The Boy Who Stood on His Desk: Artificial Intelligence and the Subversion of Human Judgement
On education, cognitive obedience and the increasingly uncertain boundary between assistance and intellectual dependence
The boy who stood on his desk
I remember a lesson from my schooldays that has remained with me for the best part of half a century. We were given what appeared to be an examination paper, containing a series of questions to be completed within a prescribed time. Like everyone else, I began at the beginning, carefully reading and answering each question before proceeding to the next. It was what we had been taught to do. Reading was something to be undertaken thoroughly, instructions were to be followed, and examinations were occasions on which one was expected to demonstrate competence through diligence and a reasonable degree of obedience.
A few minutes into the exercise, a boy closed his paper, stood up and climbed onto his desk. This was not entirely unfamiliar behaviour at my school. I had previously witnessed chairs being thrown in the direction of teachers, such was the quality of my early education. Privilege, I suppose, comes in many forms. On this occasion, however, rather than reprimanding the boy, the teacher began to applaud. He then instructed the rest of us to stop what we were doing and read the final question. At the bottom of the paper was an instruction to disregard everything that had preceded it, put down our pencils and stand on our desks. The boy had understood the assignment. The rest of us had merely followed it.
It was a simple exercise, perhaps rather infantile in retrospect, but I have returned to it repeatedly when considering the relationship between education, cognition and intellectual independence. We had been taught to approach written information in an orderly manner, processing what was immediately before us and responding accordingly. Yet our competence in following the procedure had prevented us from recognising the structure of the task itself. We were so occupied with answering the questions that we had neglected to establish whether the questions required answering at all. There is an important qualification, of course. The boy who stood on his desk was also following an instruction, albeit one the rest of us had overlooked. His behaviour was not necessarily evidence of superior intelligence or some extraordinary capacity for independent thought. The lesson was rather more subtle: competent execution and sound judgement are not necessarily the same thing.
There is an interesting connection here with the psychology of reading. Skilled reading is neither a simple matter of identifying words sequentially nor an exercise in freely guessing what comes next. Research on eye movements, word recognition and linguistic predictability has demonstrated that readers make use of contextual information, and that predictable words are generally processed more efficiently than unexpected ones. The extent to which these effects reflect active prediction, facilitated recognition or other mechanisms remains a subject of theoretical debate (Wong et al., 2025). Our expectations assist comprehension, but they can also influence the information to which we attend. The relevance to my classroom experience is not that reading from beginning to end is somehow wrong, but that an efficient and well-established cognitive procedure can become so familiar that its appropriateness escapes examination. We become skilled at carrying out the process without necessarily questioning the assumptions upon which it rests.
I have recently found myself wondering whether artificial intelligence presents a rather more sophisticated version of that classroom exercise. It can provide us with information, explanations, recommendations and conclusions, often in a form that appears sufficiently coherent to discourage further examination. We can scrutinise individual statements, check references and look for obvious factual errors, but the more consequential question may concern whether we have accepted the structure of the problem before deciding whether that structure deserves our confidence. My concern is not that AI necessarily produces poor judgements, nor that human judgement possesses some inherent superiority. It is that the increasing availability of apparently competent machine-generated reasoning may change when, why and how people decide to exercise judgement of their own.
The distinction between thinking and judging
Human beings have always made use of external resources to reduce cognitive effort. Writing, books, calculators, statistical software and increasingly sophisticated information systems have allowed us to undertake activities that would otherwise exceed our individual capacities. Risko and Gilbert (2016) describe cognitive offloading as the use of physical actions and external resources to reduce the information-processing demands placed upon the individual. Such behaviour is neither unusual nor necessarily undesirable. It would be absurd to suggest that someone using a calculator has surrendered mathematical understanding merely because they have chosen not to perform every calculation mentally. The relevant question is whether the person retains sufficient understanding to recognise what is being calculated, why the calculation is appropriate and whether its result deserves acceptance.
Artificial intelligence introduces an interesting complication because it can undertake activities that resemble the evaluative processes for which we have traditionally relied upon human expertise. A statistical package can return the results of an analysis, but a generative AI system can also explain those results, interpret their implications, recommend subsequent analyses and suggest conclusions suitable for publication. The user may no longer be outsourcing only the calculation. They may also be outsourcing substantial parts of the judgement that determines what the calculation means. This is not necessarily problematic when the user possesses sufficient expertise to evaluate the recommendation, but it becomes considerably more concerning when the recommendation substitutes for the expertise that the user has yet to acquire.
I think it is useful to distinguish cognitive offloading from what might be called judgement offloading. The former involves delegating some aspect of information processing; the latter involves delegating part of the evaluation by which information is accepted, rejected or acted upon. These processes are not entirely separable, and the distinction is proposed here as a conceptual refinement rather than an established psychological classification. Nevertheless, it captures something that is easily overlooked in discussions about AI. Asking a system to explain a statistical result is not necessarily equivalent to asking it whether that result should be believed, what conclusions should be drawn or whether the original research question was worth asking. The latter activities involve questions of evidence, relevance, uncertainty and purpose. They require more than the production of a plausible answer.
There is a further distinction between making a judgement and retaining formal responsibility for one. An academic may approve an AI-generated assessment, a clinician may sign off an AI-assisted diagnosis, and a manager may endorse a recommendation produced by an algorithm. In each instance a human being remains nominally responsible for the decision, but this tells us remarkably little about the independence of the judgement involved. The individual may have scrutinised the recommendation thoroughly, or they may have accepted it because it appeared authoritative, aligned with their expectations or simply saved time. The signature remains the same. The cognitive process behind it does not.
When assistance becomes substitution
Evidence from education offers good reasons to take this distinction seriously. Bastani et al. (2025) conducted a randomised field experiment involving nearly a thousand secondary-school mathematics students in Turkey who were given access to different forms of AI assistance. One system resembled a relatively unrestricted GPT-4 interface, while another was designed to provide structured tutoring through teacher-informed hints rather than simply supplying completed answers. Both forms of AI assistance improved performance during practice. Compared with students receiving no AI assistance, those using the unrestricted system achieved a 48% improvement in practice performance, while those using the structured tutor achieved an improvement of 127%. However, when students subsequently completed an examination without AI assistance, those who had used the unrestricted system performed approximately 17% worse than students who had received no AI assistance. The more carefully designed tutoring system largely eliminated this detrimental effect, although it did not produce a corresponding improvement in unaided examination performance.
The importance of this finding lies in the distinction between immediate performance and acquired competence. Students were able to produce better answers while receiving assistance, yet some subsequently demonstrated weaker independent performance than their peers. The technology had helped them complete the immediate task without necessarily supporting the acquisition of the skills required to undertake a similar task independently. It would be inappropriate to interpret this as evidence of permanent intellectual deterioration. The experiment examined short-term learning in a particular educational setting, and the findings cannot simply be generalised to every subject, age group or form of AI assistance. Nevertheless, it demonstrates that successful performance in the presence of AI cannot be treated as sufficient evidence of learning. The appearance of competence may conceal a failure to develop the understanding that education is intended to produce. Particularly revealing was the observation that students' perceptions of their own learning did not necessarily correspond with their subsequent performance (Bastani et al., 2025).
A second study brings the question much closer to professional judgement. Goulas et al. (2026) conducted a preregistered randomised experiment involving 1,339 teachers in Greece who were asked to evaluate student work accompanied by deliberately inaccurate grading recommendations. Crucially, the recommendations were attributed either to a human colleague or to an AI system. When an unduly harsh grade was presented as AI-generated, the teachers' grading fairness gap was approximately 22% larger than when the recommendation was attributed to a human. This measure concerned the deviation between the teacher's revised grade and an objectively determined benchmark; it did not mean that 22% more teachers made incorrect decisions. The corresponding difference between AI-labelled and human-labelled recommendations was not statistically significant when the initial recommendation was excessively lenient. The researchers also found that perceptions of the algorithm's competence and responsibility helped account for the greater willingness to accept an unjustifiably harsh recommendation.
This is particularly revealing because the underlying recommendation was held constant while its apparent source was manipulated. The teachers were not necessarily responding to better evidence or a more convincing argument. They responded differently because of where they believed the recommendation had originated. The effect was also asymmetric, suggesting that trust in AI does not operate as a simple, uniform tendency to accept machine-generated advice. The content and direction of the recommendation matter, as do the expectations that individuals bring to the decision. The experiment was conducted using grading scenarios rather than observing routine marking practices, so its findings should not be treated as a direct estimate of the frequency of such behaviour in schools or universities. Nevertheless, it demonstrates something that ought to concern educational institutions: professional expertise does not guarantee that an individual will exercise independent judgement when presented with apparently authoritative automated advice.
These findings have implications beyond the familiar debate about students using AI to complete assignments. Consider the lecturer using AI to evaluate an essay, the supervisor asking AI to comment upon a dissertation, or the institution employing algorithmic systems to identify students requiring additional support. None of these activities is necessarily inappropriate, and in many circumstances AI may improve consistency or reveal problems that a human assessor has overlooked. Yet the presence of a qualified professional at the end of the process should not be mistaken for evidence that the recommendation has been independently evaluated. Indeed, it is plausible that repeated positive experiences with automated systems could encourage reduced scrutiny of future recommendations, although establishing such a process would require longitudinal evidence. This is a possibility deserving investigation rather than a demonstrated inevitability.
The clinician and the machine
Clinical practice provides a particularly useful comparison because medical judgement carries consequences that are often immediate, tangible and potentially irreversible. Clinicians are trained to interpret evidence, recognise uncertainty and consider alternative explanations. We might therefore expect professional expertise to offer considerable protection against inappropriate dependence on AI. The evidence suggests that this protection is imperfect.
Dratsch et al. (2023) investigated automation bias in mammography, examining how purported AI-generated BI-RADS assessment suggestions influenced radiologists' interpretations. Twenty-seven radiologists, representing different levels of experience, reviewed mammograms while receiving recommendations from what appeared to be an AI-based decision-support system. Importantly, this was a simulated AI interface displaying recommendations and heat maps prepared in advance, rather than the live output of a trained diagnostic model. This design allowed the researchers to control the accuracy of the advice presented to participants. Incorrect recommendations significantly affected performance across experience levels, although less experienced radiologists showed greater susceptibility in some conditions. The study demonstrates that the apparent authority of automated advice can influence clinical interpretation even when the practitioner possesses relevant professional expertise.
The phenomenon itself is not new. Parasuraman and Manzey (2010), reviewing research on the human use of automation, described how automation-related complacency and bias can affect novices and experts alike. Their analysis indicated that these effects involve attentional processes as well as decision-making biases, and cannot reliably be eliminated through simple instructions, practice or training. This provides an important theoretical foundation for understanding contemporary AI. The problem is not uniquely attributable to large language models. AI enters an established psychological landscape in which human beings can become excessively reliant upon systems that appear competent, consistent or authoritative.
More recently, Qazi et al. (2026) conducted a randomised clinical trial involving 44 physicians in Pakistan who had completed 20 hours of AI-literacy training. Participants evaluated six clinical vignettes, with one group receiving access to error-free ChatGPT-4o recommendations and the other receiving recommendations containing deliberately introduced errors in three of the six cases. Physicians were free to consult, modify or reject the AI advice. Nevertheless, those exposed to erroneous recommendations demonstrated poorer diagnostic reasoning performance than those receiving error-free advice. Mean diagnostic reasoning accuracy was 73.3% in the erroneous-advice group compared with 84.9% in the control group, with an adjusted difference of 14.0 percentage points. The adjusted estimate differs from the simple subtraction of the two reported means because it derives from the study's statistical analysis. The investigation provides evidence that erroneous LLM recommendations can adversely influence diagnostic reasoning despite previous AI-literacy training, although it does not establish whether either group would have performed better or worse without AI assistance altogether.
The broader clinical literature offers additional support. Zeng et al. (2026), in a systematic review and meta-analysis of erroneous clinical AI and automated decision support, examined 46 reports representing 41 independent cohort clusters. Their analysis found that clinicians were substantially more likely to reach correct decisions when automated recommendations were correct rather than incorrect. In four heterogeneous real-world studies involving computerised electrocardiogram interpretation, the pooled estimate suggested that 26.6% of erroneous automated interpretations were neither recognised nor corrected, although the confidence interval was wide and between-study heterogeneity was substantial. The authors emphasised that human oversight should be treated as an error-interception capability requiring empirical evaluation rather than as a guarantee of safety arising merely from the presence of a professional.
There is an important methodological caution here. Comparing performance with correct and incorrect automated advice does not, by itself, establish how much an individual's judgement has deteriorated relative to working without assistance. The clinical studies also differ in design, task, technology and outcome measures, and some involve conventional automated decision support rather than generative AI. These limitations matter, particularly when considering whether findings can be transferred directly to educational settings. Even so, the evidence challenges an assumption that seems to underlie much of the discussion about responsible AI adoption: that retaining a human decision-maker is sufficient to preserve the independence and quality of the decision. It may preserve responsibility, but the exercise of judgement is a separate empirical question.
The danger of a plausible answer
One of the more distinctive properties of contemporary generative AI is its capacity to produce explanations that are coherent, confident and readily intelligible. This is frequently an advantage. Explanations can make unfamiliar concepts accessible, provide alternative perspectives and help users identify weaknesses in their understanding. Yet a plausible explanation may also reduce the motivation to examine how a conclusion was reached. A recommendation need not be demonstrably false to influence judgement inappropriately. It may simply establish the assumptions around which subsequent reasoning takes place.
Lee et al. (2025) surveyed 319 knowledge workers, examining 936 examples of tasks undertaken with generative AI assistance. They found that greater confidence in AI was associated with less self-reported critical-thinking engagement, whereas greater confidence in one's own abilities was associated with more. Their qualitative findings suggested that AI was also changing the character of critical thinking, shifting activity towards information verification, response integration and the management of AI-assisted tasks. The study relied upon self-report and does not demonstrate a causal reduction in cognitive ability. Nor should reduced cognitive effort automatically be interpreted as a detrimental outcome; considerable value may lie in achieving a task with less unnecessary effort. Nevertheless, the findings raise an important possibility: as AI becomes increasingly integrated into intellectual work, users may spend less time constructing an initial judgement and more time evaluating a judgement already supplied to them.
This change may appear relatively inconsequential, particularly if the AI-generated recommendation is usually correct. But the order in which information is encountered can influence subsequent evaluation. Someone who formulates an interpretation independently and then compares it with an AI recommendation occupies a different cognitive position from someone who begins with the recommendation and subsequently attempts to evaluate it. In the latter case, the machine-generated response may establish the initial frame within which alternatives are considered. The individual can still disagree, but disagreement may now require departing from an apparently reasonable position rather than developing a position independently. This is a theoretical interpretation consistent with the broader literature on cognitive bias and automation; the studies considered here do not establish that the sequence always has this effect.
My concern is not simply that people might stop thinking. That seems an improbable and rather crude account of human adaptation to technology. More interesting is the possibility that people continue to think, sometimes quite extensively, but do so within a framework whose assumptions have already been established elsewhere. The output may be scrutinised, corrected and refined without its underlying premises ever becoming the object of serious examination. This is a potential form of judgement displacement that may be difficult to detect by looking only at the quality of the final answer.
There is an obvious counterargument. Human reasoning has never been entirely independent of external influence. Academic judgement is shaped by disciplinary traditions, peer review, textbooks, institutional expectations and the perceived authority of established scholars. A student accepting a professor's explanation without question is hardly demonstrating greater intellectual independence than a student accepting an AI-generated explanation. Nor should we overlook the possibility that AI may challenge existing authority, allowing individuals who lack conventional credentials or institutional influence to explore arguments that might otherwise remain inaccessible to them. The capacity to interrogate an apparently authoritative human judgement with AI may be just as valuable as the capacity to interrogate an AI judgement with human expertise.
That counterargument is persuasive, and it prevents the discussion from collapsing into another defence of human exceptionalism. The problem is not that AI possesses an illegitimate claim to intellectual authority while human experts possess a legitimate one. Authority should be earned through the quality of evidence, reasoning and demonstrated reliability, regardless of its origin. The concern is that AI can generate the appearance of such authority cheaply and at scale, while providing a finished interpretation before the recipient has necessarily developed a basis for evaluating it. The same criticism may be directed at human institutions, but the speed, accessibility and apparent fluency of AI make the problem particularly relevant.
What education is supposed to produce
The educational implications become more significant when we distinguish the completion of academic work from the development of intellectual competence. Universities have traditionally assessed learning through the production of essays, reports, examinations, presentations and other observable outputs. Such outputs are imperfect indicators of understanding, but they have served as practical evidence that students can acquire, organise and communicate knowledge. Generative AI complicates this relationship because the production of a competent output may no longer require the same degree of intellectual engagement from the person submitting it. This does not render the output worthless, but it weakens the assumption that its quality provides direct evidence of the student's independent capabilities.
There is a temptation to respond by concentrating upon the security of assessment, returning to invigilated examinations, oral defences or other mechanisms intended to establish individual authorship. These approaches may have a legitimate role, particularly where independent competence must be demonstrated. Yet the question of whether a student produced an answer independently is not identical to the question of whether the student has developed sound judgement. An individual can memorise disciplinary conventions, reproduce accepted interpretations and follow prescribed analytical procedures without demonstrating much capacity to question their assumptions. Conversely, a student may use AI extensively while developing considerable understanding through comparison, criticism and revision. Treating the absence of AI as evidence of intellectual independence would therefore be as misguided as treating the presence of AI as evidence of intellectual dependence.
The mathematics experiment reported by Bastani et al. (2025) is instructive here because it demonstrated that the design of AI assistance matters. A system that encourages students to work through a problem, consider hints and attempt solutions can support a rather different relationship with learning from one that simply supplies the completed answer. This distinction accords with a more general educational principle: assistance should be evaluated not merely according to what a learner can achieve while receiving it, but according to what the learner subsequently becomes capable of doing independently. There is nothing inherently virtuous about struggling unnecessarily with a task, but there is considerable educational value in developing the competence to recognise when an answer is wrong, when a procedure is inappropriate and when the question itself requires reconsideration.
For higher education, I would argue that judgement deserves more explicit attention as an educational outcome. This need not involve another elaborate framework of graduate attributes or an institutional exercise in converting intellectual development into measurable indicators. Universities already claim to value independent thought, critical enquiry and ethical reasoning. The difficulty is establishing what these aspirations mean when increasingly capable systems can undertake many of the activities through which they have traditionally been demonstrated. It may be more useful to distinguish between the ability to execute a procedure, the ability to evaluate its output and the ability to question whether the procedure should have been undertaken in the first place.
The last of these brings me back to the boy who stood on his desk. The lesson was not that established procedures should always be disregarded, nor that unconventional behaviour is inherently evidence of superior judgement. Indeed, after the exercise, I suspect many of us developed the habit of looking at the final question before starting another examination paper. The teacher may simply have replaced one procedure with another. What made the incident memorable was the recognition that conscientious effort could coexist with a failure to understand the larger task. We had demonstrated the competence to carry out instructions without first establishing which instructions were relevant.
The limits of human oversight
One response to the risks of AI-assisted decision-making is to insist that people remain in control. Human oversight has become a familiar principle in discussions of responsible AI, particularly in education, medicine and other areas involving consequential decisions. It is an understandable requirement, but one that deserves closer examination. Oversight can mean anything from independently evaluating the evidence and considering alternative explanations to approving a recommendation that appears reasonable. These activities are not cognitively equivalent, although institutional procedures may treat them as such.
The clinical and educational findings considered here suggest that oversight should be assessed in terms of what people actually do when confronted with erroneous advice. Can they identify the error? Do they recognise when the recommendation rests upon an inappropriate assumption? Does the perceived authority of the system change their willingness to disagree? Most importantly, can they construct and defend an alternative judgement without relying upon the system whose recommendation they are supposed to evaluate? These questions concern the quality of human oversight rather than its formal presence.
There is also a potential difficulty with AI-literacy programmes that concentrate primarily on teaching users to recognise hallucinations, verify references and formulate effective prompts. Such skills are useful, but they may not address the more fundamental issue of judgement displacement. The physician who has learned to check an AI-generated diagnosis may still be influenced by the diagnosis before beginning that evaluation. The academic who verifies an AI-generated interpretation may still overlook the assumptions upon which the interpretation depends. Training people to check an answer is not necessarily equivalent to developing their capacity to establish what constitutes an appropriate answer in the first place. The findings of Qazi et al. (2026) are particularly relevant here, although their study does not establish which specific elements of AI-literacy training succeed or fail in protecting professional judgement.
This is not an argument for excluding AI from professional practice or education. Evidence from Goh et al. (2024), for example, demonstrates the complexity of human–AI collaboration. In their randomised clinical trial involving 50 physicians, those provided with access to GPT-4 did not significantly outperform physicians using conventional resources. The model, when evaluated independently, achieved a higher score than the conventional-resources physician group, although that comparison does not establish the effect of deploying it independently in real clinical practice. The capacity of a technology to produce good answers does not automatically translate into improved performance when people use it. Equally, the fact that people sometimes defer inappropriately to AI does not mean that independent human judgement is invariably preferable. The central problem is how to combine machine capability with human understanding in ways that improve decisions without obscuring responsibility or diminishing the ability to recognise error.
I would therefore resist framing the issue as a competition between human and artificial intelligence. The question is not which possesses the superior judgement in the abstract, but how judgement should be organised when different sources of intelligence contribute to the same decision. Human beings are inconsistent, susceptible to bias and often remarkably poor at evaluating evidence that conflicts with their expectations. AI may correct some of these weaknesses while introducing others. A sensible approach must acknowledge both possibilities and avoid treating either human expertise or machine competence as sufficient grounds for unquestioning trust.
Towards a psychology of judgement displacement
The evidence reviewed here suggests a potentially useful direction for further research. Cognitive offloading provides an established framework for understanding how individuals delegate aspects of information processing, while automation-bias research demonstrates that apparently authoritative recommendations can influence decisions inappropriately. What remains less clearly established is how repeated interaction with systems capable of producing complete interpretations and recommendations changes the initiation, construction and evaluation of human judgement over time.
I propose that judgement displacement might be investigated as a process in which an externally generated evaluation increasingly becomes the starting point for an individual's reasoning, rather than one source of evidence considered alongside an independently formulated position. I use the term descriptively, rather than claiming to introduce a wholly new phenomenon or a previously unrecognised construct. It is not intended to describe every instance of AI-assisted decision-making, nor does it imply that accepting external advice is necessarily irrational. Appropriate reliance upon a more reliable source may itself demonstrate sound judgement. The proposed distinction concerns whether individuals remain capable of recognising when reliance is inappropriate, and whether repeated exposure to externally generated recommendations changes their tendency to formulate independent assessments before encountering advice.
An empirical approach might compare participants required to make an initial judgement before receiving AI assistance with those who encounter an AI recommendation before forming their own position. Additional conditions could require participants to justify agreement or disagreement, consider alternative interpretations or explicitly evaluate the assumptions underlying the recommendation. AI advice could be manipulated for accuracy, apparent authority and expressed confidence, while subsequent performance could be assessed without assistance. Such a design would allow researchers to distinguish appropriate reliance from inappropriate deference and examine whether the sequence in which recommendations are encountered affects judgement. Longitudinal assessment would be necessary before making stronger claims about persistent changes in independent competence.
There is also an important distinction between an incorrect final judgement and a compromised decision-making process. A person may independently arrive at an incorrect conclusion after carefully considering the evidence, while another may accept a correct AI recommendation without understanding why it is correct. If the sole outcome measure is accuracy, the latter appears superior. In some circumstances, particularly where immediate consequences are substantial, that may be an entirely reasonable practical assessment. Yet education has a different responsibility. It must consider not only whether the individual reaches the correct answer today, but whether they develop the competence to make and evaluate judgements tomorrow.
This creates an interesting tension between the immediate benefits of AI assistance and the longer-term educational value of independent reasoning. A system that consistently provides better recommendations may improve current performance while reducing opportunities to develop the expertise required to scrutinise future recommendations. That possibility remains a hypothesis rather than an established account of long-term cognitive change. Nevertheless, the experimental evidence on learning and automation bias provides sufficient grounds to investigate it seriously. The danger may not be that people lose the ability to think, but that the circumstances in which they consider independent thought necessary gradually change.
The final question
I began with a schoolroom exercise that rewarded a boy for doing something the rest of us regarded as absurd. He stood on his desk while his classmates continued diligently answering questions that, as it turned out, did not require answers. For years I regarded it as a mildly amusing lesson in reading instructions properly. I now think it raises a more interesting question about the relationship between competence and judgement, particularly in a world increasingly populated by systems capable of producing convincing interpretations on our behalf.
I remain broadly optimistic about artificial intelligence. I see no convincing reason to suppose that human beings possess a permanent monopoly on intelligent reasoning, and I have little sympathy for arguments that defend existing professional activities simply because they have traditionally been undertaken by people. AI may help individuals question established authority, expose weak arguments and develop intellectual confidence in circumstances where conventional institutions have provided little encouragement. These are substantial benefits, and they deserve recognition alongside the risks.
Yet optimism about the capabilities of AI should not require indifference to the psychological consequences of relying upon it. If education is concerned with developing people who can exercise judgement, then it must pay attention to the conditions under which that judgement is formed. If clinical practice requires accountable decisions, it must establish whether professional oversight amounts to genuine evaluation rather than the formal approval of a plausible recommendation. The same principle applies wherever machine-generated interpretations increasingly influence what people believe and how they act.
Perhaps the lesson I remember from school was not really about reading instructions at all. It was about the ease with which competent people can become absorbed in the immediate demands of a task while overlooking its wider purpose. Artificial intelligence may help us escape that limitation by offering perspectives we would otherwise never have considered. Equally, it may provide such convincing accounts of what we ought to do that we become less inclined to ask whether doing it was sensible in the first place.
The difficulty is that there is no final question at the bottom of the page, and no teacher waiting to applaud when we discover it.
We will have to decide for ourselves when to stand on the desk.
References
Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), Article e2422633122. https://doi.org/10.1073/pnas.2422633122
Dratsch, T., Chen, X., Mehrizi, M. R., Kloeckner, R., Mähringer-Kunz, A., Püsken, M., Baeßler, B., Sauer, S., Maintz, D., & Pinto dos Santos, D. (2023). Automation bias in mammography: The impact of artificial intelligence BI-RADS suggestions on reader performance. Radiology, 307(4), Article e222176. https://doi.org/10.1148/radiol.222176
Goh, E., Gallo, R., Hom, J., Strong, E., Weng, Y., Kerman, H., Cool, J. A., Kanjee, Z., Parsons, A. S., Ahuja, N., Horvitz, E., Yang, D., Milstein, A., Olson, A. P. J., Rodman, A., & Chen, J. H. (2024). Large language model influence on diagnostic reasoning: A randomized clinical trial. JAMA Network Open, 7(10), Article e2440969. https://doi.org/10.1001/jamanetworkopen.2024.40969
Goulas, S., Megalokonomou, R., & Sotirakopoulos, P. (2026). Why do experts miss AI's errors? Evidence from a randomized labeling experiment. PNAS Nexus, 5(6), Article pgag146. https://doi.org/10.1093/pnasnexus/pgag146
Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The impact of generative AI on critical thinking: Self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (Article 1121, pp. 1–22). Association for Computing Machinery. https://doi.org/10.1145/3706598.3713778
Parasuraman, R., & Manzey, D. H. (2010). Complacency and bias in human use of automation: An attentional integration. Human Factors, 52(3), 381–410. https://doi.org/10.1177/0018720810376055
Qazi, I. A., Ali, A., Khawaja, A. U., Akhtar, M. J., Sheikh, A. Z., & Alizai, M. H. (2026). Automation bias in large language model-assisted diagnostic reasoning among physicians trained in AI literacy—A randomized clinical trial. NEJM AI, 3(5), Article AIoa2501001. https://doi.org/10.1056/AIoa2501001
Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676–688. https://doi.org/10.1016/j.tics.2016.07.002
Wong, R., Reichle, E. D., & Veldre, A. (2025). Prediction in reading: A review of predictability effects, their theoretical implications, and beyond. Psychonomic Bulletin & Review, 32(3), 973–1006. https://doi.org/10.3758/s13423-024-02588-z
Zeng, D., Lei, J., & Li, Z. (2026). Human oversight of erroneous clinical AI and decision support: A systematic review and meta analysis. npj Digital Medicine. Advance online publication. https://doi.org/10.1038/s41746-026-03408-5