Negotiating generative AI in online and distance higher education: lecturers' task-selective adoption in Yogyakarta, Indonesia

Generative artificial intelligence (AI) adoption among university faculty is already substantial, although institutional preparedness remains uneven. Shata and Hartley (2025) found that 66.4% of the 294 faculty members they surveyed had previously used generative AI. Meanwhile, Lee et al. (2024) reported that fewer than one-quarter of university educators believed that their institution had adequately equipped them to respond to AI, while more than three-quarters expressed a need for further support. An analysis of institutional policy documents shows the same asymmetry: guidance is being written, but slowly, unevenly and usually after practice has already changed (McDonald et al., 2025). This gap between fast-moving individual practice and slow-moving institutional governance is the empirical starting point for this study.

A fairly consistent body of research documents how university teachers respond to generative AI. Framed largely through technology acceptance and unified acceptance models, this literature finds that perceived usefulness is the strongest driver of adoption among academic staff and students alike (Bamasoud and Mohammad, 2025; Onal et al., 2025). A more qualitative strand documents the underside of that usefulness: early-adopter instructors in the Middle East accept generative AI for assessment-related work only selectively (Khlaif et al., 2024); Palestinian university educators describe a similarly cautious, task-dependent engagement shaped by uncertainty about authorship and originality (Hamamra et al., 2024, 2026); and educators elsewhere report AI-related technostress and a paradoxical dependency, recognizing the risks of overreliance yet continuing to rely on the tool under sustained workload pressure (Alhur et al., 2025; Khlaif et al., 2025). Together, these studies converge on ambivalent, task-differentiated adoption rather than wholesale acceptance or rejection.

What this literature has not yet done is examine that ambivalence among lecturers, as distinct from students, in Southeast Asian open and distance learning (ODL) systems, using evidence that triangulates a moderately sized survey against in-depth interviews. Indonesian scholarship has so far concentrated on practical guidance materials for institutions (Hanjani et al., 2025) or on student-facing outcomes, leaving lecturers' own professional and ethical reasoning comparatively unexamined, even though lecturers are the ones who must decide where the line between assistance and substitution sits in their own work. Technology Acceptance Model (TAM) and its extensions explain a good deal of why lecturers start using generative AI; they explain comparatively little about how lecturers subsequently draw and defend boundaries around that use once ethical stakes, authorship and institutional silence enter the picture (Bearman et al., 2023; Feng and Yu, 2025).

This study responds to that gap through a convergent mixed-methods design that combines a 120-lecturer survey with eight in-depth interviews, conducted among public and private higher education institutions in Yogyakarta, Indonesia, a province where post-pandemic higher education has become substantially more digitally mediated. The design allows perceived usefulness to be examined alongside explicit statistical comparisons by age, academic rank and institution type, as well as interview evidence on how lecturers reason about ethical risk and institutional guidance. From this evidence, the study develops negotiated professional adoption as a conceptual interpretation of how lecturers engage generative AI: not as passive technology acceptors, but as professional actors who differentiate use by task, weigh ethical and epistemic risk against efficiency gains and do so largely without institutional guardrails. Specifically, this study examines how lecturers in Yogyakarta perceive, use and set boundaries around generative AI in academic work, and where survey and interview findings converge, diverge, or complement one another.

Recent research consistently identifies perceived usefulness, more than ease of use, as the main driver of lecturer adoption. Higher education staff and students value AI chiefly for efficiency gains (Bamasoud and Mohammad, 2025; Hazzan-Bishara et al., 2025; Onal et al., 2025), a pattern reported across Saudi and Turkish universities and systematic reviews (Uluma, 2024; Bamasoud and Mohammad, 2025). However, Khlaif et al. (2024) found that Middle Eastern early adopters' acceptance of generative AI for assessment depended on performance expectancy, effort expectancy and social influence, with clear boundaries around appropriate tasks. Quantitative usefulness research therefore explains whether adoption occurs but treats it as a single decision; this study instead examines task-specific usefulness ratings across seven academic tasks.

Technology Acceptance Model (TAM) (Davis, 1989) and Unified Theory of Acceptance and Use of Technology (UTAUT) explain acceptance in terms of perceived usefulness, perceived ease of use, social influence and facilitating conditions. These constructs remain relevant because usefulness and facilitating conditions predict educators' acceptance of AI (Buabeng-Andoh and Baah, 2020; Vanderschaaf et al., 2021; Or, 2024; Hazzan-Bishara et al., 2025). However, they explain less well what lecturers do when a useful, accessible technology also creates academic risk. Unlike learning management systems or search engines, generative AI directly produces text, analysis and argument on which professional authority rests (Bearman et al., 2023). Thus, TAM and UTAUT remain necessary but must be extended with ethical, institutional and epistemic dimensions.

Qualitative research documents ethical unease alongside favorable attitudes toward generative AI. Palestinian educators used ChatGPT for lower-stakes preparation but regarded scholarly writing as ethically fraught (Hamamra et al., 2024), a pattern echoed in later research on unclear ethical boundaries (Hamamra et al., 2026). Concerns about preserving pedagogical and academic quality also created strain beyond technical difficulty (Khlaif et al., 2025). Although conducted in medical education, Alhur et al. (2025) examined educators' AI dependency, overreliance and continued use under workload pressure, making its findings relevant to professional and ethical ambivalence. As this literature is concentrated in the Middle East and North Africa, this study examines task-graded ambivalence in a Southeast Asian ODL context.

Institutional governance of generative AI lags behind individual adoption almost everywhere it has been studied. Higher education institutions across the Association of Southeast Asian Nations (ASEAN) report limited formal structures for monitoring or regulating AI use even as adoption accelerates (Buragohain and Chaudhary, 2025) and Indonesia's national guidance, issued through Belmawa Dikti (2024), coexists with implementation that remains fragmented and uneven at the institutional level (Hanjani et al., 2025). This gap is not simply an administrative delay: where institutional guidance is silent, lecturers construct their own boundaries around acceptable use, compounding authorship ambiguity and epistemic risk. Whether the perceived need for institutional AI regulation differs by age, academic rank, or institution type remains underexamined and the present study examines these group differences directly.

Recent studies in the Asian Association of Open Universities Journal (AAOUJ) show how generative AI is reshaping the learner and educator experience in the region. In-service distance-learning teachers in Hong Kong integrated AI chatbots more intensively than pre-service teachers, with self-selected training most strongly associated with use (Wong et al., 2026). Postgraduate learners at an open university in developing Asia primarily demonstrated foundational critical AI literacy and emphasized the preservation of human qualities in AI-mediated education (Lim et al., 2025). An editorial likewise argues that ODL must be reimagined around intelligent technologies rather than merely adopting them (Li et al., 2025). Together, this scholarship frames the lecturer-side tension examined here and supports testing whether training needs differ by age and rank.

The literature suggests that TAM/UTAUT constructs remain necessary but insufficient for generative AI's academic integrity and governance implications. Perceived usefulness and ease of use explain initial adoption but not why usefulness varies across tasks or why ease of use does not produce unconditional acceptance. Three additional interpretive dimensions are required: ethical ambivalence, explaining restrictions despite usefulness; institutional readiness, treated here as an interpretive domain encompassing ethical guidance, authorship policy, institutional regulation and training support rather than as a standalone measured construct; and academic legitimacy/epistemic risk, explaining greater caution toward scholarly writing and research than presentations or teaching materials. Supplementary material, Table S2 maps how these interpretive dimensions relate to and extend the relevant TAM/UTAUT constructs.

Building on this synthesis, the study conceptualizes lecturers' engagement with generative AI as negotiated professional adoption: an emergent, contextually grounded interpretation, not an established or externally validated theory, describing a process in which lecturers neither accept nor reject generative AI, but continuously weigh its pedagogical benefits against ethical risk, academic legitimacy and professional responsibility, on a task-by-task basis and largely in the absence of institutional guidance. This framing directly informs the study's research questions: (1) how do lecturers in Yogyakarta perceive the usefulness of generative AI across different academic tasks; (2) do these perceptions differ systematically by age, academic rank, or institution type; (3) what ethical and institutional considerations do lecturers raise when explaining their own use; and (4) where do the survey and interview findings converge, diverge, or complement one another in explaining adoption. No formal hypotheses are proposed, since the study is exploratory and was not designed a priori to test directional predictions.

This study used a convergent mixed-methods design (Schoonenboom and Johnson, 2017; Creswell, 2009), in which quantitative and qualitative data were collected over broadly the same period and analyzed separately using methods appropriate to each data type. They were merged at the interpretation stage to identify convergence, divergence and complementarity. The two strands were treated as of equal analytic status: the survey was not designed to test hypotheses generated from the interviews, nor were the interviews designed merely to illustrate survey results; each strand addressed the research questions independently before being brought together in the joint display reported in the Findings. This design was selected because generative AI adoption combines a measurable behavioral component with an interpretive component that survey items alone cannot capture (Bozkurt et al., 2020).

The study was conducted among lecturers at public and private higher education institutions in the Special Region of Yogyakarta, Indonesia. Yogyakarta was selected because the researchers had access to diverse institutions that had integrated online and distance modes after the pandemic. Site selection was based on feasibility rather than theoretical uniqueness and is treated as a boundary condition in the Limitations. The province nevertheless provides a relevant ODL-adjacent setting, with institutions varying in size, resources and disciplinary mix within one regional policy environment.

Quantitative data were collected through an online survey administered via Google Forms over two weeks, from 9 to 23 December 2024. Participants were recruited using purposive convenience sampling. Eligible participants were active lecturers employed at public or private higher education institutions in Yogyakarta, aged 30 years or older, holding at least a master's degree and a recognized academic functional rank. No minimum length of teaching experience was imposed. Of 126 responses received, six incomplete questionnaires were excluded, leaving 120 valid responses for analysis. Respondents were heterogeneous in age, academic rank, institution type and discipline (Table 1, Panel A). Doctoral-degree holders comprised 76.7% of the sample (n = 92) and master's-degree holders 23.3% (n = 28); 65.0% (n = 78) were affiliated with private universities and 35.0% (n = 42) with public universities.

The questionnaire combined self-developed items with items adapted and modified from prior instruments on educational-technology adoption and generative AI use in higher education. It covered respondent characteristics, patterns of generative AI use, perceived usefulness, role substitution, training need, institutional regulation, salient benefits and concerns and open-response reflections. The perceived-usefulness items were adapted from TAM/UTAUT-based instruments validated in comparable educational-technology contexts, most directly Buabeng-Andoh and Baah's (2020) integration of UTAUT and TAM for learning-management-system adoption and Vanderschaaf et al.’s (2021) study of student information-technology adoption, with wording modified to refer specifically to generative AI tasks; item phrasing was cross-checked against Or's (2024) meta-analytic synthesis of TAM constructs. The research team developed the single-response concern item and ethics-related open-response items, informed by the task-dependent concerns reported among Middle Eastern educators (Hamamra et al., 2024; Khlaif et al., 2025). The training-need and institutional-regulation items were developed separately with reference to the governance gap documented in ASEAN higher education (Buragohain and Chaudhary, 2025) and Indonesia's national guidance (Belmawa Dikti, 2024). The main adapted and literature-informed measures are supported by matching references. Supplementary material, Table S3, provides a detailed item-by-item mapping of the questionnaire, including each item's source and whether it was self-developed, adapted, or modified.

Perceived usefulness was assessed at two levels. A four-item measure covering academic efficiency, teaching-material development, scholarly-writing support and idea exploration showed acceptable reliability (Cronbach's α = 0.79). Seven additional task-specific items assessed usefulness across academic activities and were analyzed individually to identify task-selective adoption. Thus, α = 0.79 applies only to the four-item measure, while Table 2 reports item-level results for the seven task-specific ratings. Before data collection, the questionnaire was reviewed by an educational-technology expert specializing in generative AI and instrument development, then piloted with 30 lecturers excluded from the final sample. Feedback improved item clarity and contextual relevance. Other questions were analyzed as individual or categorical measures, so scale-level reliability was not applicable. Institutional readiness was not operationalized as a standalone psychometric construct or scale; it was interpreted from the separate training-need and institutional-regulation items, together with interview evidence concerning institutional guidance and formal training support.

Qualitative data were obtained through semi-structured interviews with eight lecturers purposively selected from survey respondents who indicated willingness to be interviewed further, with selection aimed at maximizing variation in age, institution type, rank, discipline and intensity of generative AI use. Interviews, conducted online or face-to-face and lasting approximately 45–60 min, addressed lecturers' experiences with generative AI broadly, while allowing them to discuss ChatGPT as the tool most familiar to them, together with perceived benefits and risks, changes in academic practice, ethical challenges and institutional readiness. The semi-structured interview guide is provided in Supplementary Appendix S1. All interviews were audio-recorded with participants' consent and transcribed for analysis; participants were assigned anonymized codes P01-P08. All eight transcripts were included in the thematic analysis; quotations were selected for thematic relevance, clarity and representativeness, and quotations from six of the eight participants are presented in Table 5.

Descriptive statistics were computed for all Likert-scale items directly from the verified dataset. A Shapiro–Wilk test was conducted on each ordinal outcome included in the group comparisons; all outcomes departed significantly from normality (W = 0.63–0.86, p < 0.001 in every case). Given these distributions, the ordinal five-point response formats and unequal group sizes, group differences were examined using the non-parametric Kruskal–Wallis H test rather than ANOVA (Creswell, 2009). Three grouping variables were examined: age, academic rank and institution type (public/Perguruan Tinggi Negeri (public higher education institutions) (PTN) versus private/Perguruan Tinggi Swasta (private higher education institutions) (PTS)). The complete H, df, p and effect-size results for all 30 comparisons are reported in Supplementary material, Table S1, while Table 3, Panel A, presents the four statistically significant comparisons. Epsilon-squared (ε²) was reported for comparisons involving more than two groups, using conventional small (≥0.01), moderate (≥0.08) and large (≥0.26) benchmarks; rank-biserial correlation was reported for two-group institution-type comparisons. No post hoc pairwise comparisons were conducted because the significant age and academic-rank effects for training need were interpreted at the omnibus level.

Interview transcripts were analyzed by the first author using descriptive thematic analysis (Morriss, 2024), which involved repeated readings of all eight transcripts, open coding, comparison of codes across participants, categorization and the progressive development of themes. Analytical memos and a coding audit trail documented code definitions, interpretive decisions and revisions to the thematic structure. A second team member reviewed the coding framework, candidate themes and supporting excerpts; differences in interpretation were discussed and resolved by returning to the original transcripts. Thematic sufficiency was judged to have been reached when the final interviews yielded no substantively new codes relevant to the research questions and representative quotations were selected for relevance, clarity and ability to illustrate recurrent and contrasting perspectives.

The two strands were integrated at the interpretation stage using a merging strategy that compared survey findings and interview themes across perceived usefulness, task-selective use, ethical ambivalence and institutional readiness. Interview evidence was used to extend and contextualize quantitative patterns rather than to test them as hypotheses, consistent with the equal-status convergent design; results are reported as a joint display in the Findings, following Fetters et al.’s (2013) convention of stating whether strands converge, diverge, complement, or expand one another.

The Ethics Commission for Health Research of the Faculty of Health Sciences, Universitas Respati Yogyakarta, Indonesia, approved this study. Before participating, survey respondents received an electronic information sheet explaining the study's purpose, procedures, voluntary participation, confidentiality and the right to discontinue. They provided electronic informed consent via a mandatory item on the first page of the questionnaire. Interview participants provided separate informed consent for participation and audio recording. Survey responses were analyzed and reported in aggregate; interview transcripts were de-identified (P01-P08). Electronic data were stored in password-protected files accessible only to the research team and retained for one year before secure deletion; no personally identifiable information is reported in this article.

Findings are reported in three parts, aligned with the study's research questions: quantitative, qualitative and integrated mixed methods.

Of 126 initial survey responses, six were excluded as incomplete, leaving 120 valid responses. Table 1, Panel A, presents the profiles of the survey respondents, and Table 1, Panel B, presents the profile of the eight purposively selected interview participants. The survey respondents represented public and private institutions and three broad disciplinary fields, while the interview participants varied in age, institution type, academic rank, discipline and intensity of generative AI use.

Respondents' task-specific ratings of generative AI usefulness varied substantially across academic tasks, indicating task-selective rather than uniform adoption. Table 2 reports the percentage of respondents who agreed or strongly agreed with each task-specific statement, together with the corresponding item-level mean and standard deviation. These ratings were analyzed separately to preserve differences across academic tasks and were not treated as a single composite scale.

Table 2 shows an overall graded pattern: universal agreement regarding academic efficiency, moderate agreement regarding presentations and teaching material development and lower agreement regarding scholarly articles, research and academic books. This item-level variation provides the empirical basis for interpreting lecturers' adoption as task-selective.

Because all ordinal outcomes included in the group comparisons departed significantly from normality (Shapiro–Wilk p < 0.001 in all cases), the Kruskal–Wallis H test was used to compare groups. Of the 30 comparisons conducted across age, academic rank and institution type, four were statistically significant. The complete results, including the non-significant comparisons, are reported in Supplementary Appendix S1, Table S1. Table 3, Panel A, summarizes the four statistically significant group comparisons and their effect sizes, while Table 3, Panel B, presents the percentage distribution of perceived training need by age and academic rank.

None of the task-specific usefulness items differed significantly by age or academic rank, indicating that the observed evaluations of AI across academic tasks were broadly shared across age groups and career stages. Institution type, by contrast, was associated with two items: a small-to-moderate effect for AI use in academic books (PTS median = 4, PTN median = 3; r = 0.29) and a moderate effect for the view that AI could eventually replace lecturers (PTN median = 4, PTS median = 3; r = −0.38). Perceived need for AI training, by contrast, differed significantly by age and academic rank, with a large effect for age (ε² = 0.43) and a moderate effect for academic rank (ε² = 0.19). Descriptively, “strongly needed” was selected by 55.0% of lecturers aged 41–50 and 44.1% of those aged over 50, compared with none of those aged 40 or younger. By academic rank, it was selected by 47.1% of Associate Professors and 42.9% of Professors, compared with 28.8% of Senior Lecturers and none of the Assistant Professors (Table 3, Panel B).

Table 4 reports respondents' single most salient positive perception and single most salient concern about generative AI.

The fact that “no significant negative impact perceived” was the most frequent response to the concern item qualifies the ethical-ambivalence narrative developed later in this paper: while a majority of respondents (80.0%) named a specific ethical or professional concern, a substantial minority did not. This finding should not be flattened into a claim that all or most lecturers experience acute ethical unease.

Beyond the training-need item, respondents were asked whether formal institutional regulation of generative AI use among lecturers was needed. Overall, 73.3% selected “needed” or “strongly needed,” and this response did not differ significantly by age, academic rank, or institution type. The perceived need for institutional regulation was therefore broadly shared, whereas the intensity of perceived training need varied significantly by age and academic rank.

Thematic analysis of all eight interview transcripts generated six interrelated themes: AI as an academic support tool; task-selective adoption; ethical ambivalence and epistemic risk; critical verification and professional responsibility; institutional governance gap; and limited formal training support. Table 5 summarizes the analytic meaning of each theme, participant coverage, coded references and representative quotations.

Task-selective boundary setting was the most widely shared and densely coded theme, followed by instrumental academic support and ethical ambivalence, thereby supporting the central finding that lecturers did not uniformly accept or reject generative AI but calibrated its use according to the academic and epistemic stakes of each task.

The interview accounts describe the same task-selective logic identified in the survey data: lecturers accept generative AI readily for preparatory and administrative tasks, while treating scholarly writing and formal research output as requiring far greater caution and verification, mirroring the usefulness gradient in Table 2. This is best interpreted as professional risk-management: lecturers calibrate their acceptance of AI assistance to the academic-integrity stakes of the task, rather than to the technology's general usefulness or ease of use.

Ethical ambivalence emerged as a second consistent pattern, but, consistent with the 20.0% “no significant negative impact perceived” survey finding, not as a uniform or totalizing concern. Several informants linked generative AI to specific, named risks, including the blurring of the boundary between academic assistance and plagiarism (P03) and changes in students' reading and information-seeking behavior (P06). Others offered a more measured, use-conditional view, holding that what mattered was how AI was used ethically while still preserving academic substance (P02). This spread, from acute concern to conditional acceptance, is consistent with the roughly four-fifths/one-fifth split between respondents who named a specific concern and those who did not.

The absence or ambiguity of institutional guidance was the theme that informants raised with the least ambiguity. Beyond P04's account in Table 5, informants linked unclear guidance to uneven understanding among colleagues and to the need for lecturers to establish their own boundaries for acceptable AI use. Participants also described formal training opportunities as limited (P07). The qualitative evidence, therefore, converges with the broadly shared survey-level need for institutional regulation and complements the survey finding that the intensity of training need varies by age and academic rank.

Table 6 presents a joint display integrating the quantitative and qualitative strands across four principal empirical domains, with an explicit statement of the relationship between strands for each domain.

Read together, the strongest and most consistent pattern across both strands is task-selective adoption grounded in perceived academic-integrity risk, not a simple usefulness-versus-resistance dichotomy. Institutional type emerges as a genuinely new and statistically supported axis of difference, discussed further below.

The steep decline in agreement from efficiency to scholarly writing and research (Table 2) confirms and sharpens a pattern already visible in prior usefulness-driven literature (Bamasoud and Mohammad, 2025; Onal et al., 2025). What this study adds is a within-sample comparison across seven distinct tasks, revealing a gradient rather than a single overall mean. This mirrors the assessment-specific caution documented among early-adopter instructors in the Middle East, who accepted generative AI more readily for lower-stakes assessment work (Khlaif et al., 2024). The mechanism, in both that study and the interview accounts reported here, appears to be epistemic risk: lecturers evaluate generative AI's utility conditional on the degree of authorship, originality and academic judgment a given task requires.

None of the task-specific usefulness items showed a statistically significant difference by age or academic rank. This absence is consistent with the AAOUJ finding that professional context and self-selected training, rather than age or demographic cohort, drove the intensity of generative AI use among distance-learning early childhood educators in Hong Kong (Wong et al., 2026): professional need, not chronological age, appears to be the more decisive variable. Perceived training need, by contrast, differed significantly by both age and academic rank, with a large effect for age and a moderate effect for academic rank (Table 3, Panel A), with direct implications for differentiated training design (see Practical Implications).

Institution type, in turn, was significantly associated with two attitudinal items, both with small-to-moderate effects: AI use in academic books and the possibility that AI could eventually replace lecturers. PTS lecturers were more favorable toward using AI for academic books than PTN lecturers. In contrast, PTN lecturers agreed more strongly that AI could eventually replace lecturers. Because the survey did not collect data on institutional AI policy, resourcing, or workload, these findings should be read as descriptive differences that require further explanation rather than as evidence of a causal mechanism. One plausible account, consistent with the institutional-readiness literature (Buragohain and Chaudhary, 2025), is that public universities' typically more bureaucratic promotion structures make the prospect of AI substituting for lecturers feel more institutionally salient, while private universities' higher individual publication pressures may make AI assistance with book-length work feel more instrumentally necessary.

The finding that 80.0% of respondents named a specific ethical or professional concern, while 20.0% reported none, indicates that ethical ambivalence is a majority pattern rather than a defining feature of the entire sample. The qualitative accounts explain both sides: informants who voiced concern most often cited the blurred boundary between legitimate assistance and plagiarism in the absence of clear regulation (P03), while those reporting no significant concern framed AI as ethically neutral, with responsibility resting on use (P02). This pattern is broadly consistent with the ambivalence documented among Palestinian university educators (Hamamra et al., 2024, 2026) and with the technostress associated with preserving academic quality while adopting generative AI (Khlaif et al., 2025), extending that literature by showing this ambivalence is unevenly distributed rather than shared by the whole population studied.

Consistent with regional evidence that ASEAN higher education institutions generally lack formal AI governance structures (Buragohain and Chaudhary, 2025) and with Indonesia's own uneven implementation of national AI guidance (Belmawa Dikti, 2024; Hanjani et al., 2025), 73.3% of respondents agreed that institutional regulation of generative AI use is needed, with no significant variation by age, rank, or institution type, the most uniformly held attitude identified in this study. The interview accounts suggest a mechanism linking this governance gap to ethical ambivalence: where institutions remain silent, lecturers construct their own boundaries around acceptable use (P04), compounding authorship ambiguity and epistemic risk across academic staff. This extends to the lecturer population the concern is that critical AI literacy in open universities in developing Asia remains largely foundational and requires deliberate institutional attention (Lim et al., 2025).

Building on the findings above, this study proposes negotiated professional adoption as an emergent, contextually grounded conceptual interpretation of how lecturers in this sample engage generative AI. It is not an established or externally validated theory but an interpretation bounded by the evidence reported here, requiring testing elsewhere. The concept describes a process in which lecturers do not simply accept or reject generative AI as a single decision, but continuously negotiate its pedagogical benefits, ethical risks, academic legitimacy and acceptable boundaries of use, on a task-by-task basis. Perceived usefulness is the only TAM construct directly examined here; perceived ease of use and facilitating conditions remain background constructs, neither operationalized nor independently tested. What TAM and UTAUT leave unexplained and what negotiated professional adoption adds, are the three dimensions set out in the conceptual framework: ethical ambivalence, institutional readiness and academic legitimacy/epistemic risk, each of which is directly evidenced above.

For universities. Institutions should move from general encouragement of AI use toward task-specific policies that distinguish between AI used for pedagogical and administrative support, and AI used as a substitute for scholarly authorship, mirroring the boundaries lecturers already draw informally. Clear authorship disclosure requirements and a documented academic-integrity mechanism for evaluating suspected misuse would directly address the regulatory uncertainty identified in the interviews.

For lecturers. The findings suggest lecturers would benefit from structured guidance on verification and output-quality evaluation, particularly for higher-stakes tasks, where the data show the greatest hesitation and the need for professional judgment, rather than generic AI-literacy training pitched at a single level regardless of task or seniority.

For ODL administrators. Given that professional context and self-directed, need-based learning were associated with more effective technology integration among distance-learning educators elsewhere in the region (Wong et al., 2026), administrators should consider differentiated, opt-in training tracks rather than uniform mandatory modules, calibrated to the training-need levels observed here by seniority (Table 3, Panel B).

For policymakers and the Ministry. Consistent with the division of responsibility implied by Indonesia's national guidance (Belmawa Dikti, 2024), the findings support ministry-level bodies setting minimum standards and a national framework, while individual institutions retain responsibility for operational policy and differentiated training, since the institution-type differences identified here suggest a single uniform national policy is unlikely to address the different pressures public and private universities face.

This study examined lecturers at public and private universities in Yogyakarta using purposive convenience sampling; the resulting sample of 120 lecturers had an unbalanced public-private ratio, and the survey's cross-sectional, self-report design means the findings cannot support causal or over-time claims about behavior change. The eight interviews deepened and explained survey patterns rather than representing the wider lecturer population, and negotiated professional adoption remains a conceptual interpretation requiring testing in other regions, disciplines and higher-education systems; these are boundary conditions on inference rather than defects in execution, and the convergence of survey, interview and joint-display evidence strengthens confidence in the central finding of task-selective adoption. Future research should pursue two directions in particular: multisite studies with more balanced public-private samples to test the institution-type differences identified here with greater statistical power, and longitudinal designs that track how lecturers' negotiation of generative AI and the validity of negotiated professional adoption itself develop as institutional policy and tool capabilities evolve.

This study examines how lecturers in Yogyakarta's higher education institutions perceive, use and set boundaries around generative AI, and finds that adoption is best described as task-selective and negotiated rather than uniform or unconditional. Lecturers overwhelmingly agree that generative AI improves academic efficiency. However, their agreement declines sharply as tasks move closer to scholarly authorship and formal knowledge production, a pattern consistently confirmed by survey and interview evidence. This pattern does not differ significantly by lecturers' age or academic rank; where age and rank matter is in a distinctly different question: how much formal training lecturers feel they need. Institution type is associated with two specific attitudinal differences that merit further investigation. Ethical ambivalence about plagiarism, dependency and declining critical reflection is a majority, though not universal, experience. It coexists with a broadly shared perceived need for institutional regulation, while interview participants described existing institutional guidance as absent or unclear.

The study's principal conceptual contribution, negotiated professional adoption, is offered as an emergent, contextually grounded interpretation of this pattern rather than a validated or generalizable theory: it describes lecturers as professional actors who continuously weigh pedagogical benefit against ethical risk, academic legitimacy and institutional silence, on a task-by-task basis.

These claims are bound to the lecturers and institutions studied in Yogyakarta and should not be read as representative of Indonesian higher education as a whole, of the Global South more broadly, or of ODL systems in general; the sample is a purposive convenience sample from a single province. The study's central contribution is nonetheless a well-evidenced empirical account, triangulated across survey and interview data, that generative AI adoption among lecturers is a continuing, task-by-task negotiation rather than a single acceptance decision and that institutional governance has not yet caught up with that negotiation. Pursuing the future research directions identified above would allow the field to determine whether negotiated professional adoption describes a pattern specific to this setting or a more general feature of how academic professionals engage generative AI under conditions of institutional uncertainty.

The authors would like to thank all lecturers who participated in this study and shared their professional experiences regarding the use of generative artificial intelligence in online and distance education environments.

The supplementary material for this article can be found online