Piloting And Validating Tailored Pre- and Post-Intervention Assessments for PRODIGI Target Groups

1 2 Please cite this report as: Azadi, T., d’Haenens, L., De Nolf, A., Ponte, C., Luna, E., Tomczyk, Ł., Donoso, V., Torres da Silva, M., Batista, S., & Żegleń, M.(2025). Piloting And Validating Tailored Preand Post-Intervention Assessments for PRODIGI Target Groups. PRODIGI, KU Leuven. Disclaimer PRODIGI is funded by the European Union, under Grant Agreement no. 101182849. Views and opinions expressed are, however, those of the author(s) only and do not necessarily reflect those of the European Union. The European Union cannot be held responsible for them. 3 PILOTING AND VALIDATING TAILORED PREAND POST-INTERVENTION ASSESSMENTS FOR PRODIGI TARGET GROUPS Work package 1 – Deliverable D1.2 Submission date: 19 December 2025 Lead beneficiary: KU Leuven Authors: Tania Azadi, Leen d’Haenens, Ans De Nolf, Cristina Ponte, Estrella Luna, Łukasz Tomczyk, Verónica Donoso, Marisa Torres da Silva, Susana Batista, Magdalena Żegleń 4 Table of contents 1. Executive Summary ........................................................................................................................................ 6 Key insights ................................................................................................................................................... 7 From the rapid review: .................................................................................................... 7 From the experts: ............................................................................................................ 7 From the cognitive interviews: ......................................................................................... 7 From the pilot study: ....................................................................................................... 7 In conclusion .................................................................................................................. 8 2. About the PRODIGI Project ......................................................................................................................... 9 Objectives ............................................................................................................................................................... 9 3. Introduction ................................................................................................................................................... 13 4. Methodology ................................................................................................................................................. 14 PRODIGI AI literacy definition and conceptualisation ........................................................................ 14 AI Literacy Scale Design ................................................................................................................................. 17 Item and Scale Development Phase ............................................................................. 17 Question and Answer Scale Formulation of the Skill Items ............................................. 19 Question and Answer Scale Formulation of the Knowledge Items ................................... 19 Belgian context .................................................................................................................................................. 20 Expert consultation ....................................................................................................... 20 Cognitive interviews with vulnerable youth .................................................................... 20 Pilot survey testing ........................................................................................................ 21 Portuguese context ............................................................................................................................... 22 Expert consultation and cognitive interviews ................................................................. 22 Polish context .......................................................................................................................................... 22 Expert consultation ....................................................................................................... 22 Ethical Considerations and Governance ........................................................................ 23 Limitations ................................................................................................................... 23 5. Findings ........................................................................................................................................................... 25 Belgian context .................................................................................................................................................. 25 Expert consultation ....................................................................................................... 25 Cognitive Interviews ...................................................................................................... 28 Pilot survey testing ........................................................................................................ 36 5 Portuguese context .......................................................................................................................................... 45 Expert consultation & cognitive interviews ..................................................................... 45 Polish Context .................................................................................................................................................... 50 Expert consultation ....................................................................................................... 50 6. Conclusion and Recommendations ...................................................................................................... 61 7. Acknowledgements ..................................................................................................................................... 62 8. References ...................................................................................................................................................... 63 9. Appendices ..................................................................................................................................................... 66 A. Rapid Review study overview ................................................................................................................. 66 B. Distribution of Correct Responses Across the 30 AI Knowledge Items: Results of the Survey in Belgium ........................................................................................................................................................... 67 C. Full overview of Initial AI Literacy Scale .............................................................................................. 68 AI Skill Items ................................................................................................................. 68 AI Knowledge Items ...................................................................................................... 69 D. yAILS Refined Scale .................................................................................................................................... 70 E. Initial sAILS ..................................................................................................................................................... 76 6 1. Executive Summary PRODIGI adopts preand post-test designs, using the Youth AI Literacy Scale (yAILS) and the Senior AI Literacy Scale (sAILS), to assess AI literacy and digital skills. Both the Youth AI Literacy Scale and the Senior AI Literacy Scale are newly developed within the PRODIGI project to ensure age-appropriate measurement and contextual relevance. In this report we document and report the development, piloting, and initial validation of PRODIGI’s preand post-intervention assessment tool for measuring AI literacy, among vulnerable populations in Belgium, Portugal, and Poland. This assessment tool forms the empirical backbone for evaluating upcoming intervention programmes in PRODIGI’s Work Package 2. First, we began with a rapid review of existing digital-skills and AI-literacy assessments. The review showed that many existing tools: • focus on self-perceived knowledge instead of demonstrated skills, • rarely address the needs of vulnerable groups, • do not include emerging competencies around AI systems, such as understanding algorithms, identifying bias, or reasoning about ethical implications. In response, PRODIGI built upon the Youth Digital Skills Indicator (yDSI)0F0F 1 as the conceptual framework and developed a new scale covering the four underlying digital skill dimensions of yDSI as well as a fifth domain: ethics & responsible AI use, extending the framework to capture contemporary AImediated challenges. Our instrument, hereafter referred to as the AI Literacy Scale, was translated into partner languages and tested by structured expert consultations and cognitive interviews with target groups to ensure comprehensibility, cultural fit, and age appropriateness. Moreover, a largescale pilot (among 1,200 youngsters in Belgium) enabled exploratory factor analysis (EFA) to examine the instrument’s structure. Instead of the originally proposed five distinct conceptual domains, the revised scale for youth is organised into four theoretically grounded and empirically coherent dimensions that capture how young people understand and use AI in practice: technical and operational, information navigation and processing, communication and content creation, and ethical and responsible use. These patterns reflect how AI tasks require integrated, not isolated skills: prompting, creating, evaluating, and ethical reasoning often co-occur. This supports a conceptual shift away from modular digital skills toward interwoven competencies shaped by human-AI interaction. This pilot identifies substantial knowledge gaps in AI-usage, as nearly half of the young participants achieved scores of 16 or fewer correct answers across 30 AI knowledge items. AI understanding varied strongly by educational track but not by gender. Qualitative co-design research in Portugal, Poland, and Belgium further shows that youth, older people, and experts repeatedly requested the need for clearer, simplified language and flagged scale items that felt too abstract, overly technical, or irrelevant to their lived experiences. This feedback was systematically integrated into substantial revisions of the scale items, enhancing clarity, age appropriateness, and contextual relevance across target groups. 1 The yDSI (Youth Digital Skills Indicator) was developed under the H2020-ySKILLS (Youth SKILLS) project. 7 Key insights From the rapid review: The rapid review showed that several existing tools for assessing media, digital, and AI literacy are inadequate. Many rely on self-reported abilities rather than measuring what people can actually do, and they overlook skills that are increasingly important in an AI-driven world, such as understanding how AI works, recognising algorithmic influence, spotting biased or unfair information, or its societal and environmental impact. These tools were also not designed for or tested among vulnerable groups, including unaccompanied minors, disadvantaged youth, or older adults, who have varied levels of access and experience with technology. Given these gaps and the rapid pace of AI development, there is a clear need for updated assessments that place a strong emphasis on ethics and the responsible use of AI while taking into account vulnerable and marginalised populations. From the experts: The expert consultations helped improve relevance and clarity as well as the content and the wording of the assessment. Experts recommended using simpler language and removing technical terms that may be unfamiliar to many young people or in at-risk situations (e.g., due to a migration background). They also suggested adding clear examples, particularly for questions concerning responsibility, fairness, and other ethical issues, to help respondents relate these concepts to real-world situations. In addition, experts noted that some questions overlapped and encouraged a clearer separation among creative, ethical, communication, and information-checking skills. Overall, they stressed that the assessment should be easy to understand across diverse cultural and linguistic contexts and should be developed in collaboration with the users to ensure it remains clear and relevant. From the cognitive interviews: Across all three countries, our interviews indicated that participants across age groups are often curious about AI, yet struggle with the way current AI-related questions are formulated. Many young people and seniors do not realise how much AI they already use in everyday apps, find technical terms (such as “algorithm”, “bias”, “deepfake”) and long, abstract sentences complex to understand, and are confused by questions that do not match their real lives or skills. The items that worked best were short, concrete, written in simple language, and tied to familiar activities such as using a chatbot for homework, writing birthday wishes, finding recipes, or checking whether a message looks suspicious. Overall, the key message is that, to measure AI literacy fairly, we need fewer questions, more precise wording, everyday examples, and, at times, additional explanation or support. From the pilot study: The pilot study with 1,200 young people helped test PRODIGI AI literacy scale for youth. The results showed that young people use AI-related skills in combination rather than as separate abilities. The first exploratory factor analysis (EFA) was conducted on split sample 1 and yielded a four-factor solution. This solution was retained as the initial measurement model for the newly developed scale because it provided the clearest and most interpretable structure, with comparatively fewer problematic cross-loadings and less conceptual overlap between factors than alternative solutions. Items clustered into four meaningful components: Factor 1 (Content Creation and Communication), 8 Factor 2 (Ethics), Factor 3 (Information Navigation), and Factor 4 (Technical/Operational).This differs from earlier models and reflects how young people actually use AI. The study also found that many participants had limited knowledge about how AI functions. Nearly half answered 16 or fewer out of 30 questions correctly. Performance differences were primarily attributable to educational track, with academic students scoring higher than those in vocational programmes, whereas age and gender had little effect. In conclusion Developing this AI literacy scale requires several important considerations that emerged from our research with experts and with participants across Belgium, Portugal, and Poland. First, the findings showed that people do not think about AI skills in separate categories. Instead, skills such as content creation, communication with AI tools, information verification, and technology operation tend to blend in practice. The scale should therefore reflect these natural “hybrid” skill areas rather than forcing them into strict boxes. Second, the language used in the scale must be familiar and straightforward. Many participants among PRODIGI vulnerable groups, across different age groups and backgrounds, found technical or academic terms confusing. Terms such as “algorithm,” “framework,” “multimodal,” and even “bias” were often unclear; therefore, the scale should include everyday alternatives, such as “unfair treatment” or “fake video,” and provide clear explanations of concepts that may be unfamiliar to some users. Third, it is important that the questions feel relevant to people’s real lives. Young people connect best with examples involving social media, homework help, or creative tasks, whereas older adults relate more to everyday activities such as finding recipes, writing messages, or checking opening hours. Questions that assume prior knowledge, such as programming skills or an understanding of plagiarism, should be included only if they are explained. Fourth, the scale must also stay focused and easy to complete. Removing overlapping questions and avoiding items that ask two things at once will help prevent confusion and reduce respondent burden. Finally, users' voices must guide the design. Throughout the research, experts often considered certain items clear, whereas participants found them confusing. In these cases, user feedback should inform decision-making. This approach helps ensure that the scale reflects the real experiences, needs, and abilities of the people it is intended to serve, particularly vulnerable groups such as refugees, older adults, and students in vocational pathways. 9 2. About the PRODIGI Project Misinformation and disinformation pose significant contemporary challenges, affecting individuals and societies alike. Their consequences are particularly harmful in high-stakes domains such as public health, where false information has been associated with risky behaviours, lower vaccination uptake, and heightened anxiety and depression. Beyond individual harms, misinformation can erode trust in journalism, science, and public institutions, thereby weakening the shared informational foundations that democratic societies rely on. These risks have intensified as AI technologies increasingly shape how content is produced and consumed. Deepfakes can convincingly fabricate audio-visual “evidence,” and widely used conversational chatbots can generate fluent, authoritative-sounding text at scale, making misleading narratives easier to create, adapt, and circulate rapidly while obscuring provenance and accountability. In this context, AI literacy becomes directly relevant to misinformation resilience not simply as technical familiarity, but as a set of competencies that help people recognise how AI systems generate and prioritise content, understand their limitations (e.g., hallucinations, bias, and opacity), and apply appropriate verification strategies. Recognising that AI and digital literacy are not standalone solutions, this project approaches AI literacy as a multifaceted capability that combines critical judgement (e.g., source evaluation, detection of manipulation, etc.) with operational skills (e.g., using tools and settings safely and effectively). This is especially important for at-risk populations who face digital disparities and may have fewer resources to verify information, making targeted AI literacy efforts a necessary component of broader responses to AI-amplified misinformation. PRODIGI investigates AI literacy and digital skills within the context of vulnerability, focusing on three demographic groups: unaccompanied minors in Belgian refugee centres, economically disadvantaged youth in Portuguese vocational education, and elderly individuals in Polish small cities and rural areas. Objectives PRODIGI’s key objectives are as follows: 1. To assess AI Literacy and Digital Skills using performance tests in three different countries: Belgium, Portugal, and Poland. This assessment involves three distinct target groups. In Belgium, we will focus on unaccompanied minors and their guardians/youth workers. In Portugal, our attention will be on children and young individuals from socio-economically disadvantaged families enrolled in professional education programmes. In this context, we will also involve their teachers/trainers. Finally, in Poland, we will focus on elderly residents of rural areas who receive assistance from pre-teachers in lifelong learning centres. This intergenerational contact will facilitate the use of digital media and promote AI literacy among older people. We aim to conduct a comprehensive evaluation of AI literacy and digital skills intervention (with a focus on misand disinformation) within these diverse groups, taking into account the unique contexts and requirements of each country. 2. Following this, we will work together with members of the target populations and individuals in caregiving roles to collaboratively improve relevant existing intervention programmes. We will then design, develop, and implement targeted interventions and assess its effectiveness through preand post-tests. 3. Furthermore, we will create a resilience toolkit tailored explicitly for the implementation of preand post-tests, enabling the measurement of intervention programme outcomes. This 16 In this report, we describe the methodology we used to develop and validate our scale, the Youth AI Literacy Scale (yAILS) as well as the Senior AI Literacy Scale (sAILS). yAILS and sAILS hold the promise to be a valid and reliable assessment tool to assess AI literacy of youth and seniors. This scale tries to address the existing inequalities in communication research and in knowledge production within the field by co-designing and co-validating the tool across the vulnerable and marginalised groups (unaccompanied minor and refugee youth as well seniors). Co-designing and co-validating the AI Literacy Scale with vulnerable youth and seniors is a deliberate effort to ensure that the tool is both inclusive and contextually relevant. By involving disadvantaged youth, refugee youth and seniors (PRODIGI target groups) in the development process, we aim to capture diverse lived experiences and to avoid reproducing dominant cultural biases that are often embedded in standardised assessments. This participatory approach empowers our young and elderly participants as co-creators of knowledge, promotes equity in communication research, and enhances the scale’s validity across different social and cultural contexts. It also aligns with ethical research practices by respecting the agency of those whose realities the scale seeks to measure (Moll et al., 2020; Mulvale et al., 2019; Peters et al., 2024). Developed using best-practice frameworks for scale development, the scale is grounded in rigorous design principles and adheres to established validation standards, including assessments of criterion, convergent, and discriminant validity. This approach addresses limitations in individual tools and enhances the scale’s applicability to diverse populations, enabling meaningful comparisons across groups. A distinctive contribution of yAILS is its explicit inclusion of the ethical and responsible use of AI, a dimension often neglected as a separate domain in current assessments. In parallel with the development of the Youth AI Literacy Scale (yAILS), the PRODIGI project has initiated the development of the Senior AI Literacy Scale (sAILS). Building on the conceptual framework, item pool, and validation strategy of yAILS, sAILS is being adapted to reflect the specific life contexts, usage patterns, and learning trajectories of older adults. Particular attention is paid to differences in cognitive load, digital experience, accessibility needs, and everyday AI encounters that are especially salient for senior populations (e.g., voice assistants, recommender systems, and AIsupported public and health services). The development of sAILS follows the same best-practice principles of inclusive and participatory scale construction that underpin yAILS. Professionals working with older adults were actively involved through interviews and cognitive testing to assess item clarity, relevance, and interpretability, thereby ensuring that the scale captures meaningful dimensions of AI literacy without relying on age-biased assumptions or deficit-oriented framings. This co-design and co-validation process aims to avoid reproducing generational biases commonly found in digital skills assessments and to foreground seniors’ experiential knowledge and practical competencies. Preliminary analyses focus on evaluating the feasibility of a shortened and accessible version of the scale that balances psychometric robustness with practical constraints related to survey length and respondent burden. Ongoing validation work examines reliability, factor structure, and measurement equivalence across age groups, laying the groundwork for future comparative analyses between youth and senior populations. Through this parallel and aligned development process, sAILS extends the PRODIGI project’s commitment to inclusive, context-sensitive, and ethically grounded measurement of AI literacy across the life course. 17 AI Literacy Scale Design We followed the best practices for developing and validating scales for health, social, and behavioural research developed by Boateng et al. (2018), which acts as the gold standard in the domain of scale development and validation. There are three phases to creating a rigorous scale. They include item development, scale development, and scale evaluation. In the following sections, we describe the procedures and steps undertaken to develop and validate the yAILS. Item and Scale Development Phase Two main activities in item development are domain identification and item generation, followed by expert and target population evaluation (Boateng et al., 2018). We applied a combined deductive and inductive approach to identify our domains or dimensions and to generate items. We rely on the four dimensions (domains) of the cross-country-validated yDSI framework, the literature we reviewed, and interviews with 15 youth (ages 13-17), including four with lower socioeconomic status, to identify the fifth dimension. That also shaped part of our co-design approach toward the yAILS item development and domain identification. Instead of introducing yet another standalone instrument, yAILS builds on the strengths of previously validated AI literacy measures to provide a more comprehensive and robust assessment tool. By integrating well-established items from existing scales, yAILS ensures broad content coverage and capitalises on the proven validity of prior instruments. We built upon a systematic review of AI literacy scales (Lintner, 2024), which identified 22 studies validating 16 scales targeting various populations, including the general population, higher education students, secondary education students, and teachers. Due to the limited number of studies focusing on children and youth, we also considered other validated scales developed for other populations. These 16 scales were included because they generally demonstrated good structural validity and internal consistency. The review by Lintner was conducted up until 2024, so to cover the literature published after 2024 we conducted a rapid literature review in accordance with established guidance for streamlined evidence synthesis (Tapia-Benavente et al., 2021). This method was particularly appropriate for the rapidly evolving field of AI literacy, where new frameworks and assessment tools are continually developed (Cotilla & Esther, 2025). Using this approach, we obtained a clear picture of what is already known and identified the key ideas that should inform the development of a new AI literacy scale, particularly for vulnerable or underserved communities. The literature search was conducted across several major academic databases, including WOS, Scopus and ACM Digital Library and relevant reliable reports, including the UNESCO AI Competency Framework, the OECD Artificial Intelligence Principles, and the European Union’s Ethics Guidelines for Trustworthy AI. Search terms combined descriptors related to artificial intelligence literacy, digital competence, algorithmic literacy, computational thinking, responsible AI, and assessment. The initial search produced 1595 records. These records were screened in two stages after deduplication (n=341). First, titles and abstracts were reviewed to exclude publications that did not address AI literacy scales, questionnaires, or assessment instruments. Full-text screening was then conducted using predefined inclusion and exclusion criteria based on the PRODIGI–AI Literacy Scale Scoping Review protocol. To be included, studies had to present the development, adaptation, or validation of an AI literacy scale or measurement tool and report empirical, quantitative, or mixed-method data. Eligible studies were required to provide at least one form of validity evidence (such as content, construct, convergent, discriminant, or criterion validity) and at least one reliability measure, including internal consistency metrics (e.g., Cronbach’s alpha), composite reliability, or test–retest reliability. 18 Publications also had to be peer-reviewed, written in English, and published from 2024 onward. Studies were excluded (n=1169) if they: • consisted solely of conceptual or theoretical discussions of AI literacy; • contained educational interventions without psychometric testing; qualitative studies that did not generate a quantitative measurement instrument; • reported on tools assessing general digital, media, or information literacy without a specific AI focus; public opinion polls lacking psychometric evidence; • or were duplicate or incomplete records, such as preprints without peer review or abstracts without full text. In total, 42 studies met the criteria and were included in the review. From these studies and supplementary grey literature, we identified 18 existing AI literacy and AIrelated competence scales. An overview of these items and thematic categories is provided in Appendix A. Items and thematic categories from each scale were extracted and mapped onto the four dimensions defined by the yDSI framework. This process identified 307 initial items later on refined to 60 items aligned with the operational, informational, communicative, and content creation dimensions of the yDSI. These items included both skill-based and knowledge-based components and represented current thinking about what young people need to understand and be able to do when interacting with AI systems. During the conceptual mapping process, it became clear that ethical and responsible engagement with AI technologies is essential to AI understanding. Constructs such as fairness, transparency, accountability, privacy, sustainability, and societal impact appeared consistently across guidance from international organisations, including the UNESCO AI Competency Framework, the OECD Artificial Intelligence Principles, and the European Union’s Ethics Guidelines for Trustworthy AI. These concepts were not fully represented in the original four-dimensional yDSI framework. Given their prominence in modern discussions of AI literacy and their relevance to safe and informed participation in digital societies, we added a fifth dimension to the developing scale: Ethical and Responsible AI Use. This new dimension added 12 items, covering the responsibilities, risks, and broader societal implications associated with AI. This dimension, although not part of the original yDSI framework, was considered crucial for capturing the specific challenges and responsibilities associated with the use of AI systems in contemporary society. 19 Question and Answer Scale Formulation of the Skill Items We also followed the wording of the questions in accordance with the yDSI guideline. The original question asks: “Please indicate how true the following statements are of you when thinking about how you use the internet and technologies such as mobile phones or computers. Reply, thinking about how true this would be of you if you had to do it now, on your own. If you do not understand what the question is asking, tick the box “I do not understand what you mean by this” To align the item with the assessment of AI-related skills, references to general digital technologies were replaced by AI tools (e.g., chatbots, image generators, voice assistants). All items are scored on a Likert-type scale measuring self-perceived skill proficiency: • Not at all true of me • Not very true of me • Neither true nor untrue of me • Mostly true of me • Very true of me • I do not understand what you mean by this • I do not want to answer Question and Answer Scale Formulation of the Knowledge Items For the knowledge items, similarly, we followed the wording of the questions in accordance with the yDSI guideline. The original question asks: “To what extent are the following statements about technologies such as the internet and mobile phones true or not true? If you are not sure, please let us know”. In order to adapt the question to our scale objective, which assesses skills in relation to AI, we changed the phrase “technologies such as the internet and mobile phones” to “AI tools, such as chatbots (like ChatGPT), image generators (like DALL·E), or voice assistants (like Siri or Google Assistant)”. Scale Evaluation Phase To assess whether the 60 items adequately measure the five dimensions of interest, we conducted two types of evaluation: expert evaluation and evaluation by our target population. Expert evaluation was conducted by expert judgments using an online survey, and target population evaluation was conducted using cognitive interviews with disadvantaged youth, refugee youth, as well as seniors. The methodology and results of expert evaluation and cognitive interviews have been provided in the next sections per each participating country. To complement these qualitative stages, a pilot study was conducted in which the initial version of the scale was administered to young people in Belgium (n=1200). This pilot provided preliminary quantitative evidence regarding item performance, internal consistency, and the emerging factor structure. It also allowed us to test administrative procedures and evaluate the instrument's ability to capture variation in literacy levels within an at-risk population. In what follows, we will detail the methodology for each of the three research contexts. 20 Belgian context Expert consultation To strengthen the development of the PRODIGI AI Literacy Scale, we collected expert feedback via an online survey. The goal was to evaluate the clarity, relevance, and appropriateness of the draft scale items and to gather suggestions for improvement before further testing with young people and other target groups. Experts were invited to participate based on their experience in AI literacy, digital literacy, media education, and educational assessment. Invitations were sent by email to five experts in digital skills, of whom three responded and agreed to participate in the survey, and included a link to the survey, along with a brief explanation of the project’s goals and the purpose of the scale. Participation was voluntary. For each item, experts were asked to rate how relevant the statement was for measuring different dimensions of AI literacy and how clear the wording was. Both ratings used a four-point scale ranging from low to high. In addition to these quantitative ratings, open-ended comment boxes enabled experts to explain their reasoning, identify ambiguous or unclear wording, suggest alternative formulations, or flag missing concepts that should be included in the scale. In total, the survey produced both structured ratings and qualitative comments. These responses were systematically reviewed to identify patterns, such as items consistently rated as unclear, items considered highly relevant, and areas in which experts recommended revisions to improve the scale's overall structure. Overall, they stressed that the assessment should be easy to understand across diverse cultural and linguistic contexts and should be developed in collaboration with the users to ensure it remains clear and relevant. A detailed analysis is provided in the next section, which presents the findings. Cognitive interviews with vulnerable youth Six cognitive interviews were conducted with five 17-year-old boys from Afghanistan, Syria, and Eritrea and one 15-year-old from Afghanistan. An overview of all informants is provided in Table 1. Interviews were conducted at Refugee Centre Fedasil Overijse, in Belgium on 26/11/2025 by two researchers. Table 1: Overview of informants Participant Gender Age Nationality Interviewed by 1 Boy 17 Afghanistan Researcher 1 2 Boy 17 Syria Researcher 1 3 Boy 17 Eritrea Researcher 1 4 Boy 15 Afghanistan Researcher 2 5 Boy 17 Syria Researcher 2 6 Boy 17 Afghanistan Researcher 2 Researcher 1 had access to a private room to conduct the interviews. She conducted the interviews in English and had a translator present for three of them. Researcher 2 interviewed in an open area accessible to both residents of the refugee centre and personnel. During these interviews, there was movement and interaction among people. This affected the concentration levels of both the children and the researcher. Despite this, three interviews could be finalised. 21 Both researchers obtained participants' consent to audio-record the interviews. The boys interviewed by researcher 1 consented to the recordings, whereas researcher 2 could not record due to the interviewees’ preference; therefore, only handwritten notes were taken during the interviews. The research team presented participants with the entire set of scale visual cards from which to choose. The cards were printed in different background colours representing the different item categories of the scale. Researchers gave participants complete agency as regards which cards they chose. This helped make participants feel at ease, which in turn facilitated the interviewing process. In total, 15 items were tested in English and 16 in Dutch, with one item tested twice in Dutch. Two of the three children interviewed first observed the cards laid out on the table, selected one, looked at it, and either kept it or discarded it. When asked why some cards were being discarded, participants consistently responded that the question was too long. Sometimes they added that the question was difficult or that they did not understand it, and therefore preferred to choose another card. Overall, the interviews showed that many young people found technical terms such as “algorithm,” “bias,” “multi-modal,” “framework,” “deepfake”, etc. difficult, struggled with long or abstract sentences, and were confused by questions that did not align with their everyday experiences or perceived skills. A detailed analysis is provided in the next section, which presents the findings. Pilot survey testing To assess how well the emerging AI literacy framework functioned in practice, we conducted a pilot survey among secondary school students (n=1200; age: 16–18) in Flanders, Belgium to explore whether the structure of traditional digital skills models could adequately capture the competencies required for interacting with AI systems. Earlier frameworks, such as the yDSI within ySKILLS, were designed for environments in which people worked with digital tools rather than intelligent agents. In those contexts, skills could be measured separately because tasks such as searching for information, navigating interfaces, adjusting settings, communicating, or creating content typically unfolded in clear, modular steps. However, AI systems break this modularity. Interacting with chatbots, multimodal generators, or recommendation engines requires users to interpret information, communicate effectively, exercise creativity, and make ethical judgments simultaneously. For example, writing an effective prompt requires both linguistic clarity and creative thinking, whereas evaluating AI outputs involves ongoing attention to fairness, accuracy, and potential biases, as well as concerns such as cognitive offloading, attention fragmentation, and reduced originality (Thaker et al, 2025). These overlapping cognitive demands mean that the skill areas traditionally treated as distinct (creativity, communication, information evaluation, and ethics) become closely intertwined during AI use (Shippee, 2024). Our pilot testing, therefore, examined how these competencies emerged together within real interactions with AI systems, revealing a more hybrid factor structure than that found in classical digital skills models. Rather than reflecting isolated task categories, the resulting domains of AI literacy capture the lived experience of engaging with generative AI, in which interpretation, co-creation, and ethical reasoning unfold simultaneously. The first exploratory factor analysis (EFA) was conducted on split sample 1 and yielded a four-factor solution. This solution was retained as the initial measurement model for the newly developed scale because it provided the clearest and most interpretable structure, with comparatively fewer problematic cross-loadings and less conceptual overlap between factors than alternative solutions. Items clustered into four meaningful components: Factor 1 (Content Creation and Communication), Factor 2 (Ethics), Factor 3 (Information Navigation), and Factor 4 (Technical/Operational). This differs from earlier models and reflects how young people actually use AI. The study also found that many 22 participants had limited knowledge about how AI functions. A detailed analysis is provided in the next section, which presents the findings. Portuguese context Expert consultation and cognitive interviews An expert review was conducted to assess the clarity, relevance, and age-appropriateness of our AI Literacy Scale for youth enrolled in a range of vocational programmes. Four expert reviewers, comprising AI specialists, teacher educators, and digital literacy practitioners, were recruited to examine each survey item and provide detailed qualitative feedback. Reviewers received the draft scale along with guidance prompts inviting them to comment on terminology, conceptual alignment with curriculum objectives, potential misunderstandings, and any perceived redundancy among items. Experts annotated the items directly and submitted written commentary. Their observations were then consolidated into an analytic memo synthesising recurring themes, including problematic jargon (e.g., “frameworks”, “multimodal prompting”), unfamiliar or abstract terms (e.g., fidedigna, enviesamento), and items requiring contextualization or definition. Reviewers also flagged ageinappropriate content (e.g., assumptions about prior programming knowledge or plagiarism) and identified overlapping material across sections. This feedback was systematically coded and used to generate item-level recommendations. The expert review served as the first stage of the co-design validation process, informing revisions prior to youth cognitive testing. The implementation and analysis of the cognitive interviews was guided by Haddon and Ponte (2012) and draws on annotations generated during cognitive testing sessions with 10 Portuguese students. The students were enrolled in a range of vocational programmes: Performative Arts, Cooking, SocioCultural Mediation, Legal Assistance, Child Care Support and Sports Training. Two of them had a Venezuelan background. Feedback from both groups contributed important perspectives on item clarity, accessibility, and perceived relevance. Throughout the analysis, illustrative examples and quotations from student participants are used to highlight recurring concerns, such as requests for more precise terminology or for identifying items that were confusing or overly complex. These insights informed the item-level recommendations and helped ensure that the scale is both valid and responsive to youth perspectives. Polish context Expert consultation To better understand how digital tools for older adults can be improved, particularly in the context of rapidly evolving technologies and AI, the Polish research team conducted a qualitative study using expert interviews. The study took place in the second half of November 2025 and was conducted online. Several well-known Polish specialists in digital education for older adults were invited to participate. These experts were selected because they have extensive experience in training older adults, advising organisations, and conducting academic research on digital inclusion. Each expert took part in a semistructured interview. This format allowed the Polish team to explore recurring themes while also giving participants the freedom to expand on their experiences and insights. With consent, all interviews were recorded and transcribed. 23 The interview data were analysed using thematic analysis, which enabled the identification of recurring topics, concerns, and recommendations (Ahmed et al., 2025). Through this process, the Polish team identified key themes, including the need for simple, clear language in diagnostic tools, the importance of avoiding overly technical or lengthy assessments that may overwhelm older adults, and the value of using real-life situations that are familiar and practical for older adults. Points of agreement among experts, as well as the most frequently mentioned challenges and solutions, were highlighted during the analysis. By bringing together experts’ hands-on experience and existing knowledge from adult education and learning in later life, the Polish team gained a clearer picture of what older adults need from digital assessment tools. The themes that emerged from the interviews informed practical recommendations to make future tools easier to understand, less cognitively demanding, and better suited to the real challenges seniors face when using digital technologies and AI in their everyday lives. Ethical Considerations and Governance The development and planned validation of the yAILS and sAILS adheres to rigorous ethical standards appropriate for research involving minors and vulnerable populations. Ethical approval for this study was obtained from the Social and Societal Ethics Committee (SMEC) at KU Leuven, Belgium (G-202510348). The research protocol, including recruitment procedures, consent materials, and data handling plans, was reviewed by the committee to ensure compliance with KU Leuven institutional and international research ethics guidelines for research involving children and youth. Special attention was given to informed consent and assent procedures. All participants under the age of 18 were asked to provide assent, while their legal guardians was required to provide informed consent too. For refugee youth and unaccompanied minors, additional safeguards was implemented in collaboration with care institutions (e.g., Fedasil) and legal representatives to ensure ethical participation and comprehension of study procedures such as having a trusted guardian/social worker present or available during the session, using age-appropriate and plain-language information sheets and verbally checking understanding (e.g., asking participants to explain in their own words what participation involved), emphasising voluntariness and the right to stop or skip questions at any time without consequences for services or support, and ensuring privacy and confidentiality by conducting interviews in a safe space and limiting access to identifiable data to authorised research staff only. Participation was voluntary, and all participants were informed of their right to withdraw at any time without any negative consequences. They were also reminded of this right during the session, and were encouraged to take breaks, skip questions, or stop the interview whenever they felt tired or uncomfortable. Beverages and small snacks were provided throughout the sessions to support participants’ comfort. Confidentiality was strictly maintained throughout the study. Personally identifiable information was not collected, and all responses was anonymised prior to analysis. Data was stored securely on encrypted servers managed by KU Leuven (e.g., ManGo), with access restricted to authorised research personnel only. Backup procedures and data retention policies complied with GDPR standards. This ethical framework ensures that the rights, dignity, and safety of all participants, especially those from vulnerable and marginalised groups, are protected throughout the research process. Limitations While the scale has been developed through a rigorous test-theoretical framework and co-designed with vulnerable populations, some limitations should be acknowledged : 24 • Context-Specific Design: Items were co-developed with youth in specific cultural and educational contexts. While efforts were made to ensure inclusivity (like a pilot test with 1200 children in 5th and 6th years of secondary school in Flanders, Belgium), the scale may require adaptation for use in other regions or populations. • Language and Literacy Constraints: Although the scale was designed for youth aged 14–19, variations in reading level, language proficiency, and digital exposure may affect item comprehension in some subgroups. • Confirmatory Factor Analysis (CFA) and Independent Validation: Finally, confirmatory factor analysis (CFA) has yet to be conducted; therefore, the stability and fit of the factor model should be tested in an independent sample before the structure is considered definitive. 25 5. Findings This section summarises our findings while testing and improving the AI Literacy Scale (yAILS and sAILS) in Belgium, Portugal, and Poland. We combined survey data from 1,200 young people with feedback from experts, students, and senior learners. In Belgium, we analysed survey responses to see how different AI skills naturally group together. In Portugal and Poland, we examined how clear, relevant, and realistic the questions were for young people and older adults with very different language and digital skills. Together, these findings indicate which parts of the scale perform well, which questions are confusing or repetitive, and how AI skills are used in everyday life. They also guide the simplification and shortening of the scale and its alignment with EU digital education goals, so that it can be used reliably with diverse groups of learners. Belgian context Expert consultation An expert review was conducted to evaluate the Youth AI Literacy Scale (yAILS), focusing on the clarity and relevance of sixty statements designed to measure AI literacy in young people. Three experts rated each item on a four-point scale for both relevance and clarity and provided optional qualitative comments. The purpose of this analysis was to assess the overall scale strength, identify items requiring revision, and determine whether the current set of statements appropriately reflects key aspects of AI literacy. Main findings Relevance The expert feedback on the relevance of the questions in the Youth AI Literacy Scale (yAILS) focused heavily on item redundancy, the categorisation of specific skills, and the practical likelihood that users would perform the described tasks. Overall, the experts rated the scale's content as highly relevant. The average relevance score across all items was 3.28 out of 4, and more than 80% of ratings fell within the two highest categories (3 or 4). This indicates that the reviewers agree that the scale successfully captures important concepts within AI literacy. Statements addressing foundational ideas (such as the possibility that AI systems may generate harmful content, mix true and false information, or provide different outputs to different users) received the highest ratings for both relevance and clarity. Clarity Clarity ratings for the Youth AI Literacy Scale (yAILS), while generally positive, showed greater variability than relevance ratings, with an average score of 2.89 out of 4 and notably more evaluations falling into the lower clarity categories. This indicates that, although the content is broadly appropriate, the wording of many items remains somewhat complex, academic, or ambiguous for the target youth audience, including teenagers, refugees, and individuals with low digital literacy. Reviewers frequently noted that vocabulary such as “integrate,” “insights,” “neutral,” and “comprehensive understanding” was overly advanced, while technical terms like “multimodal,” “frameworks,” and “programming languages” were likely to be misunderstood. Items with the lowest clarity scores commonly included abstract phrasing, lengthy or multi-part sentence structures, or 32 Item comprehension analysis based on participant feedback How participants categorised the difficulty of the items during the cognitive interviews is reflected in the sub-sections below (Table 3). 1. Items considered difficult to explain (Participants understood the general idea but could not articulate or paraphrase it) • Items in English: S17, K24, S23, S15, S28, S27, K2 These items were cognitively demanding although participants could grasp part of the meaning, they struggled to explain the concept, give examples, or clarify what the item was asking. This suggests abstract wording, layered ideas, or unfamiliar terminology. • Items in Dutch: K3, K29, S17 o K3: In this item the term “patroon” was identified as difficult. Although the participant claimed to understand the question, it remained unclear for Researcher 2 if this was really the case due to difficulties to rephrase the question and to give concrete examples to illustrate item comprehension. In a later conversation with an Okan teacher, this item was flagged as very confusing and abstract for students from Okan class. The recommended simplifying the item, avoiding as much as possible abstract terms and using basic, everyday language and shorter sentences. o K29: Apart from having difficulties to grasp the meaning of specific terms such as “zorgt” [causes] and “milieu” [environment], the participant did not seem to grasp the question. This was probably due to a lack of awareness about the impact of AI on the environment. o S17: In this item, the terms “een evenement organiseren” [organise an event] were unknown. The supervisor suggested modifying this example by “een feestje maken” [have a party]. After the clarification, and despite understanding the meanings of most individual words in this item, the question remained difficult. 2. Items not understood at all (Participants indicated they did not understand the meaning of the item.) • Items in English: K1, S23, S6 These items exceed the participants’ current knowledge of AI or use terminology that is too technical for this age/language group. These items may require simplification, additional context, or examplebased framing. • Items in Dutch: K11, K20, S4, S10, S18, S29 o K11: The term “voorordelen” (bias) was unknown to the participant, and it was difficult to explain in simple, concrete terms. Researcher 2 got the impression that despite her explanation, the item was not understood. o K20 despite understanding the meaning of most words in the question, except “meningen” (opinions), the participant was unable to grasp the meaning of the question. 33 o S4: In this item the terms “App” and “AI” caused difficulties and got somehow confused with each other. It is unclear if this confusion stems from the participant’s little familiarity with AI and/or to the low Dutch proficiency level of the participant. o S10: The participant indicated that he was not sure if he understood the question, even after explaining the meaning of the terms that were flagged as difficult, namely “betrouwbare bronnen”[reliable sources] and “vergelijken” [compare]. o S18: Although the participant with the highest command of Dutch mentioned that this item was clear to him, Researcher 2 got the impression that the participant was using his previous knowledge of social media platforms to answer the questions (e.g. he knew how to get likes on a post). During the conversation with the researcher, it was not evident that the participant grasped if nor how AI impacts social media. S29: In this item, the term “dagelijkse leven” [daily life] was flagged as difficult. The rest of the terms were familiar to the participant with the highest command of Dutch. Despite understanding the words, the item proved complex and too abstract. It was not clear for the participant what the aim of the question was. 3. Items easy to understand (Participants could clearly explain the item and felt confident answering) • Items in English: K16, K1, S4, K4, K5, K24, S21 These items were generally short, concrete, and written in plain language. Note: K1 and K24 appear in multiple categories, suggesting inconsistent comprehension across participants, possibly due to differences in English proficiency or background knowledge. This may indicate the need to refine clarity or provide clearer examples. • Items in Dutch: K5, K7, K17 o K5: In terms of vocabulary, the question was easy to understand by the participant with the highest Dutch proficiency level. However, another participant who tested the same question was unable to understand the question without additional help. Given that most of the words in the item were known to both participants and the question was short, the level of difficulty for the second participant may stem from either his lack of familiarity with AI and/or his lower Dutch proficiency level. o K7: The participant indicated that this question was easy to understand. o K17: The participant indicated that this was an “OK” question, simple and not too long 4. Items with irrelevant or confusing images (The visual did not match or support the meaning of the item) • Items in English: S15 Two participants indicated that the accompanying image did not help them interpret the item and may have caused confusion. The visual needs revision to better reflect the intended meaning or should be removed. • Items in Dutch: Not observed. No comments were made regarding the visuals. 5. Items understandable in general, but with difficult words (Participants understood the sentence overall but flagged specific terms as unclear) • Items in English: S14, S20 34 These items contain technical or academic English vocabulary that challenged participants (e.g., “integrate,” “evaluate,” “fairness,” “prompts”, “tools”, etc.). Simplifying the language or providing examples may improve comprehension. o Items in Dutch: K21, S3K21: In this item, terms difficult to understand were “reageren” [react] and “onaardig” [unpleasant]. The Centre supervisor suggested changing this term by “lief”. Despite clarifying these terms, the item remained difficult to understand, which suggests limited knowledge of AI among the participant. o S3: In this item, the term “prompts” was flagged as unknown. The wording of the question should be simplified (e.g., provide a concrete example, such as ChatGPT or Gemini, instead of “an AI tool”), and the item should be shortened. 6. Items flagged as too long (Length of the item made it harder to follow) • Items in English: S27 This item’s length contributed to difficulty, especially for young people with lower literacy levels. Shortening the item or breaking it into simpler parts may enhance clarity. • Items in Dutch: S8, S19 o S8: This item was identified as loo long and, apparently, it was a bit difficult to grasp. o S19: The question was perceived as too long and difficult to grasp. These terms flagged as difficult: “inhoud” [content], tools and “aanpasssen”[modify] Overall interpretation • Items that were short, concrete, and written in plain language tended to be understood better. • Items containing technical vocabulary, abstract concepts, or lengthy phrasing generated the most difficulty. • Inconsistent classification across participants for certain items (e.g., K1, K24) indicates that language proficiency plays a major role in comprehension. • Visual support needs careful alignment with item concepts to avoid confusion. 35 Table 3: Overview of cognitive interview results Category Definition Items Reported by Participants (English version) Items Reported by Participants (Dutch version) Interpretation/ Implications Difficult to explain Participant understands the general idea but cannot explain it, paraphrase it, or give examples. S17, K24, S23, S15, S28, S27, K2 K3, K29, S17 Items may contain abstract or multi-layered concepts; wording may require simplification or added examples. Not understood at all Participant cannot interpret the meaning of the item. K1, S23, S6 K11, K20, S4, S10, S18, S29 Items likely contain unfamiliar technical concepts; require substantial simplification or reframing. Easy to understand Participant clearly understands the meaning and can explain it confidently. K16, K1, S4, K4, K5, K24, S21 K5, K7, K17 Items use simple, concrete language. Note: K1 and K24 show inconsistent understanding across participants—consider clarifying wording. Image not relevant/ confusing The visual does not match or support the meaning of the item. S15 n/a Replace or revise image to align with item concept; irrelevant visuals create confusion. Understandable but contains difficult words Participant grasps the main idea but flags specific vocabulary as unclear. S14, S20 K21, S3 Replace difficult terminology (e.g., integrate, evaluate, fairness) with simpler alternatives. Too long/ overly complex Participant reports the item is difficult to follow due to length. S27 S8, S19 Shorten item and simplify structure; break into smaller parts if needed. 36 Conclusions These findings will guide refinement of item wording particularly simplifying terminology, increasing contextual examples, and assessing whether certain items require scaffolding or restructuring before formal translation and broader testing. • Most items tested in both languages were difficult for participants to understand. • Because participants were allowed to choose which questions to test, it became clear that many items were perceived as (too) long and were therefore immediately excluded. This should be considered when interpreting the table above, which does not flag many Dutch items as lengthy. • Target-language proficiency appears to play a crucial role in participants’ comprehension of the questions. Higher proficiency is associated with a greater likelihood of understanding the intended meaning. • Items referring to social media require particular attention. Respondents tended to conflate AI with social media, often drawing on their prior social-media knowledge and skills rather than their AI-related skills when providing answers. Pilot survey testing To develop and validate the scale, the research team conducted a series of exploratory factor analyses (EFAs) using survey data from approximately 1,200 young people in Belgium using the initially developed scale. To assess the stability of the results, the dataset was randomly split into two groups of 600 participants each, and the analysis was repeated on both subsamples. A third EFA was then conducted using the full dataset. Before running these analyses, several standard checks confirmed that the data were appropriate for this type of statistical modelling. The Kaiser–Meyer–Olkin (KMO) measure showed that the sample size was adequate, and Bartlett’s test of sphericity indicated that the items were suitably correlated (p < .001). The team used principal axis factoring with oblimin rotation, which allows the identified factors to be related to one another. Factors were retained when their eigenvalues exceeded 1, in accordance with widely used criteria. Because AI literacy is a still developing concept, the team used a fully exploratory approach. At this stage, the priority was to discover which underlying dimensions naturally emerged from the data, rather than to test a pre-defined structure. For this reason, a confirmatory factor analysis (CFA) has not yet been conducted. This means that the stability of the factor structure in completely new samples remains to be formally tested. To partially address this limitation, the team strengthened the internal reliability of the findings by repeating the EFA on two split samples and then on the full sample. This internal replication increases confidence in the robustness of the four-factor solution. Nevertheless, future studies, ideally using new participant groups or longitudinal data, should conduct CFA or similar techniques to confirm and refine the factor structure. This step will be taken in the next phase of the PRODIGI project. EFA on the first half (n=600) produced a four-factor solution (accounting for 59.8% of the variance). Results from split sample 1 The first exploratory factor analysis (EFA) conducted on split sample 1 yielded a four-factor solution, which served as the refined model for the youth developed scale. Items were broadly clustered into 37 four interpretable components. This four-factor solution was selected because it offered the most interpretable and parsimonious structure with acceptable/simple loading patterns and theoretical coherence. However, the structure still showed some cross-loadings typical of an initial EFA, with some expected cross-loadings due to the oblimin rotation, as shown in Table 4. Table 4: Factor loadings (Split sample 1) Factors 1 2 3 4 AITECH1 .159 -.089 -.009 .757 AITECH2 .009 -.017 .013 .874 AITECH3 .098 -.023 .017 .806 AITECH4 .081 .238 .000 .547 AITECH5 -.120 .240 -.402 .249 AITECH6 .195 .010 -.148 .525 AIINF1 .068 .122 -.282 .447 AIINF2 -.029 .039 -.624 .279 AIINF3 .086 .008 -.655 .089 AIINF4 .077 -.122 -.752 .140 AIINF5 .172 -.119 -.623 .195 AIINF6 .116 -.165 -.602 .301 AICREA1 .654 .091 .104 .192 AICREA2 .125 .707 .078 .029 AICREA3 .027 .755 .001 .012 AICREA4 .324 .128 -.297 .127 AICREA5 .539 .229 .055 .129 AICREA6 .188 .621 -.006 .032 AICOM1 .561 -.026 -.119 .231 AICOM2 .449 .010 -.335 .118 AICOM3 .776 .129 -.056 .026 AICOM4 .614 .242 -.007 .076 AICOM5 .553 -.192 -.164 .073 AICOM6 .649 -.014 -.150 .094 AIETH1 .071 .417 -.381 .092 AIETH2 -.014 .522 -.479 .020 AIETH3 .271 .020 -.738 -.163 AIETH4 -.070 .254 -.666 -.031 AIETH5 .193 .142 -.475 -.023 AIETH6 -.148 .557 -.444 .022 Extraction Method: Principal Axis Factoring. Rotation Method: Oblimin with Kaiser Normalization. Total variance explained: 59.78 38 a. Rotation converged in 15 iterations. b. Sample adequacy: KMO = .951 c. Bartlett’s Test of Sphericity: χ²(df = 435) = 7876.972, p .001 Factor 1: Content Creation + Communication Strong positive loadings on both content creation and communication items define factor 1. Several creative-production items load highly on this factor (e.g., AICREA1, AICREA5), and this pattern is reinforced by consistently high loadings from the communication items (e.g., AICOM1, AICOM3, AICOM4, AICOM6). This factor thus represents a combined competence domain in which creative ideation, content generation, and effective articulation or prompting are integrated. Factor 2: Ethics Substantial positive loadings from multiple ethical items anchor factor 2. Several ethics items load strongly on this dimension (e.g., AIETH1, AIETH2, AIETH6), indicating that this factor reflects a coherent ethical judgment domain. A few creative items (e.g., AICREA2, AICREA3, AICREA6) also show moderate loadings, suggesting that creativity and ethical reflection are not fully independent in this sample. However, the core of this factor is clearly ethical competence. Factor 3: Information Evaluation Factor 3 is characterised by strong negative loadings from all information-evaluation items (AIINF2– AIINF6), with AIINF1 also contributing more moderately. The consistent pattern of negative loadings indicates a well-defined latent dimension representing information literacy and evaluation skills, particularly the ability to assess the trustworthiness, detect bias, and evaluate the reliability of AI outputs. Although a few items from other domains exhibit slight cross-loadings, the information items dominate this factor. Factor 4: Technical/Operational Skills Factor 4 is primarily defined by high loadings from the technical/operational items (AITECH1– AITECH6). Most technical items load strongly and cleanly (e.g., AITECH1, AITECH2, AITECH3), indicating that operational skills form a distinct dimension in this subsample. Some items from other domains show small cross-loadings, but the technical items clearly anchor this factor. Results from split sample 2 An independent EFA on the second half (n=600) replicated a similar four-factor structure, indicating internal consistency across subsamples. The exploratory factor analysis on the second subsample also produced a four-factor solution. While minor cross-loadings were present, as expected in oblique rotations, the resulting factors were clearly interpretable and showed coherent internal structure. An overview of these loadings can be found in Table 5 on the next page. 39 Table 5: Factor loadings (Split sample 2) Factors 1 2 3 4 AITECH1 .443 -.022 .491 .246 AITECH2 .395 -.004 .547 .275 AITECH3 .432 .075 .483 .230 AITECH4 .249 .299 .451 .180 AITECH5 -.166 .358 .533 -.082 AITECH6 .434 -.015 .539 .188 AIINF1 .136 .165 .630 .081 AIINF2 -.057 .042 .763 -.165 AIINF3 -.082 .054 .744 -.207 AIINF4 .024 -.120 .807 -.142 AIINF5 .182 -.097 .728 -.082 AIINF6 .164 -.038 .731 -.044 AICREA1 .775 .113 -.030 -.010 AICREA2 .100 .819 .013 .042 AICREA3 -.018 .811 .009 -.080 AICREA4 .341 .194 .196 -.155 AICREA5 .758 .172 -.082 -.024 AICREA6 .356 .495 -.056 -.068 AICOM1 .764 .049 .127 -.020 AICOM2 .503 .054 .271 -.098 AICOM3 .760 .069 .020 -.126 AICOM4 .754 .083 -.007 -.064 AICOM5 .623 -.217 .103 -.065 AICOM6 .655 .084 .103 -.076 AIETH1 .247 .328 .068 -.402 AIETH2 .166 .438 .080 -.433 AIETH3 .268 -.053 .298 -.537 AIETH4 .077 .166 .224 -.505 AIETH5 .329 .016 .128 -.484 AIETH6 .022 .432 .102 -.456 Extraction Method: Principal Axis Factoring. Rotation Method: Oblimin with Kaiser Normalization. Total variance explained: 62.54 a. Rotation converged in 21 iterations. b. Sample adequacy: KMO = .956 c. Bartlett’s Test of Sphericity: χ²(df = 435) = 8650.826, p .001 40 Factor 1: Content Creation + Communication Factor 1 is defined by consistently high loadings from both creative-production and communicationoriented items. Several creative items show strong associations (e.g., AICREA1, AICREA5), alongside communication items that load robustly on this factor (AICOM1, AICOM3, AICOM4, AICOM6). This pattern indicates that, for respondents, generative creativity and practical expression or prompting form a unified competence domain. Factor 2: Content Creation + Ethics blend Factor 2 is primarily anchored by additional creative items, particularly AICREA2 and AICREA3, which load strongly on this dimension. Ethical items also contribute moderately (e.g., AIETH1, AIETH2, AIETH6), suggesting that some aspects of responsible creation are conceptually integrated with creative skills. This factor, therefore, reflects a secondary creative dimension with an ethical component, indicating that respondents tend to associate generative work with considerations of fairness and responsibility. Factor 3: Information Evaluation + Technical Factor 3 shows a clear concentration of information-evaluation items, with all AIINF items loading strongly and consistently on this factor (e.g., AIINF2-AIINF6). Notably, the technical/operational items (AITECH1-AITECH6) also load substantially on this same dimension. This indicates that informationseeking, verification, and assessment skills are closely connected to operational or tool-handling skills in this subsample. Rather than forming separate domains, these abilities appear to constitute a broader “information-technical” competence cluster. Factor 4: Ethics Factor 4 is characterised by strong primary loadings from all ethical items (AIETH1-AIETH6), which consistently load most strongly on this dimension. Although there are small cross-loadings elsewhere, the ethical items are clearly anchored in this factor. This suggests that ethical judgment regarding fairness, bias, and trustworthy use of AI constitutes a distinct underlying dimension in this subsample. Results from the combined sample The combined-sample EFA (N=1200) confirmed the four-factor model. With the larger sample size, item loadings stabilised and revealed four interpretable competence dimensions. An overview of these factor loadings is available in table 6 on the following page. 41 Table 6: Factor loadings (full sample) Factors 1 2 3 4 AITECH1 .352 -.042 -.539 .325 AITECH2 .272 .005 -.592 .358 AITECH3 .344 .040 -.518 .318 AITECH4 .183 .288 -.435 .252 AITECH5 -.142 .328 -.536 -.040 AITECH6 .357 -.008 -.520 .197 AIINF1 .116 .155 -.611 .113 AIINF2 -.037 .066 -.771 -.117 AIINF3 -.038 .065 -.726 -.158 AIINF4 .017 -.094 -.819 -.139 AIINF5 .145 -.083 -.744 -.067 AIINF6 .135 -.079 -.763 -.031 AICREA1 .783 .091 .051 .067 AICREA2 .079 .785 .035 .079 AICREA3 -.035 .829 .002 .006 AICREA4 .359 .184 -.240 -.087 AICREA5 .700 .195 .074 .032 AICREA6 .266 .577 .040 -.008 AICOM1 .734 .003 -.144 .026 AICOM2 .515 .032 -.285 -.075 AICOM3 .822 .088 .020 -.059 AICOM4 .728 .150 .023 -.007 AICOM5 .649 -.218 -.098 -.053 AICOM6 .704 .030 -.096 -.033 AIETH1 .248 .396 -.152 -.290 AIETH2 .150 .513 -.174 -.340 AIETH3 .314 .008 -.361 -.457 AIETH4 .052 .253 -.341 -.408 AIETH5 .346 .101 -.191 -.363 AIETH6 .003 .538 -.181 -.337 Extraction Method: Principal Axis Factoring. Rotation Method: Oblimin with Kaiser Normalization. Total variance explained: 60.82% a. Rotation converged in 36 iterations. b. Sample adequacy: KMO = 963 c. Bartlett’s Test of Sphericity: χ²(df = 435) = 16187.942, p .001 48 Some attitudinal items also needed clearer grounding in students’ experiences. One question suggested that speaking politely to AI tools helps avoid problems, but students did not understand what problems this referred to, since being rude to Siri or ChatGPT has no obvious consequences in their experience. Including an example or a clearer explanation made the intended message easier to grasp. Another item stated that “There is no risk in getting emotional support from an AI,” which students understood literally and could misinterpret. Rewriting it to “There is risk in getting emotional support from an AI” made the message clearer and avoided confusion caused by the original structure. These adjustments showed that even when an item covers a relevant topic, it must be framed in a way that makes sense to students’ daily lives. Overall, aligning the scale’s content with students’ actual experiences (such as technology use, schoolwork, and social interactions) helps ensure that the measure captures what it intends to assess. Students respond more accurately when they recognize the situations described. The cognitive testing phase was essential for revealing when an item did not fit naturally with students’ experiences, allowing the design team to update wording, examples, or context to improve relevance and comprehension. Redundancy and overlap in items The expert review also identified several items in the draft scale that repeated the same ideas or overlapped in what they measured. Redundancy can make a questionnaire feel longer than necessary and reduce the usefulness of the results, so one aim of the co-design process was to streamline the list of questions. In some cases, two items placed in different sections were essentially asking about the same skill. For example, one question about asking clear, specific questions when using an AI chatbot was almost identical to another question about writing clear prompts for AI. Because both targeted the same ability, keeping both would add little value. The plan is to retain the clearer, more general version and remove the duplicate. A similar pattern appeared in questions about verifying information produced by AI tools. Several items asked students to confirm that AI information was accurate or trustworthy, and experts noted that these items did not differ in meaningful ways. One question concerned assessing whether information is “true and trustworthy,” while another concerned evaluating accuracy by comparing it with other sources. In yet another section, a question about double-checking AI answers repeated the same idea. Too many versions of the same concept can confuse respondents or elicit repetitive responses. The recommendation is to combine these into a single, well-phrased question about factchecking AI outputs, unless different aspects require separate measurement, such as accuracy versus bias. There was also a draft item that sought to measure two distinct abilities simultaneously: generating ideas for AI projects and developing those projects. Experts pointed out that these are separate skills, and some students might be comfortable coming up with ideas but not with building something technical, or vice versa. Splitting the item into two clearer questions, or rewording it to focus on just one task, will give more precise information and make it easier for students to respond. Although concerns about redundancy mainly came from experts, student feedback supported this point as well. During testing, a few students commented that some items felt repetitive, or they referred back to an earlier answer when a similar question appeared. This suggests that students do notice overlap, even if they do not always express it directly. Reducing repetition helps keep students engaged and makes each question feel meaningful. 49 Conclusions and recommendations Cognitive testing combined with expert review proved essential in validating and improving the AI literacy scale. The co-design approach ensured that the instrument is both theoretically sound and practically accessible to the target youth audience. Based on the synthesis of findings, the following actions for finalizing the scale were recommended. Revise wording for clarity First, wording should be simplified by replacing technical terms with everyday language or adding brief explanations where needed (e.g. describing AI as computer tools that answer questions or create content, and rephrasing bias as treating some people unfairly). Items should be written in language that reflects how 12–15-year-olds speak in everyday situations. Moreover, abstract or ethical concepts should be supported with short, concrete examples to help respondents easily imagine real-life situations. Ensure age-appropriate content A second important finding concerns the need to remove or adapt materials that assume knowledge beyond students’ age or experience. Items requiring advanced technical or programming knowledge should be omitted or included only when students are expected to have that background. For unfamiliar concepts such as plagiarism or deepfakes, short explanations should be added directly in the item or as a brief note (e.g. explaining plagiarism as “copy-pasting someone else’s text and saying it is your own”). Working with educators can help ensure that all concepts used in the survey are covered in the curriculum. Overall, the goal is to ensure that all students, regardless of academic track, can understand each question well enough to provide a thoughtful answer. Increase relevance and engagement Third, it is important to ground items in situations that feel authentic and engaging to students. In this regard, questions should refer to AI tools young people already know, such as chatbots, image generators, or social media recommendations. For less engaging topics, such as AI’s environmental impact, interest can be increased by presenting a short fact or using a simple true/false format. Ongoing involvement of young people in reviewing items is essential to ensure they feel relevant. If an item cannot be made relatable, it should be revised, introduced with a brief explanation, or reconsidered for inclusion. Eliminate redundancies Fourth, the interviews indicated the need to consolidate survey items by merging similar questions and removing redundancies duplicates. Each item should assess a distinct aspect of AI literacy, and the revised scale should be reviewed again to ensure no unintended overlap remains. This streamlining will shorten the survey, improve the response experience, and reduce the risk of contradictory responses due to similar questions. The resulting item set will be concise, comprehensive, and non-repetitive. Verify again if necessary. Finally, the interview findings suggest conducting a second round of cognitive pre-testing or piloting after making revisions, ideally with a different group of students. This step will confirm whether the changes have effectively addressed the issues (for example, students now understand “tendencioso” instead of “enviesado”, or they correctly handle the revised plagiarism item) and check that no new confusions have arisen. 50 Ongoing expert input is key to keep content aligned with the AI literacy framework’s goals and ensure its reliability and validity after updates. In summary, the co-design process has greatly enhanced the AI literacy scale. By focusing on youthfriendly language, age-appropriate content, real-world relevance, and concise coverage of topics, the improved tool will more accurately assess students’ AI literacy levels. The integration of expert and student insights – including their disagreements – provided valuable guidance: areas of consensus highlighted clear priorities for change, while differences (such as assumptions made by adults versus actual student knowledge) emphasised the importance of incorporating learners’ perspectives in the design. The finalized scale, as shown in this report, embodies these collective insights. It functions as a tool that is both educationally robust and genuinely co-created with its end users, ensuring it will produce valid results during future field testing. This analysis utilises qualitative feedback from expert reviewers of the Portuguese AI literacy scale (Comments from Portugal expert review) and annotations from cognitive testing sessions with Portuguese students (cognitive testing spreadsheet data). Key examples and quotations from these sources are cited to support the findings, such as expert notes on terminology explanations and student feedback on confusing terms or items. These combined insights informed the development of item-level recommendations, ensuring the scale's validity and its responsiveness to youth. The initial version of the yAILS, tested through expert feedback, cognitive interviews, and the school survey, is presented in Appendix C, and the refined version is presented in Appendix D. Polish Context Expert consultation The dynamic development of digital technologies and the rapid spread of artificial intelligence-based tools present both opportunities and new challenges for older people. Contemporary research emphasises that effective digital inclusion of older people requires not only access to technology, but also the reduction of cognitive, social and emotional barriers that prevent them from fully participating in the digital world (Tomczyk & Kielar, 2025a). In the Polish context of technological education for older adults, there is a growing need to redefine teaching models so that they respond to the rapid pace of change and the diverse levels of competence of older adults (Tomczyk & Kielar, 2025b). At the same time, studies conducted in diverse urban and rural environments show that digital exclusion is structural in nature and affects the well-being, sense of agency and opportunities for social participation of seniors (Diana et al., 2025). In this context, diagnostic tools that allow for a reliable assessment of the digital competence of older people are of particular importance. The latest work on educational goals in the area of digital inclusion emphasises the need to simplify language, embed content in real-life situations, and adapt the pace and form to the specificities of late adulthood (Tomczyk & Edisherashvili, 2024). 51 The results of the qualitative part of the study presented in this report confirm this need: experts emphasise the importance of simplicity, clarity and practicality of tools, while pointing out that their excessive length and technicality can lead to cognitive overload in older people. The study was conducted in the second half of November 2025 using in-depth, semi-structured online interviews with prominent Polish experts in the field of digital education for older adults – practitioners with many years of experience in training, consulting and academic work. The material was subjected to thematic analysis, focusing on identifying common problem areas and areas of consensus among specialists. This approach made it possible to capture both the dominant patterns of thinking about the digital competences of older people and practical guidelines for the design of diagnostic tools. By combining an empirical perspective (backed by extensive professional experience) with current knowledge in the field of andragogy and geragogy, this part of the report presents the directions for necessary modifications to the tool and indicates the principles that should shape future solutions supporting the digital inclusion of seniors in the effective use of AI or increasing resistance to manipulation and disinformation mediated by new media. 1. Positive aspects of the tool Basic geragogical assumptions emphasise that older people learn more effectively when the content is close to their everyday experiences and does not require specialised vocabulary. Andragogy, on the other hand, indicates that adults (including seniors) need a sense of competence and comprehensibility of materials in order to enter the learning process without frustration. The respondent notes that some of the questions in the tool meet these minimum conditions: they are general, based on everyday phrases, and do not present the user with unnecessary cognitive barriers. These types of elements are a friendly starting point, consistent with adult learning theory. "I didn't mark which questions could be left out, but I think the most general ones are still suitable. Sometimes there were more general questions, and they are still okay because they use normal, everyday phrases." (Respondent 1, female, extensive experience in training seniors) 52 In the process of adult education, favourable conditions are extremely important: visual comfort, clarity and aesthetics. Barriers such as a too-small font size or a chaotic layout can completely block access to the content. Respondent 2 points out that the tool has graphic qualities that can act as a facilitator, although she emphasises differences in the level of preparation across different groups of seniors, which is consistent with geragogical theory on the heterogeneity of the 60+ population. "On the positive side, the tool is really well designed graphically. The font size is also fine. Overall, I have one thought that sums it all up: this tool is great, but for advanced users – for students. I can see that it could even be introduced for seniors in Krakow, but not for other groups." (Respondent 2, male, extensive experience in training seniors in the use of AI) From a geragogy perspective, it is important to eliminate elements that overload perception, and one of the most significant burdens in working with older people is an excess of material and an overly long list of questions. Respondent 3, as a practitioner of senior education, notes that although the form of the answers may be user-friendly, their potential is negated by the excessive number of tasks. Her comment points to the need to apply the principle of "less is more", which is key in teaching people in late adulthood. "In my experience, such a number of survey questions will discourage them, even though the forms of answering are simple." (Respondent 3, female, extensive experience in senior education, university employee) 2. Negative aspects of the research tool 2a. Too difficult language and specialist vocabulary Geragogical studies have repeatedly confirmed that seniors need the language of reality, not the language of technology, especially when learning something completely new. Exceeding the limits of comprehensibility impedes learning because, unlike younger people, older adults cannot compensate for their lack of digital intuition. Respondent 1 points out a fundamental barrier: the very word "AI" is unclear to many of them, which means that the tool needs to be thoroughly simplified before it can even be measured. "There are names there that are closely related to artificial intelligence and tools. And, gosh, how is a senior citizen supposed to figure that out? What's more, for some seniors, even the abbreviation 'AI' will be a bit of a mystery." (Respondent 1, female, extensive experience in training seniors) In the next excerpt, the same person emphasises that overly difficult questions are not only a problem for seniors; even educated adults may have difficulty interpreting them. This is an important point from an andragogical perspective: a tool cannot place demands on its users that exceed their competencies. "Can we, as adults who consciously use these tools, clearly identify which tool provides incorrect or less clear answers? I think that even for some of us, such a test would be quite interesting." (Respondent 1, female, extensive experience in training seniors) From a geragogy perspective, digital inequalities should be taken into account, especially between urban and rural seniors. Overly difficult language and question structure can completely exclude a significant proportion of respondents. The respondent points out that many questions are formulated 53 in a non-intuitive way, which contradicts the principles of adult education, which assume comprehensibility as a condition for cognitive engagement. "Most rural seniors will not understand them, and that is my biggest hesitation. For them, the content of the questions will be too complicated, even the structure of the questions will be incomprehensible. In my opinion, is the biggest problem with this tool." (Respondent 2, male, extensive experience in training seniors in the use of AI) In senior education, clarity is extremely important. Older people are less likely to guess meanings or contexts. The introduction of unclear or imprecise concepts, especially in English, violates the basic principles of geragogy. The respondent points to specific examples that could undermine seniors' sense of understanding and confidence. "I have doubts — perhaps due to a lack of knowledge — about the content of certain questions and understanding the essence of the question: slide 8 = fair/unfair mistakes?; slide 9 — treats everyone fairly; slide 12 — examples of how... implied positive/negative changes?" (Respondent 3, female, extensive experience in educating seniors, university employee) 2b. Questions that are too detailed and demanding One of the key principles of geragogy is that learning should be functional, i.e. related to everyday life practices. Overly detailed questions requiring technical knowledge not only violate this principle but can also cause competence anxiety, which blocks further cooperation. The respondent notes that some questions require an understanding of topics that most seniors have never needed to learn. "For me, this tool is too detailed. It asks very precise questions that a senior citizen cannot answer without basic knowledge. There was something about creating emails or email accounts, and not every senior citizen is aware of this." (Respondent 1, female, extensive experience in training senior citizens) Respondent 2 also points to a typical pattern of technology use among older people that is highly simplified, indicating that the tool does not correspond to their actual behaviour, a fundamental flaw from the perspective of senior education. “Seniors who use email accounts have them, but do they use them consciously [...] Their use ends with them being able to log into their account, usually on one computer, because they have their password saved there.” (Respondent 1, female, extensive experience in training seniors) Basic andragogical assumptions emphasise that teaching adults must begin with learners' prior experiences. Respondent 2 points out that the questions in the tool are detached from seniors' everyday practices and therefore devoid of cognitive meaning. She points out how important it is to base questions on specific activities that are close to everyday life. "In questions such as 'I know how to use tools to plan a project together', you have to write something specific. [...] They must have a particular task that relates to their lives, for example: 'I know how to create a recipe for chicken soup using an AI tool' or 'I know how to write wishes for my grandson'." (Respondent 2, male, extensive experience in training seniors in the use of AI) 54 According to the basic principles of geragogy, every educational activity must be firmly anchored in the realities of an older person's life. Pointing to school tasks or exercises familiar to young people is grossly inadequate. The respondent clearly emphasises this inconsistency: "The final part of the survey contains questions related to education — writing homework, creating a poster. In my opinion, these can be removed, since the respondents are seniors." (Respondent 3, female, extensive experience in educating seniors, university employee) 2c. Overload, repetitiveness, unclear logic Excessive material is one of the major barriers to learning for older adults. This is due to natural cognitive processes and ageing. The respondent reports her personal experience of reviewing the tool, which points to its tiring and repetitive nature. "For me, there is too much material. It is all too long. By the twelfth page, the twelfth question, when I was reviewing this PDF, I had the impression that it was always the same thing, just in different colours." (Respondent 1, female, extensive experience in training seniors) Based on accumulated experience, the research tool should not only be simple but also methodologically reliable. Respondent 2 points out that many tools cannot fulfil their function because most seniors will not understand their content. This indicates that a complete redesign is required, not merely cosmetic changes. "In my opinion, the entire questionnaire needs to be thoroughly redesigned. [...] In its current state, seniors will not understand the questions. Maybe they will understand one, two, three." (Respondent 2, male, extensive experience in training seniors in the use of AI) From the perspective of basic research methodology, the structure of educational material should be consistent, predictable and logical. The respondent points out that the tool contains duplicates and unclear distinctions, which creates an unnecessary burden for older people. "Such a number of survey questions will discourage them. I suggest that certain questions — and there are several that sound similar — be presented as one." (Respondent 3, female, extensive experience in educating seniors, university employee) "I am wondering about questions I.2, II.1 and III.1 in the skills section – to me, they are identical." (Respondent 3, female, extensive experience in educating seniors, university employee) 3. Opinions on the ideal tool In geragogy, there is a principle of prioritising short, clear materials. The more minimalist the tool, the greater the chances of engagement. The respondent indicates the optimal form: one sheet, clear and organised. "The ideal tool would be short and clear. One sheet, printed on both sides at most." (Respondent 1, female, extensive experience in training seniors) Respondent 2 emphasises that the tool must limit the number of questions to an absolute minimum. According to andragogy, the less content there is, the greater the concentration and motivation of adult learners, and this effect is even more substantial in seniors. 55 "If we want reliable results, the fewer questions, the better. I would recommend 12-14 questions, really basic ones." (Respondent 2, male, extensive experience in training seniors in the use of AI) Respondent 3 again draws attention to the need to reduce duplicates, which is consistent with both the geragogical postulate of transparency and the andragogical principle of cognitive efficiency. "I suggest that certain questions, and there are several that sound similar, be presented as one." (Respondent 3, female, extensive experience in educating seniors, university employee) 4. Adapting the tool to the experiences of seniors According to geragogy, older adults often use technology intuitively and unconsciously, without ascribing to it the status of "learning". The respondent describes the situation, noting that understanding this mechanism is crucial for designing competency tests. "Sometimes they use elements of AI, but they do not feel that it can be classified as such [...]" (Respondent 1, female, extensive experience in training seniors) At the same time, she emphasises that the lack of explanations from loved ones deepens the technological gap, which is consistent with the theory of social learning in older age. "For many seniors, artificial intelligence is a complete abstraction. [...] Grandchildren and children do not explain to them what AI is because it is obvious to young people." (Respondent 1, female, extensive experience in training seniors) From a geragogical perspective, the design of tools must be based on an assessment of the group's level of digitalisation. The respondent points out that a significant proportion of seniors, especially in rural areas, lack even basic devices, which significantly shapes the form of the tool. "From my training in rural areas, I can see that the level is really low [...] Maybe 20% have smartphones." (Respondent 2, male, extensive experience in training seniors in the use of AI) Respondent 3 discusses the problem of excessive diversity in the response categories. Seniors often do not notice subtle semantic differences. The multitude of options makes it difficult to make decisions. "It would be difficult for me to distinguish between helpful, true, useful and reliable information. This can determine the answer." (Respondent 3, female, extensive experience in educating seniors, university employee) 5. Suggestions for test design From the perspective of Respondent 1, questions rooted in the realities of seniors' lives help to break down abstraction. The respondent notes that seniors need examples that are close to their everyday decisions, not models. "If these were simple statements or short situations that seniors could relate to, it would be easier for them to understand." (Respondent 1, female, extensive experience in training seniors) R 1 also indicates that the scale should be as simple as possible, because too many options make it difficult to make a decision, which is confirmed by theories of declining information processing speed in late adulthood. 56 "The scale cannot be long, because many people will start to wonder whether they know a lot or a little less." (Respondent 1, female, extensive experience in training seniors) Respondent 2 points out that seniors have difficulty recognising false content, which is one of the key areas of digital competence for people aged 65+. According to the principles of andragogy, questions must reflect real cognitive threats. "Most rural seniors watch videos [...] the important question is whether they can distinguish between real and fake recordings." (Respondent 2, male, extensive experience in training seniors in the use of AI) The statements also emphasise the need for interpersonal support – seniors often learn most effectively in the company of a supportive person, which fits in perfectly with the principles of geragogy regarding the role of the educator and carer. "I imagine my groups and think that 30% of seniors will need an assistant to help them use the tool." (Respondent 2, male, extensive experience in training seniors in the use of AI) Respondent 3 points out that excessive complexity in the response categories is contrary to both adult learning theory and the practice of working with seniors, for whom too many semantic nuances can lead to confusion. "It would be difficult for me to distinguish between helpful, true, useful and reliable information." (Respondent 3, female, extensive experience in educating seniors, university employee) Conclusions Results from the expert review of the sAILS should be understood as an attempt to compare the practical observations of educators working with seniors with the current, empirically confirmed state of knowledge on learning in late adulthood. Contemporary research clearly indicates that effective education of older people requires consideration of three co-occurring areas: age-related cognitive changes (Zuber et al., 2019), sensory limitations typical of ageing (Germine et al., 2019) and the specific motivations of adult learners, including a strong focus on practicality and immediate usefulness of content (Lee et al., 2019). The table therefore presents a summary of how respondents (based on many years of experience) diagnose the compliance or non-compliance of diagnostic tools with regard to theoretical foundations. Based on the literature on digital competence and technological education for older adults, it is evident that individuals aged 65 and over are particularly susceptible to information overload, difficulties in grasping abstract technological concepts, and challenges in processing complex digital messages (Grotto & Buja, 2025; Nedeljko et al., 2022). Research further emphasizes that effective digital education for this group, and the assessment of its outcomes, should adopt a geragogical approach, characterized by short learning modules, the use of plain and colloquial language, integration into everyday practices, and strong, ongoing instructional support (Gates and WilsonMenzfeld, 2022; Lee, 2023). The respondents’ comments regarding overly technical language, insufficient practical examples, and the excessive length of the tool implicitly corroborate these recommendations. A distinct and increasingly important line of research focuses on the vulnerability of older adults to misinformation and media manipulation. 57 The literature suggests that this vulnerability does not stem from diminished ability, but rather from differences in cognitive processing and media socialisation compared to younger generations, which can make it more challenging for older individuals to evaluate the credibility of visual and digital content (Brashier & Schacter, 2020; Guess et al., 2019). Consequently, questions addressing AI, deepfakes, or the broader impacts of technology must be formulated in an exceptionally simple and transparent manner. Otherwise, the instrument risks capturing confusion or subjective perceptions rather than actual knowledge, which may not accurately reflect real conditions. The table below translates the empirical findings into concrete design recommendations concerning the structure, language, length, and internal logic of the diagnostic tool. Table7: Overview of feedback Area/topic Respondent 1 (F, senior training) Respondent 2 (M, senior training in AI) Respondent 3 (F, senior education, university) Practical implications (geragogy/andragogy) Positive aspects of the tool General questions based on everyday language are understandable for seniors and can remain in the tool. Very good aesthetics, legible font – visually, the tool is user-friendly, but rather for advanced users. The simple form of the answers is an advantage, but is lost when there are too many questions. Keep: good graphic design, large letters, simple language and general questions. Reduce: the number of items so that the positives "come to the fore". Language and terminology Overly specialised terms (AI, tools, programming languages) – seniors do not understand the terms. Even adult users of the tools have difficulty with some of the questions. Rural seniors will not understand most of the questions – difficult wording, complex sentence structures. Unclear English terms ("fair/unfair mistakes", "treats everyone fairly", etc.) raise doubts even among educators. Use colloquial and unambiguous language, avoid AI jargon. In accordance with geragogy – minimise cognitive load, in accordance with andragogy – language must be transparent and understandable to the recipient. Level of difficulty and detail of questions The tool is too detailed and requires technical knowledge (emails, mailboxes) that seniors often do not have. The questions are detached from the real tasks of seniors; concepts such as "project" are meaningless to them, they need practical examples (recipes, wishes). Questions about homework, posters and school activities are inappropriate and should be removed. Design questions as simple life scenarios (wishes, bills, recipes). Avoid school contexts; take into account the actual level of digital competence. Length and repetitiveness of the tool Too much material, tiring document, feeling of "always the same, just in different colours". The questionnaire needs a thorough overhaul – in its current form, seniors will only understand a few questions. Too many questions, duplicates ("identical" items in skills), suggest combining similar questions. Limit the number of questions, remove duplicates, reduce the number of pages. Rule: short, unambiguous, unique items. Proposed target length "One, maximum one double-sided printed sheet". 12–14 basic questions to ensure that the results are reliable and not "sloppy". Combine similar questions into one, reduce excess. Goal: a short tool – approx. 1–2 pages, 12–14 wellthought-out items. Tool structure (logic) First, general questions about information retrieval, then resistance to manipulation, and finally AI and details. First, the basics: awareness of what AI is, whether the senior citizen uses AI at all (e.g. ChatGPT/Gemini). No clear logic – many questions seem practically identical. Arrange the tool from general to specific: (1) awareness of AI, (2) simple uses, (3) critical thinking/manipulation, (4) more advanced aspects (if any). Adapt to seniors' experiences Seniors often use AI (Spotify, YouTube) unconsciously; AI is an abstraction to them because no one explains it to them. Low digitisation in rural areas – many people have keypad phones, and only a small percentage use smartphones. It is difficult for seniors to distinguish between subtle categories such as "helpful", "true" and "reliable". Start from their actual experiences (recommendation streams, search engines, instant messengers), explain AI using examples. Keep technicality to an absolute minimum. Simplify response categories. 64 Helsper, E. J., Schneider, L. S., J.A.M, A., & van Laar, E. (2020). The youth Digital Skills Indicator: Report on the conceptualisation and development of the ySKILLS digital skills measure. https://doi.org/10.5281/zenodo.4608010 Law, N., Woo, D., de la Torre, J., & Wong, G. (2018). A global framework of reference on digital literacy skills for indicator 4.4.2 (UIS Information Paper No. 51). UNESCO Institute for Statistics. Lee, J., Sung, J., & Choi, M. (2019). The factors associated with subjective cognitive decline and cognitive function among older adults. Journal of Advanced Nursing, 76(2), 555–565. Portico. https://doi.org/10.1111/jan.14261 Lee, S. (2023). Internet Use and Well-Being of Older Adults Before and During the COVID-19 Pandemic: Findings from European Social Survey. Journal of Gerontological Social Work, 67(1), 96–113. https://doi.org/10.1080/01634372.2023.2217682 Lintner, T. A systematic review of AI literacy scales. npj Sci. Learn. 9, 50 (2024). https://doi.org/10.1038/s41539024-00264-4 Long, D., & Magerko, B. (2020). What is AI literacy? Competencies and design considerations. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems (CHI ’20) (pp. 1–16). Association for Computing Machinery. doi:10.1145/3313831.3376727 Moll, S., Wyndham-West, M., Mulvale, G., Park, S., Buettgen, A., Phoenix, M., Fleisig, R., & Bruce, E. (2020). Are you really doing “codesign”? Critical reflections when working with vulnerable populations. BMJ Open, 10(11), e038339. doi:10.1136/bmjopen-2020-038339 Mulvale, G., Moll, S., Miatello, A., Robert, G., Larkin, M., Palmer, V. J., Powell, A., Gable, C., & Girling, M. (2019). Codesigning health and other public services with vulnerable and disadvantaged populations: Insights from an international collaboration. Health Expectations, 22(3), 284–297. doi:10.1111/hex.12864 Nedeljko, Assist. M., Bogataj, Prof. DDr. D., Perović, Assist. Prof. Dr. B. T., & Kaučič, B. M. (2022). Digital literacy during the coronavirus pandemic in older adults: Literature Review and Research Agenda. IFACPapersOnLine, 55(39), 153–158. https://doi.org/10.1016/j.ifacol.2022.12.027 OECD. (2024). Explanatory memorandum on the updated OECD definition of an AI system, OECD Artificial Intelligence Organisation for Economic Co-operation and Development, & European Commission. (2025, May). Empowering learners for the age of AI: An AI literacy framework for primary and secondary education (Review draft). OECD & European Commission. Papers, No. 8. OECD Publishing, Paris. https://doi.org/10.1787/623da898-en. Peters, S., Guccione, L., Francis, J., Best, S., Tavender, E., Curran, J., Davies, K., Rowe, S., Palmer, V. J., & Klaic, M. (2024). Evaluation of research co-design in health: A systematic overview of reviews and development of a framework. Implementation Science, 19, 63. doi:10.1186/s13012-024-01394-4 Potter, W. J. (2019). Media literacy (9th ed.). SAGE Publications. Roger, A., Aïmeur, E., & Rish, I. (2023). Towards ethical multimodal systems. ArXiv (Cornell University). https://doi.org/10.48550/arxiv.2304.13765 Shippee, M. (2024). Beyond the 4c’S: Framing Our Understanding of Future Skills in the Era of the Fourth Industrial Revolution. SSRN Electronic Journal. https://doi.org/10.2139/ssrn.4692130 Spencer, R., McCormack, M. J., Drew, A. L., Gowdy, G., & Keller, T. E. (2021). (Not) minding the gap: A qualitative interview study of how social class bias can influence youth mentoring relationships. Journal of Community Psychology, 50(3). https://doi.org/10.1002/jcop.22737 Spurava, G., & Kotilainen, S. (2022). Young People as Empirical Experts of Participatory Research in the Age of Information Disorders. Seminar.net, 18(1). https://doi.org/10.7577/seminar.4740 65 Tapia-Benavente, L., Vergara-Merino, L., Garegnani, L. I., Ortiz-Muñoz, L., Loézar Hernández, C., & VargasPeirano, M. (2021). Rapid reviews: definitions and uses. Medwave, 21(1), e8090. https://doi.org/10.5867/medwave.2021.01.8090 Thaker, V., Rawal, S., Shukla, H., Dave, I., Jani, A., & Thaker, M. (2025). The dual impact of artificial intelligence on human cognition: A comprehensive review. Journal of International Commercial Law and Technology, 6(1), 34–45. Tomczyk, Ł., & Edisherashvili, N. (2024). Learning Objectives in Older Adult Digital Education - Redefining Digital Inclusion. International Journal of Cognitive Research in Science, Engineering and Education (IJCRSEE), 12(3), 507–520. https://doi.org/10.23947/2334-8496-2024-12-3-507-52 Tomczyk, Ł., & Kielar, I. (2025a). Empowering the elderly in the information society: Redefining digital education for Polish seniors in the age of rapid technological change. Educational Gerontology, 51(10), 1122– 1137. https://doi.org/10.1080/03601277.2024.2439908 11. Tomczyk, Ł., & Kielar, I. (2025b). Neutralising External and Internal Barriers in the Digital Inclusion Process for Seniors - Finding Ways to Effectively Shape Digital and Media Competences among Older People. Technology, Knowledge and Learning, 30(4), 2231–2248. https://doi.org/10.1007/s10758-024-098137 3. UNESCO. (2011). Media and information literacy curriculum for teachers. UNESCO. https://unesdoc.unesco.org/ark:/48223/pf0000192971 UNESCO. (2013). Global media and information literacy assessment framework: Country readiness and competencies. UNESCO. https://unesdoc.unesco.org/ark:/48223/pf0000224655 Walker, K. C., & Larson, R. W. (2012). Youth Worker Reasoning about Dilemmas Encountered in Practice: ExpertNovice Differences. Journal of Youth Development, 7(1), 23-May. https://doi.org/10.5195/jyd.2012.149 Willis, G. B. (2005). Cognitive interviewing: A tool for improving questionnaire design. Thousand Oaks, CA: Sage. Yang, W. Artificial intelligence education for young children: why, what, and how in curriculum design and implementation. Comput. Educ. 3, 100061 (2022). Zuber, S., Ihle, A., Loaiza, V. M., Schnitzspahn, K. M., Stahl, C., Phillips, L. H., Kaller, C. P., & Kliegel, M. (2019). Explaining age differences in working memory: The role of updating, inhibition, and shifting. Psychology & Neuroscience, 12(2), 191–208. https://doi.org/10.1037/pne0000151 66 9. Appendices A. Rapid Review study overview Scale Doi/URL 1 AI Literacy Test; 31 items https://doi.org/10.1016/j.caeai.2023.100165 2 AI-CI—AI literacy concept inventory assessment ; 20 items https://doi.org/10.1007/s40593-024-00398-x 3 AILQ—AI literacy questionnaire; 32 items https://doi.org/10.1111/bjet.13411 4 AILS—AI literacy scale; 12 items doi.org/10.1111/bjet.13411 5 AI self-efficacy scale (AISES) doi.org/10.1007/s10639-023-12015-w 6 EVT Instrument doi.org/10.1186/s40561-023-00284-4 7 ChatGPT Literacy Scale https://doi.org/10.1007/s12144-024-05723-0 8 GSE-6AI https://doi.org/10.3389/feduc.2024.1293437 9 Digital literacy scale in the artificial intelligence era https://doi.org/10.3837/tiis.2023.08.016 10 Intelligent TPACK https://doi.org/10.1016/j.chb.2022.107468 11 The Artificial Intelligence Literacy Scale for Middle School Students https://10.9708/jksci.2022.27.03.225 12 MAILS https://doi.org/10.1016/j.chbah.2023.100014 13 MAIRS-MS https://doi.org/10.1186/s12909-021-02546-6 14 Pinski & Belian’s Instrument https://hdl.handle.net/10125/102649 15 SAIL4ALL https://doi.org/10.31235/osf.io/bvyku 16 SNAIL https://doi.org/10.1016/j.chbr.2023.100338 17 UNESCO AI competency framework for students https://doi.org/10.54675/JKJB9835 18 Empowering Learners for the Age of AI: An AI Literacy Framework for Primary and Secondary Education https://ailiteracyframework.org 67 B. Distribution of Correct Responses Across the 30 AI Knowledge Items: Results of the Survey in Belgium Knowledge Frequency Percent Valid Percent Cumulative Percent Valid .00 6 .5 .5 .5 1.00 16 1.3 1.4 2.0 2.00 14 1.1 1.3 3.2 3.00 11 .9 1.0 4.2 4.00 22 1.8 2.0 6.2 5.00 18 1.5 1.6 7.8 6.00 25 2.0 2.3 10.1 7.00 27 2.2 2.4 12.5 8.00 36 2.9 3.2 15.8 9.00 26 2.1 2.3 18.1 10.00 23 1.9 2.1 20.2 11.00 36 2.9 3.2 23.4 12.00 62 5.0 5.6 29.0 13.00 44 3.6 4.0 33.0 14.00 49 4.0 4.4 37.4 15.00 52 4.2 4.7 42.1 16.00 71 5.8 6.4 48.5 17.00 53 4.3 4.8 53.2 18.00 87 7.1 7.8 61.1 19.00 60 4.9 5.4 66.5 20.00 70 5.7 6.3 72.8 21.00 90 7.3 8.1 80.9 22.00 70 5.7 6.3 87.2 23.00 55 4.5 5.0 92.2 24.00 44 3.6 4.0 96.1 25.00 25 2.0 2.3 98.4 26.00 12 1.0 1.1 99.5 27.00 6 .5 .5 100.0 Total 1110 90.4 100.0 Missing System 118 9.6 Total 1228 100.0 68 C. Full overview of Initial AI Literacy Scale AI Skill Items Dimension AI Skill Item Technical and operational 1. I know how to use an AI-based tool (e.g. a voice assistant or ChatGPT) to complete a specific task or activity. 2. I know how to write clear and effective prompts or questions to get useful results from a generative AI tool (e.g. ChatGPT, DALL,E). 3. I know how to improve the answers I get from an AI tool by changing or refining my prompts or questions. 4. I know how to choose the most suitable AI tool or app for what I want to do. 5. I know how to check and adjust the privacy or security settings when using an AI tool (e.g. disabling data sharing or deleting chat histories) 6. I know how to use AI tools to make my work or study tasks more efficient. Information navigation and processing 7. I know how to find and select reliable information or content with the help of AI tools. 8. I know how to check whether information created or suggested by AI is true and trustworthy. 9. I know how to recognize when AI-generated content is incomplete, misleading, or biased. 10. I know how to verify AI-generated information by evaluating its accuracy and checking it against credible sources. 11. I know how to integrate information from AI tools with insights from other sources to form a comprehensive understanding. 12. I know how to distinguish when to accept, revise, or reject AI-generated content based on its quality Content creation and production 13. I know how to do creative writing or storytelling using an AI tool (e.g., ChatGPT or Gemini). 14. I know how to design and develop projects or applications that use AI tools. 15. I know how to select useful tools (e.g., frameworks, programming languages) to program an AI. 16. I know how to use AI in assignments while respecting rules on originality and plagiarism. 17. I know how to use AI to create a poster for an event I am organizing. 18. I know how to use AI (e.g., Gemini or ChatGPT) to make a social media post gain more attention Communication and interaction 19. I know how to ask clear and specific questions when I use an AI chatbot (e.g. ChatGPT, DALL·E) so I get helpful answers. 20. I know how to recognize when an AI tool gives a wrong or confusing answer and ask it again in a clearer way. 21. I know how to work with an AI tool to plan or create something together, such as a story, design, or project. 22. I know how to use AI tools to help me write or reply to messages, posts, or emails when communicating with other people. 23. I know how to recognize when AI tools don’t think or feel like people when I talk or write to them. 24. I know how to use and talk with different kinds of AI systems (e.g. ChatGPT or Google Assistant) Ethical and responsible use 25. I know how to test an AI app (like an image-recognition tool) by giving it different examples to see if it makes fair or unfair response. 26. I know how to compare how an AI tool works for different kinds of people to check if it treats everyone fairly. 27. I know how to double-check answers from an AI tool (e.g. ChatGPT) using trusted sources to see if they are correct and safe to use. 28. I know how avoid sharing personal details when using an AI tool to keep my information safe. 29. I know how to find real examples of how AI has changed people’s jobs, environment, education, or daily life. 30. I know how to distinguish who should take responsibility when an AI system causes a problem and explain my reasons. 69 AI Knowledge Items Dimension AI Knowledge Item Technical and operational 1. AI systems follow rules made by people and a combination of math and logic to make choices. 2. AI systems get better by practicing with lots of data and learning from their mistakes. 3. Tools like ChatGPT make new things by finding patterns in data, but they don’t really understand what they say. 4. AI can think and make its own choices without help from people. 5. AI always works perfectly, and it cannot make mistakes. 6. Once an AI is made, people can’t change or fix it anymore. Information navigation and processing 7. The answers given by AI tools are always correct, so there is no need to double-check them. 8. Different people can get different results from AI tools even when they ask the same question. 9. AI-generated content can sometimes mix true and false information. 10. It does not matter how you phrase your question, as AI tools will always give the same quality of answer. 11. AI tools are completely free from human bias. 12. Everything shown or suggested by AI tools online can be trusted as neutral and reliable. Content creation and production 13. Companies can use AI to create ads based on the data I share online. 14. When using AI, the tools and content should be adapted to the specific context and audience (e.g., creating something for children versus for adults). 15. Multimodal prompting means giving an AI both words and images to help understand a question. 16. AI can create completely original content, without using existing information. 17. Using AI tools to do homework or assignments is the same as doing it myself 18. Using AI to make content has nothing to do with fairness, credit, or using other people’s work responsibly. Communication and interaction 19. An AI system can be asked to give answers as if it were talking to a different kind of person (for example, someone younger or with different opinions). 20. AI tools can influence the opinions I see online. 21. An AI app may respond unpleasantly to my prompts. 22. I should talk to AI tools as if it were a person, being polite, so I don’t have problems. 23. If I create a “Deepfake” of someone just for fun, it cannot hurt them. 24. There is no risk in getting emotional support from an AI. Ethical and responsible use 25. AI systems can sometimes create harmful or inappropriate content. 26. Human workers help build, train, and look after AI tools. 27. AI should be made and used in ways that are fair, clear, and safe for everyone. 28. AI tools always make fair and correct decisions without mistakes. 29. AI tools never cause any social or environmental problems. 30. Training and running AI systems does not use much energy or harm the environment. 70 D. yAILS Refined Scale Below, you can find the reformulated AI Skills and Knowledge items, for measuring AI literacy among vulnerable youth (yAILS) and seniors (sAILS). The items in this scales are updated based on results from the pilot survey, feedback from experts, and interviews with participants in all countries. For each question, we looked at how well it worked in the survey across Belgian Schools, how people understood it during the cognitive interviews, how experts rated the questions and how strongly it related to the skill or knowledge it was meant to measure. Most changes were made to make the questions more straightforward to understand and more relevant to everyday use of AI. Technical or unclear terms were replaced with simple language and familiar examples. Examples of these changes can be found in table 5. Table 2: Rewording in Final Scale Original wording Revised wording Privacy or security settings Turn off data sharing Deleting chat histories Delete my chat Fair or unfair responses Does not treat anyone badly Responsibility and accountability Who is to blame, and why Societal impact Real examples from daily life Environmental impact Using lots of water or electricity Some questions that covered very similar ideas were combined, while others that required advanced technical knowledge were removed. We also adjusted how the questions were grouped to better reflect how people actually use AI, for example, when creating content or communicating with others. Table 3: Unclear Distinctions Removed distinction How it was handled Task vs. activity Combined into one idea Tool vs. app Used “AI” or “AI app” consistently Prompt vs. question Used “question” only In addition, examples were added to clarify what is meant by AI tools or apps such as Siri or ChatGPT, and questions about responsible use were rewritten to focus on real-world impacts, such as privacy, fairness, and environmental effects. Overall, these changes were made to ensure that the questions are clear, meaningful, and suitable for our three participant groups. 71 Refinement of Youth AI Literacy Scale (Skill Items) Dimension Original item Final Scale Item Changes made Technical and operational I know how to use an AIbased tool (e.g., a voice assistant or ChatGPT) to complete a specific task or activity. 1. I know how to use AI (for example, Siri or Chat GPT) to do my homework From general “AI-based tool for any task” to concrete school-related use (“do my homework”); simpler wording I know how to write clear and effective prompts or questions to get useful results from a generative AI tool (e.g. ChatGPT, DALL,E). 2. I know how to ask clear questions to AI (e.g. Chat GPT) to get helpful answers Write clear and effective prompts” → “ask clear questions”; removes technical term prompt. I know how to improve the answers I get from an AI tool by changing or refining my prompts or questions. 3. I know how to change my questions to make AI give me a better answer Keeps same skill but expressed in everyday language (“change my questions”). I know how to choose the most suitable AI tool or app for what I want to do. 4. I know how to pick the best AI app (e.g. DALL.E or Snapchat AI) for what I want to do. “Choose the most suitable AI tool” → “pick the best AI app”; adds familiar apps (Snapchat AI). I know how to check and adjust the privacy or security settings when using an AI tool (e.g. disabling data sharing or deleting chat histories). 5. I know how to turn off data sharing when using AI From multiple privacy actions (check/adjust settings) to one concrete action (turn off data sharing). 6. I know how to delete my chat with AI. Original focus on efficiency → revised focus on a specific operational action (delete chat). Information navigation and processing I know how to find and select reliable information or content with the help of AI tools. 7. I know how to find and select information I can trust with the help of AI Reliable information” → “information I can trust”; maintains meaning with simpler phrasing. I know how to check whether information created or suggested by AI is true and trustworthy. 8. I know how to check if the information from AI is correct Removes abstract phrasing (“true and trustworthy”), focuses on correctness. 72 I know how to recognize when AIgenerated content is incomplete, misleading, or biased. 9. I know how to explain the difference between searching with AI and searching with a search engine like Google “Incomplete, misleading, or biased” → “incomplete, confusing or one-sided.” I know how to verify AIgenerated information by evaluating its accuracy and checking it against credible sources. 10. I know how to check if AI’s answers are correct by comparing them with trusted sources. Keeps same verification skill; simpler, more direct phrasing. I know how to integrate information from AI tools with insights from other sources to form a comprehensive understanding. 11. I can put AI answers and other sources together to help me understand something completely. “Integrate information” → “put together”; meaning unchanged. I know how to distinguish when to accept, revise, or reject AIgenerated content based on its quality. 12. I know how to check when an AI answer is good, needs fixing, or should not be used. “Accept, revise, or reject” → “good, needs fixing, or should not be used.” Content creation and production + Communication & Interaction I know how to do creative writing or storytelling using an AI tool (e.g., ChatGPT or Gemini). 13. I know how to make different kinds of creations (like stories, music, drawings or ideas) using AI From writing/storytelling to multiple creative formats (music, drawings, ideas). I know how to use AI in assignments while respecting rules on originality and plagiarism. 14. I know how to use AI for schoolwork without copying other people’s work Original technical project/app development removed; replaced by plagiarism-aware school use. I know how to use AI (e.g., Gemini or ChatGPT) to make a social media post gain more attention. 15. I know how to use AI to make a social media post get more likes or views. New emphasis on social media performance; not present in original programming-focused items. I know how to work with an AI tool to plan 16. I know how to communicate with AI and work together with it to Combines collaboration, communication, and co-creation into one simplified statement. 73 or create something together, such as a story, design, or project. create something (like a story, design or a project) I know how to use AI tools to help me write or reply to messages, posts, or emails when communicating with other people. 17. I know how to use AI to help me write or answer messages, posts or emails when communicating with others. Minor wording changes; maintains same communication support skill I know how to recognize when AI tools don’t think or feel like people when I talk or write to them. 18. I know how to notice that AI doesn’t think or feel like a real person when I talk or write to it. Simplifies recognition that AI does not think or feel like humans. Ethical and responsible use I know how to test an AI app (like an imagerecognition tool) by giving it different examples to see if it makes fair or unfair response. 19. I know how to try out an AI app with different examples to check if it does not treat anyone badly. “Fair/unfair responses” → “does not treat anyone badly.” I know how to find real examples of how AI has changed people’s jobs, environment, education, or daily life. 20. I know how to find real examples of how AI can affect the planet, like using lots of water or electricity. Replaces fairness comparison with environmental impact examples (water, electricity). I know how to double-check answers from an AI tool (e.g. ChatGPT) using trusted sources to see if they are correct and safe to use. 21. I know how to explain how AI changes daily easier but also how it can cause harm if used the wrong way Combines benefits and harms into one accessible explanatory skill. I know how avoid sharing personal details when using an AI tool to keep my information safe. 22. When using AI, I know how to avoid sharing personal information like my name or address to keep me safe. “Avoid sharing personal details” → concrete examples (name, address) and personal safety focus. I know how to find real examples of 23. I know how to find real examples of how AI has From broad societal impacts to specific job loss examples.