The ways of teaching and learning have been transformed throughout history. Various technologies have been used to achieve effectiveness in these processes, and the 21st century presents new alternatives with artificial intelligence. Potential is found in intelligent tutoring systems and massive open online courses (Feng and Law, 2021), as well as in adaptive learning, teaching assessment, virtual classroom (Huang et al., 2021), and e-learning of students, mainly in engineering (Prahani et al., 2022). Nemorin et al. (2023) identify AI possibilities for niche market expansion. Ouyang and Jiao (2021) states that the AIED development trend supports the personalization of student learning, to lead to iterative development of personalized, data-driven and learner-centered learning. AI can help in identifying teaching tasks and teaching content accurately, where teaching strategies can be linked to learning profiles (Lin, 2022). AI can be applied to social networking sites and chatbots, expert systems for education, mentors and intelligent agents, machine learning, personalized educational systems, and virtual educational environments to develop professional competences (Tapalova and Zhiyenbayeva, 2022). AI presents potentials for education, but it is also important to consider related risks.

The risks when applying AI in education can span a wide spectrum. Debate and controversy have been generated among teaching communities and corporate AI giants (Al Braiki et al., 2020), where increased ethical risks (Lameras and Arnab, 2023) and bias are identified when using AI tools and techniques to model and inform instructional decisions and predict learning outcomes (Dieterle et al., 2022; Ibarra-Vazquez et al., 2023). Of note are concerns regarding personal data and learner autonomy (Nguyen et al., 2023), as well as in the governance of AI risk in human-centered education (Li and Gu, 2023). Risks are also related to equity, accountability, transparency, and ethics (Khosravi et al., 2022); as well as in the misuse of AI due to algorithm bias and lack of governance that inhibit human rights and lead to labor, gender, and racial inequality (Yang et al., 2021). Other risks include teacher labor automation and global platform infrastructures embedded in education (Rensfeldt and Rahm, 2023). The risks are important points for practical applications and must be addressed through educational research.

Researching the potential of using AI in education can establish a link with pedagogical innovation processes and technological applications. Indeed, AI research in education has increased in recent years (Kaban, 2023; Vázquez-Parra et al., 2024b), with a high number of institutions and researchers with publications indicating a new trend for the artificial intelligence in education (AIED) community (Bittencourt et al., 2023). Chen et al. (2020) identified the main AI research topics as intelligent tutoring systems for special education; natural language processing for language teaching; educational robots for AI education; educational data mining for performance prediction; among others. AI, as a field of study, identifies innovations and resulting developments that have culminated in computers, machines and other artifacts with human-like intelligence, characterized by cognitive abilities, learning, adaptability, and decision-making capabilities (Chen et al., 2020; Ibarra-Vazquez et al., 2024). Practical applications include Benhadj et al. (2019) who integrated serious games into classroom management on a large scale, leading to improved discipline, motivation, and participation in the classroom. Also Andersen et al. (2022) used block-based programming in small groups to solve teacher-given tasks. Practical application research sheds light on knowledge transfer and it is also important to locate opportunities for further educational research.

The AIED research field presents opportunities to further increase its educational application possibilities. There is an urgent need for research and development in teacher preparation as well as in the philosophy of technology in education to bridge the gap between AI and education (Pham and Sampson, 2022). There are still doubts about the ability of AI to monitor students' behavior and direct learning, improve the efficiency of the education system, provide grades and reviews, reduce dependency on teachers, and improve social interaction (Al-Tkhayneh et al., 2023). For example, in the socioemotional domain of students, Lai et al. (2023) found that AI had a negative impact on adolescents' social adaptability, and that it is significantly negatively correlated with social adaptability and family support. In the same vein, Xie et al. (2022) highlight the need to target interventions according to the relationship between psychosocial factors and social adaptability to enhance the positive influence of AI and promote the development of social adaptability. Possibilities are located for systems to be implemented, models, standards, and invisible evaluations of learning (Rahm and Rahm-Skågeby, 2023). Opportunities increase when looking for studies on practical applications in communities in the Global South.

Science clubs are educational settings where problem-solving is promoted through intentional, non-formal, school-based learning. In informal education, Burke (2023) analyzed STEM clubs to develop a community-responsive approach to out-of-school club programming. Larina et al. (2023) worked in the health field with student science clubs at a medical university to prepare students for future professional activities, locating motivational factors for learning. Similarly, Ilyenko et al. (2023) studied motivation with medical students participating in science clubs and found a high demand for the experience of research work, the expansion and deepening of knowledge in the specialty, as well as the positive impact of participation in science clubs on the further training of future clinicians. Sewry et al. (2023) also found positive attitudes from science club work that integrated kitchen chemistry practices, with improvements in conceptual understanding and improved student performance. Science clubs are presented as mobilisers of high ability thinking through challenging learning environments.

In this research, complex thinking is conceptualized as a cross-cutting macro-competency that encompasses four interrelated sub-competencies—systemic, critical, innovative, and scientific thinking. Together, these enable learners to address problems from diverse perspectives and develop integrated solutions to multifaceted, real-world issues (Sanabria-Z et al., 2024). Systemic thinking refers to the ability to recognize, model, and interpret the interconnections within complex systems (Cruz-Sandoval et al., 2023b). Critical thinking is understood as the capacity to challenge underlying assumptions, evaluate arguments, and examine evidence (Patiño et al., 2023). Innovative thinking denotes the ability to generate original and contextually appropriate solutions to problems (Suárez-Brito et al., 2024). Lastly, scientific thinking is defined as the process of posing hypotheses, designing empirical investigations, and interpreting findings to construct evidence-based knowledge (Cruz-Sandoval et al., 2023,a).

Training in scientific and complex thinking involves generating different and motivating scenarios. Their development in time might be influenced by different factors such as gender (Brage-del-Río et al., 2025; González-Gallego et al., 2025) For instance, Medina-Vidal et al. (2023) identified and substantiated the existence of a gender gap in the development of perceived complex thinking competency and its sub-competencies since women's perceived competency levels diminished as they progressed through their formative process. In the quest to develop scientific thinking and attract more individuals to science and technology studies, Lövheim (2014) worked with science clubs to make science more interesting and fun for students and to change young people's attitudes toward these topics. Also, Romaniuk et al. (2023) integrated science clubs and placed the development of medical students' research skills, systematization of information, and science communication. In the field of humanitarian sciences, Ermachkov et al. (2018) applied practices and methodological approaches in the organization of students' scientific activities, aiming to improve the quality of students' scientific research. Buenestado-Fernández et al. (2023) analyzed participants in science clubs in Mexico and found that science education activities with a gender perspective are necessary in non-formal education spaces such as science clubs.

Building on these insights, the inclusion of gender as an analytical lens in this study is grounded in evidence documenting persistent differences between men and women in STEM participation and competency development. Social and cultural factors, such as gender stereotypes linked to technical skills, unequal access to role models or mentors, and societal expectations regarding career paths, can shape both engagement and self-perception in STEM learning environments (Hernández-Pérez et al., 2024; Beroíza-Valenzuela and Salas-Guzmán, 2024). In non-formal contexts such as science clubs, prevailing norms may favor male participation, for example by framing activities around tools traditionally coded as masculine or through competitive formats that may limit female engagement. The literature identifies strategies to foster more inclusive environments, including the design of collaborative projects with real-world applications, balanced representation among facilitators, and the explicit discussion of gender bias during activities (Archer et al., 2025). Furthermore, Beroíza-Valenzuela and Salas-Guzmán (2024) highlight that educational resources, cultural norms, and institutional support mechanisms play a decisive role in reinforcing or reducing gender gaps in STEM engagement. Considering that science clubs are central to fostering both scientific and complex thinking competencies, integrating a gender perspective enables us to examine how inclusivity—or its absence—may influence the development of these competencies across diverse student populations. From this perspective, how can innovative spaces are encouraged in science clubs that work with AIEDs that promote high capacities?

This study employed the multiple case study methodology (Yin, 1993) to examine at an individual and collective level the development of complex thinking and scientific thinking among 83 students participating in AI-themed science club activities. A case study can be defined as an integrated and bounded system, emphasizing its nature as an object rather than a process (Stake, 2007). It entails a specific and intricately operational entity that can encompass individuals, institutions, characteristics, and interrelationships (Yin, 2006). In the context of this study, a case includes a unique configuration of students, mentors, technologies, and pedagogical strategies within each club. The rationale for selecting science clubs as the unit of analysis lies in their hybrid nature: they combine project-based, exploratory, and interdisciplinary learning, making them ideal micro-environments for studying how learners engage with abstract and complex domains such as AI.

In this study, each science club is considered a bounded case comprising a specific configuration of students, mentors, AI tools, thematic focus, and pedagogical strategies. While the present analysis focuses primarily on quantitative data derived from standardized instruments, the case study framing is grounded in the recognition that each club operates within a unique educational micro-context. These contextual differences—including thematic variation, participant composition, and implementation setting—inform the comparative interpretation of results. The multiple case study approach thus enables the identification of patterns across diverse contexts while preserving the integrity of each case as a distinct analytical unit. We acknowledge that this iteration of the research does not incorporate qualitative coding or ethnographic elements; however, rich contextual descriptions and standardized cross-site protocols provide the comparative depth characteristic of multi-case designs. Future work will integrate interviews, open-ended responses, and observational field notes to further align methodological practice with a fully mixed-methods multiple case study approach.

The comparative dimension of the study was carefully designed to allow for both cross-case analysis and contextual sensitivity. While students in each club engaged in different AI-themed projects, all were guided by a shared pedagogical structure emphasizing problem-solving, creative design, and collaborative exploration. The thematic variation across clubs was not a limitation but a feature: it enabled a richer understanding of how students activate scientific and complex thinking in diverse yet comparable educational contexts.

Each science club integrated hands-on, project-based activities directly related to the thematic focus and AI tools under study. For example, in “Artificial Intelligence Applications: Data Science for Maternal Health and Cancer Care” (OAX5), students learned data analysis in R and ArcGIS, applying geospatial methods to explore environmental determinants of health. In “AI-Powered Drug Discovery” (GTO2), participants worked with computational models to identify potential new uses for existing drugs. “Bits and Atoms: Quantum Computing and Machine Learning” (MTY3) provided an introduction to quantum computing concepts alongside practical machine learning exercises, while “Sensory Expansion” (MTY5) focused on programming algorithms to detect luminescent properties of materials for diagnostic purposes. In “Adventures in AI” (GDL1), students explored everyday applications of machine learning, including self-driving laboratory simulations, and in “Untangling the Neurons of Artificial Intelligence” (GDL2) they designed and trained artificial neural networks inspired by brain processes. These case-specific activities illustrate the diversity of contexts in which students engaged with AI concepts and developed higher-order thinking skills.

The present study was guided by four research questions aimed at understanding how students engage with complex and scientific thinking through their participation in AI-focused science clubs. The participant group consisted of 83 students-−42 women and 41 men—enrolled in six AI-themed clubs implemented in four different Mexican cities. These clubs provided a range of thematic areas, such as computing, medicine, and robotics, allowing for an exploration of potential variations in the development of competencies across contexts. Specifically, the study sought to explore:

These questions were addressed through a multiple case study design, which allowed for in-depth analysis of both individual experiences and contextual variations across six different science clubs. By focusing on these dimensions, the study aimed to generate insights into how informal, project-based learning environments can foster key transversal competencies in diverse student populations.

Prior to any data collection activities, the entirety of science club participants enrolled in AI-themed clubs were invited to participate in the study and were provided with detailed information about the nature of the research, its potential impact, and the intended use of the collected data. Participants were explicitly informed about their voluntary participation and were assured of the confidentiality and anonymity of their responses. While all the science club participants took part in the club activities, this study presents results of 83 high school and university students having consented to participate in the study distributed in six groups or cases. This collection of cases, each representing a unique and complex phenomenon with practical educational applications of AI, has been analyzed to identify the scientific skills and complex thinking competency of the participants. Each group or case was presented with a series of learning activities including a game-base activity and a questionnaire to collect data regarding their complex thinking competency.

Data collection was semi-simultaneous, involving the participation of a team of six trained researchers and research assistants. The participant recruitment process began with obtaining formal permission from the coordinators of each science club to invite both participants and instructors to join the study. Once approval was granted, instructors were contacted via e-mail and invited to an online briefing session where the research objectives, data collection instruments, and participant activities were presented in detail, following the main points of the research protocol. Recruitment of club participants took place at the onset of the club activities, with invitations extended directly during the first session. To ensure methodological consistency across sites, the research team co-designed a standardized protocol and held virtual coordination meetings prior to fieldwork to align procedures, ensuring that the same instruments were applied in the same order and manner in each location. This distributed data collection structure was necessary due to the clubs being held concurrently across different cities (i.e., Guadalajara, Monterrey, Oaxaca, and Guanajuato) and separate locations within the city. While data collection did not occur at the exact same time in all four cities, it occurred during the same time span, which made it impossible for the same researchers to be in-person in all groups. Despite the logistical complexity, a consistent protocol was used to ensure coherence in data gathering. In each case, students interacted with AI content under similar pedagogical conditions, enabling comparisons while respecting the integrity of each unique context. In this sense, the analyzed cases serve as microcosms, allowing for an in-depth exploration of their unique contexts and the development of their activities. Thus, the comparability of the cases is grounded in their shared structure and theoretical alignment, while the value of the comparison lies in highlighting how varied educational contexts mediate students' engagement with AI and the development of higher-order thinking skills. Table 1 presents the science club name and the description of the covered thematic areas, the utilized technologies, and sociodemographic data to describe the participants.

E-Complexity is a Likert-scale questionnaire developed to assess students' perception of their development in the complex thinking competency and its four sub-competencies: critical thinking, scientific thinking, innovative thinking, and systemic thinking (Sotelo et al., 2023). The instrument consists of 25 items rated on a five-point Likert scale, ranging from “Strongly disagree” (1) to “Strongly agree” (5). E-Complexity was designed through a comprehensive three-stage validation process, including theoretical, design, and content validation by subject-matter experts. A theoretical review of existing instruments revealed the need for a more integrative tool focused specifically on the structure of complex thinking. Additionally, the instrument has undergone rigorous statistical validation (Cruz-Sandoval et al., 2023). Its internal consistency, evaluated using Cronbach's alpha, achieved a coefficient of 0.93, indicating excellent reliability across its items (Vázquez-Parra et al., 2024a). Furthermore, Vázquez-Parra et al. (2024b) confirmed its construct validity through confirmatory factor analysis within a structural equation modeling framework, which supported the theoretical four-subcompetency structure and reinforced the instrument's credibility for measuring these latent constructs.

The analytical framework of qualitative multiple case studies often extend to include descriptive statistics analysis. This statistical dimension serves as a complementary tool to summarize and present key characteristics and trends within and across cases. While qualitative in nature, this incorporation of descriptive statistics offers a quantitative lens through which patterns, variations, and emergent themes can be elucidated, adding a layer of empirical rigor to the comprehensive qualitative analysis. In this multiple case-study research, descriptive statistical analyses were performed using the R software and RStudio. We emphasized measures of central tendency, specifically the mean and standard deviation. To enrich our research, we also employed tools such as Boxplots, Principal Component Analysis (PCA), Biplot visualization, ANOVA, and the t-test.

The Boxplot, or whisker diagram, graphically displays data distribution by quartiles, emphasizing its spread, symmetry, and potential outliers. PCA, on the other hand, is a dimensionality reduction technique that converts a set of correlated variables into uncorrelated principal components. The first of these components captures the most variability in the data, followed by the second, and so on. This analysis is visualized with the Biplot, projecting both observations and variables onto a plane defined by the principal components. ANOVA analysis is used to evaluate differences between the means of three or more groups, determining if these variations are statistically significant. Lastly, the t-test allows us to determine if there are significant differences between the means of the cases. All analyses were conducted in R version 4.5.1 (2025-06-13) and RStudio version 2025.5.1.513 (“Mariposa Orchid”), using the following packages: dslabs, dbplyr, tidyverse, ggplot2, devtools, and ggbiplot.