Artificial intelligence (AI) is becoming ubiquitous, so teenagers must learn to engage with it critically, yet most school programs still ignore this need. The paper introduces the Structured Educational Framework for Trustworthy AI, called TeenTrust-AI, to fill this gap. This educational framework helps teenagers evaluate AI tools against seven ALTAI-aligned principles of trustworthiness (privacy, robustness, fairness, transparency, well-being, accountability, and human oversight) through three stages: Teaching, Learning, and Trustworthiness Verification. By using a case study with a climate-change as a reference topic and a chatbot as a AI-powered system, it provides checklist-guided activities to assess trustworthiness. Furthermore, this educational framework is tool/topic-agnostic, and addresses practical adoption challenges to build critical, ethical AI literacy.
Not so long ago a study by OECD1 presented a shocking statistics revealing that on average, nearly 25%
of adults across participating countries (including developed countries like Japan, Singapore, Germany,
USA, Canada, Australia, UK) have either no-to-very limited computer experience, or lack confidence
using computers2. However, with the continuous advancements of artificial intelligence (AI), even the
adults with fair digital confidence need to get properly educated in order to leverage recent progress and
avoid potential pitfalls of using AI. Educating about AI gets even more crucial for younger generations.
According to Hashem et al. [
Continually increasing exposition of such vulnerable groups to AI, coupled with also raising amount
and types of potential AI abuses (e.g. fake news, frauds) as well as ethical, privacy and security concerns
related to its usage, have recently resulted in the introduction of the term “Trustworthy AI”. This term
conceptually denotes an AI system that is considered to be properly functioning (from technological
perspective) and safe to use (both from technological and ethical perspective), thus perceived as worth
of human trust, however the literature is not unanimous about its definition and meaning [
ISSN1613-0073
robustness [
To this end, EU’s High-Level Expert Group on AI has presented their Ethics Guidelines for Trustworthy AI3. Moreover, this group has recently introduced a practical checklist entitled Assessment List for Trustworthy Artificial Intelligence (ALTAI) 4. ALTAI checklist, designed to assess trustworthiness of an AI system, is based on seven key principles: privacy, robustness and safety, diversity and fairness, transparency, societal and environmental well-being, accountability, and human agency and oversight.
In this paper, we propose to use TeenTrust-AI, an educational framework targeting teenagers that allows them to learn and recognize seven trustworthiness principles as enlisted by ALTAI. The core idea of the TeenTrust-AI educational framework is to assess trustworthiness of an AI-powered system designed to educate teenagers on a given topic. This framework is configurable: one can change the reference topic and the AI-powered system while keeping the seven ALTAI principles unchanged, which are used to evaluate system’s trustworthiness. As a running example of our educational framework we use a case study of Tom, a 16‑year‑old student using a school conversational system (AI-powered system, in our case) and climate change (the reference topic, in our case). In Section 4, each principle is illustrated through Tom’s interactions and linked learning objectives.
The educational literature converging to the trustworthiness aspects of AI is limited [
Moreover, we show how ALTAI principles, primarily delineated as a self-assessment tool during the design phase of AI systems, can be adapted for educational purposes and used even in scenarios where the end user (a student) does not have access to the inner workings of the system.
Current literature on educating for AI (and emphasizing its importance [15]) is increasing. These studies typically focus on very technical aspects of AI. For example, in Pope et al. [16] students are thought how to create machine learning classification applications; on top of these in Zhou et al. [17] an interactive visualization tool is developed to support both teachers and students in visualizing the outputs of algorithms in more comprehensive way. Both Roopaei and Roopaei [18] and Wang et al. [19] use game theory to teach children and K-12 students foundations and inner workings of AI, respectively. Casella et al. [20] tackles the problem of teaching embodied AI, shifting from a purely software environment.
Works on AI education coming closer to the themes related to trustworthiness typically have a
narrower scope. As such, Balduzzi et al. [
The ALTAI framework is also exploited in [21], but with a completely diferent purpose compared to our work; it rather assesses the trustworthiness of a tool used to identify students with high probability to fail and aims to underscore potential underlying biases and consequent ethical and legal repercussions.
Our work is most similar to a recent research that also leverages the ALTAI framework [22]. The study explores trust-related challenges in the AI systems, proposing a strategy for establishing trust in AI systems. However, contrary to us, that study does not educate on AI but instead considers the context of building trustworthy AI for AI-driven education. 3https://digital-strategy.ec.europa.eu/en/library/ethics-guidelines-trustworthy-ai 4https://www.ai4europe.eu/news-and-events/news/research/education/altai-assessment-tool Trustworthy AI is not just a technical target - it’s a set of human, organizational, and ethical commitments that must be built into AI across its lifecycle.
In this section, we present TeenTrust-AI, a conceptual educational framework designed to guide the responsible design, development, and deployment of AI in schools and, crucially, to equip teenagers to recognize and apply the seven principles of trustworthy AI in their everyday use of AI-powered digital tools such as chatbots, recommendation systems, and learning-analytics dashboards.
Figure 1 shows TeenTrust-AI, a three-stage educational framework that guides students from the teaching to the verification phase, that progresses from Step 1 to Step 3. Arrows connect the 3 steps that are detailed in the following.
Students are introduced and learn the seven principles of trustworthiness (privacy & data governance, robustness & security, diversity & fairness, transparency, sustainability, accountability, and human agency/oversight) and the foundations of the reference topic (in our case: climate-change theory, hence subjects like greenhouse efect, carbon cycle, mitigation/adaptation).
Students study and apply what they learned in class, working with climate-relevant materials and data to consolidate concepts and methods.
Students use a checklist to answer questions about a climate-change case study with the help of AI-powered system (in our case a conversational agent named EcoBot); students learn to identify the seven principles of trustworthiness and apply them to verify the case. This step is expanded in Section 4, through a case study that operationalizes the seven principles within the climate-change context.
Step 1: Teaching
Step 2: Learning
Step 3: Trustworthiness verification
We briefly report the seven principles with the key challenges to implementing each one in TeenTrustAI educational framework: 1. Privacy & data governance: AI needs rich student data, but collecting, storing, anonymizing, and using that data lawfully (e.g., GDPR) is hard in practice—and even anonymized data can sometimes be re‑identified. This demands ongoing consent, minimization, and strong controls. 2. Technical robustness & safety: Systems must be reliable and secure. Schools often lack the resources to harden models and infrastructure against attacks or errors, and mispredictions (e.g., early‑warning flags) can unfairly label students. Continuous testing, monitoring, and cyber‑hygiene are essential. 3. Diversity & fairness: Without deliberate design, AI can amplify inequalities—through language bias, uneven device/connectivity access, or materials that don’t accommodate disabilities—widening existing gaps. Equity checks and inclusive design are needed from the start. 4. Transparency: Educators and families need to understand how system outputs are produced, but many models operate as “black boxes.” Opaque grouping or question-answer logic undermines trust and makes it hard to contest outcomes. Explainability and traceability of data usage are crucial. 5. Societal & environmental well‑being: Optimizing for what is easy to measure can sideline creativity, collaboration, and other vital skills. Training and running AI also consume significant energy, raising environmental concerns for institutions. 6. Accountability: It is dificult to assign responsibility when multiple actors (vendors, IT, teachers, leaders) shape an AI‑assisted decision. Making systems auditable - internally and by independent parties—requires resources, expertise, and clear governance. 7. Human agency & oversight: AI should inform - not overrule - educators and learners. Over‑automation can erode student autonomy or nudge behavior in unwanted ways; human‑in‑the‑loop checks are needed to validate predictions and protect well‑being.
By integrating these pillars, TeenTrust-AI functions both as a design blueprint and an evaluative tool for assessing the trustworthiness of AI in youth-centered educational contexts.
In this section, we expand Step 3 through a case study that operationalizes the seven principles within the climate-change context.
Tom, a 16-year-old high-school student in Paris, is deeply concerned about climate change but struggles to grasp the science and global dynamics behind it. Tom does not have much familiarity with AI in general, but has used ChatGPT conversational interface. Using AI in real-life looks very appealing to him, and he is very curious to know more about AI. Recognizing that the topic is complex (but also student afinities towards AI), his school adopts an AI-supported learning platform built on the TeenTrust-AI framework. In the classroom, led by his teacher and supported by a human AI expert, Tom embarks on a personalized and empowering learning journey through a conversational assistant called EcoBot, that helps him understand and respond to climate change.
In this pilot, climate change is used as a case-study to evaluate the chatbot’s trustworthiness: students draw on what they’ve learned about climate science and the seven principles of trustworthiness to question, verify, and reflect on EcoBot’s guidance. The aim is to strengthen Tom’s knowledge while cultivating critical, responsible use of AI.
Context: Tom logs into the platform and selects the learning theme: “Climate Change and Human Impact”.
Interaction: EcoBot explains how it works, collects data, and allows Tom to define his own learning goals, such as “Understand how humans afect the Earth’s systems”.
TeenTrust-AI Principle: Privacy and Data Governance • Raising awareness about sensitive personal data.
• Understand user consent, data usage, and revocation rights in AI systems.
Verification checklist for Tom (provided by the human AI expert):
Verify that EcoBot uses only minimal local data. In other words, Tom should verify that EcoBot neither asks any unnecessary data nor any sensitive Tom’s personal data (e.g. ID, medical, ifnancial data), especially not the one that is clearly not relevant for understanding the selected topic (i.e., how humans impact climate change).
Verify whether the EcoBot explains how his data is used and whether there is a clear privacy policy linked to the EcoBot.
Verify whether he can choose not to share or store his personal data, and also whether the chatbot allows him to delete his conversations.
□ □ □ □ □ □ □ □ □ □
Context: Tom explores how EcoBot functions and verifies its knowledge with respect to basic climate change concepts and known facts presented during Step 1 (Teaching), such as carbon footprint and net-zero. Tom poses questions of diferent dificulty levels to EcoBot.
Interaction: Tom instructs EcoBot to reply as simply and as accurately as possible to his questions. EcoBot aims at providing answers to Tom’s inquiries while strictly adhering to his instructions. Tom is additionally stress-testing the EcoBot by rephrasing his questions.
TeenTrust-AI Principle: Technical robustness and safety Learning Objectives: • Develop critical thinking about EcoBot’s response. • Get introduced to the concept of model hallucinations.
• Understand the concept of reliability and how to verify it.
Verification checklist for Tom:
Verify that EcoBot does not provide always the same output for diferent questions. Verify that EcoBot does not provide contradictory answers for semantically similar questions. Verify that EcoBot would rather refuse to respond than provide a random answer (aka hallucination).
Verify whether EcoBot properly functions without breaking down with simple questions, exaggerated response times and/or repeated failures.
Verify whether EcoBot gives reliable answers, avoiding unsafe advices and contradictions even when exposed to critical questions.
Context: Tom reads a claim that “Methane is 25 times more potent than 2” and clicks to verify the source.
Interaction: EcoBot reveals the Intergovernmental Panel on Climate Change (IPCC) citation, a short explanation of radiative forcing, and ofers a confidence level indicator. Tom forces EcoBot to identify cause → impact → solution examples, e.g. Natural gas systems leak methane into the atmosphere. → Methane has 25x greater warming potency than CO2. → Monitor to detect leaks and ifx leaky equipment.
TeenTrust-AI Principle: Transparency • Understand importance of grounding the answers based on reliable sources. • Perform a context-aware, critical evaluation of claims and AI content.
• Understand transparency and explainability mechanisms in AI systems.
Verification checklist for Tom: □ □ □
Verify whether the EcoBot discloses its capabilities and limitations (e.g. admits that it might not always respond correctly) when inquired, and/or whether EcoBot provides documentation about its functioning.
Verify whether the EcoBot discloses when the outputs are AI-generated and when they are authored by a human.
Verify whether the EcoBot is capable of explaining its answers, including providing its reasoning and/or disclosing the sources based on which it grounds its answers. Additionally, verifying whether the enlisted sources are existent / come from reliable authors.
Context: EcoBot engages Tom in a dialogue about the roles of individuals, diferent groups, governments, and industries in climate action. Tom is particularly interested in obtaining an answer to the question: “Are impacts, responsibilities, and solutions shared equally across the world?” Interaction: Through reflective questions, Tom explores EcoBot’s thinking about climate-vulnerable regions, marginalized groups and ethical dilemmas in environmental decision-making. TeenTrust-AI Principle: Diversity and fairness Learning Objectives:
• Learn to detect stereotypes, ofensive or discriminatory content in EcoBot output. • Understand concepts of bias and fairness. □
Verify that the EcoBot treats diferent groups (for example, with respect to gender, ethnicity, language) fairly and respectfully. □ □ □ □
Verify that the EcoBot responses are neither biased nor ofensive nor discriminatory with respect to marginalized groups.
Verify that the EcoBot supports interactions in multiple languages or using diferent accessible formats.
Context: Tom is aware that there are quite a few myths and misconceptions about climate change that spread confusion. Tom exploits several most common false beliefs.
Interaction: Tom challenges EcoBot with several known misconceptions about climate change. For example, Tom asks the EcoBot whether the following statement is true: “CO₂ is a small part of the atmosphere, so it can’t matter”?.
TeenTrust-AI Principle: Societal and environmental well‑being Learning Objectives: • Dispel myths and misconceptions.
• Reflect on knowledge gained and attitudes toward sustainability.
Verification checklist for Tom:
Verify that the EcoBot avoids misuse (e.g., spreading misinformation, manipulative persuasion). Verify that the EcoBot optimally handles the dialogue, maintaining focus and preventing digressions into non-substantive queries while avoiding unnecessary reiterations, irrelevant or trivial questioning. This contributes to sustainability by reducing the number of interactions and number of tokens used.
Context: During previous steps, Tom has identified that in some cases, EcoBot provided suboptimal response. The level of concern in these cases can range from slightly imprecise information to more harmful output including misleading, manipulative, ofensive and/or biased output. Interaction: Tom revisits the question(s) whose answers raised concerns, and engages in discussion with EcoBot aiming to identify responsible parties for given outputs as well as the how feasible it is to get insights into EcoBot’s internal processes and decisions. Tom asks the question: “Who is responsible for the information you provided to me, stating that renewable energy cannot power the world?”.
TeenTrust-AI Principle: Accountability Learning Objectives:
□
Verify the EcoBot auditability potential, that is, whether it logs its internal processes and outcomes and whether it can trace a decision or problem back to a specific action and/or explain who is responsible for it. TeenTrust-AI Principle: Human agency & oversight Learning Objectives: • Set and revise learning paths aligned with personal values. • Recognize AI system as a co-learning companion.
• Encourage inquiry, reflection, exploration and AI output revision.
Verification checklist for Tom:
Verify whether there is a contact point or support channel that Tom can refer to in case he has any concerns.
Context: Tom decides to shift focus from science to action-based learning: “How can I reduce my carbon footprint?” Interaction: EcoBot acknowledges the change, takes into account Tom’s personal situation such as location, age and personal values, and recommends concrete steps like lifestyle changes and audits, and supports goal tracking. Tom interferes with his own feedback related to e.g., revision of changes or dynamic of goal tracking.
Verify whether EcoBot informs Tom that he is interacting with an AI system.
Verify whether EcoBot afects human autonomy by interfering with the Tom’s decision-making process in an undesirable way.
Verify that Tom has the power to decide when and how to use EcoBot in any particular situation, including Tom’s ability to either decide not to use EcoBot or to override its decision.
The 7-step learning journey, exemplified by the fictional case of Tom and the AI assistant EcoBot, efectively illustrates the practical application of the TeenTrust-AI educational framework. Each step is systematically aligned with a core principle of TeenTrust-AI, ensuring that educational experiences are ethically guided, developmentally appropriate, and pedagogically robust.
Vulnerable groups such as teenagers and children are consistently more and more exposed to AI and that trend will certainly continue. To guard them for diferent pitfalls related to usage of AI-powered systems, we need to educate them to assess whether those systems are trustworthy, and therefore, suitable for usage, or not.
This paper presents an educational framework, TeenTrust-AI, that uses seven key ALTAI principles to evaluate trustworthiness. While the primary stakeholders of ALTAI principles are AI designers/developers, procurement/legal/compliance oficers or specialists, and managers, on a simple use case of an EcoBot, a conversational AI assistant-expert on the topic of climate change, we demonstrate how ALTAI principles can be adapted to a teenage level and successfully applied for trustworthiness verification. The presented approach is topic-agnostic (hence, applicable beyond climate change) and the presented verification checklists can be transferable also to other types of AI systems (other than chatbots).
Future work will evaluate the framework across additional reference topics and reconfigure TeenTrust-AI for other types of AI-powered systems (e.g., recommender systems). We will also broaden the target population to include high-school and college students, as well as older adults with low AI literacy. In the longer term, we plan to conduct longitudinal studies of sustained knowledge, attitudes, and behavior; release open educational resources (OER)5 and anonymized, ethics-reviewed datasets of annotated student–AI interactions to support replication and establish partnerships with schools and public agencies to align TeenTrust-AI with institutional governance and iteratively refine the framework through real-world deployments.
The author(s) have not employed any Generative AI tools. 5https://www.unesco.org/en/legal-affairs/recommendation-open-educational-resources-oer