, in parallel, supplies the personalised, adaptive, and inexhaustibly patient instruction that a single human tutor cannot scale to deliver.

The human supplies what the machine cannot: disciplinary presence, embodied authority, and the affective registers of mentorship from which younger learners draw the will to persist through difficulty. The machine supplies what the single human cannot: depth of personalisation at scale, infinite patience, and the capacity to meet each learner at her precise point of confusion without the social cost of admitting confusion in the first place. Neither alone resolves the problem the critics have identified. Together, they do.

This combination — the human tutor and the artificial intelligence in tandem — is not a marginal improvement on prevailing pedagogy but something approaching a revolution in the conditions under which mass education is conducted (Mollick 2024, 178-80; VanLehn 2011, 215).

The Bespoke Principle

The strongest rebuttals to the critic’s arguments are those that emphasise the principle of bespoke learning and the augmentative use of AI to address the structural mismatch in mainstream education. The principle of bespoke learning states that students vary widely in prior knowledge, pace, confidence, and preferred modes of explanation, and that conventional classrooms — obliged to teach to a median learner — systematically underserve those who depart from that median (Bloom 1984, 6-7; VanLehn 2011, 198-99). In this scenario, artificial intelligence is not undermining learning but compensating for an antecedent failure of personalisation that the educational literature has long recognised but conventional classroom architectures have been unable to remedy at scale.

Consider a classroom of one hundred students: thirty are markedly ahead of the curriculum, fifty are within the band for which the lesson is calibrated, and twenty are markedly behind. Critics would assert that the introduction of an AI tutor threatens the integrity of learning, since it permits the advanced students to outsource their thinking and the behind students to receive answers without struggle. However, if the AI tutor is configured adaptively — posing Socratic questions to the advanced students that exceed the curriculum, providing the median students with targeted retrieval practice calibrated to their individual error patterns, and supplying the behind students with patient and unembarrassed re-explanation of the prerequisites they lack — artificial intelligence does not undermine learning but extends it. This approximates what Bloom (1984, 4) famously identified as the two-sigma advantage of one-to-one tutoring over group instruction, an advantage long understood to be educationally transformative but historically unattainable at scale.

An illustration

I might be permitted, at this juncture, to offer a brief autobiographical illustration of what such bespoke instruction can yield. Over the course of approximately two weeks, I employed a generative model to extract what might be described as a cognitive profile of my own modes of comprehension — the analogies that landed, the structures that did not, the sequences in which prerequisites needed to be supplied, the registers in which difficult material became tractable rather than opaque — and I used the resulting map to acquire, to a working depth, the substance of a finance curriculum that conventional pedagogy would have spread across many months. I do not advance this as decisive evidence; the experiment is an n of one, and a self-selected one at that. However, it is illustrative of the order of magnitude of the gain that becomes possible once instruction is aligned to the learner rather than the learner to the instruction. What was previously available only to those whose families could afford private tuition of the highest calibre — a teacher who continually re-calibrates the explanation to the particular learner before her — is now, in principle, available to anyone with access to a competent generative model.

The Deeper Question — Access

The most serious stakes of this debate are not registered until the question is reframed. The conventional question — whether artificial intelligence undermines genuine learning — implicitly presupposes that conventional learning is the standard against which any new technology must be assessed. However, this presupposition itself merits interrogation. Conventional pedagogy is calibrated to a statistical mean that has never accurately reflected the heterogeneity of human cognition (Rose 2016, 8-10). For a substantial minority of learners — those who are dyslexic, those who are on the autism spectrum, those whose attention is structured differently from the median, those whose intellectual capacities are obscured rather than revealed by standardised modes of instruction — the conventional classroom has functioned not as a vehicle of learning but as an instrument of systematic misdiagnosis. These learners are not, as the prevailing rhetoric has too often implied, slow or incapable. Their cognitive architectures are mismatched with a pedagogical apparatus designed for an average that almost no individual fully instantiates (Eide and Eide 2011, 22-24; Armstrong 2010, 9).

For the overwhelming majority of human history, learning was acquired through methods that were necessarily idiosyncratic — observation, apprenticeship, narrative, ritual, play, and the patient transmission of knowledge from elder to junior in forms calibrated to the particular learner before them (Gray 2013, 33-36). The standardised classroom of the modern nation-state is, by historical standards, an extraordinarily recent and extraordinarily narrow innovation. It was designed, in the main, to confer a basic literacy and numeracy upon the largest possible number of children at the lowest possible per-pupil cost; it was not designed, and could not coherently have been designed, to honour the cognitive diversity of those it processed. The question is not whether the introduction of artificial intelligence threatens an educational regime that has otherwise served all children well. It is whether the continuation of an educational regime calibrated to the statistical mean remains defensible once the means of its replacement have become available.

Free education was the great equaliser of access: it ensured that every child could enter the classroom. Bespoke, AI-augmented education, properly administered, is the equaliser of fit: it ensures that every child, once inside the classroom, encounters an instructional environment calibrated to her own cognitive constitution rather than to a statistical fiction. The first was the equality of form; the second is the equality of substance — and it is, accordingly, the unfinished work of educational justice.

A final rebuttal would posit that all students living in a digitally mediated society have, voluntarily or involuntarily, implicitly consented to the integration of computational tools into their cognitive lives. The cognitive environment of the contemporary learner already includes search engines, autocomplete, spell-check, and recommendation systems, and there is no coherent position from which to admit these instruments while excluding generative AI (Selwyn 2019, 89). In the contemporary classroom there is an implicit pedagogical contract between the learner, the educator, and the cognitive environment in which both are situated. Artificial intelligence, to which a portion of cognitive scaffolding has been delegated under conditions of governance, may participate in the formation of understanding, while the educator’s epistemic authority — from which curricular judgement, ethical formation, and the assessment of genuine learning continue to issue — remains the locus of the relationship.

Conclusion

This essay has presented the critics’ arguments that artificial intelligence undermines genuine learning by permitting students to substitute machine-generated output for their own cognitive labour, and by inducing a cognitive dependency that may atrophy the faculties education is meant to cultivate — a dependency rendered more acute in the present generation by the dopaminergic hijacking of attention through short-form media. I have rebutted the primary contention by arguing that the shortcut behaviour at issue is more accurately diagnosed as a symptom of pre-existing structural weaknesses in mainstream education than as a pathology produced by AI itself. I have rebutted the dependency contention, in its strongest form, by arguing that the disciplinary preconditions for productive AI use, where they are absent in younger learners, are properly supplied by the human tutor in tandem with the machine — an architecture that is not a marginal improvement on prevailing pedagogy but a revolution in the conditions under which mass education becomes possible.

Finally, I have argued that the deepest stake of the debate is not the integrity of conventional learning but the question of access: whether the bespoke instruction long available to the affluent can now be extended to every child, including those neurodivergent and dyslexic learners whose cognitive architectures have been ill-served by an educational regime calibrated to a statistical mean. Artificial intelligence in education does not undermine genuine learning. It is, on the contrary, the principal means by which the unfinished work of educational equality — the equality of fit, rather than the equality of mere form — can at last be brought within reach.

If your organisation is grappling with how AI augments rather than replaces human capability, we would welcome the conversation.

References

Armstrong, T. (2010) Neurodiversity: Discovering the Extraordinary Gifts of Autism, ADHD, Dyslexia, and Other Brain Differences. Cambridge, MA: Da Capo Lifelong Books.
Bjork, E.L. and Bjork, R.A. (2011) ‘Making Things Hard on Yourself, but in a Good Way: Creating Desirable Difficulties to Enhance Learning’, in Gernsbacher, M.A., Pew, R.W., Hough, L.M. and Pomerantz, J.R. (eds.) Psychology and the Real World. New York: Worth Publishers, pp. 56–64.
Bloom, B.S. (1984) ‘The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One Tutoring’, Educational Researcher, 13(6), pp. 4–16.
Carr, N. (2010) The Shallows: What the Internet Is Doing to Our Brains. New York: W.W. Norton.
Cotton, D.R.E., Cotton, P.A. and Shipway, J.R. (2024) ‘Chatting and Cheating: Ensuring Academic Integrity in the Era of ChatGPT’, Innovations in Education and Teaching International, 61(2), pp. 228–239.
Eide, B.L. and Eide, F.F. (2011) The Dyslexic Advantage: Unlocking the Hidden Potential of the Dyslexic Brain. New York: Hudson Street Press.
Gerlich, M. (2025) ‘AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking’, Societies, 15(1), pp. 1–28.
Gray, P. (2013) Free to Learn. New York: Basic Books.
Haidt, J. (2024) The Anxious Generation. New York: Penguin Press.
Holmes, W., Bialik, M. and Fadel, C. (2019) Artificial Intelligence in Education: Promises and Implications for Teaching and Learning. Boston, MA: Center for Curriculum Redesign.
Kasneci, E. et al. (2023) ‘ChatGPT for Good? On Opportunities and Challenges of Large Language Models for Education’, Learning and Individual Differences, 103, 102274.
Lembke, A. (2021) Dopamine Nation. New York: Dutton.
Mollick, E. (2024) Co-Intelligence: Living and Working with AI. New York: Portfolio.
Rose, T. (2016) The End of Average. New York: HarperOne.
Selwyn, N. (2019) Should Robots Replace Teachers? AI and the Future of Education. Cambridge: Polity Press.
VanLehn, K. (2011) ‘The Relative Effectiveness of Human Tutoring, Intelligent Tutoring Systems, and Other Tutoring Systems’, Educational Psychologist, 46(4), pp. 197–221.
Vygotsky, L.S. (1978) Mind in Society: The Development of Higher Psychological Processes. Cambridge, MA: Harvard University Press.

Does the use of artificial intelligence in education undermine genuine learning?