GPTs and the Choice Architecture of Pedagogies in Vocational Education

Abstract

Generative pre-trained transformers (GPTs) have rapidly entered educational contexts, raising questions about their impact on pedagogy, workload, and professional practice. While their potential to automate resource creation, planning, and administrative tasks is widely discussed, little empirical evidence exists regarding their use in vocational education (VE). This study explores how VE educators in England are currently engaging with AI tools and the implications for workload and teaching practice. Data were collected through a survey of 60 vocational teachers from diverse subject areas, combining quantitative measures of frequency, perceived usefulness, and delegated tasks with open qualitative reflections. Descriptive statistics, cross-tabulations, and thematic analyses were used to interpret responses about the application and allocation of work given by teachers to GPTs. Findings indicate cautious but positive adoption, with most educators using AI tools infrequently (0–10 times per month), yet rating them highly useful (average 4/5) for supporting workload. Resource and assessment creation dominated reported uses, while administrative applications were less common. The choice architecture framing indicates that some GPTs guide teachers to certain resources over others and the potential implications of this are discussed. Qualitative insights highlighted concerns around quality, overreliance, and the risk of diminishing professional agency. The study concludes that GPTs offer meaningful workload support but require careful integration, critical evaluation, and professional development to ensure they enhance rather than constrain VE pedagogy.

1. Introduction

Introduction to GPTs and Their Potential

The attraction of large, or neural, language models (LLMs), but commonly known as generative pre-trained transformers (GPTs), lies in their ability to generate quality text in response to user prompts. Pre-trained on vast datasets, they underpin a range of chatbot applications capable of diverse tasks. Commercially prominent examples include ChatGPT (developed by OpenAI, San Francisco, CA, USA); TeacherMatic (first developed by Innovative Learning Technologies Limited, Avallain, St. Gallen, Switzerland), Gemini (developed by Google, Mountain View, CA, USA); Claude (developed by Anthropic, San Francisco, CA, USA) and Deepseek (developed by Deepseek, Hangzhou, China); while alternative platforms based on similar deep-learning mechanisms also exist to process, sequence, and analyse data for complex problem-solving [1]. These technologies are primarily online, though some can now run offline; for example, transformer networks are embedded in Acrobat PDFs, and open-source alternatives found on community developer websites such as https://huggingface.co/spaces (accessed on 25 September 2025), which develop free alternatives that can often be operated on desktop central processing units without internet access, showing great advances in the short lifespan for LLMs.

Educators have quickly recognised the potential of GPTs to automate time-consuming tasks such as marking, planning lessons or courses, and administrative duties [2], with specialised platforms such as Teachermatic developed specifically to assist teachers with their varied tasks. Offloading tasks such as designing lesson plans or differentiating learning tasks to be more individualised has led to speculation about the extent to which these models can save teachers’ time and improve efficiency [3]. However, ethical concerns arise where biases embedded in training data [4] may influence outputs, with implications for fairness and equality in education [5]. Similarly, the ability of LLMs to generate credible student essays has disrupted conventional forms of assessment, raising concerns about plagiarism detection and academic integrity [5,6].

These debates acquire particular importance in vocational education (VE), often referred to as the skills sector in England, or technical education. VE emphasises procedural knowledge supported by spatial, technical, and psychomotor skills, as well as creative and experiential learning. Despite being chronically underfunded and frequently overlooked [7] compared with schools and higher education [8], VE must remain innovative and responsive to new technologies. Teachers in this sector are therefore at the forefront of negotiating how tools like GPTs might be used not only to enhance teaching and learning but also to improve efficiency and productivity in challenging working conditions [8,9] that extend globally around issues of shortages in teacher recruitment and retention [10] and how technologies may represent one intervention to supporting overburdened teachers and improving students’ learning experiences [9,11]. Therefore, while situated in the English vocational context, the paper draws from a number of international sources that illustrate how issues such as high workload, the cognitive offloading of tasks to AI, and the potential automation of teaching with high satisfaction tasks, such as resource creation, are distributed to chatbots or GPT platforms. Simultaneously, there is potential premise in these platforms to help personalise learning experiences and reduce administrative tasks that contribute to teacher attrition, while VE globally needs revitalisation through the exploitation of digital technologies.

The English vocational sector’s engagement with digital innovation has been influenced by reports such as FELTAG (2014), which recommended that substantial portions of FE teaching should move online [8]. This reflects wider global considerations of how to improve vocational education [11] in order to align it with changes in real world working practices common to the 4th Industrial Revolution, which optimises data, robotics, and automation [12]. Although predating COVID-19, the report highlighted the need to future-proof the sector through digital literacies. Teacher digital competence is central to this agenda, as recognised in models such as TPACK [13] and the Technology Acceptance Model (TAM) [14], which emphasise the interplay of technical, pedagogical, and content knowledge, as well as ease of use in technology adoption. Similarly, the SAMR framework identifies four levels of technology use, ranging through the ways in which technology can impact teaching and learning tasks via the substitution, then enhancement of traditional methods to the augmentation of them, before modification becomes a layer in the redesign, and finally redefinition as the ultimate stage. Critics [15] note how SAMR prioritises product over process, a problematic emphasis in VE where experiential learning, learning by accident or discovery, and a reflection and embracement of mistakes can be essential [3,12]. Teachers in VE therefore globally face different challenges: integrating technology, reflecting the real world with ever shifting dynamics, digital competencies, and innovative pedagogy [11,16,17].

While GPTs offer accessibility and potential cost savings, they also bring challenges of hype, inflated expectations, and a susceptibility to generating errors [18]. Ethical concerns about bias and inequality persist, even as platformisation makes such tools increasingly embedded in educational environments [19]. Beyond technical issues, the sector faces systemic pressures: high workloads, teacher retention and recruitment crises, and poor funding [7,20]. Policymakers have suggested that AI might alleviate teacher workload through resource creation, curriculum planning, personalised feedback, and administrative support [21]. Yet, this raises questions about teacher agency and professionalism, since excessive reliance on automation may risk eroding educators’ autonomy and hindering their professional development [12,21].

Where we can identify the application of AI as solution, as it is so often presented, we can also identify tangible systemic problems [22]. Workload inevitably emerges as a key area where GPTs may provide tangible support. Teachers have a broad range of tasks and responsibilities that go beyond the classroom and pedagogical planning and include vast amounts of administration and data analytics about student learning and attendance [23]. GPTs may be able to streamline these processes [3], freeing time for more creative and relational aspects of teaching that would arguably give more satisfaction that prevents teacher attrition [10]. However, because GPTs are stateless and lack contextual awareness of classroom dynamics, their outputs often require adaptation. Teachers’ agency is thus preserved through the refinement and contextualisation of AI-generated resources. Importantly, studies suggest that the usefulness of GPTs depends on teachers’ subject expertise and ability to craft effective prompts [24], with weaker outcomes for those lacking these skills [25,26].

This paper therefore investigates how vocational educators are currently using GPTs, particularly with respect to workload management. We ask how these technologies influence teachers’ professional development, decision-making, and delegation of tasks to AI tools. Although our sample spans varied English VE contexts, this diversity highlights the need for a more unified framework for effective GPT use in the sector. Such a framework could guide prompt engineering practices and ensure that AI integration enhances rather than undermines teachers’ professional roles. An important argument towards our conclusion is for teaching professionals to share repositories of their GPT-generated resources and materials, with explanations of the prompts given and objectives sought, so a bank of those is available across different subject professionals to access and re-use. This goes some way towards a unified framework, since it highlights the need for the adjustment of resourced outputs initially produced by GPTs.

2. Materials and Methods

This study presents findings from a survey of 60 vocational educators regarding their use of artificial intelligence (AI) tools in vocational educational settings. The authors wished to develop an understanding of how vocational educators were starting to use AI in their settings and contexts. This involved asking about the types of tools understood as being ‘AI’, their application and purpose, their frequency of use and their perceptions of usefulness in helping with workload. Other contextual information about the educators’ length of teaching experience and subject specialism within teaching was drawn in to try to ascertain whether newly qualified or more experienced teachers had differing views of the uses of GPTs. Finally, an open question was added to the survey in order to elicit views and values about the emergence of AI into education (‘Is there anything else you would like to share about your experience with using AI Tools in education?’). The reason for these areas to be covered in the survey was based on two things. Firstly, the corresponding literature about AIE and claims around its potential impact on workload, as well as automation generally arising from this fourth industrial revolution, which linked to enquiries around application of the platforms. Secondly, enquiries that were framed around the tools influence on the ways in which they teach, the purposes of the tools and their perceived usefulness were linked to previous research [27] undertaken during practical workshops were teacher participants speculated two years before about the potential influence of these tools on teachers’ workloads and subject specialisms. Both approaches were designed to help us to understand the types of tasks teachers were attributing it to, in order to understand if AI is potentially enabling more pedagogical creativity for teachers or whether it supported them with the bulk of their administrative tasks. The survey was designed by the writers without being validated but was designed on the interests pursued and few other straightforward instruments that were available in publication.

Quantitative analysis was used to identify patterns, and a reading of the qualitative data was applied through a deductive approach. A deductive (a priori) coding approach was applied, using pre-identified codes, explained below, derived from the research questions and relevant literature. The data were then reviewed against these codes, with illustrative extracts selected to demonstrate the extent to which participants’ perspectives aligned with or challenged the framework. The codes included 1. ‘instances of a criticality by teachers towards AIE platforms’ and 2. ‘comments on the quality of AI output.’ Within the general survey approach of investigating how AIE was influencing teachers and their workload, we anticipated that our questions would stimulate responses that touched upon these themes. This was done in order to try to identify insights about the quantity of tasks most given to GPTs by these teachers and whether some comprise a majority, thereby seeking to identify the main uses of the platform by these participants but also how they perceive the generated responses.

2.1. Limitations of the Study

We acknowledge the limitations in this research design. The survey was not subject to validity checks, such as reliability or construct validity, and was designed under pragmatic constraints in order to capture perspectives during a window of opportunity when teachers were likely to be able to reflect on their uses of AI across the college year. This presents us with opportunities for future methodological improvements to the research design, where the survey would be subject to scrutiny from participants for content validity. While more extensive psychometric testing (e.g., factor analysis, test–retest reliability) was beyond the scope of this study, future research could further strengthen construct validity through confirmatory analysis and broader piloting.

2.2. Participants

The small sample size of 60 VE teachers was purely down to who responded to the survey. Survey responses were drawn from requests to participate from VE workplaces, through a learning technologies news bulletin and across social media channels by the authors and peers. This was not meant to be demonstrative or generalizable to VE educators nationally or further afield but a contextual insight from who was available to us and responded to our call for participants. Further, we note that users of social media or readers of a learning technologies newsletter may be more likely to be innovative adapters and users of technologies in their teaching than is probably normally representative of VE staff nationally, but we count this as an advantage of the study rather than a limitation, since they may be more able to yield insights through the opportunity for open questions. In all, the sample was closer to a convenience sample than a random one, since we engaged with who was available for the short window of five months during which we wished to collect the data.

Participants were sought from as wide a field of representation in VE as possible. We launched the survey in February and closed it in July, hoping to maximise the potential for respondents during what was a working period of high intensity but which entered into a lighter period of post-assessments. This means responses are likely to be those with an interest in technology or with a workload light enough to reply to questionnaire requests, with either potentially leading to bias within the results they respond with. Ultimately, participants generally ranged from across VE subject specialisms, including subject experts from Carpentry and Joinery, Maths, Counselling, Sports, English and ESOL (English as a Second Language), Business, Engineering, English, Computing, Bakery and catering, Travel and Tourism and Special Educational Needs support. This gives a varied and realistic representation of the typical VE subjects taught in English college systems. The range of teaching experience is presented in Figure 1 as drawn from the survey itself and shown here as an illustration of the teaching backgrounds of participants. As can be seen, the 60 participants have a balanced representation across early-, mid- and late-career experiences, but we state that these results are not reflective of generalizable output.

Survey data was collected via a Microsoft Forms survey and was a mixture of information gathering questions, a Likert scale ranking and an open question giving qualitative insights into one combined dataset. Given the high workload, we were fixed on creating a straightforward questionnaire that did not demand much of teacher’s time but allowed for depth and reflection. The questions are shown below:

• What is your role at the College and subject specialism?
• How many years teaching experience do you have?
• How often do you use artificial intelligence tools (such as TeacherMatic, Gemini, ChatGPT, etc.) for work each month?
• Which is the AI tool that you use most frequently?
• On a scale of 1–5 Stars (1 being ‘Not at all useful’ and 5 being ‘Extremely useful’), how would you rate these tools in terms of supporting your workload?
• Which functions of these AI tools do you use most frequently? (Select up to three.)
• How often do you need to adapt the output created by AI?
• In general, do you think AI tools are worthwhile for educators?
• Is there anything else you would like to share about your experience with using AI tools in education?

3. Results

As mentioned, participants provided details on their teaching subject and experience, their frequency of AI tool usage, usefulness ratings, task types assigned to AI, and reflective comments. This straightgforward manner allowed us to gather this information and present it below.

3.1. Key Quantitative Findings and Descriptive Results

Overall sample

• Total responses: 60

3.1.1. AI Use Frequency (Monthly)

Most teachers are low-to-moderate users, as the modal category is low-use (0–5/month), illustrated in Figure 2. There are a few heavier users, but they are a small minority. The majority of participants (21) reported using AI tools 0–5 times per month. This was followed by 14 users reporting 6–10 uses, with smaller groups using tools more intensively (11–20 or 21+ times). These occurrences in Figure 2 suggest a cautious adoption phase, with a few outliers who are heavy users. There appeared to be no connection between heavier users and duration of teaching experience.

3.1.2. Usefulness Ratings (1–5)

Despite some low instances of use per month, teachers clearly rated the usefulness of AI tools highly for supporting their workload, with an average score of 4.03 out of 5, shown in Figure 3. The most common ratings were four and five stars, indicating that many educators found these tools beneficial.

• Q. How helpful do teachers find AI for supporting workload?

Teachers who rated usefulness generally considered AI tools helpful for workload (mean = 4/5), shown in Figure 3. There is variation (some low ratings), but the distribution skews positive. Higher usefulness ratings showed that those participants mostly used it for three main purposes with assessment purposes (such as creating assessment tests) the most common, followed by lesson planning, and resource creation, such as worksheets or presentations. Fewer of the respondents reported using the platforms to take on administrative tasks, such as report writing or tracking, etc. Overall tasks delegated to GPTs are shown below.

3.1.3. Tasks Delegated to GPT/AI Systems

As shown in Figure 4, the most common task that was assigned to AI was resource creation (48 mentions were made in responses in the survey, equating to 40% of participants who responded), followed by assessment creation (35 mentions), and lesson planning (27 mentions).

Further analysis now follows, based on the following.

• Cross-tabulation of usefulness by frequency of AI use
• Correlation between teaching experience and AI adoption
• Participant case examples and illustrative quotes

3.1.4. Cross-Tabulation and Group Comparisons

Mean usefulness (with respondent count) by frequency:

• 0–5—n = 16, mean usefulness ≈ 4.00, sd ≈ 1.06
• 6–10—n = 12, mean usefulness ≈ 4.08, sd ≈ 0.67
• 11–20—n = 3, mean usefulness ≈ 4.33, sd ≈ 1.15
• 21+—n = 1, mean usefulness = 5.0
• Never used—n = 1 (no usefulness rating)

Mean usefulness is high across frequency bands. Heavier users show slightly higher mean usefulness, but the heavy-user groups are small, so the usefulness rating is limited to infer correlation.

3.1.5. Influence of Extent of Teaching Experience on Perceptions of GPTs

The data was incomplete in the survey results from these categories, that is, there was missing data in years variable for enough cases to make a stable correlation unreliable. Naturally, correlations are exploratory only, based on modest sample size and with some variables having missing values. The correlation between use frequency and usefulness is the most interpretable result here and consistent with intuition as greater use frequency points to values gained (usefulness). Greater years of teaching experience alone do not predict how useful respondents find GPTs for their workload. The distribution of teaching experience showed a broad spread, with clusters in the 0–5 years, 6–10 years, 11–20 years, and 20+ years ranges. When cross-tabulated with usefulness ratings (1–5 scale), there was no strong linear correlation between experience length and perceived usefulness. The Pearson correlation coefficient hovered very close to zero (≈−0.05), indicating no statistical relationship in a predictive sense. Teachers with 0–5 years’ experience reported average usefulness scores just above 4.0, consistent with the overall mean. Teachers in the 6–10 year band tended to rate usefulness slightly higher (≈4.2), with less variation, suggesting a confident uptake among mid-career teachers. Teachers in the 11–20 year band reported a wider spread of scores, including the lowest usefulness ratings in the dataset (2 s and 3 s), pulling the mean down slightly. Teachers with 20+ years of experience surprisingly converged towards higher usefulness ratings again (≈4.1–4.3), although this group was smaller in size. There is no clear experience-based stratification: both early-career and late-career teachers appear positive about AI’s usefulness. The mid-career group (11–20 years) displayed more scepticism and variability. Qualitative nuances in the spread of responses suggest that mid-career teachers may show greater caution, while early- and late-career groups are more optimistic.

3.1.6. Illustrated Anonymized Comments from Open Questions

The below is a selection of samples from the open question, with more of these qualitative responses drawn into the discussions towards the end of the paper. What largely stood out in the responses were observations about the poor quality of the materials sometimes created and the need to modify or adapt what was produced. Sample quotes are shown below.

• “The end result is only ever as good as the prompts you put into whichever AI generator you are using.”
• “Some features of TeacherMatic are used more frequently than others. It’s good that there is a ‘favourites’ option.”
• “There is a general expectation that they should be better than they are. (e.g., produce a scheme of work or lesson plan perfectly first time).”
• “Please do not utilise A.I in education just to save time on teaching: use it to enhance learning experiences.”
• “They are useful in the current teaching environment. But they shouldn’t have to be if the workload was balanced correctly. If we continue to have to use it, will it de-skill teachers [?]. Will we run the risk of lesson be created by AI and the teacher not knowing or understanding how or if it meets the needs of learners [?]. Meaning that lesson are used inappropriately.”
• “Often it is sold as reducing your workload but I’m not sure. Often the quality or robustness of the product it gives you requires more work to make it effective. I worry the impact it will have on student teachers and the lessons they will lose in their early career as they use AI.”

As stated, these contributions give some insights about the importance of ‘quality’ overall, and that this concern supersedes simply using such systems to help alleviate workload. We can see ‘quality’ referred to in each answer, from the first, which indicates the need for decent prompts being used, to the subjective identification of what constitutes ‘quality’ on a personal level in the second example, and that these can easily be clustered together.

Certain functionalities and features (rubric/assessment generators) of the systems are highly valued, but there are considerable caveats about pedagogy and ethics, as well as observations around the approaches with use of such systems. The last three statements all highlight criticality about the parallels between the systems and the environments they are being introduced into. Heavy workload in education is a commonly understood issue (statement 5), but the final quote (statement 6) points to the inferiority of these systems against the value of a teacher, and where those have recently come into the profession, they may often not understand the fit with their students or what came before the common infiltration of GPT systems into educational workplaces. Nevertheless, the survey results point to the general usefulness of AI for easing workload, albeit with caveats around quality. Adoption levels of AI are uneven, and teaching experience does not necessarily align to usefulness or effectiveness.

3.1.7. Choice Architecture and GPTs Interpretation of the Data

This section is included to explore the qualitative statements with interpretations and where possible link these to the statistical analysis. We have used the heuristic of the choice architecture as an overarching thematic code to read the qualitative statements, in order to explain the ways in which a choice architecture appears present in the ways that the platforms are used as shown by this data. This is evident in various ways. The most common uses of GPTs is to generate lesson resources, with PowerPoints and quizzes cited as part of that. Teachers use AI for tasks that give immediate time savings without high risk. Tasks that involve grading, administration, or personalised feedback demand trust of the user and potentially alignment with institutional policies, as well as careful scrutiny, so they tend to have less uptake in the ranges of uses by teachers.

Teachers also reported that they used GPTs to develop lesson plans. Most uses were for teaching/learning resource creation, though creating these is unlikely to be the source of teacher workload that leads to fatigue and attrition as opposed to long meetings or excessive administrative work. GPTs seemingly guide responses to certain normative pedagogical choices, for instance direct instruction and grouped work, which might suppress teachers and students’ creativity and innovation and, one would think, not be the best means of personalising learning approaches as proposed in much of the literature [3]. This determinism means that innovation according to individual needs may be stifled by the implementation, or prescriptivism, of the tools. Innovative approaches may appear in courses, which require practical elements, including experiential learning, or, for example, where teachers want to advocate for risk-taking and experimentation, allowing for mistakes and problem-solving through so-called ‘inquiry-based approaches. These are potentially inhibited by the choice architecture of the tools, as noted by an English teacher in their response to the open question: “It [TeacherMatic] has a good variety of resources that can be drawn from it; in terms of what it does, it tends to induce teachers to teach in certain ways by narrowing down the planning it suggests, and the resources almost always need some tweaking.”

GPT systems appear to foreground certain functions in the user interface (e.g., template-based generators, “favourites”), lowering the cognitive and time cost for those tasks. This indicates the ways in which the tools are presently used by this participant sample points to a kind of desire path: tools that help to cut some corners, as shown by the data in Section 3.1.4, where we see that even the lightest users (0–5 per month) give AI a strong average usefulness rating of 4.0 out of 5. This suggests that teachers, regardless of frequency, perceive value when they do use AI. By making content generation fast, visible, and low-risk, teachers may then be “nudged” or compelled towards these features without necessarily exploring a full pedagogical range or alternative options. Over time, this knowledge could dissipate. This was evident in responses to the open question, where some participants recognised that innovation was necessary, was guided by experience of prompting (“Understanding that the output is only ever as good as the input is crucial to using AI effectively.”), and did not always result in the common pedagogical design resulting from resource or planning creation (that is, PowerPoints or presentations and grouped work). “Please use as innovation for learning and teaching to enhance experience and engagement. I recently visited one of our English classes where the teacher felt that students do not understand what it was like before technology. We organised something through Nichesss where students had a discussion with Jane Austen to ask what it was like before technology—by using technology. A fantastic bridge, and great lesson.” The example here shows innovation as extending beyond the formulaic design of standard PowerPoint delivery of content, which appears evident in the responses. This is a feature borne out by other studies in vocational education [28], where 77% of teachers relied on platforms such as TeacherMatic to produce resources, which they cited as quizzes and PowerPoints.

It is plausible to suggest that without rich teacher education, knowledge exchange, and continued professional development, mediation between teachers and students with their peers could diminish at the cost of the creation of staid resources. Alternatively, GPTs may provide ideas and examples for experiential learning, and it is possible that in the future other technologies emerge from this interregnum period in GPT development that can supplement those ideas into practice. There already exist virtual and immersive reality, in the form of smartglasses, which have been shown to enhance mobile assessment of vocational education, but this is reliant on highly imaginative teachers [29].

Another concern about the ways that teachers are utilising the platforms is in what we may see as an early-stage novelty effect bias towards visible and tangible outputs. AI platforms like TeacherMatic may push more familiar outcomes to the surface on interfaces, for instance, when they are used to generate lesson resources, worksheets, or presentations as default, high-visibility options. Teachers under workload pressure take the easiest, lowest-barrier use case. This is natural: teachers want products that seemingly alleviate processes and improve efficiency in doing so. It is shown in a lower uptake in lesson planning and administrative tasks that may result in less visually gratifying results. What is highly regarded tend to be resource outputs that can be put to use (quizzes, rubrics, worksheets). Teachers are more likely to adopt features that yield immediate, concrete artefacts, such as creating resources or lesson ideas, which feels like a win. This is also a type of prominence bias, where the most visible, “one-click” features shape adoption of the tool: in this case with 40% stating that they use GPTs to create lesson resources. The architecture of AI tools emphasises ready-to-use, tangible outputs that subtly channels user engagement away from back-end or abstract tasks, where workload may be higher and more nuanced. As for the question of lowering workload, this is hard to discern, though on the surface level, it would appear that the infrequent uses reported show staff are not taking the opportunities seemingly available to offload work to GPTs. Again, this is sometimes put down to resources and refinement of them, as one Maths teacher notes: “I’m not sure if its usefulness is lacking due to my subject. The quizzes it produces are simply adequate and more well though tout resources, particularly those in line with the principals of mastery, are ready made elsewhere on specialist sites, of which there are an abundance.” This modification, entailing extra work, is echoed by others, such as this Counselling Teacher: “I find TeacherMatic extremely useful for giving me ideas and provides me with an advance in creating resources and assessment I need when preparing my lessons, more so when I have limited time. Another effective function is the Rubric, however just lately I have had to make significant amendments to make it sufficient.” While another Maths teacher was critical of this particular platform: “Doesn’t seem to work effectively with Maths—was looking to see whether it could generate examples with or without solutions but only seemed able to draw in definitions. Could generate problems to an extent, but incomplete without graphical elements (e.g., ‘use the graph to define’ without a graph).” These conflicting views seem to point to the need for a better understanding of the affordances that can be gained from better prompt engineering, harnessing the tools to get them to work for users’ specific needs. This, however, also depends on the development capacity of the tools to capture more nuance and, to another extent, improvements in teacher education, such that “the end result is only ever as good as the prompts you put into whichever AI generator you are using.” Elsewhere, CPD workshops around these issues have been developed that yield positive results in being critically informed about AI, curriculum mapping to identify spaces where it can be integrated, while also exchanging and sharing knowledge in participatory sessions regarding the ways in which GPTs can enhance teaching and assessment [3,30]. These affordances will potentially lead to time-saving and reduced workload if more advanced practitioners can support this knowledge with their experience.

Yet others in the survey reported categorically on its time-saving potential, as shown by a Functional Skills and English teacher: “TeacherMatic has made 2 h jobs 20 min jobs; now I can use my time for more student facing work.” However, this is in the context of one tool being licenced across a college, where others find having to explore the full range of GPTs available to be time-consuming in itself: “There is so much and it’s hard to know what is best to use from a security, cost and output point of view. Significant time has to be spent exploring different tools.” This, again, might result in a choice architecture where users go to the centralised technologies available from Big Tech, since they are most prominent and available (and this statement came from a Digital Innovation coach, who uses Microsoft’s Copilot most commonly). Finally, another statement from an Arts lecturer is distinctly more negative about the impact of these tools, and they point to the issues prevalent in working cultures, rather than technologies: “They are useful in the current teaching environment. But they shouldn’t have to be if the workload was balanced correctly. If we continue to have to use it, will it de-skill teachers? Will we run the risk of lesson be created by AI and the teacher not knowing or understanding how or if it meets the needs of learners?” This statement underlines the importance of the choice architecture aspect of these technologies, with which this paper contributes. Teachers determine how technologies such as GPTs will be deployed in ways that best alleviate pressures on them or improve practices, but currently the indicators appear to be that the priority is for time-saving rather than raising quality (or personalising learning). AI GPTs themselves eventually become a choice architecture in the sense that we are gradually depending on them (over ourselves) to give us solutions for everything, from heavy workloads to improved personalization of learning and better resources and richer feedback.

4. Discussion

Given the uneven levels of adoption of AI, it was interesting to see the range of tools being deployed as cited in the surveys. For the most part, this was predicated on TeacherMatic and ChatGPT, which is unsurprising, since the former has licences in many college institutions, while the latter is prolifically used as an early marketized tool. Other mentions were given to Copilot, Perplexity, Gemini, and Notebook. A choice architecture reading of the data shows that these common ‘Big Tech’ tools are being used in the majority, so that where institutions and professional bodies, as well as Government, are heedlessly advocating for the embracing of ‘AI’, the main beneficiaries to this tend to be the market established companies whose products are visibly available, or built into existing platforms, like portable documents or browsers. This is problematic, since there are more varieties of tools available being overlooked. The fact that nearly 70% of respondents mentioned only one or two mainstream systems suggests a narrowing of the ecosystem, meaning institutional procurement and default settings strongly shape practice. In short, adoption is less about teacher exploration than about the availability and visibility of dominant commercial tools.

While the data does point towards teachers deploying GPTs mainly for designing lesson resources and this seemingly being of value, it could be argued that they appear to diminish subject specialism skills that would otherwise be honed in a labour-intensive manner through the organic creation of original resources. Survey results confirm this dominance, with resource creation accounting for around 40% of reported uses, double or triple the proportion of any other task type. Instead, something replaces the act of creativity: a refinement at best, as it is reported that these resources need careful scrutiny [24,25,30] because they often contain mistakes that can go unchecked straight into teaching. Refinement or checking and modifying resources produced by GPTs, one could infer, is not always going to be practiced. We argue that certainly early-career teachers may benefit from creating resources to develop knowledge of how to communicate and teach the content knowledge of their subject specialism. For example, beginning teachers typically need to design and develop schemes of work for an extended syllabus, that may contain twelve weeks of lessons and activities mapped to a curriculum. This is exhaustive and could be automated and processed by an interface such as TeacherMatic, yet while this or other GPTs might create a standard twelve week lesson plan, they could also problematically overlook much of the classroom environment, knowledge of students’ needs and preferences, the diversity of students in the group, targets, and other variations that are unexpected or hard to predict or control. This concern is reflected in survey comments, where one in four respondents explicitly noted that AI-generated resources required “careful checking” or “major editing” before classroom use. The implication is that while workload may be reduced in principle, the erosion of teacher creativity and accuracy risks is a significant trade-off.

It may alternatively be argued that early-career teachers would benefit from the agency and undertaking the ‘grunt work’ involved in planning a scheme of work to give them closer oversight of the scaffolding of the subject and its curriculum, the nuances of lessons, and a personalised understanding of their students’ learning needs that an interface may not be able to provide, given it may generate generalised responses. However, many teachers recognised that the first results were never appropriate and there is evidently skill in evaluating the resulting resources: “It’s a great tool to take the initial edge of bulk admin tasks. I find it easier being the editor and creator than the doer (that’s the Ai’s job).” This statement has interesting comparison with other studies, which point to “pedagogical atrophy as human pedagogic processes emulate and become increasingly dependent on AI” [21], meaning the teacher’s agency becomes diminished with increased reliance on AI tools. However, the participant here notes the need for modification (editing) of produced materials, indicating a need for close scrutiny and quality control of outputs—a demanding redefinition of teacher’s pedagogical knowledge and subject agency, when reusing materials. ‘Atrophy’ may be contentious, but it can be indicated in Section 3.1.4, where it appears that early-career and late-career teachers are slightly more positive about AI’s usefulness. This could reflect a generational factor: more experienced teachers are often established in their pedagogical routines and may be more critical about AI’s alignment with professional values, or they may be under different workload pressures compared with early-career or near-retirement colleagues. Early-career teachers may be using it for cognitive offloading, helping them get through, which may not necessarily help them to hone their skills (hence ‘atrophy’). The stability of high usefulness ratings among the least and most experienced teachers suggests that perceived usefulness is less about years in the profession, and more about personal orientation toward technology and workload relief.

Others [30,31,32] more fervently take the view that automation through AI can certainly alleviate teachers’ workload and support greater personalization and focus on students, as the AI takes the work of ‘repetitive tasks’ (with mentions of ‘grading, attendance tracking, and student progress monitoring’). Yet this is not borne out by the data: assessment marking and feedback together accounted for less than 15% of reported use, while administrative functions were cited by only 10%. In practice, teachers are not yet leaning on AI to handle repetitive tasks at scale, suggesting a mismatch between potential and actual application. This is a critical finding, since it points to the lack of take-up by teachers to exploit GPT tools to undertake the repetitive tasks but the propensity for them to utilise them for creative or fulfilling teaching tasks, such as learning resources.

The perception of usefulness surrounding these tools is striking, since the data shows that even light users rate usefulness highly. Teachers using AI just 0–5 times per month still awarded an average usefulness rating of 4.0 out of 5, while those using it 6–10 times nudged this slightly higher (4.08). Even among heavier users, the trend continued upwards, reaching 4.33 and even 5.0 in isolated cases, despite very small sample sizes. This consistency shows that usefulness is broadly recognised, irrespective of frequency of use. Indeed, two-thirds of all respondents gave AI tools a usefulness rating of four or five stars, underlining strong positive perceptions despite reservations about accuracy and oversight. AI is being used primarily in creative and planning tasks, rather than decision-making, marking, or grading, with administrative tasks featuring less frequently. The data shows that GPT platforms hold strong potential for resource creation.

The implication is that barriers to adoption are not primarily about scepticism of value, but rather about habit, institutional framing, or workload patterns that limit experimentation. With relatively modest nudges, teachers could be steered into deeper adoption, where usefulness tends to increase with frequency of use.

Recommendations

It is important to note that this paper is published at an interregnum, that is, a threshold of what the human imagination puts to those systems and what GPT systems can produce. In a short period of time, capacities change and grow, and systems tend to improve. This means that the results are determined mainly by the understanding of the teachers of what is possible and can be delegated: workload for nuanced tasks may well end up being syphoned to AI.

The survey findings suggest that adoption of AI tools in vocational teaching is currently concentrated around a narrow set of market-dominant systems (notably TeacherMatic and ChatGPT), with nearly 70% of respondents naming only one or two such tools. Despite this limited range, usefulness ratings were consistently high, even among light users: those using AI just 0–5 times per month reported a mean usefulness of 4.0, rising to 4.33 among moderate users, with two-thirds of all respondents awarding a rating of four or five stars. At the same time, task distribution points to an imbalance: resource creation accounted for around 40% of use, far outstripping applications in assessment, feedback, or administration (together less than 25%). This highlights both the enthusiasm for AI’s potential and the uneven ways in which it is currently being channelled into practice.

These findings carry several implications. Strategically, the fact that usefulness ratings rise with frequency of use suggests that teachers may benefit from nudges, training, or structured experimentation to cross from occasional into regular adoption, where compounding benefits are more likely to emerge. From an AI literacy perspective, the data showing that one in four teachers flagged the need to “carefully check” or heavily edit GPT outputs underlines that refinement and modification are not peripheral but central skills [24]. This means AI literacy frameworks in teacher education should focus less on uncritical adoption and more on the iterative process of quality assurance and adaptation. Finally, the dominance of resource creation as a use case, paired with concerns about diminished subject specialism, signals the need for professional knowledge sharing. Teachers could benefit from structured exchanges or open repositories where not only resources but also prompts, refinements, and editing strategies are shared transparently, especially for vocational contexts where subject-specific expertise is critical.

In sum, the data show that AI in vocational teaching is viewed as highly useful but unevenly adopted and unevenly applied. By encouraging more regular experimentation, embedding modification as a key literacy skill, and fostering professional collaboration around resource refinement, institutions can help ensure that AI adoption supports, rather than supplants, the professional and pedagogical expertise of teachers.

5. Conclusions

The study presents the findings from a literature review around vocational education and a survey of how AI is being used by vocational teachers in the English college sector. Perceptions of usefulness were apparent in results, even if these were not always aligned to desired outcomes. The study identifies potential offloading of work to GPT systems, yet it is questionable whether this usually results in work tasks becoming automated that teachers wish to relinquish, since it mostly involves the creative tasks that give them oversight of lessons and module plans. This potentially points to a problem, that teachers (and especially early-career teachers) need to know to look for issues in resources provided by GPTs, rather than wilfully accept the results will always be to a good standard. There may be issues with this going forward, as at this juncture there is a strong focus generally on identifying what is created by AI, particularly with a view to what mistakes are present. In the future, it may be less easy to note these as we become more accustomed to its outputs, arguably more trusting of it, and less critical.

We assert that manual creation of these resources is often necessary to teacher’s professional development overall but that GPT tools definitely hold values to teachers with more innovative flair who can apply them to better uses and with a more sophisticated understanding of prompts. This is likely a challenging study to undertake, since so much of a teacher’s work is unseen and hidden, unpredictable, given to serendipity or happenstance, and involves human interactions at micro levels, which GPTs, at least at this interregnum in its evolution, will never be able to automate or accommodate for. This is something we endorse as we strive to argue that education and learning are human-crafted experiences, accentuated by technologies but not necessarily determined by them.