Using LLMs without trust: the adoption–trust paradox as institutional governance failure in academic ecosystems

Abstract Large language models have rapidly made their way into higher education. Still, the integration of these tools remains contested. This study presents the results from a survey of 80 higher education professionals, most of them in Nordic countries, examining LLM adoption, trust, academic integrity, and institutional preparedness. The picture that came back is contradictory. In this self-selected sample, 82.5% already use LLMs in their teaching work. At the same time, they do not really trust what these tools produce. Trust in LLM outputs scored only 2.45 out of 5. Concerns about privacy scored considerably higher (3.82/5) and so did concerns about academic integrity (3.42/5). The respondents are fairly confident that students use LLMs regularly (4.19/5), but they have little confidence in telling student work from LLM-generated output (2.58/5). The lowest score across all areas (mean 1.99/5) was found for use of LLMs as a tool for grading students. When asked about challenges, 47.5% named academic integrity, indicating vulnerability within institutional assessment practices rather than individual misconduct. The data are consistent with a self-reinforcing cycle. Educators in this sample report clear benefits and assume widespread student use, while trusting the outputs little, and continued use does not appear to build that trust. We call this the adoption-trust paradox. The survey did not measure institutional governance directly, but Nordic policy studies report institutional responses ranging from comprehensive AI guidelines to little more than references to existing integrity norms. Read against that background, the same pattern repeated itself across every section of the survey. High adoption, low trust, and educators left to draw the lines themselves. The sample is modest and weighted toward the social sciences and humanities, but the consistency of the pattern makes it hard to dismiss, and it supports the case for an institutional response. Article PDF 1 Introduction LLMs have gone from experimental technology to standard tools in higher education in a very short time. They now affect how academics write, prepare assessments, develop curricula, and support students (Kasneci et al. 2023; Peláez-Sánchez et al. 2024; Pireci Sejdiu and Sejdiu 2025). Easy access and perceived efficiency gains have accelerated the spread. So have rising expectations that graduates will work with AI as a matter of course (Shi et al. 2026). The questions this raises go well beyond pedagogy. Who authored a given text? How can it be verified? Under what conditions is academic work produced and assessed? These are socio-technical questions. The answers depend on how the technology interacts with the institutional, organizational and social arrangements it sits within (Coeckelbergh and Gunkel 2024). All of this puts pressure on higher education institutions. They need to find ways of using LLMs in teaching and learning that do not compromise academic standards, integrity or trust (Foltynek et al. 2023). In this study, academic work covers the full range of professional activities in higher education, including teaching, research, supervision, assessment and related administrative tasks. Nothing in the history of educational technology has spread this fast. Within months of the public launch of ChatGPT in November 2022, students and educators across disciplines were using LLM-based tools, most of the time without any formal institutional policy or shared norms for responsible use (Kasneci et al. 2023). Previous waves of educational technology went through institutional procurement and piloting first. LLM adoption skipped that entirely. It was bottom-up, driven by what individual users found useful rather than by decisions made at institutional level (Ghimire and Edwards 2024). The consequence is that everyday practice now runs ahead of governance and educators are left to figure out the ethical and pedagogical implications more or less on their own. Studies from outside the Nordic region paint a mixed picture. There is optimism about AI in higher education, but it comes with reservations. Misuse, bias, and threats to academic integrity keep surfacing as concerns (Hsain and El Housni 2024; Tigerstedt and Fabricius 2025; Yigci et al. 2025). Interestingly, students who actually use AI tools tend to develop more positive attitudes over time (Acosta-Enriquez et al. 2024). Yet in public debate, the conversation keeps narrowing. Institutions and commentators often treat LLM challenges as a student discipline problem, specifically cheating and plagiarism, and then reach for detection tools and assessment redesign as fixes (Alberth 2023). That framing misses the larger picture. What makes the Nordic case worth examining separately is the starting position. Trust in institutions is high, digital infrastructure is mature, and there are strong traditions around student agency and academic freedom (Jóhannesdóttir et al. 2025). You would expect that to make LLM adoption smooth. In some ways it does. But the same trust-based governance that lets individual educators experiment freely also means there is less urgency around creating formal rules. A multi-country Nordic study showed the consequences. The institutional responses ranged from comprehensive AI guidelines at one end to little more than a reference to existing integrity norms at the other, with nothing specific about LLMs (Jóhannesdóttir et al. 2025; Erhardt et al. 2025). In practice, each educator ends up drawing the line themselves. That is exactly the kind of institutional uncertainty this study sets out to document. The cheat-and-detection framing may risk overlooking concerns that recent socio-technical and qualitative work has raised. To begin, discussions around trust in AI can become disconnected from the institutional conditions from which trust is a meaningful concept, leading to what Gill KS (2020) terms ‘ethics washing’, such that ethics becomes an overlay on technology, transparent enough for the technology sales cycle not to be hampered by ethical concerns, but without any sense of real accountability structures. Second, this framing disregards more profound questions raised by LLMs regarding the nature of authorship and knowledge production and the role of educators that adoption rates and attitude scales fail to capture (Pireci Sejdiu and Sejdiu 2025). Recent work on LLMs in the science system drew a convergent conclusion, highlighting how these tools are already reshaping scientific communication, peer review, and the needed skills of researchers, changes for which regulatory and institutional frameworks have not yet kept pace (Fecher et al. 2025). This widespread informal adoption remains underexplored in terms of institutional governance, policy and support structures, particularly in Nordic higher education (Jóhannesdóttir et al. 2025; Erhardt et al. 2025), and a more systematic examination is warranted. Despite that, there has been a growing volume of work around institutional responses to generative AI, for instance assessments of existing policies and guidelines outlining governance frameworks (Ghimire and Edwards 2024; Erhardt et al. 2025; Azevedo et al. 2025). Across the broader AI governance landscape, numerous scholars characterize accountability as a sociotechnical structure composed of authority, standards, processes and forums of answerability, rather than one isolated compliance mechanism (Novelli et al. 2024). However, little is known about the intersection of everyday LLM use and the trust educators have in these tools, especially in contexts where institutional governance has not yet caught up with practice. Across the literature, attitudes and adoption rates are most commonly measured separately with results showing more positive attitudes among those who report higher levels of familiarity (Acosta-Enriquez et al. 2024; Yigci et al. 2025). That evidence comes mainly from student samples and it measures attitudes and intention to use rather than trust in what the tools produce. Whether familiarity predicts trust among educators, and what happens when it does not, is less researched. This study fills this gap by approaching the co-occurrence of high adoption and low trust directly, interpreting the combination as a failure in institutional governance instead of individual attitudes. Such a gap, specifically our limited comprehension of how widespread use of LLMs relates to trust among educators under weak institutional governance, was addressed by the research team conducting a survey to elicit professional perspectives on the outlooks, reflections and concerns around usage of LLMs in higher education. There were both closed- and open-ended questions in the survey. This paper reports results from the closed-ended items of the survey, summarized using descriptive statistics, based on responses from HE professionals at higher education institutions. The present study addresses four research questions. It begins with a descriptive study of the adoption of large language models among HE professionals: the extent of their familiarity, how extensively they use them, and the professional tasks to which they are applied. Second, it investigates educators’ perceptions of how large language models benefit their daily practice along with the perception of associated risks and trustworthiness. Third, it explores educators’ foremost worries about academic integrity, assessment practice, and students using large language models. Fourth, it considers what the response patterns, particularly the perceived challenges, suggest about institutional readiness for large language model integration in terms of governance arrangements and practical support. Together, these questions seek to move beyond the predominant focus on adoption rates and attitudes and instead examine the conditions under which LLMs are used, trusted and governed in higher education settings. Instead of a purely technological or pedagogical perspective, then, the study is grounded in a trustworthiness and governance perspective (Zicari et al. 2021). This perspective conceives of AI integration as a socio-technical question, where the concept of technological capability is inseparable from the tacit practices, institutional settings and shared understandings in which it is deployed (Gill SP 2023). This perspective highlights that the responsible integration of AI technologies depends not only on the properties of the tools themselves but also on the institutional conditions under which they are deployed, including the availability of policies, clear delineation of roles and responsibilities, and the presence of mechanisms for oversight and accountability. The study seeks to characterize how LLMs are situated within higher education, not as fully institutionalized technologies but as unavoidable yet insufficiently governed systems, by capturing and examining the perspectives of HE professionals on their use for educational purposes. This work advances discussion about trustworthy AI in education by shifting attention from individual tool use toward institutional uncertainty, systemic risk and the conditions required for responsible and sustainable integration of LLMs in higher education. It moves beyond the dominant adoption-and-attitude literature to make explicit the structural and institutional dimensions of LLM integration, pointing toward governance arrangements that may facilitate trustworthy use in practice. 2 Methods 2.1 Study design This study employed a descriptive online survey design comprised of closed-ended and open-ended questions focusing on how HE professionals use LLMs, their assessment of trust in LLM outputs, academic integrity, and institutional preparedness. The survey, which includes 41 questions on how educators use LLMs, is divided into five sections: (1) Background information, (2) Use and familiarity with AI, (3) Trust and reliability, (4) Use of LLMs in teaching and learning, and (5) Future perspectives. Most items use a 5-point Likert scale (1 = strongly disagree, 5 = strongly agree). Several items use multiple-choice or checkbox formats. Before distribution, the survey instrument was pilot-tested using a small sample of higher education professionals, and items were reviewed for clarity, relevance and comprehensibility. The survey instrument in its entirety can be found in Appendix 1. Due to the relatively rapid and still evolving nature of LLM adoption throughout higher education, a descriptive approach was considered appropriate. The study seeks to establish baseline patterns of use, trust and institutional response at a particular point in the integration process, rather than to test causal relationships or predict outcomes. The survey also contained several open-ended questions such as ethical use of LLMs by students, future trends in LLM technology and significant challenges to higher education given increasing LLM integration. Open-ended responses were collected and will be analyzed and reported in a separate publication. We used descriptive statistics to summarize the closed-ended items. Data were collected through an online survey aimed at higher education professionals. The survey ran on the Eduix E-lomake platform from 14 January 2025 to 27 May 2025. The survey was anonymous and did not collect sensitive personal data. All data processing follows the guidelines of the Finnish National Board on Research Integrity (TENK) for non-medical research involving human participants. Participation was voluntary. In line with GDPR, data are stored on institutional servers at Arcada University of Applied Sciences, and only the research team has access. Data will be retained in accordance with GDPR and the Finnish Data Protection Act, and not beyond what is needed for research integrity and verification. We recruited participants through institutional networks, professional mailing lists and social media, mainly LinkedIn. The survey was in English and reached professionals directly involved in teaching, research or educational administration in higher education. The only inclusion criterion was affiliation with a higher education institution. The recruitment strategy aimed for broad reach within Nordic professional networks but also attracted respondents from other European countries and the United States. Because the survey was distributed through open channels, a formal response rate could not be calculated. 2.2 Analysis We summarized closed-ended items using descriptive statistics, reporting frequencies, percentages, means and standard deviations. Each Likert-scale item was treated as a single-item measure. We did not construct composite scales, so internal consistency estimates like Cronbach’s alpha do not apply. For multiple-choice items, we report the count and percentage of respondents who selected each option. We collected demographic and contextual information, including academic role, years of experience and institution type, to help interpret the findings. Given the exploratory nature of the study and the sample size (n = 80), we did not apply inferential statistical tests. We did calculate an exploratory Spearman rank correlations for selected variable pairs to look at potential associations between familiarity, experience, trust and concerns about academic integrity (Appendix 2). Claude (Anthropic) was used for language editing of the manuscript. No content was AI-generated. 3 Results Each Likert-scale item was used as a single-item measure of an aspect of LLM attitudes or behavior. This approach was taken as the survey items were not designed to measure latent constructs through multi-item scales, but rather to capture distinct aspects of LLM use in higher education. In line with this, internal consistency estimates (such as Cronbach’s alpha) were not applicable since composite scales were not constructed. Single-item measures have greater measurement error than scales, yet are commonly used in exploratory survey research and appropriate when items cover well-differentiated topics (Allen and Seaman 2007). As shown in Appendix 2, exploratory Spearman correlations were computed post hoc to examine whether anticipated, theoretically grounded patterns such as a positive association between familiarity and trust were present in the data. Spearman’s rank correlation was used due to all variables involved being ordinal (Likert 1–5 or ordered categorical). Fisher’s z-transformation was employed to compute 95% confidence intervals. These correlations are presented independently in Appendix 2 to keep them separate from the main descriptive results. 3.1 Respondent characteristics Table 1 summarizes the characteristics of the 80 respondents. The sample consists predominantly of experienced educators: 51.2% have more than 10 years of experience in higher education, and 55.0% hold a doctoral degree. The largest professional group is lecturers and instructors (45.0%), followed by professors, deans, or heads of department (25.0%). Respondents are roughly evenly distributed between universities (53.8%) and universities of applied sciences (45.0%). The gender distribution is balanced (46.2% women, 46.2% men). The sample is dominated by respondents from Nordic countries, with Finland (40.0%) and Norway (22.5%) together accounting for 62.5% of all responses. The most represented fields of study are Social Sciences (36.2%) and Humanities (31.2%). 3.2 LLM adoption and use patterns Respondents reported high familiarity with LLMs (mean 3.98 on a 1–5 scale, SD 1.04). A large majority (82.5%, n = 66) reported having used LLMs in teaching and teaching-related activities. Only 14 respondents (17.5%) did not use LLMs at all (Table 2, Panel A). Among those who use LLMs, GPT-based services dominate: 57 respondents (71.2%) reported using GPT, followed by Copilot (48.8%), Gemini (18.8%), and Claude (15.0%). Sixteen respondents (20.0%) reported using other services (Table 2, Panel B). The most common areas of application were writing assistance (67.5%), preparing questions and exams (57.5%), and brainstorming (56.2%). Use for student assessment was notably low at 13.8% (Table 2, Panel C). 3.3 Perceived benefits, impact, trust, and reliability Respondents perceive significant benefits from LLMs in their professional activities (mean 3.93, SD 1.17) and report a positive impact overall (mean 3.71, SD 1.18). The perceived negative impact is low (mean 2.30, SD 1.16). Views on whether LLMs foster creativity are moderately positive (mean 3.15, SD 1.27) (Table 3, Panel A). Despite high adoption rates, respondents report low trust in LLM outputs (mean 2.45, SD 1.01). Perceptions of LLM competence in respondents’ own fields are moderate (mean 3.01, SD 1.17). Concerns about privacy are pronounced (mean 3.82, SD 1.15), and concerns about integrating LLMs into practice are substantial (mean 3.56, SD 1.14). The technical difficulties are reported at a relatively low level (mean 2.63, SD 1.12). Concerns about the effect of LLMs on trust between students and teachers score moderately high (mean 3.28, SD 1.13) (Table 3, Panel B). 3.4 Teaching, assessment, and student use A majority of respondents have adjusted their teaching or assessment strategies in response to LLMs (mean 3.59, SD 1.18), and most explicitly mention their own LLM use to students (mean 3.68, SD 1.07). Students are generally made aware of the instructor’s use of LLMs (mean = 3.45, SD = 1.09). Respondents strongly believe that students frequently use LLMs (mean = 4.19, SD = 0.86), with the highest mean score in the survey. At the same time, concerns about student LLM use are substantial (mean 3.79, SD 1.13). Encouragement of student LLM use is moderate (mean 3.19, SD 1.26) and views on whether LLMs improve student learning are neutral to skeptical (mean 2.81, SD 1.12). Two items are particularly relevant for the assessment challenge. Respondents report that students rarely reveal their LLM use (mean 2.76, SD 1.07), and the respondents themselves express low confidence in distinguishing student-produced work from LLM-generated content (mean 2.58, SD 1.19). The use of LLMs for assessing student work is minimal (mean 1.99, SD 1.12), the lowest mean in the survey (Table 4). 3.5 Skills, pedagogical approaches, and challenges When asked which skills are essential in an AI-enhanced educational environment, respondents rated critical thinking as most important (n = 52), followed by information literacy (n = 49), ethical reasoning and decision-making (n = 47), problem-solving (n = 46), collaboration and communication (n = 46), and creativity and innovation (n = 44) (Table 5, Panel A). The most favored pedagogical approaches were project-based learning (n = 46), collaborative learning (n = 34), flipped classroom (n = 26), inquiry-based learning (n = 24), and peer-to-peer learning (n = 17) (Table 5, Panel B). The most frequently selected challenges were ensuring academic integrity (47.5%), addressing errors or bias in AI outputs (40.0%), maintaining critical thinking and creativity (40.0%), adapting curriculum and assessment methods (38.8%), ensuring equal access to AI resources (36.3%), and preparing students for an AI-augmented workforce (36.3%) (Table 5, Panel C). Taken together, the survey data tell a consistent story across all five domains. Respondents know these tools well (mean familiarity 3.98) and most of them use them (82.5%), a share that partly reflects who chose to respond. They see real benefits (mean 3.93) and they assume their students use LLMs regularly (mean 4.19). But they do not trust what the tools produce. Trust in LLM outputs scored only 2.45 on average. Privacy worries are strong (mean 3.82) and respondents have little confidence in their ability to spot LLM-generated student work (mean 2.58). Using LLMs for student assessment scored lowest of all items (mean 1.99). When asked what skills matter most, respondents pointed to critical thinking, information literacy and ethical reasoning. The most frequently named challenge was academic integrity (47.5%). These patterns appear together, not in isolation, and we examine what that means in the Discussion below, where we introduce the adoption–trust paradox. 4 Discussion Combined, the results show that HE professionals do not primarily experience LLMs as innovative instruments or disruptive technologies but rather as a source of what we have named institutional uncertainty, defined by low trust in outputs (mean 2.45), high perceived privacy risk (mean 3.82), low confidence in detecting work sourced from LLMs (mean 2.58) set against an institutional policy environment that Nordic studies describe as thin and uneven (Jóhannesdóttir et al. 2025; Erhardt et al. 2025). We characterize institutional uncertainty as when educators adopt and routinely use a technology they do not trust, without adequate institutional policies, governance structures or assessment frameworks in place to manage the risks of that use. Here, the empirical and the interpretive parts of our contribution should be kept apart. What the data show directly is the descriptive co-occurrence of high adoption, low trust in outputs, low detection confidence and pronounced privacy concern. The survey did not include items asking respondents about institutional governance, policies or support structures. The governance-failure reading is our inference. It combines these co-occurring patterns with policy research documenting the absence or unevenness of institutional rules for LLM use (Jóhannesdóttir et al. 2025; Erhardt et al. 2025; Ghimire and Edwards 2024). On that combined reading, the condition resembles a structural absence of the policies, assessment frameworks, verification and accountability that responsible use of LLMs in academic work would require. LLMs are common but insufficiently trusted, necessary but poorly governed (Lelescu et al. 2025). Core academic concepts including integrity, authorship and critical thinking are not rejected but renegotiated in the context of AI-mediated practices. But the true gap is thus not between users and non-users, but rather between widespread practice and absent governance, a policy vacuum, as recounted across institutional contexts (Ghimire and Edwards 2024; Azevedo et al. 2025). These co-occurring patterns become what we call an adoption–trust paradox (Fig. 1) as described above. The terms ‘adoption–trust paradox’ and ‘institutional uncertainty’ are analytical concepts we introduced through this study to describe the patterns in the data. These were not drawn from earlier theoretical work but instead arose from descriptive results. This situation is not unique to our sample, but the specific combination of high adoption, low trust, and absent governance reported here adds empirical specificity to what has largely been discussed at a conceptual level in the literature. Several recent reviews and policy analyses have identified the need for institutional governance of AI in higher education (Lelescu et al. 2025; Azevedo et al. 2025), but our data show what this governance gap looks like from the perspective of individual educators who must handle it on a daily basis. The results across all five sections of the survey display a consistent picture. Educators in this sample show very high levels of adoption, recognize the benefits of LLMs, and assume widespread use by students. At the same time, there is low trust in outputs, limited confidence in detecting LLM-assisted work and serious concerns about privacy, integrity, and institutional support. Such patterns do not fall easily into either hope or opposition. Instead, the data suggest a condition where use outpaces trust and individual behavior outstrips institutional response. The major patterns and their implications are discussed in the next subsections, followed by a comparison with international findings and a note on limitations of our work. This is not simply a pattern of either resistance or enthusiasm. Instead, the data convey a profile that could be labeled as pragmatic adoption under conditions of unresolved uncertainty, in which educators continue to use LLMs out of necessity while harboring significant reservations about their reliability, governance and the long-term implications of these tools for academic work. 4.1 The adoption–trust paradox High adoption and low trust is, therefore, the key empirical observation of the study. We use ‘paradox’ to indicate that high uptake and low willingness to rely on LLM outputs simultaneously coexist in high-stakes academic work. We use the term in a pragmatic sense. High use combined with low trust is not, by itself, paradoxical. People routinely use technologies they do not fully trust when the benefits are immediate and the risks feel manageable. What makes the combination noteworthy is the expectation it fails to meet. The prediction that familiarity builds trust comes from studies of students, which measured attitudes and intention to use rather than trust in outputs (Acosta-Enriquez et al. 2024; Yigci et al. 2025). Whether that prediction extends to educators, who answer for assessment integrity and teaching quality in ways students do not, had not been tested directly. Our data indicate that it does not extend and ‘paradox’ refers to this gap between the expected and the observed pattern in high-stakes academic work. Exploratory correlations are consistent with this interpretation. LLM familiarity showed no association with trust (rho = − 0.01, p = 0.91). Years of experience showed a weak negative correlation with trust (rho = − 0.28, p = 0.015, 95% CI − 0.47 to − 0.06), tentatively suggesting that more experienced educators report somewhat lower trust. The upper bound of that interval lies close to zero and the estimate comes from a small sample with three parallel comparisons, so it should be read as indicative at most (Appendix 2). Neither variable was associated with academic integrity concerns. These patterns are consistent with the reading that increased exposure does not resolve the trust deficit in this sample. This pattern deviates from findings in several international studies of students. Here increased familiarity and experience with AI tools tend to be associated with more positive attitudes and greater confidence (Acosta-Enriquez et al. 2024; Yigci et al. 2025). In our sample, frequent use does not appear to resolve trust-related concerns. Educators use LLMs because they perceive benefits (mean 3.93) and because student use is assumed to be widespread (mean 4.19), but they do so under conditions of persistent doubt about the reliability and integrity of the outputs. This gap matters because it cannot be resolved by rhetorical appeals to “trustworthy AI” alone. As Krüger and Wilson (2023) say, trust in AI risks becoming a verbal claim that substitutes for, rather than reflects, the institutional and knowledge-based conditions that would make trust warranted (Krüger and Wilson 2023). This suggests that educators keep using LLMs not because they trust them, but because they feel they have to and because everyone around them does too (Ghimire and Edwards 2024). Figure 1 shows the adoption–trust paradox as a feedback loop in six stages. It begins with educators who adopt LLMs out of perceived practical necessity. The reasoning tends to be straightforward. Students are already using these tools, so opting out feels unrealistic (stage 1). In the model, high adoption (stage 2) brings repeated exposure to outputs that are unreliable or hard to verify (stage 3), and this exposure is observed alongside low trust in outputs (stage 4). Low trust in turn co-occurs with the absence of clear institutional rules for LLM use reported in the policy literature. We call this institutional uncertainty (stage 5). We do not claim a direct causal link here. The connection between low trust and institutional uncertainty rests on descriptive co-occurrence across three survey sections, Trust and Reliability, Teaching and Learning, and Skills and Challenges. The model is a conceptual summary, not a tested causal chain. Many institutions have still not put clear policies or guidelines in place for responsible LLM integration (stage 6) (Erhardt et al. 2025; Ghimire and Edwards 2024). Without that guidance, educators fall back on personal judgment. The cycle restarts at stage 1. In the model, the loop then returns to stage 1. LLM use spreads across courses and among colleagues without anyone steering it (Pireci Sejdiu and Sejdiu 2025), peer expectations grow, and personal use starts to feel unavoidable. We present this account as hypothesis-generating, not as a demonstrated mechanism. If it holds, adoption alone will not improve trust, and the policy and uncertainty gaps in stages 5 and 6 of Fig. 1 are the plausible points of institutional intervention. Our data cannot confirm that reading, but they are consistent with it. We read this paradox as a self-reinforcing loop, and we stress that this is an interpretation. In the model, peer norms and productivity pressure sit alongside rising adoption, while trust lags where institutions offer few ways to verify LLM outputs or hold them to account. Figure 1 lays this out as a model built from the descriptive patterns in this study. The survey shows co-occurring patterns. It does not test how strong or directional the proposed relationships are. That is a work for future research, using inferential methods or longitudinal designs. The qualitative analysis of open-ended responses from this survey is still to come and may shed light on what connects the stages. The adoption–trust paradox we found does not match what the international literature reports. Acosta-Enriquez et al. (2024) studied college students at scale and found that more experience with ChatGPT went hand in hand with stronger intent to use it and more responsible use. Their reading is that familiarity breeds a kind of practical trust. Yigci et al. (2025) found the same. Students exposed to LLM-based chatbots came away with better attitudes and higher acceptance. Among the educators in our sample, none of that holds. The exploratory correlation between LLM familiarity and trust was effectively zero (rho = − 0.01, 95% CI [− 0.23, 0.21]). More experienced educators tended to report somewhat lower trust rather than higher, a weak association whose confidence interval approaches zero (rho = − 0.28, 95% CI − 0.47 to − 0.06). Why the difference? Educators occupy a different position. They are accountable for assessment integrity, teaching quality, and the credibility of what the institution puts out. That makes them more alert to what LLMs get wrong. Students, by contrast, often treat the tools as a way to get work done faster. The Nordic setting may play a role too, diverse as it is. Where institutional trust and academic autonomy run deep, a technology that muddies verification and accountability may land harder than in systems already built around surveillance-based quality checks. Put together, the data and these contextual factors say the same thing. Trust in LLMs does not come from knowing the tools better. It comes from institutional structures that make verification, accountability and recourse real. Dahlin (2025) reached a comparable conclusion in ethnographic work on other AI settings, where explainability on its own did not produce trust. 4.2 Academic integrity as an institutional challenge Ensuring academic integrity was the most frequently selected challenge in the survey (47.5%). At the same time, respondents report low confidence in distinguishing student-produced work from LLM-generated content (mean 2.58) and note that students rarely reveal their LLM use (mean 2.76). This combination points to a structural vulnerability in current assessment practices. The problem is not only that students may use LLMs inappropriately, but also that the assessment infrastructure lacks the capacity to respond to situations in which LLM-assisted work is widespread and difficult to detect (Perkins 2023; Foltynek et al. 2023; Bittle and El-Gayar 2025). The concern extends beyond individual student misconduct. At the sample level, academic integrity (47.5%), adapting curriculum and assessment methods (38.8%), and maintaining critical thinking and creativity (40.0%) were the three most frequently selected challenges. The multiple-selection format does not allow us to confirm whether the same individuals selected all three, but the clustering of high selection rates suggests these are experienced as interconnected institutional concerns rather than isolated issues. This cluster of responses suggests that academic integrity is experienced less as a disciplinary matter and more as a systemic issue affecting institutional credibility and the value of qualifications. Recent research supports this interpretation. Bittle and El-Gayar (2025) identify a shift in the literature from viewing generative AI as a cheating tool to recognizing it as a systemic challenge requiring redesigned assessment and institutional policy. Foltynek et al. (2023), writing on behalf of the European Network for Academic Integrity, recommend that institutions develop explicit AI use policies and integrate AI literacy into curricula rather than relying primarily on detection. Our findings are consistent with this direction. The concern expressed by respondents is less about individual student behavior and more about the capacity of existing assessment structures to function meaningfully when the boundary between human and machine-generated work is increasingly unclear. 4.3 The assessment and detection gap The data reveal a notable gap in assessment-related practices. The use of LLMs for assessing student work is the lowest-scoring item in the survey (mean 1.99). This low adoption rate may partly reflect deliberate professional caution, as evaluating student learning is a core educator responsibility where the consequences of error are high. Under the EU AI Act, AI systems used for educational assessment and access decisions may be classified as high-risk, requiring specific transparency and human oversight measures (European Parliament and Council of the European Union 2024). Specifically, AI systems used to determine access to or admission to educational institutions, to evaluate learning outcomes, or to assess the appropriate level of education are classified as high-risk under Annex III, point 3, requiring conformity assessments, transparency obligations, and human oversight. The low adoption of LLMs for student assessment observed in this study (mean 1.99) may, therefore, reflect not only professional caution but also an emerging regulatory reality that most institutions have yet to address through formal governance structures. Respondents have adjusted their teaching and assessment strategies to some degree (mean 3.59), but the low confidence in detecting LLM-generated content (mean 2.58) indicates that these adjustments have not yet produced workable solutions. Educators are therefore operating in a space where they know the rules have changed but lack reliable tools or norms to respond effectively. The moderate encouragement of student LLM use (mean 3.19), combined with high concern about student use (mean 3.79), further illustrates the ambivalence that characterizes this transitional period. This ambivalence, encouraging student use while lacking confidence in detection, is one of the most practically significant findings of this study. It suggests that educators are caught between competing imperatives: the expectation that they prepare students for AI-augmented work, and the professional obligation to ensure that assessment outcomes reflect genuine student learning. Without institutional frameworks that clarify how LLM-assisted work should be evaluated, assessed, and documented, individual educators bear the full weight of this tension. 4.4 Trust, privacy, and perceived competence Privacy stands out as the most pronounced concern among the trust-related items (mean 3.82). This is consistent with broader European debates about data protection in educational technology contexts (Drachsler and Greller 2016; Tsai et al. 2020; Lelescu et al. 2025). LLMs are perceived as moderately competent in respondents’ own fields (mean 3.01), but this moderate score does not translate into trust. The gap between perceived competence and actual trust (3.01 vs. 2.45) suggests that, in this sample, trust involves more than technical performance. We interpret this gap as indicating that trust also depends on governance, transparency and institutional backing, conditions that the policy literature describes as underdeveloped in most higher education institutions (Jóhannesdóttir et al. 2025; Erhardt et al. 2025). This finding aligns with broader research on the privacy paradox in educational technology, in which users express strong concern about data protection while continuing to use tools that collect and process personal data (Tsai et al. 2020). In the context of LLMs, the privacy concern is compounded by uncertainty about how input data is processed, stored, and potentially used to train future model iterations. For educators who may enter student names, assignment content, or research data into LLM interfaces, these concerns are not abstract but relate to concrete professional responsibilities under GDPR and institutional data protection policies. 4.5 Skills and pedagogical responses Respondents identify critical thinking (n = 52), information literacy (n = 49), and ethical reasoning (n = 47) as the most essential skills for an AI-enhanced educational environment. The survey did not use the term ‘AI literacy,’ which has gained traction in recent literature (Shi et al. 2026). The skill priorities reported here, critical thinking, information literacy, and ethical reasoning, can be understood as constituent elements of what the broader literature describes as AI literacy, rather than as a single unified competency. These skill priorities are consistent with the concerns expressed in other survey items. If LLM outputs cannot be fully trusted, then the ability to evaluate, contextualize, and critically assess AI-generated content becomes paramount. The preference for project-based learning (n = 46) and collaborative learning (n = 34) as pedagogical approaches suggests that educators see active, process-oriented methods as better suited to an environment where the production of text alone is no longer a reliable indicator of student competence. The convergence of these skill priorities with the concerns expressed in other sections of the survey suggests that educators are already reasoning about what a post-LLM pedagogy might require. The emphasis on process-oriented and collaborative approaches can be understood as an attempt to shift assessment and learning design toward activities where LLM-generated outputs are less likely to substitute for genuine student engagement. This pragmatic pedagogical reasoning, however, is occurring in the absence of institutional coordination, which means that individual educators are independently arriving at similar conclusions without the benefit of shared frameworks or institutional support. 4.6 Limitations Several limitations should be noted. The sample size (n = 80) is modest, and the convenience sampling strategy (distribution through institutional networks, internal forums, and LinkedIn) limits generalizability. Recruitment through these channels likely over-represents educators who already take an interest in LLMs. The adoption rate of 82.5% should therefore be read as a description of this self-selected group, probably an upper bound, rather than as a prevalence estimate for higher education professionals in general. The sample is heavily weighted toward Nordic countries (Finland 40.0%, Norway 22.5%), and the findings may reflect characteristics specific to Nordic higher education systems, such as high levels of institutional trust, strong traditions of student autonomy, and relatively early exposure to digital teaching tools. A descriptive comparison of Nordic and non-Nordic respondents (Appendix 2, Table 7) showed near-identical trust in LLM outputs (2.44 vs 2.47) and similar adoption (81% vs 88%), with somewhat higher academic integrity concern among non-Nordic respondents (3.87 vs 3.31), suggesting that the non-Nordic minority does not drive the overall pattern. Respondents are predominantly experienced educators with doctoral degrees, which may not represent the full range of HE professionals. The sample is also weighted toward Social Sciences (36.2%) and Humanities (31.2%), and LLM adoption patterns, perceived benefits, and trust may differ in STEM, health sciences, and IT disciplines. A response rate could not be calculated due to the distribution method. All data are self-reported, with the usual limitations regarding social desirability and recall accuracy. The descriptive design does not allow causal inferences. The open-ended survey responses, which will be analyzed and reported separately, may provide additional depth and context to the patterns identified here. Additionally, the use of single-item Likert measures for constructs such as trust and perceived competence limits the precision of the measurement. While single-item measures are appropriate for the exploratory purpose of this study, future research should consider developing validated multi-item scales for key constructs, particularly trust in AI outputs and perceived institutional preparedness. The exploratory correlations reported in Appendix 2 should be interpreted with caution, given the sample size and the number of comparisons. The sample does not support stratified or adjusted analyses, so confounding of the experience-trust association by discipline or country cannot be fully ruled out, although the association held within the Nordic subsample and was negative within every discipline (Appendix 2, Table 6). 5 Implications The findings of this study have several implications for HE institutions, educators, particularly within Nordic countries, and policymakers addressing the integration of large language models (LLMs). 5.1 Implications for institutional governance First, the results suggest that LLM integration cannot be treated as a purely pedagogical or technical matter. The widespread framing of academic integrity as an institutional risk indicates that governance structures, assessment frameworks, and quality assurance mechanisms must be reconsidered at the organizational level (Perkins 2023). Reliance on individual educators to manage LLM-related challenges, in the absence of coherent institutional policies, creates fragmentation and uneven practice. Institutions, therefore, need explicit strategies that clarify acceptable use, responsibility, and accountability, rather than leaving interpretation to individual discretion. These strategies should be developed through inclusive processes that involve educators, students, administrators, and quality assurance bodies, rather than being imposed top-down as compliance measures. Recent empirical work on how AI governance is experienced from the inside, based on industry interviews about the EU AI Act, similarly finds that top–down compliance frameworks alone do not produce meaningful trust or accountability when operational realities are not accommodated (McCormack et al. 2025). The adoption–trust paradox documented in this study underscores an argument made in the AI governance literature, that governance frameworks are unlikely to be effective unless they are perceived as practically relevant and responsive to the realities of everyday academic work (Novelli et al. 2024; McCormack et al. 2025). Institutions that develop AI governance in isolation from the educators who use these tools risk producing policies that are ignored or circumvented in practice. 5.2 Implications for teaching and assessment Second, the redefinition of critical thinking observed in the data has direct implications for teaching and assessment design. If critical thinking increasingly involves evaluating, contextualizing, and interrogating AI-generated content, assessment practices must move beyond output-focused evaluation toward processes that make reasoning, judgment, and decision-making visible. This shift requires pedagogical support and shared models rather than isolated experimentation by individual instructors. Concrete approaches might include portfolio-based assessment, oral examinations, process documentation, and collaborative projects where individual contributions are traceable. The low confidence in distinguishing student work from LLM output (mean 2.58) underscores the urgency of this shift; if detection is unreliable, assessment must be redesigned to make the use or non-use of AI tools less consequential for the validity of the evaluation. 5.3 Implications for staff competence and professional development Third, the strong sense of responsibility, combined with a perceived lack of preparedness, highlights the need for structured professional development. Training should not focus solely on tool use but also address epistemic questions, ethical reasoning, and the limits of LLMs. Without institutional investment in competence-building, educators remain positioned as responsible actors without adequate resources to fulfill that responsibility. Such professional development initiatives should be embedded in ongoing institutional practice rather than offered as one-time workshops. Given the pace of change in LLM capabilities, static training programs risk becoming outdated before they are completed. Institutions should consider establishing communities of practice, peer learning networks, and regular review cycles that allow educators to share experiences and adapt their approaches as the technology and the regulatory landscape evolve. 5.4 Implications for trustworthy AI in higher education Finally, from a trustworthy AI perspective, our findings, read alongside the policy literature, point to a mismatch between widespread LLM use and limited institutional oversight. In this context, trustworthiness is better understood as a socio-technical property that emerges from governance, transparency, and shared understanding, rather than as a feature of the technology itself. Closing this gap requires institutional policies that are developed in step with everyday academic practice, rather than control mechanisms added after informal adoption has already taken place. The Z-Inspection® framework (Zicari et al. 2021) provides one model for how such assessments might be conducted, emphasizing multi-stakeholder involvement and systematic evaluation of socio-technical requirements. Applying similar approaches to educational AI systems could help institutions move from reactive to proactive governance. 5.5 Implications for AI regulation and the EU AI Act The findings also carry implications for the regulatory landscape. Under the EU AI Act (European Parliament and Council of the European Union 2024), AI systems used in educational contexts for assessment, access decisions, or evaluation of learning outcomes are classified as high-risk, triggering requirements for transparency, human oversight, and conformity assessment. Our data suggest that most higher education institutions are far from meeting these requirements. The minimal use of LLMs for student assessment (mean 1.99) may partly reflect an intuitive recognition of these stakes, but the broader pattern of informal, ungoverned adoption across other academic activities suggests that institutions have not yet developed the governance infrastructure needed to comply with emerging regulatory expectations. As the EU AI Act moves toward full implementation, higher education institutions will need to assess which of their AI-related practices fall within scope and develop appropriate documentation, oversight, and accountability mechanisms. 6 Conclusion This study set out to explore how HE professionals make sense of large language models through their own accounts, focusing on institutional, pedagogical, and ethical dimensions rather than technological performance. Within this predominantly Nordic sample, weighted toward the social sciences and humanities, LLMs are neither resisted nor fully embraced. They are used under conditions of persistent uncertainty. Educators experience LLMs as unavoidable components of contemporary academic practice, while institutional governance, policy and shared frameworks lag behind, a gap documented in the Nordic policy literature rather than measured by our survey items (Jóhannesdóttir et al. 2025; Erhardt et al. 2025). Rather than framing LLM-related challenges as problems of student misconduct or individual competence, the response patterns indicate concerns about assessment validity and the future meaning of core academic concepts such as authorship and critical thinking. These concerns persist across levels of experience and frequency of use, and the exploratory correlations point, weakly and tentatively, toward lower rather than higher trust among more experienced educators. The adoption–trust paradox identified in this study offers a framework for understanding this condition. It suggests that the problem is not simply one of slow adoption or resistance to change, but of a self-reinforcing cycle in which use and distrust coexist. Our interpretation, which combines the survey patterns with the policy literature, is that the absence of institutional governance helps sustain this cycle. Breaking this cycle requires deliberate institutional action to establish clear policies and assessment practices that are viable in an AI-mediated academic environment. Understanding the paradox as institutional governance failure, an interpretation rather than a directly measured finding, reframes the response required, from behavioral adjustment to structural reform. By highlighting institutional uncertainty, this study contributes to the literature on AI in higher education by shifting attention from adoption rates and attitudes to the conditions under which AI technologies become normalized in academic environments. Responsible and sustainable integration of LLMs will depend less on increased familiarity with tools and more on collective governance, pedagogical clarity, and institutional capacity to redefine academic practice. Future research should examine how different institutional strategies shape responses to LLM integration over time and how governance frameworks can better align with the realities of everyday academic work. Longitudinal and comparative designs would be particularly valuable for tracking how the adoption–trust paradox evolves as institutions develop more formal governance structures. The qualitative analysis of the open-ended responses from this survey, currently in preparation, will provide additional insight into the reasoning and experiences behind the patterns reported here. Data availability The survey data generated during this study are not publicly available due to privacy considerations. The anonymous survey collected demographic and professional information from a small sample (n = 80) where combinations of country, institution type, academic role, and field of study could potentially enable re-identification of individual respondents. Aggregated data supporting the findings are reported in the tables within this article. Requests for additional aggregated data may be directed to the corresponding author. References Acosta-Enriquez BG, Arbulú Ballesteros MA, Huamaní Jordan O et al (2024) Analysis of college students’ attitudes toward the use of ChatGPT in their academic activities: effect of intent to use, verification of information and responsible use. BMC Psychol 12:255. https://doi.org/10.1186/s40359-024-01764-z Alberth (2023) The use of ChatGPT in academic writing: a blessing or a curse in disguise? TEFLIN J Publ Teach Learn Engl 34:337–352. https://doi.org/10.15639/teflinjournal.v34i2/337-352 Allen IE, Seaman CA (2007) Likert scales and data analyses. Qual Prog 40:64–65 Azevedo L, Robles P, Best E, Mallinson DJ (2025) Institutional policies on artificial intelligence in higher education: frameworks and best practices for faculty. New Dir Adult Contin Educ 2025:70–78. https://doi.org/10.1002/ace.70013 Bittle K, El-Gayar O (2025) Generative AI and academic integrity in higher education: a systematic review and research agenda. Information (Basel) 16:296. https://doi.org/10.3390/info16040296 Coeckelbergh M, Gunkel DJ (2024) ChatGPT: deconstructing the debate and moving it forward. AI Soc 39:2221–2231. https://doi.org/10.1007/s00146-023-01710-4 Dahlin E (2025) Trust in AI. AI Soc 40:6089–6095. https://doi.org/10.1007/s00146-025-02429-0 Drachsler H, Greller W (2016) Privacy and analytics: it’s a DELICATE issue a checklist for trusted learning analytics. In: Proceedings of the sixth international conference on learning analytics & knowledge—LAK‘16. ACM Press, Edinburgh, United Kingdom, pp 89–98. https://doi.org/10.1145/2883851.2883893 Erhardt C, Kullenberg H, Grigoriadis A et al (2025) From policy to practice: the regulation and implementation of generative AI in Swedish higher education institutes. Int J Educ Integr 21:21. https://doi.org/10.1007/s40979-025-00195-6 European Parliament and Council of the European Union (2024) Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 (EU AI Act) Fecher B, Hebing M, Laufer M et al (2025) Friend or foe? Exploring the implications of large language models on the science system. AI Soc 40:447–459. https://doi.org/10.1007/s00146-023-01791-1 Foltynek T, Bjelobaba S, Glendinning I et al (2023) ENAI recommendations on the ethical use of artificial intelligence in education. Int J Educ Integr 19:12. https://doi.org/10.1007/s40979-023-00133-4 Ghimire A, Edwards J (2024) From guidelines to governance: a study of AI policies in education. In: Olney AM, Chounta I-A, Liu Z et al (eds) Artificial intelligence in education. Posters and late breaking results, workshops and tutorials, industry and innovation tracks, practitioners, doctoral consortium and blue sky. Springer, Cham, pp 299–307. https://doi.org/10.1007/978-3-031-64312-5_36 Gill KS (2020) AI & Society: editorial volume 35.2: the trappings of AI agency. AI Soc 35(2):289–296. https://doi.org/10.1007/s00146-020-00961-9 Gill SP (2023) Why thinking about the tacit is key for shaping our AI futures. AI Soc 38:1805–1808. https://doi.org/10.1007/s00146-023-01758-2 Hsain A, El Housni H (2024) Large language model-powered chatbots for internationalizing student support in higher education. https://doi.org/10.48550/arXiv.2403.14702 Jóhannesdóttir S, Kousa P, Korkala S et al (2025) “Do as your teacher tells you!” How is AI use regulated in Nordic higher education institutions? In: Coman A, Vasilache S (eds) Social computing and social media. Springer, Cham, pp 194–218. https://doi.org/10.1007/978-3-031-93539-8_14 Kasneci E, Sessler K, Küchemann S et al (2023) ChatGPT for good? On opportunities and challenges of large language models for education. Learn Individ Differ 103:102274. https://doi.org/10.1016/j.lindif.2023.102274 Krüger S, Wilson C (2023) The problem with trust: on the discursive commodification of trust in AI. AI Soc 38:1753–1761. https://doi.org/10.1007/s00146-022-01401-6 Lelescu A, Sava S, Grosseck G, Malita L (2025) Exploring trust in generative AI for higher education institutions: a systematic literature review focused on educators. Humanit Soc Sci Commun 12:1961. https://doi.org/10.1057/s41599-025-06253-1 McCormack L, Bendechache M, Lewis D, Huyskes D (2025) Trust and transparency in AI: industry voices on data, ethics, and compliance. AI Soc. https://doi.org/10.1007/s00146-025-02654-7 Novelli C, Taddeo M, Floridi L (2024) Accountability in artificial intelligence: what it is and how it works. AI Soc 39:1871–1882. https://doi.org/10.1007/s00146-023-01635-y Peláez-Sánchez IC, Velarde-Camaqui D, Glasserman-Morales LD (2024) The impact of large language models on higher education: exploring the connection between AI and Education 4.0. Front Educ 9:1392091. https://doi.org/10.3389/feduc.2024.1392091 Perkins M (2023) Academic Integrity considerations of AI large language models in the post-pandemic era: ChatGPT and beyond. J Univ Teach Learn Pract. https://doi.org/10.53761/1.20.02.07 Pireci Sejdiu N, Sejdiu S (2025) The quiet transformation of higher education in the AI era. Open Res Eur 5:249. https://doi.org/10.12688/openreseurope.20715.1 Shi Y, Yu K, Dong Y, Chen F (2026) Large language models in education: a systematic review of empirical applications, benefits, and challenges. Comput Educ Artif Intell 10:100529. https://doi.org/10.1016/j.caeai.2025.100529 Tigerstedt C, Fabricius S (2025) Understanding generative AI’s role in higher education: a teacher perspective on responsible integration of AI in business education. In: Ahram T, Karwowski W, Etinger D (eds) Human systems engineering and design (IHSED2025): future trends and applications, vol 198, pp 387–395. https://doi.org/10.54941/ahfe1006792 Tsai Y-S, Whitelock-Wainwright A, Gašević D (2020) The privacy paradox and its implications for learning analytics. In: Proceedings of the tenth international conference on learning analytics & knowledge. ACM, Frankfurt Germany, pp 230–239. https://doi.org/10.1145/3375462.3375536 Yigci D, Eryilmaz M, Yetisen AK et al (2025) Large language model‐based chatbots in higher education. Adv Intell Syst 7:2400429. https://doi.org/10.1002/aisy.202400429 Zicari RV, Brodersen J, Brusseau J et al (2021) Z-Inspection®: a process to assess trustworthy AI. IEEE Trans Technol Soc 2:83–97. https://doi.org/10.1109/TTS.2021.3066209 Acknowledgements We thank Emilie Francois Mathez for the contribution to conceptualization and methodology. During the preparation of this manuscript, the authors used Claude Opus 4.6 (Anthropic) for language editing and proofreading. The authors reviewed and edited all content and take full responsibility for the published work. Author information Authors and Affiliations Contributions Writing of the manuscript was led by P.K. All authors contributed to conceptualization and methodology. C.T. and A.G. curated and prepared the data, and A.G. performed the statistical analyses. C.T. led the project administration and supervision. All authors reviewed and approved the final manuscript. Corresponding authors Ethics declarations Conflict of interest The authors declare no competing interests. Additional information Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Electronic supplementary material Below is the link to the electronic supplementary material. Appendices Appendix 1: Survey instrument The survey consisted of 41 items distributed across five sections. Likert-scale items used a 5-point scale (1 = Strongly disagree, 2 = Disagree, 3 = Neutral, 4 = Agree, 5 = Strongly agree). Open-ended items (marked with “–” in the Response options column) were collected but are not analyzed in this paper. Items marked with an asterisk (*) were mandatory. Q# | Question | Format | Response options | |---|---|---|---| I. Background information | ||| 1 | Age | Single choice | 18–24/25–34/35–44/45–54/55–64/65 or above | 2 | Gender | Single choice | Man/woman/non-binary/other/prefer not to say | 3 | Highest level of education completed | Single choice | Institutional degree/Bachelor’s/Master’s/Doctoral/prefer not to answer/other | 4 | Primary current academic role | Single choice | Professor, Dean, HoD/Lecturer, Instructor/Researcher/Administrator/other | 5 | Years of experience working within higher education | Single choice | Less than 1 year/1–3 years/4–6 years/7–10 years/more than 10 years | 6 | Type of higher education institution | Single choice | University/UAS or University college/other | 7 | Department or field of study | Single choice | Humanities/natural sciences/social sciences/other | 8 | Location of higher education institution (country) | Open text | – | II. Use and familiarity with AI | ||| 9 | I am familiar with LLMs | Likert 1–5 | Strongly disagree to strongly agree | 10a | Have you used any Large Language Models (LLMs) in your teaching and teaching-related activities? (Yes, if you have used them more than 10 times) | Yes/no | Yes/no | 10b | If yes, what LLM services have you used at least once? (Select all that apply) | Multi-select | GPT/Claude/Gemini/Copilot/other | 10c | If no, please feel free to share why not and proceed to Section III | Open text | – | 11 | In which areas have you used LLMs or generative AI tools? (Select all that apply) | Multi-select | Coaching students/assessing students/preparing questions or exams/brainstorming/pedagogical planning curriculum/creating visual content/creating online lectures with avatars/case studies/data analysis/writing assistance/administrative tasks/other | 12 | I perceive significant benefits from using LLMs in higher education | Likert 1–5 | Strongly disagree to strongly agree | 13 | Using LLMs has impacted my professional activities positively | Likert 1–5 | Strongly disagree to Strongly agree | 14 | Using LLMs has impacted my professional activities negatively | Likert 1–5 | Strongly disagree to strongly agree | 15 | Please add any further information or thoughts on the benefits or impact of using LLMs in your professional activities | Open text | — | III. Trust and reliability | ||| 16 | I have concerns about integrating LLMs into higher education practices | Likert 1–5 | Strongly disagree to strongly agree | 17 | I have faced technical difficulties when using LLMs | Likert 1–5 | Strongly disagree to Strongly agree | 18 | I am concerned about privacy when I use LLMs | Likert 1–5 | Strongly disagree to strongly agree | 19 | I think the use of LLMs in education affects fairness positively | Likert 1–5 | Strongly disagree to strongly agree | 20 | I have concerns that using LLMs in higher education will negatively affect the trust between students and teachers | Likert 1–5 | Strongly disagree to strongly agree | 21 | I trust the outputs generated by the LLMs I have used | Likert 1–5 | Strongly disagree to strongly agree | 22 | I think the use of LLMs in education fosters creativity | Likert 1–5 | Strongly disagree to strongly agree | 23 | I perceive that LLMs can perform competently within my field | Likert 1–5 | Strongly disagree to strongly agree | 24 | Other thoughts related to LLMs and trust and reliability? | Open text | — | IV. Use of LLMs in teaching and learning | ||| 25 | I have adjusted my teaching and/or assessment strategies to account for the availability of LLMs | Likert 1–5 | Strongly disagree to strongly agree | 26 | My students are aware that I use LLMs in my teaching or research | Likert 1–5 | Strongly disagree to strongly agree | 27 | I explicitly mention my use of LLMs to my students and ensure they are aware of how I use LLMs in my teaching or research | Likert 1–5 | Strongly disagree to strongly agree | 28 | I believe my students frequently use LLMs for assignments and projects | Likert 1–5 | Strongly disagree to strongly agree | 29 | I have concerns about students using LLMs in their coursework | Likert 1–5 | Strongly disagree to strongly agree | 30 | I encourage my students to use LLMs for their coursework | Likert 1–5 | Strongly disagree to Strongly agree | 31 | The use of LLMs improves students’ learning | Likert 1–5 | Strongly disagree to strongly agree | 32 | The use of LLMs impacts academic integrity in my courses | Likert 1–5 | Strongly disagree to strongly agree | 33 | The use of LLMs influences fairness in education | Likert 1–5 | Strongly disagree to strongly agree | 34 | Students in my courses reveal when and how they use LLMs for their assignments and projects | Likert 1–5 | Strongly disagree to strongly agree | 35 | I am confident in distinguishing between student-generated work and LLM-assisted work | Likert 1–5 | Strongly disagree to Strongly agree | 36 | I use LLMs for assessing student-submitted work | Likert 1–5 | Strongly disagree to strongly agree | 37 | What practices or tools can help ensure that students use LLMs ethically, while also improving their learning? | Open text | – | 38 | Other thoughts related to LLMs in higher education? | Open text | – | V. Future perspectives | ||| 39 | To maintain human oversight of LLM technology, how do you anticipate the use of advanced LLM technologies will impact student skill development and necessitate pedagogical changes in higher education over the next five years? Provide brief explanations for: (a) Critical thinking skills, (b) Information literacy, (c) Problem-solving abilities, (d) Creativity and innovation, (e) Collaboration and communication skills, (f) Ethical reasoning and decision-making | Open text (6 sub-items) | – | 40 | What pedagogical approaches do you believe will be most effective in developing these skills in an AI-enhanced educational environment? (Select top 3) | Multi-select (max 3) | Project-based learning/flipped classroom/collaborative learning/inquiry-based learning/adaptive learning/gamification/peer-to-peer learning/other | 41 | In your opinion, what are the core challenges that higher education faces with the increasing integration of LLMs and other AI technologies? Provide brief explanations for: (a) Ensuring academic integrity, (b) Adapting curriculum and assessment methods, (c) Addressing potential errors or biases in AI outputs, (d) Maintaining critical thinking and creativity skills, (e) Ensuring equal access to AI resources, (f) Preparing students for an AI-augmented workforce, (g) Other | Open text (7 sub-items) | – | Appendix 2: Exploratory correlations Rights and permissions Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. About this article Cite this article Kringen, P., Gallucci, A., Hildt, E. et al. Using LLMs without trust: the adoption–trust paradox as institutional governance failure in academic ecosystems. AI & Soc (2026). https://doi.org/10.1007/s00146-026-03323-z Received: Accepted: Published: Version of record: DOI: https://doi.org/10.1007/s00146-026-03323-z