Prioritizing benefits, costs, and contextual and individual factors in researchers’ adoption of generative artificial intelligence: a multi-criteria decision-making analysis

GAI in academic research

GAI is increasingly embedded across all main phases of academic research handled by researchers including discovery and exploration, writing and reporting, and data analysis. Tools like ChatGPT and Bard quickly identify research gaps, suggest hypotheses, and assist with methodological planning, significantly shortening the preliminary exploration phase (Dashti et al., 2023; Firat, 2023; Tsufim & Pomerleau, 2024). Literature review platforms, such as AI Scholar and Elicit, rapidly summarize and categorize extensive scholarly databases, enhancing researchers’ efficiency despite some limitations around accuracy (Alshami et al., 2023; Bond et al., 2024). GAI tools also automate data analysis tasks including data cleaning, statistical interpretation, visualization, and thematic coding, facilitating quicker insight generation, though they still require human oversight to ensure validity (Hassani & Silva, 2023; Semrl et al., 2023).

Further along the research workflow, GAI writing assistants support researchers by drafting, editing, and refining manuscript content, especially beneficial for non-native English speakers, but vigilance is needed to prevent subtle bias or inaccuracies (Ariyaratne et al., 2023; Kassem & Michahelles, 2023). Although traditional GAI referencing tools have limitations such as citation inaccuracies, recent plugins like Consensus and Scholar GPT, integrated with databases such as Google Scholar and JSTOR, offer improved reliability, though manual verification remains essential (Aiumtrakul et al., 2023; Khalifa & Ibrahim, 2024). Overall, while GAI greatly enhances efficiency and productivity, researchers retain a critical role in verifying outputs to maintain scholarly rigor and accuracy throughout the research process (Alasadi & Baiz, 2023; Korinek, 2023).

Benefits

Several studies highlight that integrating GAI into the research workflow enhances productivity and task efficiency by enabling researchers to achieve more work in less time, with reduced effort and fewer resources required (Bond et al., 2024; Bringula, 2023). These studies report quality and speed gains in academic writing and manuscript drafting, particularly for structured scientific prose and for non-native authors (Dergaa et al., 2023). Moreover, the knowledge acquired from GAI tools enhances researchers’ abilities in problem-solving and informed decision-making. GAI also increases the accessibility of advanced research techniques to researchers, regardless of their technical skills or institutional resources (Jarrah, Wardat & Fidalgo, 2023; Park, 2023). Additionally, the use of GAI fosters the development of researchers’ digital literacy and technical proficiency (Leong, 2023; Pigg, 2024). GAI tools also facilitate knowledge sharing and dissemination, supporting collaboration and effective communication within the academic community (Zohouri, Sabzali & Golmohammadi, 2024).

Barriers to adoption

Despite reported benefits, integrating GAI tools into research has generated societal and institutional concerns, often causing resistance from researchers and stakeholders. Key worries include ethical issues related to AI-driven decisions and transparency of use, accuracy risks from biases or errors in automated processes, cognitive risks arising from researchers’ overreliance on AI, and fears of job displacement due to automation. Public perceptions of AI-assisted research as morally questionable, less trustworthy, and inferior to human-led research reflect doubts about GAI’s capability to replicate human creativity and uphold ethical standards (Khogali & Mekid, 2022; Niszczota & Conway, 2023). At the institutional level, scholars emphasize potential declines in critical thinking and originality, risks of plagiarism, and threats to data privacy and security, stressing the importance of maintaining academic rigor (Rahman & Watanobe, 2023; Salvagno et al., 2023). Studies also reveal fragmented and cautious adoption among researchers despite awareness of tools like ChatGPT, with practical usage limited due to insufficient training and clear institutional guidelines (Abdelhafiz et al., 2024; Dorta-González et al., 2024; Iyengar et al., 2023). Addressing these barriers through institutional support is essential for promoting trust and effective integration of GAI into academic research workflows.

MCDM techniques in GAI adoption

MCDM covers canonical problem types including choice (selecting one or a few best alternatives), ranking/prioritization (ordering alternatives from best to worst), and sorting/classification (assigning alternatives to predefined categories) (Roy, 1996). Major methods include value/utility models such as the Analytic Hierarchy Process (AHP), ideal-solution approaches such as the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) and VIKOR (Multi-criteria Optimization and Compromise Solution), and outranking methods such as Elimination and Choice Translating Reality (ELECTRE) (Cinelli et al., 2020). Multi-attribute group decision-making (MAGDM) extends MCDM to multiple experts, who provide evaluations that are then combined (by aggregation or explicitly within the model) to yield a group ranking/choice. Ideal-solution methods such as TOPSIS and VIKOR are widely employed to produce group rankings in multi-attribute group settings, complementing weighting-oriented methods such as AHP/BWM (Opricovic & Tzeng, 2004; Shih, Shyur & Lee, 2007; Wan, Wang & Dong, 2013).

MCDM methods have been employed across various domains to systematically rank and prioritize factors influencing the adoption of GAI. In the supply chain sector, Sharma & Rathore (2024) utilized Delphi and Analytic Hierarchy Process (AHP) alongside sentiment analysis, highlighting ethical practices, societal trust, and transparency. Nguyen, Nguyen & Nguyen (2024) implemented a probabilistic fuzzy hybrid model combining AHP and Combined Compromise Solution (CoCoSo) in the telecommunications industry, emphasizing managerial competencies, technology readiness, and security. Wang et al. (2023) applied Delphi, Analytic Network Process (ANP), and fuzzy TOPSIS in the construction industry to focus on strategic implementation factors such as digitalization, innovation, and productivity.

Within the educational domain specifically, several studies have leveraged MCDM methods to prioritize factors influencing Artificial Intelligence adoption in the learning/teaching context. Gupta & Bhaskar (2020) for instance used the AHP for prioritizing factors influencing teachers’ adoption of AI-driven educational technologies for teaching and learning. Similarly, Wiangkham & Vongvit (2024) employed the Weighted Sum Method (WSM) and explainable AI framework to identify key drivers for ChatGPT adoption by students in higher education. Priya et al. (2022) employed ISM and DEMATEL, focusing primarily on organizational readiness for AI integration. Yet, as highlighted by Batista, Mesquita & Carnaz (2024) most studies so far have focused on student-facing applications of GAI (e.g., as learning aids or classroom tools) and on developing ethical guidelines for its use, rather than on faculty’s own research-oriented use of these tools. Empirical work examining how faculty adopt GAI in their scholarly research remains sparse, and no studies have employed structured decision models to prioritize the factors influencing faculty adoption of GAI for their academic research tasks. Addressing this gap is crucial to establishing best practices for adopting GAI in faculty research. Using structured decision-making methods to identify and rank the main factors influencing GAI adoption can help institutions clearly understand faculty needs. This approach enables universities to offer targeted support to researchers, promoting responsible and effective integration of GAI into their academic work.

Factor extraction

Screening and study selection

The database queries resulted in a total of 732 records, 274 from Web of Science and 458 from Scopus. Following the PRISMA guidelines, these records underwent initial filtering based on predefined inclusion and exclusion criteria, such as publication year, document type, language, relevance of titles, duplicate entries, and restricted access. This filtering step removed 462 ineligible records, leaving 270 studies for title screening. Subsequent title screening excluded an additional 122 studies, resulting in 148 reports advancing to the eligibility assessment stage. At the eligibility stage, another 119 reports were excluded due to their focus on learning and teaching applications solely. Ultimately, 29 studies satisfied all inclusion criteria and advanced to the full analysis stage, as depicted in the PRISMA flow diagram (Fig. 2).

Factor prioritization (ranking)

Reliability and convergence

Sensitivity testing

To assess the sensitivity of the rankings to the equal-weights assumption, a leave-one-expert-out jackknife analysis was performed. This test is a resampling method that removes one data column at a time and reruns the analysis (Quenouille, 1956; Tukey, 1958). Each expert was sequentially excluded by setting their weight to zero and renormalizing the remaining weights to sum to one. For each such perturbed weight set, TOPSIS and VIKOR rankings were recomputed, and the outcomes were summarized using the median Spearman’s ρ against the baseline and the degree of Top-5 factor preservation.

Primary studies profile

Table 6 summarizes the methodological profile and focus of the 29 primary studies. The studies utilized diverse research methods, including qualitative, quantitative, mixed methods, and review-based methodologies. Qualitative methods included case studies, grounded theory, thematic analysis, and descriptive frameworks. Quantitative studies predominantly used survey methods analyzed through structural equation modeling (SEM), alongside experimental and observational designs to examine faculty and researcher adoption patterns, perceptions, and performance comparisons between GAI and human writing. Mixed-method approaches combined qualitative insights from interviews with quantitative survey data, offering a holistic exploration of GAI adoption.

Regarding the research context, the reviewed studies collectively address various aspects of GAI integration in higher education and research. Contextual focuses include faculty adoption behaviors, perceptions toward ChatGPT, academic writing practices, and peer-review processes. Studies also explored ethical issues, collaboration dynamics between humans and AI, coping mechanisms, and accuracy-related concerns such as hallucinations.

Thematic analysis results

Addressing RQ1 the thematic analysis identified 18 factors across 29 studies grouped under three main themes: Costs, Benefits, and Contextual & Individual. As shown in Fig. 3, per-document coverage favors costs (dark blue) across most studies, while benefits (green) are lower and more variable, and contextual & individual factors (light blue) are least emphasized.

Figure 4 presents the frequency distribution of the coded factors. “Costs” account for the largest share (51%), representing perceived barriers, and risks that may hinder researchers’ adoption of GAI. “Benefits” make up 28% of the coded data and reflect enabling factors and positive outcomes, that encourage adoption. “Contextual and Individual” factors contribute the remaining 21%, encompassing external conditions and personal characteristics shaping researchers’ adoption decisions. The figure demonstrates that cost-related considerations dominated the coding, followed by benefits and contextual/individual influences.

Figure 5 illustrates how often each thematic factor appears across the reviewed documents. Ethical Concerns and Performance Risk were mentioned most frequently, each appearing in all analyzed documents (100%). These were closely followed by Task Efficiency and Facilitating Conditions, representing 97% and 93% of the documents respectively. Cognitive Risk was also frequently mentioned, appearing in 66% of documents, followed by Accessibility (62%) and Knowledge Acquisition (59%). Knowledge Sharing was discussed in just over half of the sources (55%), while AI Literacy and Regulatory Concerns each appeared in nearly half (48%). Less frequently noted were Experience and Effort Expectancy, both at 45%. Hedonic Motivation (38%), Social Influence (28%), and Demographics (14%) appeared less often, with Innovativeness, Self-Efficacy, and Job Security being the least frequently mentioned, each at 10%. Although the frequency of thematic factors reflects their prominence in analyzed studies, a slightly lower frequency does not necessarily mean reduced importance. Differences in study objectives, research designs, or specific tasks can influence which factors are emphasized. For instance, Facilitating Conditions appeared in 93% of documents, slightly below the highest frequency, but this might reflect that certain studies were experimental, focusing specifically on testing GAI for tasks like abstract writing or citation generation, rather than directly examining enabling conditions or support systems.

The distribution of factors based on coded segment frequency is shown in Fig. 6. Ethical Concerns dominated with the highest frequency, accounting for 22.3% of all coded segments. Performance Risk followed at 17.0%, with Efficiency and Facilitating Conditions close behind 16.3% and 15.9%, respectively. Effort Expectancy accounted for a moderate 6.7% of segments. Social Influence was cited less frequently at 4.2%, while Knowledge Acquisition (2.8%), Cognitive Risk (2.3%), Knowledge Sharing (2.0%), and Accessibility (2.0%) were referenced sparingly. Hedonic Motivation appeared in 1.9% of segments, and the least frequently represented factors were AI Literacy (1.4%), Experience (1.3%), Regulatory Concerns (1.2%), Self-Efficacy (1.1%), Demographics (0.8%), Innovativeness (0.4%), and Job Security (0.3%).

The direct co-occurrences of the identified factors within the analyzed segments are illustrated in Fig. 7. The map represents thematic relationships with a minimum frequency of five or more co-occurrences. Node size and font reflect how frequently each theme occurred, with larger nodes indicating higher overall occurrence. Line thickness represents the strength of specific relationships between pairs of themes, with thicker lines indicating more frequent co-occurrences. Ethical Concerns is the most prominent node, indicating substantial overlap with various themes, especially Performance Risk (38 segments). The strongest single co-occurrence observed is between Efficiency and Effort Expectancy (33 segments). This emphasizes the frequent association of GAI’s productivity benefits with users’ perceptions regarding learning curve and the degree of effort required to use effectively use these tools. Other notable connections include Efficiency with Facilitating Conditions (19 segments), underscoring institutional support’s role in enhancing users’ perceived efficiency. Contextual and individual factors such as AI Literacy, Knowledge Acquisition, and Experience show moderate to fewer connections, suggesting they primarily serve as moderating influences.

Contextual and individual themes

The thematic analysis identified several specific contextual and individual factors influencing the adoption of GAI tools by researchers in academic contexts. Seven sub-codes were identified (Demographics, innovativeness, self-efficacy, facilitating conditions, prior experience, AI literacy, and social influence) as illustrated in Table 7, along with the supporting studies. Several studies indicated that demographic factors of age and gender affect adoption level. These studies reported that age works as a moderator that strengths or weakens intentions to adopt GAI for instance, younger individuals generally have greater familiarity and more positive perceptions toward GAI tools. Another possible moderating factor is gender where male researchers showed higher intentions and usage rates, especially under conditions of high effort expectancy. Personal innovativeness was found to significantly influence adoption intentions and moderate related adoption factors. Additionally, researchers with higher self-efficacy were more inclined to adopt GAI tools. Facilitating conditions, including clear institutional policies, tailored training and support programs, and collaboration among institutions, policymakers, and developers, were highlighted as essential for effective adoption. Prior experience and familiarity (Experience), along with digital and (AI Literacy), were also noted as influential factors affecting researchers’ attitudes and usage. Figure 8 shows the distribution of contextual/individual factors based on document frequency. Facilitating Conditions appeared most frequently, mentioned in 93% of documents. AI Literacy was the second most common factor at 52%, followed by Experience at 45%. Social Influence was discussed in just over a quarter of the studies (28%), while Demographics, encompassing age and gender, appeared in only 14%. Self-Efficacy and Innovativeness were the least frequently mentioned individual-level factors, each present in only 10% of the documents.

Benefits

The thematic analysis identified five key benefits (knowledge acquisition, hedonic motivation, knowledge sharing, accessibility, task efficiency) influencing researchers’ adoption of GAI tools as shown in Table 8. Knowledge acquisition involves gaining new information through personalized and adaptive learning strategies. These features enhance researcher engagement and support decision-making. Hedonic motivation was another identified benefit, reflecting users’ enjoyment and entertainment from interacting with GAI tools, which subsequently enhances their intention to use it academically. Knowledge sharing and collaboration emerged prominently, including global academic collaboration facilitated by GAI, improved interdisciplinary partnerships, and peer-connectedness within academic communities. Additionally, accessibility benefits were identified, highlighting GAI’s ability to facilitate broader resource access 24/7 and promoting inclusivity and equity. Task efficiency presents GAI capability to streamline research by automating research stages of discovery, analysis, drafting and writing, thus reducing time and effort. Figure 9 presents the distribution of benefit-oriented factors based on document frequency. Task Efficiency emerged prominently, appearing in 97% of the analyzed documents. Accessibility was cited in 62% of documents, followed closely by Knowledge Acquisition (59%) and Knowledge Sharing (55%). Hedonic Motivation was the least frequently mentioned benefit, appearing in 38% of the documents.

Costs

The thematic analysis identified six primary cost-related factors (job security concerns, effort expectancy, cognitive risk, regulatory concerns, performance risk, ethical concerns) that can hinder researchers’ adoption of GAI tools in academic contexts shown in Table 9. Job security emerged as a concern due to the perceived risk of job displacement and automation. Effort expectancy, referring to the effort required by users, impacts adoption intentions, indicating tools perceived as more intuitive and effortless are more likely to be adopted. For researchers using GAI, this refers to efforts required to learn, operate, and supervise GAI tools in academic research, including the time, cognitive load, and additional work needed to integrate and verify their outputs. Cognitive risk includes concerns about compromised critical thinking and analytical skills resulting from overreliance on AI-generated content. Regulatory concerns encompass worries about insufficient oversight, lack of institutional guidelines, and potential copyright issues related to AI-generated content. Performance risks address issues of content accuracy, AI-generated hallucinations, and AI’s limitations in evaluating complex tasks. Finally, ethical concerns emphasize risks related to unethical usage, plagiarism, over-reliance, and issues involving data privacy, security, transparency, algorithmic bias, and explainability of AI decisions.

Figure 10 illustrates the distribution of cost or risk-related factors based on document frequency. Both Performance Risk and Ethical Concerns were universally mentioned, appearing in all analyzed documents (100%). Cognitive Risk followed with significant frequency at 66%, while Regulatory Concerns and Effort Expectancy were cited in approximately half of the documents, at 48% and 45%, respectively. Job Security appeared least frequently, being cited in only 10% of the analyzed sources.

MCDM results

To answer RQ2 we used TOPSIS and VIKOR to assess and rank pool of factors identified in the thematic analysis stage (Table 10). Consolidated ranking of adoption factors based on TOPSIS (CC_i) and VIKOR (Q_i), showing scores, individual method ranks, combined ranks, and overall rank (with ties as ranges). Ethical Concerns ranks first in both methods, emphasizing its critical role. Both methods closely agree on the top factors: Task Efficiency is second, while Facilitating Conditions and Performance Risk alternate between third and fourth positions. However, some differences appear in middle rankings; for instance, Knowledge Acquisition ranks sixth in TOPSIS but seventh in VIKOR, swapping positions with Cognitive Risk. Regulatory Concerns moves from tenth in TOPSIS up to eighth in VIKOR, and Accessibility similarly advances from eleventh to ninth. Effort Expectancy shows the greatest disagreement, ranking eighth in TOPSIS (closeness = 0.637) but dropping notably to twelfth in VIKOR (Q = 0.163). The overall ranking was obtained by averaging the TOPSIS rank and the VIKOR rank for each factor. Factors were then sorted by the mean of ranks, with ties shown as ranges. This yielded Ethical Concerns 1st, Task Efficiency 2^nd, Facilitating Conditions and Performance Risk joint 3–4, and AI Literacy 5^th.

Similarly, Fig. 11 shows that TOPSIS and VIKOR ranks largely coincide, with the largest rank divergence at Effort Expectancy. Despite these variations, the Spearman rank correlation coefficient remains high at ρ = 0.963 (about 96.3% agreement), reinforcing the reliability and consistency of the rankings as illustrated in Table 11.

A leave-one-expert-out test showed extreme stability. For TOPSIS and VIKOR alike, removing each of the 81 expert columns produced a median Spearman correlation of ρ = 0.998 with the baseline order; the Top-5 factors (Ethical Concerns, Task Efficiency, Facilitating Conditions, Performance Risk, and AI Literacy) remained unchanged in all reruns as illustrated in Table 12.