To skill or never-skill: first-hand perspective of AI’s growth and turmoil

My insight into higher education was at a midpoint, when AI was developing rapidly. Across higher education, it increasingly offered answers that appeared more sophisticated and convincing than those from a typical Google search, while potentially reducing the time required for tasks such as generating alternative search terms. Whilst I never used AI to generate my ideas or produce content in my assignments (I remained mindful of the University’s guidance on appropriate AI use), I noticed it had become embedded in others’ daily routines, where ‘thinking’ became obsolete, as AI could do it more quickly, without the frustration that comes with learning. In this reflection, I consider how I developed these skills before AI became deeply embedded in education, what that experience tells me about never-skilling, and how this may shape my professional practice as a soon-to-be educator.

Going back in time. During my senior years of secondary school, I remember a select few classes where I felt frustrated and overwhelmed. Maths, English and Chemistry come to mind. Whether that was getting mathematical equations correct through countless rounds of trial and error, polishing my fictional writing skills through countless pieces of feedback from my teacher on grammar, structure, and storytelling, or being unable to skip the process of titration to get the correct concentration. Whilst I don’t use algebra, fictional writing and titrations in my everyday life, that is not really the point. I remember many of the mistakes more vividly than the answers I got right; I learned not to take shortcuts because, quite simply, those shortcuts were not available to me. As a student teacher, I now have a greater appreciation not only for the resilience of my teachers, but also for my own resilience in continuing to try.

Cognitive psychology helps explain why this struggle may have mattered. The concept of desirable difficulties suggests that learning that is harder and more error-prone may lead to stronger long-term learning than learning that feels easier in the moment (Binks, 2026; Bjork & Kroll, 2015). In that sense, some of the frustration I experienced was not separate from the learning. It was part of it. This also reflects the importance of Vygotsky’s Zone of Proximal Development (ZPD) in education. As learners develop, we should continually re-evaluate what they can do independently, what they can achieve with support, and how much challenge is appropriate to keep moving their learning forward (Eun, 2019). So why am I telling you this? Fast-forward a couple of years, and I am in first year of university, learning what felt like an entirely new language: Psychology.

Unlike at school, studying psychology at university required me to either build upon or develop a whole new set of academic skills. A few include learning how to conduct a literature search for peer-reviewed sources using techniques like Boolean searches, how to write scientific reports, how to construct question-focused essays with arguments supported by evidence and subsequently critiqued, and how to reference accurately using APA 7. I remember receiving feedback on my first essay. Whilst it achieved a strong 2:2, it diverged too much from the question at hand, specifically into a tangent on the history of operant conditioning. But aren’t we all susceptible to going off script? Nonetheless, I also remember receiving feedback in my fourth year, after four years of APA7 referencing, and somehow I still had not grasped that multiple citations within the same brackets had to be ordered alphabetically. Whilst a small technicality, it brings me back to my former point about the importance of mistakes.

This brings me to the concept of never-skilling (Ke et al., 2026), in which a skill is never truly acquired because something else performs enough of the work that the learner can bypass developing it themselves (e.g., cognitive offloading). Statistics was probably an area where I came closest to understanding its appeal. I remember submitting an assignment in which I conducted the totally wrong statistical analysis and receiving a poor grade as a result. At the time, the feedback was frustrating because I had misunderstood something quite fundamental to the field. Looking back, however, it also makes me reflect on how statistics was taught. Perhaps the practical labs, where we often replicated pre-written statistical analysis scenarios in blocks, did not prepare me well enough to select the appropriate analysis for statistical exams, as these exams utilised interleaving techniques for question presentation (e.g., t-test, ANOVA, Regression, etc.) rather than blocking (e.g., t-test, t-test, Regression, Regression, etc.) like we were taught, as with most exams; thus the learning process should have reflected that of interleaving also.

Does this, then, create pressure to drift towards never-skilling because sitting with uncertainty and discomfort is difficult? As AI became increasingly visible during the latter half of my degree, that question became increasingly relevant among students around me. Why struggle with deciding which ANOVA is appropriate when a tool can appear to make that decision for you? But that was not how I learned it. Instead, I returned to the university-provided statistical analysis flowchart, worked through the decision-making process, and gradually became better at identifying which analysis was appropriate. Yet, if that scaffold was taken away, I suspect I would have struggled. Have I therefore truly developed the skill of independently selecting statistical analyses, or have I become skilled at navigating the flowchart? The flowchart itself was not the problem. In many ways, it was exactly what I had needed at that stage of learning. Instead, my reliance on it made me question how much of the decision-making process I had actually learned for myself.

Reflecting further back, scaffolding was hardly new to me. During the senior phase of school, many lessons used model answers and structured approaches to help us prepare for external examinations in April and May. At times, I felt that the focus became less about learning the subject and more about learning how to perform successfully in the examination. After all, we wanted to get into university, didn’t we? But university felt different. There was considerably more independence, and I still remember one of my law professors making the distinction particularly clear: their job is not to teach us how to pass the exam, but to teach us the law. At the time, I was puzzled because didn’t we want to do well? How could I do well if I’m not taught what they expect? Only time would tell.

I think something important lies in that distinction. A model answer can help someone understand what a strong response looks like, just as my statistical flowchart broke a difficult decision into smaller questions and helped me recognise which analysis was appropriate. These supports make learning possible when the learner cannot yet perform the task independently. But success with a scaffold does not necessarily mean you have acquired the underlying skill. That raises a broader question. At what point does support help us to develop a skill, and at what point does the support begin to perform so much of the thinking that we become dependent on it? Perhaps the purpose of a good scaffold is not simply to help us arrive at the correct answer, but eventually to make itself unnecessary.

The difference with AI may not be that it is entirely separate from the scaffolds we have always used, but that it can take on far more of the cognitive work. A model answer still requires the learner to construct their own responses. A statistical framework still requires them to identify the relevant variables and follow the reasoning through. AI, however, can potentially generate the finished product itself, often in a way that feels convincing and fluent. This makes it difficult to see where the support tool ends, and where the learner’s reasoning begins.

The same question becomes more consequential when the learner is no longer preparing for an examination but instead preparing for professional practice. If a trainee psychologist can use AI to help produce a 5P formulation, the finished formulation may tell us relatively little about how much of the reasoning belongs to the trainee. Can they consider alternative interpretations? Can they recognise when a convincing formulation does not fit the needs and circumstances of the client in front of them? Do they really understand the ethical implications of how that formulation was produced? Perhaps this is where the flowchart question becomes particularly useful. If the scaffold is removed, what remains? In clinical training, the concern may not simply be whether trainees can use AI effectively, but whether they have developed enough knowledge and reasoning to know when not to trust it.

This matters because students may encounter AI before they develop the underlying skills independently. Sok et al. (2025) found that, among the 315 high school students surveyed, many favoured using AI to complete schoolwork and expressed little concern about a potential reduction in critical thinking. Given the study's recency, some of these students may enter professional training in fields where independent judgment is necessary and critical reasoning is especially important. Ke et al. (2026) raise similar concerns, particularly about medical students’ clinical uncertainty and the appeal of AI doing the formulation work for them.

As I move into teaching, this leaves me considering how we support students without removing the very struggle through which skills are developed. I don't think the answer is to reject AI altogether, but to be more deliberate about when to introduce it and what students should be able to do for themselves first. Mistakes, uncertainty and trial and error are not always signs that learning has gone wrong; often, they are part of how learning happens. If we remove those experiences too early, we risk producing learners who can access answers without fully understanding how those answers were reached.

I leave you with this final question: How can we question an AI chatbot’s thinking if we have yet to think for ourselves?

References

Binks, S. (2026). Why Desirable Difficulties ‘Work’: A Review of the Evidence From Cognitive and Educational Psychology and Some Caveats for the Health Professions Education Field. Journal of Evaluation in Clinical Practice, 32(1). https://doi.org/10.1111/jep.70349

Bjork, R. A., & Kroll, J. F. (2015). Desirable Difficulties in Vocabulary Learning. The American Journal of Psychology, 128(2), 241–252. https://doi.org/10.5406/amerjpsyc.128.2.0241

Eun, B. (2019). The zone of proximal development as an overarching concept: a framework for synthesizing Vygotsky’s theories. Educational Philosophy and Theory, 51(1), 18–30. https://doi.org/10.1080/00131857.2017.1421941

Ke, Y., Jin, L., Ong, J. C. L., Thirunavukarasu, A. J., Car, J., Cheung, C. Y., Tham, Y. C., Ting, D. S. W., Ong, M. E. H., Compton, S., Narayan, A., Keane, P. A., Wong, T. Y., Bates, D. W., Tan, P., & Liu, N. (2026). AI-induced never-skilling in medical education. Nature Medicine. https://doi.org/10.1038/s41591-026-04438-y

Sok, S., Heng, K., & Pum, M. (2025). Investigating High School Students’ Attitudes Toward the Use of AI in Education: Evidence from Cambodia. Sage Open, 15(3). https://doi.org/10.1177/21582440251353575

Scott Macmillan, GMBPsS, is an Assistant Psychologist at Illuminated Thinking.