Paper deep dive
Designing Safe and Accountable GenAI as a Learning Companion with Women Banned from Formal Education
Hamayoon Behmanush, Freshta Akhtari, Ingmar Weber, Vikram Kamath Cannanure
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 95%
Last extracted: 4/10/2026, 4:00:31 AM
Summary
This paper presents a remote participatory design study involving 20 women in Afghanistan to explore the development of safe and accountable Generative AI (GenAI) as a learning companion. In contexts where women are banned from formal education, GenAI serves as a critical, albeit risky, tool for self-learning. The study identifies key design directionsâsuch as safety-first interaction, context-grounded support, and pedagogical integrityâand demonstrates that participatory design processes can significantly increase participants' perceived agency and educational aspirations.
Entities (5)
Relation Signals (3)
Hamayoon Behmanush â conductedstudywith â Afghanistan
confidence 100% ¡ We present a remote participatory design study with 20 women in Afghanistan
GenAI â usedby â Women in Afghanistan
confidence 95% ¡ women in Afghanistan... turn to online self-learning and generative AI (GenAI) to pursue their educational and career aspirations
Participatory Design â increased â Perceived Agency
confidence 90% ¡ envisioning the future with GenAI through participatory design was positively associated with significant increases in participants' aspirations... perceived agency
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:In gender-restrictive and surveilled contexts, where access to formal education may be restricted for women, pursuing education involves safety and privacy risks. When women are excluded from schools and universities, they often turn to online self-learning and generative AI (GenAI) to pursue their educational and career aspirations. However, we know little about what safe and accountable GenAI support is required in the context of surveillance, household responsibilities, and the absence of learning communities. We present a remote participatory design study with 20 women in Afghanistan, informed by a recruitment survey (n = 140), examining how participants envision GenAI for learning and employability. Participants describe using GenAI less as an information source and more as an always-available peer, mentor, and source of career guidance that helps compensate for the absence of learning communities. At the same time, they emphasize that this companionship is constrained by privacy and surveillance risks, contextually unrealistic and culturally unsafe support, and direct-answer interactions that can undermine learning by creating an illusion of progress. Beyond eliciting requirements, envisioning the future with GenAI through participatory design was positively associated with significant increases in participants' aspirations (p=.01), perceived agency (p=.01), and perceived avenues (p=.03). These outcomes show that accountable and safe GenAI is not only about harm reduction but can also actively enable women to imagine and pursue viable learning and employment futures. Building on this, we translate participants' proposals into accountability-focused design directions that center on safety-first interaction and user control, context-grounded support under constrained resources, and offer pedagogically aligned assistance that supports genuine learning rather than quick answers.
Tags
Links
- Source: https://arxiv.org/abs/2604.07253v1
- Canonical: https://arxiv.org/abs/2604.07253v1
Trouble viewing inline? Open PDF directly â
Full Text
85,010 characters extracted from source content.
Expand or collapse full text
Designing Safe and Accountable GenAI as a Learning Companion with Women Banned from Formal Education Hamayoon Behmanush â behmanush@cs.uni-saarland.de Saarland Informatics Campus, Saarland University SaarbrĂźcken, Germany Freshta Akhtari Computer Science Faculty, Parwan University Parwan, Afghanistan Ingmar Weber Saarland Informatics Campus, Saarland University SaarbrĂźcken, Germany Vikram Kamath Cannanure Saarland Informatics Campus, Saarland University SaarbrĂźcken, Germany GenAI as a Substitute for Missing Community Support. F1 Privacy and Surveillance Concerns Companionship Under Constraints F2 Contextual Mismatch Pedagogical Challenges Co-Design with GenAI, Driving Aspirational Change F3 F2.2 F2.1 F2.3 AI-Facilitated Virtual Learning Spaces (F1) D1 Safety-First Interaction Design (F2.1) Supporting Learning Under Household Constraints (F2.2) Safeguarding Learning Integrity with Reasoning (F2.3) Aligning Localization with Realistic Opportunities (F1, F2.2) D2 D3 D4 D5 Remote Participatory Design with and for Women in Afghanistan RQ1 : Envisioning GenAI as Self-Learning Companion RQ2: Aspirational Change RQ3 : Accountability-Focused Design Directions Figure 1: Overview of our participatory design study with women in Afghanistan to envision GenAI as a learning companion for online self-learning. The figure summarizes our findings (F1âF3) and accountability-focused design directions (D1-D5), building on participantsâ lived experiences where they are banned from formal education. Abstract In gender-restrictive and surveilled contexts, where access to for- mal education may be restricted for women, pursuing education involves serious safety and privacy risks. When women are ex- cluded from schools and universities, they often turn to online self-learning and generative AI (GenAI) to pursue their educational and career aspirations. However, we know little about what safe and accountable GenAI support is required in the context of surveil- lance, household responsibilities, and the absence of learning com- munities. We present a remote participatory design study with 20 women in Afghanistan, informed by a recruitment survey (n = 140), examining how participants envision GenAI for learning and employability. Participants describe using GenAI less as an infor- mation source and more as an always-available peer, mentor, and source of career guidance that helps compensate for the absence of learning communities. At the same time, they emphasize that this companionship is constrained by privacy and surveillance risks, contextually unrealistic and culturally unsafe support, and direct- answer interactions that can undermine learning by creating an illusion of progress. Beyond eliciting requirements, envisioning the future with GenAI through participatory design was positively asso- ciated with significant increases in participantsâ aspirations (p=.01), â This work has been accepted at ACM Conference on Fairness, Accountability, and Transparency 2026 as a full paper. Please cite the peer-reviewed version. perceived agency (p=.01), and perceived avenues (p=.03). These outcomes show that accountable and safe GenAI is not only about harm reduction but can also actively enable women to imagine and pursue viable learning and employment futures. Building on this, we translate participantsâ proposals into accountability-focused design directions that center on safety-first interaction and user control, context-grounded support under constrained resources, and offer pedagogically aligned assistance that supports genuine learning rather than quick answers. CCS Concepts ⢠Human-centered computingâParticipatory design;⢠Com- puting methodologiesâArtificial intelligence;⢠Security and privacyâPrivacy protections;⢠Applied computingâInter- active learning environments. Keywords Gender-Restrictive Contexts, Online Self-Learning, Participatory Design, GenAI Learning Companion, Accountable GenAI ACM Reference Format: Hamayoon Behmanush, Freshta Akhtari, Ingmar Weber, and Vikram Ka- math Cannanure. 2026. Designing Safe and Accountable GenAI as a Learning Companion with Women Banned from Formal Education. https://doi.org/ X.X arXiv:2604.07253v1 [cs.CY] 8 Apr 2026 Behmanush et al. 1 Introduction Inclusive and quality education is a fundamental human right [70, 72]; however, women in Afghanistan have experienced a major contraction in access to formal education, particularly since 2021, pushing many to pursue learning through informal, online, and self- directed pathways instead [10,57,71]. These pathways increasingly rely on mobile phones and online platforms, but access remains uneven, as connectivity is often costly and unstable, devices may be shared, and learning resources in local languages remain limited [10, 15,35,71]. In addition, womenâs learning is shaped by household responsibilities, restricted mobility, and the loss of in-person peers, mentors, and institutional routines that would normally structure study [1,15,62,71]. In this setting, digital tools are not simply optional supplements to formal education; they are among the few remaining infrastructures through which women can continue learning and pursue career aspirations. Within this shift toward online self-learning, GenAI tools are increasingly framed as always-on educational support that can tutor, scaffold educational tasks, and guide learning trajectories [22,34,55]. In principle, GenAI can support self-learners by pro- viding personalized feedback [55,69], generating tasks for practice, and helping learners organize their study trajectories [28,56,58,77]. However, prior work highlights risks central to accountability in learning with unreliable outputs, shallow engagement, over- reliance, and opaque data practices [32,37,75,78]. For learners with limited prior educational opportunities, these risks are amplified by hallucinated content and unsupportive learning communities [32,33,59]. Given these challenges, we still know little about how women facing systematic educational exclusion envision GenAI as a learning companion that can responsibly support learning and employability under the constraints of privacy and surveillance. Moreover, although prior work has largely examined GenAI design and evaluation in formal educational settings [50,53,81], compar- atively little research has focused on designing with women in gender-restrictive contexts, where contextual mismatch and safety risks can render always-on support consequential. In such contexts, accountability is fundamentally about exposure: what traces GenAI use leaves and what risks those traces can trigger, as surfaced in our participatory design sessions. Using participatory design (PD) as a methodological framework for designing inclusive EdTech and GenAI tools, this study ad- dresses the following research questions: ⢠RQ1: How do women in gender-restrictive contexts en- vision GenAI as a learning and employability companion, and what constraints shape its use? â˘RQ2: How does participation in the participatory design of a GenAI learning companion associate with learnersâ aspirations, including perceived agency and perceived av- enues? â˘RQ3: What accountability-focused design directions emerge for GenAI learning companions in gender-restrictive con- texts? This paper advances the understanding of accountable GenAI for education, particularly in gender-restrictive contexts, where privacy, surveillance, and infrastructural constraints fundamen- tally shape what constitutes safe and useful support. We highlight how women navigating systemic exclusion effectively use GenAI for self-learning while facing risks of exposure and misalignment, and underscore accountability gaps that are often overlooked in well-resourced contexts. Methodologically, we complement qualita- tive participatory design insights with pre-/and post-quantitative evidence that participatory, future-oriented design can shift learn- ersâ reported aspirations, including perceived agency and avenues. Drawing on participantsâ ideas, we articulate accountability-focused design directions for GenAI learning companions that (1) rebuild missing learning community support through AI-facilitated and level-based virtual learning spaces; (2) adopt safety-first interac- tion designs by minimizing trace and maximizing learner control (e.g., anonymous use, rapid deletion); (3) support learning under household constraints with flexible and bilingual microlearning; (4) protect learning integrity by providing step-by-step reasoning over direct-answer interactions, and (5) balance localization to support effective learning with employability skill-building toward realistic working opportunities. 2 Related Work 2.1 Safety, Accountability, and Learning Under Constraint As GenAI becomes more embedded in educational practice, ques- tions of safety and accountability have become central to its respon- sible use [21,24,27,52]. Prior work frames accountability as more than model performance: learners and institutions should be able to understand when GenAI is involved, what kinds of assistance it is providing, and what forms of oversight, contestation, and rem- edy are available when harm occurs [24,27,45,52]. In education, these concerns are especially salient because GenAI can generate plausible but incorrect explanations, reproduce bias, and encourage forms of reliance that support task completion without necessarily supporting durable understanding [6, 7, 32, 37, 75, 78]. At the same time, work on situated and responsible AI argues that accountability cannot be understood only through transparency, correctness, or post-hoc governance [21,24,38]. In unequal and high-risk settings, accountability is also shaped by the social condi- tions under which systems are used. For learners studying under constrained conditions, this includes whether GenAI use leaves traces, what those traces may reveal, whether support is usable on shared or low-performance devices, and whether learners can control disclosure, retention, and visibility of their interactions [25,32,35,55]. In this sense, accountability is not only about ex- plaining model behavior; it is also about reducing exposure and giving learners practical control over when, how, and how safely support can be used. These concerns are particularly important in constrained self- learning environments, where educational access is shaped by un- stable connectivity, limited local-language resources, device sharing, household responsibilities, and reduced access to peers, mentors, and institutional routines [1,10,15,33,35,71]. Under such con- ditions, learning support must be evaluated not only by answer quality, but also by whether it is contextually safe, realistic, and pedagogically appropriate. Educational studies further show that GenAI assistance can improve short-term performance while weak- ening independent problem-solving or confidence calibration when Designing Safe and Accountable GenAI with Women Banned from Formal Education help is poorly calibrated [6,7,65]. Taken together, this literature suggests that accountability in constrained learning contexts must encompass learner control, privacy and exposure management, con- textual safety, and support for genuine learning rather than quick answers alone. 2.2 GenAI and Participatory Design in Education GenAI, particularly large language models, are now used in educa- tional settings to provide conversational tutoring, generate exam- ples and practice items, scaffold problem-solving, support writing and planning, and assist with programming tasks [22,29,34,55,80]. A recurring argument is that such tools can approximate elements of individualized tutoring by offering responsive, on-demand sup- port at scale [3,12,69]. In parallel, emerging work explores how tailored GenAI interventions may support learners who have his- torically had limited access to high-quality instruction, particularly when integrated into broader ecosystems of human support and locally appropriate resources [10, 51, 63]. Participatory Design (PD) provides a complementary method- ological foundation for shaping these tools with, rather than merely for, the people most affected by them [26,46]. PD emphasizes mu- tual learning between researcher and participants, treats partici- pantsâ values as design material, and explicitly attends to power and inequality, often with an orientation toward social justice and long-term structural change [11,68]. Within educational AI and GenAI, PD has been used to co-design learner-facing agents and creativity or learning-support tools, commonly through workshops and iterative prototyping paired with qualitative assessments of usability, perceived agency, and fit with learning goals [3,17,48,76]. Despite this growth, PD-for-GenAI in education still dispro- portionately centers formal institutions and comparatively well- resourced learner populations. Far fewer studies foreground women whose educational and employment opportunities are constrained by restrictive gender norms and socio-political instability, even though these contexts fundamentally shape both design constraints and what benefit means in practice. Moreover, while prior PD stud- ies provide rich qualitative accounts of engagement and agency, the quantitative assessment of downstream outcomes, including shifts in aspirations and future envisioning with GenAI, remains rare [30, 41, 76]. This gap motivates approaches that combine par- ticipatory methods with measurable outcomes, while incorporating the accountability requirements of exposure control and safety constraints discussed above. 3 Methodology This study is part of a broader research project [8] that seeks to adapt GenAI to support marginalized learners in gender-restrictive and sociopolitically unstable contexts. All research activities were approved by our universityâs Ethical Review Board (ERB). 3.1 Participants and Recruitment We partnered with Code to Inspire (CTI) 1 , a coding school that supports girls in Afghanistan with online and programming ed- ucation, to recruit participants for our PD study. To complement recruitment through the coding school and minimize the risk of overlooking less visible subgroups, we additionally used snowball sampling [23] in our recruitment process. In February 2025, we ad- ministered a brief bilingual (EnglishâPersian) demographic survey that requested informed consent, recorded demographic charac- teristics, and documented participantsâ access to and use of online learning and GenAI. The survey was shared through the coding schoolâs channels and via snowball referrals. In total, we received 143 responses (CTI:í=90; snowball sampling:í=53). Three participants did not provide consent to participate in the study. After the demographic survey, we offered four separate 60-minute introductory sessions in AprilâMay 2025 to familiarize the survey respondents with the PD process, answer questions, and encourage participation [47]. The sessions were conducted via Zoom in Persian by two researchers from Afghanistan (one female and one male) and provided a combined overview of PD, followed by an open Q&A. No data were collected during these sessions. From the pool of consenting survey respondents who attended the introductory session, we randomly selected 20 participants to invite to the PD sessions. To help mitigate participation barriers [47], we provided mobile data support to participants with connectivity challenges and engaged with a local gatekeeper to facilitate one participantâs involvement. The gatekeeperâs involvement was limited to the ini- tial consent stage and served only to facilitate the participantâs attendance at the PD session. He did not attend the PD session, did not participate in the discussion, and did not set any rules or constraints on the participantâs contributions. We then conducted five PD sessions, each with four participants, in MayâJune 2025. We provided each PD participant withâŹ20 in compensation, which we believe was appropriate given the session length and local economic conditions. 3.2 PD Processes After the introductory sessions, we conducted five PD sessions on Zoom. Each session included four participants, was co-facilitated by two researchers, and lasted approximately three hours. We in- tentionally adopted a small-group format with two facilitators per session to reduce pressure, support safer participation, and encour- age equitable turn-taking. Facilitators used careful prompting to encourage participation without pressuring participants to respond and periodically created space for quieter participants. The PD pro- cess was structured so that participants first articulated their own current practices, challenges, and desired forms of support in open discussion, and then used those reflections to critique and reshape researcher-prepared storyboard concepts. We used the storyboards as prompts for participants to identify misalignments, propose ad- ditions or removals, and articulate alternative interaction features and deployment choices grounded in their lived realities. 3.2.1 Pre-Surveys and Open Discussion. PD sessions began with a pre-survey that recorded informed consent, further demographic 1 https://w.codetoinspire.org/ Behmanush et al. Consent and demographics. Access to and use of online learning and GenAI. PD process overview. Q&A and encouraging for active involvement in PD. Recruitment Surveys Introductory Sessions Participatory Design (PD) Pre-surveys and Open Discussion Reflections on Storyboards Post-surveys Consent and further demographics. Aspirations-based questions. Discussion on GenAI opportunities and challenges for learning and employability. Discussion on pre- designed storyboards. Reflecting on pre- designed storyboards. Satisfaction and perceived impact of PD. Aspirations-based questions. Figure 2: Overview of our study design. The figure illustrates the sequence of activities, beginning with recruitment sur- veys, followed by introductory sessions to familiarize partic- ipants with the participatory design (PD) process, then PD sessions including pre-PD surveys and open discussions, re- flections on pre-designed storyboards, and concluding with post-surveys. information, and responses to an adapted aspiration-based scale [9,61] to assess changes in participantsâ aspirations associated with attending the PD sessions and envisioning their future with GenAI. The pre-survey was followed by an open discussion struc- tured around four prompting questions: participantsâ most recent questions posed to a GenAI tool, opportunities that generative AI offers for learning programming and supporting employability in their context, challenges with current tools, and the required features and capabilities of a GenAI companion to support learn- ing programming and employability. The open discussion led to participant-generated scenarios that both surfaced challenges with existing tools and envisioned a companion with the desired capabil- ities. These participant-generated scenarios served as the baseline against which researchersâ pre-designed storyboard concepts were later discussed, helping participants identify where the proposed companion aligned with, failed to reflect, or should be revised to better match their constraints, learning practices, and safety needs. At the end of each PD session, participants individually or in pairs translated their scenarios into storyboards. To reduce time pres- sure and support safer participation, they were allowed up to two hours after the session to finalize and submit their storyboard. An example storyboard received from participants is shown in Fig- ure 5 in Appendix D. Open discussions were audio-recorded with participantsâ consent, and facilitators took detailed notes. 3.2.2 Storyboards. Following the open discussion, participants evaluated four storyboards, each depicting a prospective GenAI companion for programming learning and employability support. Participants assessed the storyboards using four prompting ques- tions: (i) their understanding of each storyboard and whether the scenarios reflected their own situations; (i) which existing features were confusing or not helpful; (i) what additions, removals, or changes they would recommend; and (iv) how the companion could be better adapted to their learning style and cultural context. Rather than treating this activity as a summative evaluation of predefined concepts, we used it as a generative design exercise for refining the companion. During each session, facilitators documented participantsâ re- sponses to specific storyboard elements, including which features they wanted to retain, which they considered unrealistic or un- safe, and which they proposed as alternatives. We then examined these discussions across sessions to identify recurring design re- quirements while maintaining traceability between participantsâ feedback and the resulting design directions. To strengthen trace- ability between the participatory design process and the final design directions, we documented storyboard feedback at the level of spe- cific elements (e.g., interaction visibility) and recorded whether participants recommended retaining, removing, or modifying each element. For example, when participants commented on storyboard concepts involving peer- or platform-based support, they empha- sized that interaction should not require identity-linked registra- tion and should remain low-visibility on shared devices. These comments directly informed design implications related to anony- mous participation, rapid trace removal, and user control over what remains visible or stored. The four storyboards varied along two dimensions. First, they differed in the scope of support provided, encompassing not only programming assistance but also employability and soft-skill sup- port for self-learners. Second, they differed in deployment modality, depicting the companion either as embedded within an existing application, implemented as a browser plugin, or provided as a stan- dalone tool. The storyboard concepts were informed by prior work on conversational agents and AI-based support in low-resource learning environments (e.g., [2,5,10,16,43,74]). A sample story- board used in the participatory design sessions is shown in Figure 3. 3.2.3 Post-Surveys. A short post-survey was conducted to assess participantsâ satisfaction and the perceived impact of the PD ses- sions. It included open-ended questions about whether the PDs were helpful and whether participants felt their ideas were heard, and what aspects they did not find helpful. Additionally, the adapted aspiration-based scale [9,61] was reused in the post-survey to mea- sure changes in participantsâ aspirations after participating in the PDs and envisioning their future with GenAI. 3.3 Data Collection We used a demographic survey administered prior to the PD ses- sions to collect background information and technology-use pat- terns. The survey included items on education level, age, preferred language of communication, employment status, and the frequency and use cases of online learning resources and generative AI. Only responses from participants who provided informed consent were included in the subsequent analyses. Demographic survey respon- dentsâ ages ranged from 18 to 34 years (M = 23.4). Detailed demo- graphics are presented in Table 4 in Appendix B. During the PD sessions, we collected pre- and post-session sur- vey responses, participant-generated scenarios and storyboards, and audio recordings of the open discussion and storyboard reflec- tion sections. Across all sessions, participants, working individually Designing Safe and Accountable GenAI with Women Banned from Formal Education Figure 3: An example pre-designed storyboard used in our participatory design sessions. Table 1: We conducted remote PD with 20 women from Afghanistan in 5 online sessions. The table below shows their demographics. IndicatorResponseN Age18â216 22â256 26â348 Educational BackgroundComputer Science (CS) 14 Non-CS6 Marital StatusSingle15 Married5 Learning MethodSDL17 SDL and Online Classes 3 or in pairs on scenario generation, produced 12 scenarios and 12 post-PD storyboard submissions, along with approximately 8 hours of audio recordings from discussions and reflections. The audio recordings were manually transcribed in the original language and then translated into English using machine translation. The first author subsequently reviewed the translated transcripts against the original-language transcripts, corrected translation errors, and refined wording where culturally specific or potentially ambiguous expressions risked losing their intended meaning. Participants in the PD sessions ranged in age from 19 to 34 years. Most identified as self-learners (85%,í=17), while the rest reported a combina- tion of self-learning and structured online classes (15%,í=3). Demographic details for PD participants are provided in Table 1. 3.4 Data Analysis We used Python to clean the demographic, pre-PD, and post-PD survey data and to compute descriptive statistics. We analyzed pre- and post-PD surveys to assess changes in participantsâ aspirations. Statistical significance was tested with pairedíĄ-tests [64,79]. In ad- dition toí-values, we report a percentile-based effect size [39]. For each scale, we first computed the median of the post-PD scores and then located this value within the pre-PD distribution, yielding the post-PD median percentile rank relative to the pre-PD distribution. We qualitatively analyzed (1) participantsâ scenarios and post-PD storyboards and (2) transcripts of audio-recorded open discussions and storyboard-reflection activities. We treated the storyboards as visual narrative data and first examined them alongside the tran- scripts to identify recurring situations, constraints, desired features, and learning-support practices across the dataset [67]. We then Behmanush et al. conducted an inductive, iterative thematic analysis [13, 14] across the full dataset. One researcher led the initial familiarization and open coding of the transcripts and storyboard materials and devel- oped a draft codebook grounded in emergent patterns in the data. This draft codebook was then iteratively reviewed with the broader research team to refine code definitions, merge overlapping codes, distinguish conceptually separate codes, and maintain traceability between raw participant accounts, intermediate codes, and higher- level themes. For example, participant requests for rapid deletion, use without registration, and non-identifying access were initially coded separately and later consolidated into a higher-level code family related to privacy, security, and anonymity. This code family contributed to the broader theme of safety, privacy, and trace con- trol and informed the corresponding design direction of minimizing trace and maximizing learner control. After the team reached an agreement on the codebook, two coders independently applied it to the full dataset. Disagreements were resolved through discussion and, where needed, further refinement of code definitions before final theme generation. Inter-coder agreement yielded a Cohenâs kappa ofí =0.72, indicating substantial agreement [40]. We then used the final codes and themes to develop our findings and artic- ulate accountability-focused design directions for self-learners in gender-restrictive contexts. Appendix A provides a concise codes- to-themes summary that maps the final themes to their associated code families. 4 Findings In this section, we present our findings in four parts: (1) how marginalized women envision and value GenAI for online self- learning and employability; (2) the constraints that shape and often limit GenAIâs role in their context; (3) how participatory design and future-oriented envisioning with GenAI relate to shifts in partici- pantsâ reported aspirations; and (4) we synthesize an accountability- focused design direction for GenAI learning companions that better align with participantsâ infrastructural, sociocultural, gendered, and linguistic realities. 4.1 GenAI as a Partial Substitute for Missing Community Support Since the formal restrictions on womenâs education and employ- ment in Afghanistan in 2021, women and girls have increasingly shifted to online self-learning as an alternative pathway [10,57]. Insights from our recruitment survey indicate that participants primarily utilize online learning to develop programming and other employability-related skills (See Figure 4 in Appendix C). However, sustaining this learning is difficult due to limited and costly connec- tivity, weak community support, household responsibilities, and a shortage of learning resources in local languages [10, 15, 57, 71]. Even when access is available, uncertainty about how to initiate and structure self-study, without the guidance of instructors, peers, or institutional routines, further hinders the effective use of online learning. Within this constrained learning environment, GenAI is already embedded in participantsâ everyday study practices, not as an oc- casional supplement but as a frequently accessed support channel. Recruitment survey data indicate intensive use of GenAI, as 56% of respondents reported using GenAI daily, and an additional 33% reported using it multiple times per week, indicating reliance pat- terns consistent with âalways-availableâ assistance. Accordingly, participants in our PD did not characterize GenAI merely as an information source; rather, they repeatedly described it as a layered form of support that can partially compensate for gaps in peers, mentorship, and career guidance, shaping how they conceptualize GenAI as a learning partner within their self-learning environment. Peer-Like Presence. The PD participants repeatedly empha- sized GenAIâs constant availability as peer-like companionship dur- ing study sessions, describing it as âat the same level, sitting besideâ (PD4_OP; PD5_3) them while they learn. This peer-like presence is envisioned as unbound by classroom schedules or physical infras- tructure, but rather as an app installed on a mobile phone or laptop that remains âusable offline and [able to] update itselfâ (PD3_1; PD3_2; PD5_OD) whenever connectivity is available, thereby pro- viding support across time and place. âAnother important feature is its [GenAIâs] unlimited accessibility at any time and place ... or as a peer, who we can imagine sitting beside us and speaking with us at the same level.â - PD4_OP Mentor-Like Support. Beyond the GenAI peer role, a mentor- ship role also emerged across the PD sessions. Participants envi- sioned GenAI as a âknowledgeable and patient guideâ (PD5_OD; PD1_3) that can walk them step by step through complex tasks, such as debugging or algorithm design, offering detailed expla- nations of complex topics in accessible formats, including text or voice, as needed. They also articulated a need for this mentor to provide âsustained, structured feedback over timeâ (PD2_1; PD4_2), including periodic summaries of what they have studied, where they performed well, and where they continue to struggle, so that they can âidentify strengths and weaknessesâ (PD2_1) and adjust their learning efforts within the self-learning process. GenAI compensates for the missing in- person learning community support. âPeer-like presence: always-available study companion. âMentor-like support: step-by-step expla- nations and feedback. âEmployability guidance: practice soft skills and career preparation support. Employability Guidance. Participants linked employability challenges to the loss of in-person classes and the realities of self- learning, emphasizing that without classrooms and peers, they have fewer routine opportunities to develop soft skills. In this con- text, GenAI was envisioned as compensating for what self-learning cannot easily provide; for instance, helping them practice âtime management or communication and professional interaction such as email writingâ (PD4_2; PD4_4; PD5_OD) or using role-play to âsimulate a discussion betweenâ them as a freelancer and âclients in Upworkâ (PD3_2; PD4_2). Importantly, they framed this employa- bility support as concrete and pathway-specific rather than generic advice, including guidance oriented to platforms like Upwork and Fiverr, assistance with âproposal writingâ (PD4_4), and preparation Designing Safe and Accountable GenAI with Women Banned from Formal Education for âtechnical and coding interviewsâ (PD4_4; PD5_OD), alongside information about relevant remote job markets (PD4_4; PD5_OD). Taken together, this peer-like presence, mentorship, and em- ployability guidance position GenAI as an accompanying tool for these self-learners in the absence of in-person learning commu- nities. However, this companionship comes with limitations that constrain its usability and effectiveness. 4.2Companionship Under Constraints: Privacy, Surveillance, and Contextual Mismatch 4.2.1 Privacy and Surveillance Concerns. While the PD participants envision GenAI as their learning companion, they also identify con- straints and tensions in its use. Participants highlighted privacy as a key concern for the use of GenAI as an online learning compan- ion. They explicitly framed âthe protection of privacy as a major challengeâ (PD2_OD) associated with these tools. They stressed that even minor privacy breaches could have severe consequences for them, considering the restrictive gender norms in the context. â[...] for girls, protecting personal information and privacy is a critical issue. Even the smallest problem in this regard could cause lots of difficulties for us.â - PD2_3 Privacy was understood not only in technical terms, but also as deeply entangled with gendered and social risks. Participants em- phasized that âpersonal security and protection of private informa- tion is extremely importantâ (PD3_3) and therefore any educational technology must be taken seriously on this front. Closely related to privacy, participants highlighted surveillance by household monitoring or even authorities as another constraint that shapes what they can ask, when they can ask it, and how visible their learning must remain. They described living under âexcessive restrictionsâ (PD2_2; PD3_OD) where âeven asking for informa- tion from this tool might cause problemsâ (PD2_2). Participants did not discuss surveillance in the abstract; instead, they described concrete, everyday risks that they faced. Given their reliance on âshared devicesâ (PD3_2), even in some cases, revealing learning content to other family members can cause serious problems. Par- ticipants emphasize that in a context where âuse of mobile phones sometimes causes problemsâ (PD5_4), the toolâs suggestion of a âdating application as a projectâ (PD2_1) does not align with local norms and could lead to tension within families. These highlight that for learners in such patriarchal contexts, the practical realities of device access, monitoring by gatekeepers, and not aligning with local norms can carry significant social risks. 4.2.2 Contextual Mismatch. In addition to privacy and surveillance challenges, participants consistently reported a persistent mismatch between existing GenAI outputs and their local educational, cul- tural, and labor market contexts. Focusing on employability, par- ticipants noted that generated content âoften does not match local needs or the skills requiredâ (PD4_OD), instead focusing on pro- fessions relevant to developed contexts. Contextual misalignment was not only noted in employability but also in everyday learning advice and suggestions. Participants described recommendations for GenAI tools, such as studying in a mixed-gender public space or working from an internet cafĂŠ, which they believe are unrealistic in their context. â[...] or recommends going outside the home to carry out an activity in a place like an internet cafĂŠ. These kinds of suggestions are not suitable for us as girls in Afghanistan.â - PD1_1 The contextual mismatch reported by participants extended into the technical sphere, where GenAI outputs were often incompatible with their constrained resource availability. Learners expressed con- cern that GenAI-generated code sometimes requires âhigh-capacity processors and powerful devicesâ (PD4_1, PD4_OD, PD5_4). This was compounded by their direct experience that the tools some- times generated codes that âcannot be executedâ (PD4_OD), which reinforced their perception that the tools do not consider their context and circumstances while supporting them. GenAI companionship is constrained by privacy, surveillance, and contextual mismatch. âPrivacy & surveillance: Small breaches and moni- toring raise real risks. âContextual mismatch: Suggestions, code, and lan- guage support often donât fit contextual realities. âPedagogical risk: Direct answers create an illusion of learning progress. Furthermore, participants described how the tools are linguisti- cally misaligned with their everyday realities, which âchallenges their learning processâ (PD2_4; PD4_1; PD1_2), such as the rising tension between the need for native language support for immedi- ate learning and the need for English proficiency for employability. While using native language makes concepts easier to grasp, learn- ers cautioned that if they "focus too much on native" (PD2_1) while lacking skills in English, they end up limiting their "own opportu- nities" (PD2_1; PD1_2) for jobs in the broader freelancing market. They expressed that current existing tools fail to adequately support learners in this context and rarely provide structured pathways that enable them to benefit from native-language explanations while systematically developing the English skills required for real-world work opportunities. 4.2.3 Pedagogical Challenges. Pedagogical concerns, particularly the risk that GenAI might create an âillusion of learningâ (PD2_OD; PD4_1; PD1_3), highlight that ârelying too much on artificial intelli- genceâ (PD2_OD; PD4_1) may discourage them from learning new subjects and prevent the real growth of their skills. Several partic- ipants described how receiving complete solutions from existing tools undermined their thinking processes. â[...] from GenAI tools in my learning process, be- cause I believe it takes away my thinking ability and weakens my critical thinking.â - PD2_1 PD participants reported that, through extensive use of these tools, they realized they were not âactually learning anythingâ(PD1_3) but instead had âthe illusion of learningâ(PD1_3). Importantly, they do not reject GenAI on this basis; rather, they raise this concern so that âlearners become awareâ(PD2_1) of this risk and technologists can develop strategies to address it. They emphasize that this should be taken seriously, as in formal learning settings, it may be mitigated Behmanush et al. by the fear of exams, whereas in their non-formal learning context, it can persist unnoticed, gradually undermining learning and skill development. 4.3 Participatory Design to Envision the Future with GenAI: Driving Aspirational Change We examined how learning aspirations of marginalized women changed by participating in PD and envisioning the future with GenAI using an adapted aspiration-based scale [9,61]. The scale captures learnersâ aspirations as perceived long-term goals that extend beyond their current circumstances [18,19,66]. It includes two subscales: Agency, which represents their perception of ability to achieve those goals [18,19,42,66], and Avenue, which refers to the viable paths they see for reaching those goals [18,19,42,66]. Participants reported higher scores on the aspirations measure after the PD sessions, as shown in Table 2. Aspiration scores increased fromí=31.5 (ííˇ=3.3) at pre-PD toí=34.8 (ííˇ=3.7) at post- PD. The result from the pairedíĄ-test showed that this difference was statistically significant,í= .01. This suggests that, following the PD and envisioning the future with GenAI, learners articulated more long-term goals for their education and perceived these goals more positively within their circumstances. Table 2: Pre- and post-PD scores withíĄ-tests. Pre-PDPost-PDíĄ -test ScaleMean SD Mean SDíĄí Agency15.7 2.017.3 2.1 2.2 0.03 Avenue15.8 1.517.5 1.9 2.9 0.01 Aspiration31.5 3.334.8 3.7 2.7 0.01 Looking at the sub-scales, agency scores increased fromí=15.7 (ííˇ=2.0) at pre-PD toí=17.3 (ííˇ=2.1) at post-PD,í= .03, indicating a statistically significant strengthening of participantsâ belief in their ability to move toward their goals. Avenue scores likewise improved fromí=15.8 (ííˇ=1.5) toí=17.5 (ííˇ=1.9), í= .01, providing evidence that learners left the PD with a sense of more paths for achieving their aspirations. A complementary percentile-based effect size analysis provided additional insight into the scale of these changes. For the aspira- tions, the post-PD median corresponded to a percentile rank around 77th of the pre-PD distribution, a shift that falls within the range of a substantial improvement in practical terms. Similarly, for the Agency and Avenue sub-scales, the post-PD medians were located around the 80th percentile of their respective pre-PD distributions, which likewise qualify as substantial changes. These patterns align with a practically meaningful effect of the PD on participantsâ aspi- rations, Agency, and Avenue. Additionally, through the post-survey, PD participants valued the session and highlighted these sessions as an opportunity to have their voices heard (í= 18), sharing their ideas through discussion (í=11), opportunity for teamwork and collaboration (í=7), and a supportive approach to their critical thinking and gaining new insights (í= 5). 4.4 Designing Accountable GenAI Companion: Toward Trusted Support in Constrained Learning Environments Designing accountable GenAI learning companions in gender-restrictive contexts is not only about providing assistance, but also about fos- tering trust among learners, where trust does not imply uncritical confidence, but rather a warranted reliance. This requires that learn- ers can anticipate the companionâs behavior and reduce exposure to harm when privacy, surveillance, and contextual mismatch make even minor failures consequential. Building on the PD participantsâ ideas, this section translates these requirements into accountability- focused design directions. 4.4.1 AI-Facilitated Virtual Learning Spaces. During PD sessions, participants envisioned the GenAI companion as a gateway to within-platform peer communities that could compensate for miss- ing learning interactions. They contrasted their circumstances with less constrained settings where learners can âparticipate in [learn- ing] community activitiesâ (PD2_4; PD3_2), ask questions, and de- velop communication and collaboration skills, emphasizing that the absence of accessible in-person communities limits opportunities for soft-skill development. To address this gap, participants pro- posed that the companion host âa shared chat room among users of the same levelâ (PD5_3), where learners share experiences and carry out group activities anonymously, with the companion providing guidance and moderation. Taken together, these accounts motivate GenAI designs that foreground AI-facilitated, level-based virtual learning spaces and structured peer collaboration, rather than re- lying solely on isolated one-to-one companions. Consistent with our finding that participants framed GenAI as a substitute learning companion (Section 4.1) and resonant with prior work on conversa- tional tutoring and learner-centered agents (e.g., [4,10,17,20,48]), this positions GenAI as a facilitator that enables collaborative prac- tice and the development of technical and soft skills for learners who lack access to in-person communities and classrooms. 4.4.2 Safety-First Interaction Design. Participants treated privacy and safety as the precondition for GenAI companionship, fram- ing privacy not as a preference but as a high-stakes issue where âeven a very small breach can cause a majorâ (PD3_3) safety prob- lem. Accordingly, they articulated interaction requirements that minimize trace and maximize learner control over what is stored, shown, and left behind on devices. Concretely, participants asked for âa simple and quick option to delete anything that has been usedâ (PD3_3; PD2_2), alongside use âwithout registrationâ (PD2_2; PD2_1; PD4_4; PD5_4) or through non-identifying IDs. These re- quests position privacy-by-design not as back-end compliance, but as an interaction-level guarantee that supports safe use in the pres- ence of surveillance constraints (Section 4.2). Safety-first design also required boundary-aware and contextually safe support, because harm can arise not only from retained data but from suggestions that violate social constraints. Participants wanted the companion to respect âthe boundaries that are non-negotiableâ (PD4_3) and to avoid guidance that could place them at risk. For example, recom- mendations to do tasks âin a shared environment alongside menâ (PD1_1) or project ideas (e.g., a âdating applicationâ) that could trigger risks if seen by family members. Designing Safe and Accountable GenAI with Women Banned from Formal Education Taken together, these accounts motivate safety-first interaction design as a combined commitment to trace minimization, learner agency, and boundary-conditioned generation which consistent with research showing how learning and technology use are contin- ually negotiated under monitoring and gendered risk [1,36,62,73], and aligned with calls for contextually safe and equitable learning technologies [44, 54, 60, 62]. 4.4.3 Safeguarding Learning Integrity Through Step-by-Step Rea- soning. Participants did not object to GenAI because it can respond quickly; they worried that direct-answer support can undermine learning integrity when no teacher or peer is available to surface misconceptions or repair fragile understanding (Section 4.2.3). They talk about a turning point after heavy use, where they realized they were not âactually learning anythingâ (PD1_3) and were left with the illusion of learning progress, forgetting how to structure pro- grams and even how to debug code. Some participants emphasized that providing complete answers âtakes awayâ (PD1_3; PD2_1; PD1_OD) their thinking ability and weakens their critical think- ing. Consistent with prior concerns about shallow engagement and over-reliance in educational GenAI use [31,37,75,78], partici- pantsâ integrity agenda centered on interactions that the companion should privilege reasoning by prompting learners to attempt in- termediate steps and by revealing support gradually, rather than collapsing tasks into final answers. They also emphasized learner control support, considering the time pressure and difficulty of enabling quick clarification when needed, while preserving deeper practice on complex problems, thereby operationalizing the com- panion as reasoning support rather than solution delivery. 4.4.4 Supporting Learning Under Household Constraints. Consider- ing the learnersâ household responsibilities, limited study time, and language and infrastructural barriers (Section 4.2), partici- pants emphasized that GenAI support must be usable in fragments rather than merely available. They described having âdozens of family responsibilitiesâ (PD2_3; PD1_2) and proposed an âenviron- ment for microlearningâ (PD2_3; PD5_4; PD4_4; PD2_1) that fits 10â15 minute sessions, preserves state, and enables resumption with lightweight recaps when uninterrupted study is not feasi- ble. Infrastructural limits further shaped accountability require- ments (Section 4.2.2), which participants asked for offline and âweak internetâ (PD2_OD; PD1_OD) resilience, including offline access with occasional updates instead of continuous connectiv- ity, and they cautioned that outputs (e.g., code) must run on low- spec devices to avoid becoming an additional barrier. They also framed bilingual support as a progression rather than a binary choice, requesting bilingual support for local-language explana- tions using familiar terms while keeping English visible to avoid âlimit[ing] opportunitiesâ (PD2_1; PD4_4) tied to remote work. To- gether, these proposals suggest that accountable companions should couple resumable microlearning with low-bandwidth/offline opera- tion, device-appropriate outputs, multimodal explanations beyond âtext-only interactionâ (PD4_3), and bilingual support that bridges toward employability-relevant English, extending prior work on low-resource learning support and the shortcomings of generic, English-dominant materials in constrained contexts [10, 44, 49]. 4.4.5 Aligning Localized Support With Realistic Work Opportunities. Attention to locally grounded, project-based learning emerged as a core expectation for GenAI support. Participants critiqued ex- isting tools for offering generic examples and career suggestions that overlook their âlocation, family conditions, and permission to travelâ (PD1_3), and instead emphasized small, contextually mean- ingful projects that build âself-confidenceâ (PD3_1) and help the âknowledge stayâ (PD2_1; PD3_4) in their minds. They framed this kind of localization as both pedagogically necessary and pro- fessionally consequential that learning activities should feel safe and usable in their everyday environment while also producing artifacts and skills that can translate to remote working opportuni- ties. Accordingly, participants requested interactional scaffolds that âsimulate a discussionâ (PD3_2; PD4_4) between a freelancer and a client, alongside practical preparation for remote-work platforms such as Upwork and Fiverr, given the constraints on in-person employment. This emphasis responds directly to contextual irrele- vance (Section 4.2.2) and operationalizes the companion as both a substitute for missing learning support and a pathway-oriented em- ployability guide (Section 4.1), aligning with broader calls to design context-sensitive systems that respect constraints while expanding realistic opportunities for women (e.g., [44, 54, 60, 62]). 5 Discussion Our findings suggest that, in restrictive learning environments, the central question for educational GenAI is not only whether it generates correct answers, but whether it affords warranted re- liance: support that learners can use safely, pedagogically, and un- der their actual conditions. This reframes educational GenAI as a socio-technical support system whose usefulness depends on social scaffolding, exposure control, contextual and pedagogical fit, and future-oriented participatory design. 5.1 Beyond Direct Answers: GenAI as a Partial Stand-In for Missing Learning Communities Prior work on GenAI in education often evaluates support in terms of explanation quality, correctness, and adaptivity [22,55,69]. Our findings show that these criteria are not sufficient in contexts where learners have lost access to peers, mentors, and institutional rou- tines. Participants valued GenAI not mainly for speed, but for conti- nuity: someone to study with, guidance that accumulates over time, and help connecting learning to work. In this sense, the relevant unit of support is not a single answer, but an ongoing relationship to study. This shifts the evaluation of educational GenAI in restrictive en- vironments. GenAI systems should be judged not only by whether they answer well, but also by whether they help learners receive longitudinal feedback and recover some of the functions that formal education would ordinarily provide. At the same time, our find- ings point to an important tension: the more GenAI approximates community, the more carefully visibility, identity, and trace reten- tion must be managed. Community-like support may be valuable, but only when anonymity, minimal retention, and user control are treated as first-order design requirements rather than optional safeguards. Behmanush et al. 5.2 Accountability as Exposure Control In high-risk learning environments, accountability cannot be re- duced to fairness, transparency, or auditability alone [24,27,45,52]. Our findings show that it must also address exposure: what remains on the device, what others can infer from use, and whether other- wise plausible suggestions become unsafe in local conditions. This broadens harm from incorrect outputs to include trace harms, such as saved histories or registration footprints, and boundary harms, where seemingly reasonable recommendations become risky be- cause they exceed local social constraints [38]. The implication is practical. Exposure control must be built into the interaction layer rather than treated only as a back-end pri- vacy principle. Anonymous or temporary access, minimal history retention, rapid deletion, and user-steerable boundaries are not peripheral features in this setting; they are conditions of safe use. Accountable GenAI, then, is not only about preventing harmful outputs but also about giving learners meaningful control over when support becomes visible, durable, and risky. 5.3 Localization as Situated Pedagogy and Opportunity Alignment Our findings suggest that guidance may be factually correct yet still unusable if it assumes uninterrupted study time, private devices, stable connectivity, or futures that are not realistically available. Localization in this setting, therefore, cannot be reduced to trans- lation only. It is better understood as the production of support that learners can actually act on under their everyday constraints [10, 32, 38]. This requires alignment at three levels. First, pedagogical align- ment: support should fit fragmented study through resumable microlearning, lightweight recaps, and reasoning-first assistance rather than immediate solution delivery [6,7]. Second, infrastruc- tural alignment: generated code, tools, and workflows should as- sume low-spec devices and weak connectivity by default. Third, opportunity alignment: examples, projects, and career guidance should connect to realistic pathways such as remote freelancing, proposal writing, portfolio building, client communication, and interview preparation rather than generic advice detached from local constraints [15, 51]. 5.4 Envisioning the Future with GenAI: PD Beyond Idea Generation Following the participatory design (PD) sessions, participants re- ported statistically significant increases in aspirations, perceived agency, and perceived avenues toward long-term goals. Although these pre/post shifts should not be read as evidence of durable, long- term change, they suggest that future-oriented participation around GenAI can have value beyond mere requirement elicitation. This interpretation is consistent with prior work showing that participa- tory processes can strengthen agency, self-efficacy, and reflective engagement [30,41,76]. It also resonates with foundational PD arguments that participation matters not only because it improves artifacts, but because it makes participantsâ values, constraints, and aspirations consequential in shaping technology [11, 26, 46]. In this context, the significance of PD lies not only in improving the contextual fit of a proposed GenAI system but also in creat- ing a rare space for collective reflection under conditions where educational and professional futures are otherwise constrained. Participants described the sessions as opportunities to be heard, exchange perspectives, and critically examine what GenAI could and could not do for them. This suggests that PD contributed not merely by eliciting preferences, but by helping participants articu- late more concrete and plausible trajectories toward learning and employability. Seen this way, PD in accountable AI should be un- derstood not only as a design methodology, but also as a process through which participantsâ constraints, aspirations, and imagined futures become materially relevant to the direction of technological design. 6 Conclusion This study examines how women banned from formal education in Afghanistan use GenAI for online self-learning and employability. Participants framed GenAI less as an information source and more as a peer-like, mentor-like, and career-guidance companion that partially substitutes for missing learning communities. Yet they emphasized that companionship can also increase exposure: traces on shared devices, surveillance, and contextually unsafe sugges- tions that can make learning a risky endeavor. Envisioning future GenAI-companions through participatory design was associated with significant increases in aspirations, perceived agency, and perceived avenues, suggesting that participatory, future-oriented methods can shift not only design requirements but also learnersâ sense of perceived opportunities. We therefore frame accountable GenAI support as exposure control in this context by minimizing trace by default, maximizing learner control (e.g., anonymous use, rapid deletion), and constraining outputs to non-negotiable social boundaries. From this perspective, educational GenAI should be evaluated by warranted reliance: whether it helps learners sustain routines and pursue realistic opportunities without increasing harm, not only by the quality of its outputs. Acknowledgments IW and VC are supported by funding from the Alexander von Hum- boldt Foundation and its founder, the Federal Ministry of Education and Research (Bundesministerium fĂźr Bildung und Forschung). We additionally thank Code to Inspire (CTI) for their valuable support with participant recruitment for this study. Positionality and Ethical Considerations Our research team is based in Germany and Afghanistan and brings experience in GenAI for education, humanâcomputer interaction for development, and digital gender gaps. Some team members have themselves experienced restrictions on education and employabil- ity, which informed our sensitivity to power imbalances between researchers and participants. This study was approved by our universityâs Ethical Review Board (ERB). Given the gender-restrictive and sociopolitically un- stable context, we implemented additional safeguards, including multi-stage informed consent, voluntary participation with the Designing Safe and Accountable GenAI with Women Banned from Formal Education option to withdraw at any time, and mobile data support for partic- ipants who requested it. We minimized the collection of identifying information, deleted audio recordings after transcription, and re- stricted access to study materials to the research team in line with ERB-approved procedures. Generative AI Usage We used ChatGPT 5 to suggest grammar and clarity edits on author- written text. All scientific content, analysis, results, and interpreta- tions were produced by the authors; no GenAI tools were used to write the actual text of the findings or analysis. References [1]Shukra Raj Adhikari, Bhawani Shankar Adhikari, and Ganga Acharya. 2024. E-learning method and University life of married female students in patriar- chal social structure in sociological Perspective. In Forum Ilmu Sosial, Vol. 51. Universitas Negeri Semarang, Semarang, Indonesia, 66â84. [2]Edmund Afoakwah, Francisco Carballo, Alex Caro, Samantha DâCunha, Stephanie Dobrowolski, and Alexandra Fallon. 2021. Dialling up Learning. [3] Abdullah Al-Abri. 2025.Exploring ChatGPT as a virtual tutor: A multi- dimensional analysis of large language models in academic support. Education and Information Technologies 30 (2025), 1â36. [4]Ashton Anderson, Daniel Huttenlocher, Jon Kleinberg, and Jure Leskovec. 2014. Engaging with Massive Online Courses. In Proceedings of the 23rd International Conference on World Wide Web (W â14). Association for Computing Machin- ery, New York, NY, USA, 687â698. doi:10.1145/2566486.2568042 [5]Noam Angrist, Peter Bergman, and Moitshepi Matsheng. 2022. Experimental evidence on learning using low-tech when school is out. Nature human behaviour 6, 7 (2022), 941â950. [6]AndrĂŠ Barcaui. 2025. ChatGPT as a cognitive crutch: Evidence from a randomized controlled trial on knowledge retention. Social Sciences & Humanities Open 12 (2025), 102287. doi:10.1016/j.ssaho.2025.102287 [7] Patrick Bassner, Ben Lenk-Ostendorf, Ramona Beinstingel, Tobias Wasner, and Stephan Krusche. 2025. Less stress, better scores, same learning: The dissociation of performance and learning in AI-supported programming education. Computers and Education: Artificial Intelligence 6 (2025), 100537. doi:10.1016/j.caeai.2025. 100537 [8] Hamayoon Behmanush. 2025. Supporting Marginalized Learners with GenAI. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, Vol. 8. AAAI Press, Washington, DC, USA, 2848â2849. [9]Hamayoon Behmanush, Freshta Akhtari, Roghieh Nooripour, Ingmar Weber, and Vikram Kamath Cannanure. 2025. Hope, Aspirations, and the Impact of LLMs on Female Programming Learners in Afghanistan. In Workshop on Artificial Intelligence with and for Learning Sciences: Past, Present, and Future Horizons. Springer, Cham, Switzerland, 11 pages. [10]Hamayoon Behmanush, Freshta Akhtari, Roghieh Nooripour, Ingmar Weber, and Vikram Kamath Cannanure. 2025. Online Learning and GenAI: Supporting Womenâs Aspirations Amid Socio-Political Instability in Afghanistan. In Pro- ceedings of the ACM SIGCAS/SIGCHI Conference on Computing and Sustainable Societies. Association for Computing Machinery, New York, NY, USA, 401â417. [11]Erling BjĂśrgvinsson, Pelle Ehn, and Per-Anders Hillgren. 2010. Participatory design and" democratizing innovation". In Proceedings of the 11th Biennial partic- ipatory design conference. Association for Computing Machinery, New York, NY, USA, 41â50. [12]Benjamin S Bloom. 1984. The 2 sigma problem: The search for methods of group instruction as effective as one-to-one tutoring. Educational researcher 13, 6 (1984), 4â16. [13]Virginia Braun and Victoria Clarke. 2006. Using thematic analysis in psychology. Qualitative research in psychology 3, 2 (2006), 77â101. [14]Virginia Braun and Victoria Clarke. 2019. Reflecting on reflexive thematic analysis. Qualitative research in sport, exercise and health 11, 4 (2019), 589â597. [15]Arianna Briganti and Nilofar Sekandari. 2025. Digital transformation for Afghan women in Central Asia: a pathway to economic empowerment for inclusive regional stability. https://w.osce.org/files/f/documents/c/e/589841.pdf [16]Lauren Bringman-Rodenbarger and Michael Hortsch. 2020. How students choose E-learning resources: The importance of ease, familiarity, and convenience. Faseb Bioadvances 2, 5 (2020), 286â295. [17]Amanda Buddemeyer, Jennifer Nwogu, Jaemarie Solyst, Erin Walker, Tara Nkrumah, Amy Ogan, Leshell Hatley, and Angela Stewart. 2022. Unwritten magic: Participatory design of AI Dialogue to empower marginalized voices. In Proceedings of the 2022 ACM conference on information technology for social good. Association for Computing Machinery, New York, NY, USA, 366â372. [18]Vikram Kamath Cannanure, EloĂsa Ăvila-Uribe, Tricia Ngoon, Yves Adji, Sharon Wolf, Kaja JasiĹska, Timothy Brown, and Amy Ogan. 2022. âWe dream of climbing the ladder; to get there, we have to do our job betterâ: Designing for Teacher Aspirations in rural CĂ´te dâIvoire. In Proceedings of the 5th ACM SIGCAS/SIGCHI Conference on Computing and Sustainable Societies. Association for Computing Machinery, New York, NY, USA, 122â138. [19] Vikram Kamath Cannanure, Justin Souvenir Niweteto, Yves Thierry Adji, Akpe Y Hermann, Kaja K Jasinska, Timothy X Brown, and Amy Ogan. 2020. Iâm fine where I am, but I want to do more: Exploring Teacher Aspirations in Rural CĂ´te dâIvoire. In Proceedings of the 3rd ACM SIGCAS Conference on Computing and Sustainable Societies. Association for Computing Machinery, New York, NY, USA, 1â12. [20] Vikram Kamath Cannanure, Sharon Wolf, Kaja Jasinska, Tim Brown, and Amy Ogan. 2025. Enabling a virtual community of practice facilitated by a chatbot for rural teachers. In Proceedings of the 5th Biennial African Human Computer Interaction Conference. Association for Computing Machinery, New York, NY, USA, 311â320. [21] Stephen Casper, Carson Ezell, Charlotte Siegmann, Noam Kolt, Taylor Lynn Curtis, Benjamin Bucknall, Andreas Haupt, Kevin Wei, JĂŠrĂŠmy Scheurer, Marius Hobbhahn, Lee Sharkey, Satyapriya Krishna, Marvin von Hagen, Silas Alberti, Alan Chan, Qinyi Sun, Michael Gerovitch, David Bau, Max Tegmark, David Krueger, and Dylan Hadfield-Menell. 2024. Black-Box Access is Insufficient for Rigorous AI Audits. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT â24) (Rio de Janeiro, Brazil). Association for Computing Machinery, New York, NY, USA, 19 pages. doi:10.1145/3630106. 3659037 [22]Eason Chen, Jia-En Lee, Jionghao Lin, and Kenneth Koedinger. 2024. GPTutor: Great personalized tutor with large language models for personalized learning content generation. In Proceedings of the Eleventh ACM Conference on Learning@ Scale. Association for Computing Machinery, New York, NY, USA, 539â541. [23]Louis Cohen, Lawrence Manion, and Keith Morrison. 2002. Research methods in education. routledge, London, UK. [24]A. Feder Cooper, Emanuel Moss, Benjamin Laufer, and Helen Nissenbaum. 2022. Accountability in an Algorithmic Society: Relationality, Responsibility, and Ro- bustness in Machine Learning. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency (FAccT â22). Association for Computing Machinery, New York, NY, USA, 864â876. doi:10.1145/3531146.3533150 [25]Nathalie DiBerardino and Luke Stark. 2023. (Anti)-Intentional Harms: The Conceptual Pitfalls of Emotion AI in Education. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency (FAccT â23) (Chicago, IL, USA). Association for Computing Machinery, New York, NY, USA, 1386â1395. doi:10.1145/3593013.3594088 [26]Betsy DiSalvo. 2016. Participatory design through a learning science lens. In Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems. Association for Computing Machinery, New York, NY, USA, 4459â4463. [27] Ravit Dotan, Lisa S. Parker, and John G. Radzilowicz. 2024. Responsible Adop- tion of Generative AI in Higher Education: Developing a âPoints to Considerâ Approach Based on Faculty Perspectives. In Proceedings of the 2024 ACM Con- ference on Fairness, Accountability, and Transparency (FAccT â24) (Rio de Janeiro, Brazil). Association for Computing Machinery, New York, NY, USA, 14 pages. doi:10.1145/3630106.3659023 [28]Louie Giray, Jonard NemeĂąo, and Jelomil Edem. 2025. Self-Directed Learning Using ChatGPT Positively Affects Student Engagement. International Journal of Technology in Education 8, 3 (2025), 667â680. [29]Manel Guettala, Samir Bourekkache, Okba Kazar, Saad Harous, et al.2024. Gen- erative artificial intelligence in education: Advancing adaptive and personalized learning. Acta Informatica Pragensia 13, 3 (2024), 460â489. [30]Christina N Harrington. 2020. The forgotten margins: what is community-based participatory health design telling us? Interactions 27, 3 (2020), 24â29. [31]Emma Harvey, Allison Koenecke, and Rene F Kizilcec. 2025. " Donât Forget the Teachers": Towards an Educator-Centered Understanding of Harms from Large Language Models in Education. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. Association for Computing Machinery, New York, NY, USA, 1â19. [32]Wayne Holmes, Fengchun Miao, et al.2023. Guidance for generative AI in educa- tion and research. Unesco Publishing, Paris, France. [33]Samia Ibtasam, Lubna Razaq, Maryam Ayub, Jennifer R Webster, Syed Ishtiaque Ahmed, and Richard Anderson. 2019. " My cousin bought the phone for me. I never go to mobile shops." The Role of Family in Womenâs Technological Inclu- sion in Islamic Culture. Proceedings of the ACM on Human-Computer Interaction 3, CSCW (2019), 1â33. [34]Jussi S Jauhiainen and AgustĂn Garagorry Guerra. 2024. Generative AI and educa- tion: dynamic personalization of pupilsâ school learning material with ChatGPT. In Frontiers in Education, Vol. 9. Frontiers Media SA, Lausanne, Switzerland, 1288723. [35]Aziz Ullah Karimy, Juma Rasuli, P Chandrasekhar Reddy, Musa Joya, Ali Juma Hamdard, and Hassan Rahnaward Ghulami. 2024. A Review on the Feasibility of AI-Supported Education Platforms in Afghanistan: Addressing Barriers to Behmanush et al. Women and Girlsâ Education. In 2024 IEEE Global Humanitarian Technology Conference (GHTC). IEEE, PA, USA, 1â8. [36]Naveena Karusala, Apoorva Bhalla, and Neha Kumar. 2019. Privacy, patriarchy, and participation on social media. In Proceedings of the 2019 on Designing Inter- active Systems Conference. Association for Computing Machinery, New York, NY, USA, 511â526. [37] Enkelejda Kasneci, Kathrin SeĂler, Stefan KĂźchemann, Maria Bannert, Daryna Dementieva, Frank Fischer, Urs Gasser, Georg Groh, Stephan GĂźnnemann, Eyke HĂźllermeier, et al.2023. ChatGPT for good? On opportunities and challenges of large language models for education. Learning and individual differences 103 (2023), 102274. [38]Michael Katell, Meg Young, Dharma Dailey, Bernease Herman, Vivian Guetler, Aaron Tam, Corinne Binz, Daniella Raz, and P. M. Krafft. 2020. Toward Situated Interventions for Algorithmic Equity: Lessons from the Field. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency (FAccTâ20). Association for Computing Machinery, New York, NY, USA, 45â55. doi:10.1145/ 3351095.3372874 [39] Matthew A Kraft. 2020. Interpreting effect sizes of education interventions. Educational researcher 49, 4 (2020), 241â253. [40]J. Richard Landis and Gary G. Koch. 1977. The Measurement of Observer Agree- ment for Categorical Data. Biometrics 33, 1 (1977), 159â174. doi:10.2307/2529310 [41]Lea Holst Laursen. 2025. Post-its, postcards, and dream clouds: exploring partici- patory design methods as avenues to capture rural youth voices. CoDesign 21, 2 (2025), 195â210. [42]Travis J Lybbert and Bruce Wydick. 2017. Hope as aspirations, agency, and path- ways: poverty dynamics and microfinance in Oaxaca, Mexico. In The economics of poverty traps. University of Chicago Press, Chicago, USA, 153â177. [43] Niall McNulty. 2025. Digital Teacher Support in South Africa: How WhatsApp Chatbots and Generative AI Are Bridging the Gap. Medium. https://medium.com/ @niall.mcnulty/digital-teacher-support-in-south-africa-730f623d319 [44]Rim Melake, Danielle Robinson, Sarah Danman, Harmonie Kobanghe Langazo, and Alicia Hammond. 2025. What Works to Advance Womenâs Digital Literacy? A Review of Good Practices and Programs. Technical Report P173166. World Bank Group, Washington, DC.https://documents.worldbank.org/en/publication/ documents-reports/documentdetail/099040225191054238 [45]Jacob Metcalf, Emanuel Moss, Elizabeth Anne Watkins, Ranjit Singh, and Madeleine Clare Elish. 2021. Algorithmic Impact Assessments and Accountabil- ity: The Co-construction of Impacts. In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT â21). Association for Computing Machinery, New York, NY, USA, 735â746. doi:10.1145/3442188.3445935 [46] Michael J Muller and Sarah Kuhn. 1993. Participatory design. Commun. ACM 36, 6 (1993), 24â28. [47]Henrietta Nanyonjo, Melissa Densmore, Sierra Van Riel, Fiona Ssozi, and Dorothea Kleine. 2025. GeJuSTA Gender-Just Co-Design Toolkit (v5.3). https: //melissadensmore.com/gejusta. Licensed under C BY-NC-SA 4.0. [48] Michele Newman, Kaiwen Sun, Ilena B Dalla Gasperina, Grace Y Shin, Matthew Kyle Pedraja, Ritesh Kanchi, Maia B Song, Rannie Li, Jin Ha Lee, and Jason Yip. 2024. " I want it to talk like Darth Vader": Helping Children Con- struct Creative Self-Efficacy with Generative AI. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems. Association for Computing Machinery, New York, NY, USA, 1â18. [49]Hellina Hailu Nigatu, John Canny, and Sarah E Chasins. 2024. Low-Resourced Languages and Online Knowledge Repositories: A Need-Finding Study.. In Pro- ceedings of the 2024 CHI Conference on Human Factors in Computing Systems. Association for Computing Machinery, New York, NY, USA, 1â21. [50] Isabella Pu, Prerna Ravi, Linh Dieu Dinh, Chelsea Joe, Caitlin Ogoe, Zixuan Li, Cynthia Breazeal, and Anastasia K Ostrowski. 2025. "How can we learn and use AI at the same time?": Participatory Design of GenAI with High School Students. In Proceedings of the 24th Interaction Design and Children. Association for Computing Machinery, New York, NY, USA, 204â220. [51]Sabit Rahim, Gul Sahar, Gul Jabeen, Sabila Khatoon, et al.2025. Harnessing generative AI: Reviewing applications, challenges, and solutions for out-of-school children in developing regions. Sustainable Futures 10 (2025), 101206. [52]Inioluwa Deborah Raji, Andrew Smart, Rebecca N. White, Margaret Mitchell, Timnit Gebru, Ben Hutchinson, Jamila Smith-Loud, Daniel Theron, and Parker Barnes. 2020. Closing the AI Accountability Gap: Defining an End-to-End Frame- work for Internal Algorithmic Auditing. In Conference on Fairness, Accountability, and Transparency (FAccTâ20). Association for Computing Machinery, New York, NY, USA, 33â44. doi:10.1145/3351095.3372873 [53] Prerna Ravi, John Masla, Gisella Kakoti, Grace C Lin, Emma Anderson, Matt Taylor, Anastasia K Ostrowski, Cynthia Breazeal, Eric Klopfer, and Hal Abelson. 2025. Co-designing Large Language Model Tools for Project-Based Learning with K12 Educators. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. Association for Computing Machinery, New York, NY, USA, 1â25. [54]Agha Ali Raza, Mustafa Naseem, Namoos Hayat Qasmi, Shan Randhawa, Fizzah Malik, Behzad Taimur, Sacha St-Onge Ahmad, Sarojini Hirshleifer, Arman Rezaee, and Aditya Vashistha. 2022. Fostering engagement of underserved communities with credible health information on social media. In Proceedings of the ACM Web Conference 2022. Association for Computing Machinery, New York, NY, USA, 3718â3727. [55] J Roe and M Perkins. 2024. Generative AI in self-directed learning: A scoping review (No. arXiv: 2411.07677). arXiv. [56]Sami Sarsa, Paul Denny, Arto Hellas, and Juho Leinonen. 2022. Automatic generation of programming exercises and code explanations using large language models. In Proceedings of the 2022 ACM conference on international computing education research-volume 1. Association for Computing Machinery, New York, NY, USA, 27â43. [57]Abdul Qahar Sarwari and Hamedi Mohd Adnan. 2024. Alternative educational activities and programs for female students banned from formal education in Afghanistan. Issues in Educational Research 34, 3 (2024), 1170â1179. [58]Md Istiak Hossain Shihab, Christopher Hundhausen, Ahsun Tariq, Summit Haque, Yunhan Qiao, and Brian Wise Mulanda. 2025. The Effects of GitHub Copilot on Computing Studentsâ Programming Effectiveness, Efficiency, and Processes in Brownfield Coding Tasks. In Proceedings of the 2025 ACM Conference on International Computing Education Research V. 1. Association for Computing Machinery, New York, NY, USA, 407â420. [59]Shaghayegh Shirzad and Mansoor Ali Darazi. 2025. AI Literacy in Education: Balancing Innovation, Ethics, and Equity in the Digital Age. Journal of New Trends in English Language Learning (JNTELL) 4, Special Issue (2025), 9 pages. [60]Summit Shrestha, Josiah Hester, Ashutosh Dhekne, Umakishore Ramachan- dran, and Alex Cabral. 2025. Bringing Context to the Underserved: Rethinking Context-Aware Design to Bridge the Digital Divide. In Proceedings of the ACM SIGCAS/SIGCHI Conference on Computing and Sustainable Societies. Association for Computing Machinery, New York, NY, USA, 647â665. [61] C. R. Snyder, C. Harris, J. R. Anderson, S. A. Holleran, L. M. Irving, S. T. Sig- mon, L. Yoshinobu, J. Gibb, C. Langelle, and P. Harney. 1991. The will and the ways: Development and validation of an individual-differences measure of hope. Journal of Personality and Social Psychology 60, 4 (1991), 570â585. [62] Sharifa Sultana, François Guimbretière, Phoebe Sengers, and Nicola Dell. 2018. Design within a patriarchal society: Opportunities and challenges in designing for rural women in bangladesh. In Proceedings of the 2018 CHI conference on human factors in computing systems. Association for Computing Machinery, New York, NY, USA, 1â13. [63] Manohar Swaminathan and Tarini Naik. 2025. Generative AI for Teachers with Vision Impairments in the Global South: A Bridge Too Far?. In Proceedings of the 27th International ACM SIGACCESS Conference on Computers and Accessibility. Association for Computing Machinery, New York, NY, USA, 1â18. [64] Rachel Szczytko, Sarah J Carrier, and Kathryn T Stevenson. 2018. Impacts of outdoor environmental education on teacher reports of attention, behavior, and learning outcomes for students with emotional, cognitive, and behavioral disabil- ities. In Frontiers in Education, Vol. 3. Frontiers Media SA, Lausanne, Switzerland, 46. [65] Myles Joshua Toledo Tan and Nicholle Mae Amor Tan Maravilla. 2024. Shaping integrity: why generative artificial intelligence does not have to undermine education. Frontiers in Artificial Intelligence 7 (2024), 1471224. doi:10.3389/frai. 2024.1471224 [66]Kentaro Toyama. 2018. From needs to aspirations in information technology for development. Information Technology for Development 24, 1 (2018), 15â36. [67]Khai N. Truong, Gillian R. Hayes, and Gregory D. Abowd. 2006. Storyboard- ing: An Empirical Determination of Best Practices and Effective Guidelines. In Proceedings of the 6th ACM Conference on Designing Interactive Systems (DIS â06). Association for Computing Machinery, New York, NY, USA, 12â21. doi:10.1145/1142405.1142410 [68]Ari Tuhkala. 2021. A systematic literature review of participatory design studies involving teachers. European Journal of Education 56, 4 (2021), 641â659. [69]Laurah Turner, Matt Kelleher, Seth Overla, Weibing Zheng, Alexander Gregath, Micheal Gharib, Andrew Zahn, Sally A Santen, and Danielle E Weber. 2025. Harnessing the Generative Power of AI to Move Closer to Personalized Medical Education. Academic Medicine 100 (2025), 10â1097. [70]UNESCO. 2025. Nurturing Safe Learning Environments to Realize the Right to Education: Brief. https://unesdoc.unesco.org/ark:/48223/pf0000396302 [71]UNICEF and UNESCO. 2025.Afghanistan Education Situation Report 2025. https://w.unicef.org/afghanistan/documents/afghanistan-education- situation-report-2025 [72]United Nations General Assembly. 2015. Transforming Our World: The 2030 Agenda for Sustainable Development. Resolution adopted by the General As- sembly on 25 September 2015 (A/RES/70/1). https://sdgs.un.org/2030agenda Seventieth session. [73]Rama Adithya Varanasi, Divya Siddarth, Vivek Seshadri, Kalika Bali, and Aditya Vashistha. 2022. Feeling proud, feeling embarrassed: Experiences of low-income women with crowd work. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems. Association for Computing Machinery, New York, NY, USA, 1â18. [74]IW Wait. 2013. Reducing instructional barriers through software virtualization. Journal of Online Engineering Education 4 (2013), 5 pages. Designing Safe and Accountable GenAI with Women Banned from Formal Education [75]Shen Wang, Tianlong Xu, Hang Li, Chaoli Zhang, Joleen Liang, Jiliang Tang, Philip S Yu, and Qingsong Wen. 2024. Large language models for educa- tion: A survey and outlook. arXiv preprint arXiv:2403.18105 (2024), 14 pages. arXiv:2403.18105 [cs.CL] [76]Yixuan Wang. 2025. Becoming a co-designer: the change in participantsâ per- ceived self-efficacy during a co-design process. CoDesign 21, 1 (2025), 52â73. [77] Michel Wermelinger. 2023. Using github copilot to solve simple programming problems. In Proceedings of the 54th ACM Technical Symposium on Computer Science Education V. 1. Association for Computing Machinery, New York, NY, USA, 172â178. [78]Lixiang Yan, Lele Sha, Linxuan Zhao, Yuheng Li, Roberto Martinez-Maldonado, Guanliang Chen, Xinyu Li, Yueqiao Jin, and Dragan GaĹĄeviÄ. 2024. Practical and ethical challenges of large language models in education: A systematic scoping review. British Journal of Educational Technology 55, 1 (2024), 90â112. [79]Aobo Yang, Nan Wang, Renqin Cai, Hongbo Deng, and Hongning Wang. 2022. Comparative explanations of recommendations. In Proceedings of the ACM Web Conference 2022. Association for Computing Machinery, New York, NY, USA, 3113â3123. [80]Shaojun Zhang, Xiangchun He, Yaxin Zhou, Ruishuang Jiang, Yuqi Han, and Xue Guo. 2024. Artificial Intelligence Helps Teachers Personalise Their Teachingâ Take ChatGPT as an Example. In Proceedings of the 2024 7th International Con- ference on Educational Technology Management. Association for Computing Ma- chinery, New York, NY, USA, 200â205. [81]Chengbo Zheng, Kangyu Yuan, Bingcan Guo, Reza Hadi Mogavi, Zhenhui Peng, Shuai Ma, and Xiaojuan Ma. 2024. Charting the future of AI in project-based learning: a Co-design exploration with students. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems. Association for Computing Machinery, New York, NY, USA, 1â19. Behmanush et al. Appendix A: Final Themes and Associated Codes Table 3. Mapping of the final themes to the associated codes derived from the thematic analysis. ThemesCodes GenAI as a learning and employability companionAI as a Necessary Alternative Challenge: Barriers to Independent Learning Challenge: Career Path Uncertainty & Lack of Experience Challenge: Interview & Work-Readiness Anxiety Feature: Career Readiness & Employability Guidance Feature: Soft Skills & Employability Support Feature: Freelancing Skills Support Feature: Tutor Feedback Mechanisms Feature: Motivational Learning Support Solution: Interview Preparation Safety, privacy, and trace controlChallenge: Privacy, Security & Data Ethics Feature: Privacy, Security & Anonymity Feature: Safe Virtual Learning Community Cultural Adaptation: Localized & Sensitive Content Feature: Offline Access to Chat Histories Resource-constrained access and usabilityChallenge: Access & Infrastructure Limitations Challenge: Platform Limitations (WhatsApp) Feature: Offline & Low-Bandwidth Functionality Feature: Low-Resource Device Support Feature: Need for Simple Prompting Feature: Multimedia Support Feature: Accessibility via Transcripts Learning Style: Micro-Learning (Short Sessions) Learning Style: Multimodal Learning (Text, Visual, Audio, Video) Learning quality, scaffolding, and evaluationChallenge: Illusion of Learning (Over-Reliance) Challenge: Inaccurate or Low-Quality Responses Challenge: Programming Difficulties and Educational Disruption Feature: Follow-up Clarifying Questions Feature: Integrated Code Execution & Debugging Feature: Learning Analytics & Progress Tracking Feature: Need for Clarity in Assessment Process Feature: Need for Simplicity & Detail Feature: Adaptation to Learning Style Learning Style: Contextualized Example-Based Learning Solution: Step-by-Step Support Solution: Personalized Learning & Knowledge Checks Use Case: Backend Programming & Design Support Language and cultural fit and contextual relevanceChallenge: Educational Restrictions & Language Barriers Challenge: Inaccurate Translation Challenge: Lack of Context Awareness Feature: Bilingual Support Feature: Age-Appropriate Content Solution: Multilingual & Terminology Support Solution: Native-Language & Personalized Learning Practical productivity and support use casesChallenge: Data Collection Methods Solution: Research & Data Collection Support Use Case: Everyday Tasks (Health, Time Conversion) Use Case: Idea Generation Use Case: Summarization & Project Q&A Designing Safe and Accountable GenAI with Women Banned from Formal Education Appendix B: Demographics of Recruitment Survey Respondents Table 4. Demographic characteristics of the 140 women who completed the recruitment survey, including age range, education level, employment status, and preferred communication language. Survey Question ResponseN Percent Age Range 18â2410172.1% 25â343927.9% Education High School6143.6% Bachelorâs Student1611.4% Bachelorâs Degree5841.4% Other53.6% Employment Status Non-Employed7956.4% Fixed-Term or Intern2820.0% Self-Employed1712.1% Other1611.4% Preferred Communication Language English4834.3% Persian (Dari)8460.0% Other85.7% Appendix C: Employment, Online Learning, and GenAI Figure 4. Demographic surveys overview showing flows between participantsâ education level, employment status, online learning usage, key challenges with online learning, and frequency of GenAI use. Behmanush et al. Appendix D: Storyboard Figure 5. An example storyboard received from PD2_4 based on the scenario she generated in open discussion.