Designing Institutional Support for Generative AI Adoption: Building on Stem Teachers' Desired and Actual Use

Practice Paper Recommended citation: Capdevila, I., Uittenhove, K. L., & Dehler Zufferey, J. (2025). Designing Institutional Support for Generative AI Adoption: Building on Stem Teachers' Desired and Actual Use. In Kangaslampi, R., Langie, G., Järvinen, H.-M., & Nagy, B. (Eds.), SEFI 53rd Annual Conference. European Society for Engineering Education (SEFI), Tampere, Finland. DOI: 10.5281/zenodo.17631245. This Conference Paper is brought to you for open access by the 53rd Annual Conference of the European Society for Engineering Education (SEFI) at Tampere University in Tampere, Finland. This work is licensed under a Creative Commons Attribution-NonCommercial-Share Alike 4.0 International License. DESIGNING INSTITUTIONAL SUPPORT FOR GENERATIVE AI ADOPTION: BUILDING ON STEM TEACHERS' DESIRED AND ACTUAL USE I. Capdevila a, K. Uittenhove b,1 , J. Dehler Zufferey c a EPFL, Lausanne, Switzerland, https://orcid.org/0000-0003-3997-9810 b EPFL, Lausanne, Switzerland, https://orcid.org/0000-0001-5450-3875 c EPFL, Lausanne, Switzerland, https://orcid.org/0000-0001-5163-807X Author Note: Authors are in alphabetical order and contributed equally to the paper. Conference Key Areas: Digital tools and AI in engineering education Keywords: generative AI, innovation adoption ABSTRACT This practice paper presents our approach to supporting faculty in the adoption of generative artificial intelligence (GenAI) in their teaching at our technical university, drawing on both empirical evidence and theoretical frameworks. First, we surveyed faculty adoption of GenAI, particularly their use for supporting student learning, along with potential factors influencing adoption. Our findings revealed a substantial gap between actual and desired use, with faculty expressing significantly greater interest in GenAI integration than their current practices reflect. Second, informed by theories on adoption of innovation and change, we use this evidence to inform targeted support initiatives, designed not only to bridge the usage gap for faculty already interested in GenAI but also to engage those who currently do not wish to use it, thus aligning faculty needs with an effective framework for innovation adoption. 1 Corresponding Author K. Uittenhove kim.[email protected] 1 INTRODUCTION In higher education, GenAI is emerging as a powerful tool that can assist with many teaching tasks and is rapidly becoming widespread. A significant application is the use of GenAI to support student learning, allowing the creation of new opportunities with well-documented learning benefits. These include providing immediate, scalable feedback (e.g., Lee & Moore, 2024), tutoring and answering student questions (e.g., Pardos & Bhandari, 2024), and stimulating active learning and metacognitive reflection (e.g., ElSayary, 2024). Despite the huge potential of GenAI to support learning, students using it spontaneously and without pedagogical guidance are at risk of not benefiting from this potential but might offload learning tasks to GenAI (e.g., Bastani et al., 2024; Fan et al., 2024). Consequently, a key question is how to promote a pedagogically effective and innovative use of GenAI tools by teachers. 1.1. Faculty Adoption of GenAI Tools Attitudes and perceptions toward GenAI’s role in education are generally positive (e.g., 2024; Kim et al., 2025). For example, in Kamoun et al. (2024), a majority saw GenAI as an opportunity for pedagogical innovation, while a minority perceived it as a threat to teachers’ jobs. Whether GenAI in education is perceived as opportunity or threat is shaped by educators’ pedagogical orientation (Cabellos et al., 2024; Choi et al., 2023). Teachers with student-centred, constructive orientations tend to view GenAI as an opportunity to enrich learning, while those with content-centred, reproductive approaches are more likely to see it as a threat. Despite generally positive attitudes and perceptions, actual use is typically low. For example, the majority of faculty surveyed by Kim et al. (2025) reported infrequent or no use of GenAI tools. Key barriers include insufficient knowledge and training to implement these tools in practice (Chounta et al., 2022; Antonenko & Abramowitz, 2023; Al-Abdullatif, 2024), and lack of clear institutional guidelines (Kim et al., 2025). For example, the majority of faculty surveyed by Kamoun et al. (2024) reported not having the needed training, support, and resources to integrate ChatGPT into their teaching practices. Training should focus on practical applications alongside theoretical knowledge and AI literacy (e.g., Al-Abdullatif, 2024), as well as constructive approaches that maximize the potential of GenAI to support student learning (e.g., Cabellos et al., 2024). This could increase AI literacy and perceived feasibility, associated with increased adoption of GenAI (Al-Abdullatif, 2024). In turn, more regular use has been associated with teachers recognizing GenAI’s educational value, compared to individuals with limited or no experience who see it more as a threat (Cabellos et al., 2024; Kaplan-Rakowski et al., 2023). 1.2. Theories on Adoption of Innovation and Change A recent systematic review by Belkina et al. (2025) examining 21 empirical studies on the actual implementation of GenAI in higher education, found that deep integration redefining learning tasks remains rare. Based on their analysis, Belkina et al. (2025) emphasize the importance of using structured pedagogical frameworks to support integration of GenAI that is innovative as well as pedagogically sound. Building on this, we argue that meaningful integration needs to be supported in a holistic approach, recognizing that teachers, while central, are not the only agents driving innovation and change. To better explore how GenAI can be implemented in practice, we draw on two well-established theories that conceptualize change not only at the individual level but also in relation to social and systemic dynamics. One such framework, the Cultural-Historical Activity Theory (CHAT), conceptualizes change as a result of dynamic interactions among six components of an activity system, identifying: the subject (teachers engaging with AI); the object (student learning), the tools (AI tools); the community (institutional and peer influences), the rules (directives and recommendations), and the division of labor (roles of the different stakeholders). The interaction and contradictions among these components create opportunities for learning and thus, change (Engeström, 2001). Complementing this perspective, Diffusion of Innovations theory (DOI) provides a lens to better understand how innovation progresses within an institution. Rogers (1983) defines diffusion as the process by which an innovation is communicated through certain channels, over time among the members of a social system. Key factors influencing diffusion include how the innovation is perceived (relative advantage, compatibility, complexity, trialability and observability), the different adopter categories (innovators, early adopters, early majority, late majority, and laggards), and the nature of information transfer, which is more effective between individuals who are alike (peers). DOI outlines five stages in the innovation-decision process—knowledge, persuasion, decision, implementation, and confirmation— which leads individuals to either adoption or rejection of the innovation. 2 CONTEXT AND PRACTICAL WORK As a technical university, we are discussing how to not merely integrate, but moreover innovate with GenAI in pedagogical practice, all while remaining grounded in an evidence-based approach. Making the bridge between collection of evidence, informing initiatives, and implementing those initiatives through tools and practices requires the collaboration of different institutional actors. The Center for Learning Sciences (LEARN) collects evidence on current education practices and influencing factors. The Teaching Support Center (CAPE) promotes evidence-based effective pedagogies for learning among teachers. The Center for Digital Education (CEDE) designs practical tools and studies their effectiveness. Meanwhile, multiple in-house research labs investigate GenAI itself and explore its innovative applications. In this practice paper, we aim to inform initiatives across our institution with the voice of a key actor: the teachers. This work-in-progress paper presents the collected evidence on the use of GenAI by our faculty in their teaching practice, as well as characteristics associated with the adoption of GenAI. We discuss the implications through the lens of adoption and innovation change (CHAT and DOI), to suggest targeted initiatives to support effective and innovative use of GenAI by teachers. 2.2. The Survey We conducted a survey during the autumn semester of 2024 inviting all teachers at our institution (N=1’473) to share their perceptions, usage, and needs regarding GenAI in education. It was administered in English and received a total of 109 responses from all 18 sections. We considered different teaching applications, including using AI tools for designing teaching material (e.g., creation of exercises, quizzes, slides, syllabi), supporting student learning (e.g., chatbots that provide feedback during assignments, answer questions, or help students monitor their learning), and evaluating student performance (e.g., identifying strengths and weaknesses in essays, assessing accuracy on tests). We uniquely distinguished actual (“You already use generative AI tools for these tasks.”) and desired (“You would like to use generative AI tools for these tasks.”) use of GenAI, reported on a 9point rating scale, where 0 represented "Not At All" and 8 represented "Very Much". Furthermore, we drew upon established technology acceptance frameworks (e.g., Davis, 1989) to identify factors that might influence both actual and desired use and to formulate specific questions relevant to GenAI adoption in teaching. We designed specific understanding and skill questions to assess variations in proficiency among a STEM teacher cohort that is presumed to have comparatively high GenAI knowledge and competence. Responses were collected on 5-point Likert scales (from “Not at all” to “Very much”, except when stated otherwise): ● Understanding: How well do you understand GenAI? (Moderately: Key principles and mechanisms of how and why GenAI works. Very well: Advanced knowledge of neural network architecture underlying GenAI). ● Skill: How skilled are you in using GenAI? (Moderately: Comfortable using multiple GenAI tools across different tasks. Very skilled: Proficient in using APIs, fine-tuning models, and developing custom applications). ● Usefulness: Can these tools be useful for supporting your teaching activities? ● Feasibility: Is it feasible to integrate them in your teaching practice? Consider your time and resources. ● Control: Can they reduce control and autonomy of teachers? ● Threat: Can they be threatening for teachers' jobs? ● Encouragement: To what extent do you feel discouraged or encouraged at the institution to incorporate GenAI tools into your teaching practice? Finally, we asked teachers what type of support they need with GenAI in education, and gave them several options to choose from, including (1) dedicated IT infrastructure; (2) policy and rules ("must do"); (3) guidelines and best practices ("should do"); (4) information and practical examples ("could do"); (5) out-of-the-box solutions; (6) workshops and training; and (7) in-person consultations. 3 RESULTS AND INSIGHTS 3.1 Actual and Desired Use of AI Tools Among the different teaching applications, the highest actual use of GenAI tools was reported for designing teaching materials (M = 2.03, SD = 2.38), followed by supporting student learning (M = 1.32, SD = 2.16), while the lowest actual use was observed for evaluating student performance (M = 0.52, SD = 1.23). Figure 1 shows a substantial gap between actual and desired use. Across different applications, 86.3% of teachers expressed a desire to use GenAI tools to a greater extent than they currently do. In this practice paper, we focus specifically on teachers' use of GenAI to support student learning. In Figure 2, we see that even though 81.7% of teachers wish to integrate GenAI to support student learning (reported desired use level > 0 “Not at all”), about half of them do not actually use these tools yet (reported actual use level 0). Fig. 1. Actual vs. Desired Use (0 corresponding to “Not at all”, 8 corresponding to “Very much”). Fig. 2. Teacher Groups (“No” corresponding to answer “Not at all”, “Yes” corresponding to all other responses). 3.2 Factors Influencing GenAI Adoption To provide adapted support, it is crucial to identify hindering and facilitating factors associated with GenAI adoption among teachers, considering both their desire to use GenAI tools and their actual usage. Therefore, we compared self-reported characteristics of teachers contrasting them on two levels: (i) between those who desire and those who do not desire to use GenAI tools to support student learning, and (ii) among those who desire to use GenAI, between those who already use GenAI to support student learning vs those who do not (see Table 1 for descriptive statistics). Missing data (M = 4.18% per variable) were predicted and imputed based on other variables, preventing a 24.77% case loss. We conducted a Bayesian multivariate analysis in R (brms) using four Markov chains with 4,000 iterations each (including 2,000 warm-up iterations) to estimate group effects on standardized characteristics, modelled as Gaussian outcomes while accounting for correlations between them. As a result, the posterior mean differences reported in the following paragraph are expressed in standard deviation units (unlike the table, which presents means in their original scales), and represent the expected differences between teachers who desire using GenAI tools and those who do not. Statistically meaningful differences were identified when the 95% credible intervals excluded zero. Model diagnostics confirmed excellent convergence (Rhat = 1.00 for all parameters) and high effective sample sizes, ensuring robust estimates. The posterior mean differences indicate that teachers who desire using GenAI tools, compared to those who do not, perceived them as more useful (1.11, 95% CI [0.66, 1.55]), had a more positive AI attitude (0.99, 95% CI [0.52, 1.45]), and found it more feasible to integrate (0.62, 95% CI [0.13, 1.09]). Smaller but still meaningful positive differences were found for skill (0.57, 95% CI [0.09, 1.05]) and encouragement (0.55, 95% CI [0.05, 1.03]). They also perceived GenAI as less threatening (-0.59, 95% CI [-1.08, -0.09]) and less detrimental to their control (-0.52, 95% CI [-1.01, -0.03]). Taking the motivational aspect out, the further analysis looks at only those teachers who desire to use GenAI tools. Among them, teachers who actually use GenAI tools for student learning reported higher skill (0.55, 95% CI [0.14, 0.92]), greater feasibility of integration (0.47, 95% CI [0.07, 0.87]), and more usefulness (0.41, 95% CI [0.04, 0.77]) than non-users. Table 1. Characteristics of Teacher Groups comparing desired and actual use of GenAI to support student learning. No Desire Desire Desire, No Use Desire, Use m sd m sd m sd m sd AI attitude 32.11 11.46 41.91 8.90 40.35 9.65 43.74 7.64 Understanding 2.58 1.02 2.23 0.96 2.02 1.04 2.47 0.80 Skill 1.16 1.26 1.74 0.97 1.50 0.96 2.03 0.91 Usefulness 1.26 1.24 2.58 1.02 2.36 0.99 2.84 1.01 Feasibility 1.37 1.42 2.08 1.14 1.80 1.10 2.42 1.11 Control 2.38 1.31 1.79 1.00 1.88 1.10 1.68 0.88 Threat 1.89 1.52 1.12 0.99 1.07 1.04 1.18 0.94 Encouragement -0.07 0.70 0.36 0.75 0.39 0.70 0.33 0.81 Note. Red highlights statistically meaningful differences. 3.3 Support Needs As illustrated in Figure 3, the most frequently requested type of support across all teachers was guidelines and best practices, indicating a clear demand for structured recommendations. Additionally, teachers who had not yet begun using GenAI expressed a particularly strong need for information and practical examples. Teachers who do not desire to use GenAI show significantly lower demand for various support formats. Fig. 3. Support needs expressed by teachers, grouped according to their desired and actual use of GenAI to support student learning. 4 IMPLICATIONS Considering adoption factors identified in the teacher survey as well as key principles from theoretical frameworks for innovation, we propose a structured set of initiatives. We aligned them with the five stages of Rogers’ innovation-decision process, ensuring support for educators at all levels of adoption. While some of these initiatives are already implemented at our institution and could be further refined, others are strategic recommendations aimed at encouraging GenAI adoption among our faculty. Stage 1, Knowledge: Institutional and Practical Guidelines. This initiative involves the co-construction of institutional guidelines bringing together multiple stakeholders. These can be complemented by video-capsules or factsheets illustrating concrete examples, best practices and testimonials of EPFL teachers. This approach increases awareness of the usefulness and feasibility of these practices and information is transferred via peers, a key aspect for adoption in both theories. Moreover, it tackles several components of the CHAT theory fostering the multi-voicedness of the activity system. Finally, this initiative responds to the support needs indicated by all EPFL teachers, independently of their desired and actual use. Stage 2, Persuasion: Informal Peer Exchanges. This initiative aims to provide a safe space for teachers to regularly meet and discuss, share different views, ideas, experiences and concerns about GenAI. To promote meaningful exchanges, the sessions are semi-structured with one teacher presentation, followed by a discussion where diverse points of view are encouraged. This initiative aligns with CHAT and DOI to foster peer interactions and contradictions to learn and innovate. While all types of adopters can learn from peers in a low-stakes context, the most sceptical and conservative (late majority and laggers) can be particularly persuaded to try new things in their practice. Stage 3, Decision: Hands-On Workshops. This initiative focuses on the development of structured, pedagogically informed workshops to provide teachers with the opportunity to explore, experiment with, and evaluate the integration of GenAI in a low-stakes environment. The workshop catalogue is designed to accommodate the different types of adopter profiles and also aligns with the different stages of the innovation-decision process. To maximize effectiveness, each workshop should incorporate three essential components: (i) hands-on activities enabling direct engagement with GenAI tools; (ii) peer discussion, to foster community, and (iii) guided reflection to support the transfer of insights into participants’ own teaching practices. This initiative enhances confidence and competence and encourages teachers that desire to use GenAI but don’t currently do it, to take the decision to adopt it in their courses. Stage 4, Implementation: Facilitating Access to GenAI Tools. Tools are one of the core elements of the CHAT, and providing access to safe, reliable tools that meet teachers' pedagogical needs is a prerequisite for adoption. By providing GenAI tools (e.g., chatbots) that can be easily customized, through retrieval-augmented generation (RAG) and system instructions that define a specific pedagogical role (e.g., tutor, role-playing partner, mentor), we can directly align these tools with teachers’ pedagogical needs, thus increasing the perception of usefulness and feasibility. Additionally, offering the option to build these applications using locally hosted models ensures that data never leave the institution, thus addressing data protection concerns that currently hinder or even prevent their implementation and use in pedagogical practice. Stage 5, Confirmation: GenAI in Education (AInEd) Laboratories. This initiative aims to create a space and framework where innovative practitioners and in-house researchers actively working on GenAI innovation can collaboratively research and experiment with the potential of this technology in supporting student learning. It is inspired by the CHAT Change laboratories (Virkkunen & Newnham, 2013) which are conceived to trigger cycles of expansive learning including simulation and concretisation. In particular, faculty who already use GenAI can benefit from this initiative to confirm and further develop adoption of GenAI in their teaching. By integrating insights from CHAT and DOI theories with empirical survey data, these initiatives offer a structured and transferable framework for understanding and supporting GenAI adoption in STEM education. Our presented evidence-to-practice workflow, survey instrument (available on OSF), and proposed support initiatives, can be directly applied to STEM education as well as other higher education institutions. 5 ACKNOWLEDGEMENTS We acknowledge the valuable contributions of Patrick Jermann, Head of CEDE, for his expert advice in designing the teacher survey, and Roland Tormey, Head of CAPE, for his insightful perspectives on the conceptualisation of this practice paper.