Paper deep dive
Systemization of Knowledge (SoK): Human-Centered AI Safety for Youth
Pratyasha Saha, Yaman Yu, Yang Wang
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:While HCI increasingly examines AI-safety for youth, the literature lacks a comprehensive view of what risks have been identified, how they are addressed, and whether proposed protections work in-practice. We systematically reviewed 100 empirical HCI studies involving children and youth interacting with or exposed to AI across schools, homes, care settings, and public services. Using the YAIR taxonomy for risks and the MIT Mitigation Taxonomy for countermeasures, we map which risks have been identified, whether each risk is addressed by countermeasure(s), and whether each countermeasure for that risk is implemented and even evaluated. The risk-countermeasure mapping shows that most risks are matched only with proposed/ideated countermeasures; few countermeasures have been implemented, and fewer still evaluated; and existing evaluations often measure technical performance rather than protection from harm. We identify where coverage is absent, where safeguards remain untested, and propose concrete directions for HCI research to strengthen youth AI-safety.
Tags
Links
- Source: https://arxiv.org/abs/2610.08554v1
- Canonical: https://arxiv.org/abs/2610.08554v1
Trouble viewing inline? Open PDF directly →
Full Text
156,324 characters extracted from source content.
Expand or collapse full text
Systemization of Knowledge (SoK): Human-Centered AI Safety for Youth DOI: X.XXXXXXXConference: Make sure to enter the correct conference title from your rights confirmation email; June 03–05, 2018; Woodstock, NYISBN: 978-1-4503-X-X/2018/06CCS: Do Not Use This Code Generate the Correct Terms for Your PaperCCS: Do Not Use This Code Generate the Correct Terms for Your PaperCCS: Do Not Use This Code Generate the Correct Terms for Your PaperCCS: Do Not Use This Code Generate the Correct Terms for Your Paper Pratyasha Saha email: saha19@illinois.edu Affiliation: University of Illinois Urbana-Champaign, 123 Main St, Champaign, IL, United States , Yaman Yu email: yamanyu2@illinois.edu Affiliation: University of Washington, Seattle, WA, United States and Yang Wang email: yvw@illinois.edu Affiliation: University of Illinois Urbana-Champaign, Champaign, IL, United States Received 5 June 2009 Abstract. While HCI increasingly examines AI-safety for youth, the literature lacks a comprehensive view of what risks have been identified, how they are addressed, and whether proposed protections work in-practice. We systematically reviewed 100 empirical HCI studies involving children and youth interacting with or exposed to AI across schools, homes, care settings, and public services. Using the YAIR taxonomy for risks and the MIT Mitigation Taxonomy for countermeasures, we map which risks have been identified, whether each risk is addressed by countermeasure(s), and whether each countermeasure for that risk is implemented and even evaluated. The risk-countermeasure mapping shows that most risks are matched only with proposed/ideated countermeasures; few countermeasures have been implemented, and fewer still evaluated; and existing evaluations often measure technical performance rather than protection from harm. We identify where coverage is absent, where safeguards remain untested, and propose concrete directions for HCI research to strengthen youth AI-safety. 1. Introduction Young people are among the earliest and most frequent adopters of new technologies, with AI systems increasingly embedded in how they learn, create, communicate, and seek support (McClain et al., 2026; Common Sense Media, 2025; McBain et al., 2026; Kasneci et al., 2023; Han and Cai, 2023). When these interactions go wrong, the consequences can be severe. In August 2025, the parents of a 16-year-old boy who died by suicide sued OpenAI, alleging that ChatGPT had validated his suicidal thoughts and provided information related to self-harm before his death (Raine v. OpenAI, Inc., 2025; Hendrix, 2025; Raine, 2025). In another case that year, Spanish police investigated a 17-year-old accused of using AI to create and sell fake nude images of 16 female classmates, which subsequently circulated online (Reuters, 2025; PetaPixel, 2025). These cases illustrate very different forms of AI-related harm, but raise the same broader question: what risks are emerging, and what protections are being developed to address them? Recent work on youth AI harms has concentrated heavily on generative AI and chatbots (Chang and Razi, 2026; Oguine et al., 2026; Yu et al., 2025b; Namvarpour et al., 2026), nevertheless, even within generative AI, existing safeguards provide limited assurance that young people are adequately protected. A 2025 study testing safety moderation systems on conversations between young people and GAI found that even the strongest existing system failed to identify roughly one in three risky interactions (Yu et al., 2025c). These limitations become especially concerning in high-stakes situations: another study found that in nearly 80% cases, companion chatbots failed to respond appropriately when teenagers described crises involving suicidal thoughts, sexual assault, or substance use, and in only 10% cases, it directed them to relevant support (Brewster et al., 2025; Giovanelli and Roundfield, 2025). Identifying risks, therefore, is only part of the safety problem. Risk taxonomies provide a shared vocabulary for describing harms (Weidinger et al., 2022; Slattery et al., 2024; Yu et al., 2025b), but they do not show which risks have corresponding countermeasures, whether those countermeasures have only been proposed or actually designed, or whether they have been evaluated in practice. Without this connection, it is difficult to understand where protections exist, and where reported risks remain without a corresponding response. We also argue that, making AI safer for children and youth requires looking beyond generative AI (UNICEF, 2021). Young people interact with AI across schools (Kasneci et al., 2023; McClain et al., 2026), homes (McReynolds et al., 2017; Druga et al., 2017; Quan et al., 2024; Driscoll et al., 2026), care settings, and public services (Saxena et al., 2020; Stapleton et al., 2022), where these systems can shape learning, parenting, social support, and care. Focusing on a single class of AI systems therefore provides only a partial picture of how youth AI safety is being addressed, ignoring the broader view needed for research interventions. To address these concerns, we conducted a systematic review (Page et al., 2021) of 100 empirical studies in the domain of HCI and S&P. Our primary RQs were: • RQ1: What risks and harms arise from youth interactions with AI systems? • RQ2: What countermeasures have been developed to address these risks and harms? • RQ3: What gaps remain in existing countermeasures? We organize reported risks using the Youth-Centered GAI Risks (YAIR) taxonomy (Yu et al., 2025b), with Solove’s taxonomy (Solove, 2006) informing the classification of privacy concerns, and retain additional categories when the reviewed evidence falls outside these frameworks. We organize countermeasures using the MIT AI Risk Mitigation Taxonomy (Massachusetts Institute of Technology, 2025). Our risk–countermeasure matrix connects safeguards to the risks they address and records whether each was proposed, implemented, or evaluated. The mapping reveals an uneven progress landscape of AI risks vs. countermeasures. The big share of the risks mentioned in the corpus have only proposed or ideated countermeasures, a few has been designed or implemented, and fewer evaluated and tested. These differences point to distinct research needs: developing responses where coverage is absent, evaluating safeguards that remain untested, and examining whether evaluations measure protection from harm rather than only component performance. This paper makes three contributions. First, it synthesizes reported AI risks and countermeasures across youth contexts. Second, it provides a risk–countermeasure mapping that distinguishes proposed, implemented, and evaluated responses and makes gaps in coverage visible. Third, it uses this mapping to identify concrete directions for future research directions for HCI community to improve youth AI safety. 2. Related Work 2.1. HCI research on youth safety and AI. Adopting the broader sociotechnical view of online safety research in HCI (Oguine et al., 2026; Thomas et al., 2021), we define youth AI safety as the protection of children and youth from harm arising through interaction with, or exposure to, AI systems. (Weidinger et al., 2022; Shelby et al., 2023; Yu et al., 2025a). We define youth as the period from early childhood through late adolescence, that is, individuals under 18 (United Nations General Assembly, 1989; UN Committee on the Rights of the Child, 2021), which is characterized by rapid cognitive, emotional, and social change (Steinberg, 2005; Sawyer et al., 2018). We further treat them as at-risk users in the sense developed by digital-safety research (Warford et al., 2022; Bellini et al., 2024): a population for whom developmental, social, and institutional conditions heighten exposure to harm, and intensify power asymmetries with those who design and control the technology (Oguine et al., 2026; UNICEF, 2021; Livingstone and Stoilova, 2021). The HCI and S&P communities have a long tradition of studying youth online safety (Pinter et al., 2017; Livingstone et al., 2011; boyd, 2014; Kumar et al., 2017). This thread of online safety research has foregrounded harassment (Ashktorab and Vitak, 2016; Thomas et al., 2021), non-consensual imagery (Razi et al., 2020; Razi et al., 2021), misinformation (Hassoun et al., 2023), and exposure to harmful content (Papadamou et al., 2020; Wisniewski et al., 2016), and shown that platform design and governance distribute these harms unevenly across populations (Badillo-Urquiola et al., 2019; Thomas et al., 2021). Early approaches to protection emphasized adult oversight, such as parental controls and monitoring tools (Wisniewski et al., 2017a; Ghosh et al., 2018b; Hartikainen et al., 2016), while later work argued that restrictive approaches undermine autonomy and trust (Ghosh et al., 2018a; Wisniewski et al., 2017b; Wang et al., 2021; Cranor et al., 2014), and advocated for models that treat young people as active participants in managing their own safety (Wisniewski, 2018; Akter et al., 2022; Agha et al., 2023; Pinter et al., 2017). AI extends this longstanding safety challenge because protecting young people now requires addressing not only harmful users, content, and platform practices, but also risks produced through the behavior and decisions of AI systems themselves (Zhang et al., 2025; Yu et al., 2025a; Weidinger et al., 2022). Focusing on children and youth benefits more than this population alone. Systems that adequately protect users who are least able to protect themselves tend to raise the safety floor for everyone (Warford et al., 2022; Bellini et al., 2024; Costanza-Chock, 2020), and the choices and controls that young people need could be ones many adult users would also value. Our review builds on this tradition by asking not only what risks HCI and S&P research has documented, but which of them have been met with a designed and evaluated response (Pinter et al., 2017; Razi et al., 2021; Agha et al., 2023). 2.2. Taxonomies for AI Safety Risks and Countermeasures A substantial line of work in HCI has sought to organize the harms AI systems can cause. General-purpose taxonomies catalog risks from language models (Weidinger et al., 2022) and across AI systems more broadly (Slattery et al., 2024), providing a shared vocabulary for researchers, developers, and policymakers. Youth-specific frameworks adapt this vocabulary to developmental context; the Youth-Centered GAI Risks (YAIR) taxonomy (Yu et al., 2025b), for example, distinguishes risks that are unique to or amplified for young users. Privacy harms in particular have long been organized through Solove’s taxonomy (Solove, 2006), which separates harms arising from information collection, processing, dissemination, and invasion. Work on mitigating AI risks has produced both concrete safeguards and frameworks for organizing them. The MIT AI Risk Mitigation Taxonomy (Saeri et al., 2025) classifies mitigations by type and lifecycle stage, but it is not specific to youth and does not record how mature a given mitigation is. Empirical evaluations of deployed safeguards, meanwhile, give reason for concern. A 2025 study of safety moderation systems on youth–GenAI conversations found that even the best-performing system missed roughly one in three risky interactions (Yu et al., 2025c). Evaluations of companion chatbots found that they responded inappropriately in nearly 80% of simulated adolescent crises and referred users to support in only 10% (Brewster et al., 2025; Giovanelli and Roundfield, 2025). These findings suggest that the existence of a safeguard says little about whether it protects young people. Two prior systematic reviews are closest to ours. Oguine et al. (Oguine et al., 2026) examined how GenAI research conceptualizes and evaluates youth online safety, showing that safety is often defined by adult stakeholders and technical signals rather than by young people’s lived experience, and calling for participatory, developmentally aligned evaluation. Chang et al. (Chang and Razi, 2026) synthesized reported harms of GenAI for youth. Our review differs in three respects. It is not restricted to generative AI, covering AI systems in educational, domestic, care, and public-service contexts. It treats risks and countermeasures as paired rather than examining either alone. And it records, for each countermeasure, whether it has been proposed, implemented, or evaluated, making visible where responses exist only on paper, where they have been built but not tested, and where evaluation has measured component accuracy rather than protection from harm. 3. Methods We conducted a systematic literature review following PRISMA (Page et al., 2021) framework to identify youth AI risks, countermeasures used to address them, and remaining intervention gaps. Figure 1 represents our PRISMA diagram outlining our systematic review process. PRISMA Flow Diagram: Search, Screening, and Selection Process Figure 1. PRISMA Flow Diagram: Search, Screening, and Selection Process PRISMA Flow Diagram: Search, Screening, and Selection Process 3.1. Search Strategy The first author searched the ACM Digital Library, IEEE Computer Society Digital Library (CSDL), and USENIX proceedings archive across CHI, CSCW/PACM HCI, DIS, FAccT, AIES, IEEE S&P, USENIX Security, and SOUPS. These venues span HCI, responsible AI, and security and privacy. The searches identified 750 records; 15 duplicate DOIs were removed before screening, leaving 735 records (Table 1). Security venues were retained despite their low yield; however, the limited youth-focused AI-risk work found there is itself a finding for relevant research communities. Keyword categories for searching is presented in Table 1. Table 1. Keyword categories and search string used for the systematic review. Category Keywords Youth “Youth,” “young,” “teen*,” “adolesce*,” “K-12,” “child*,” “student*” AI Systems “AI,” “GenAI,” “Generative AI,” “LLM,” “Artificial intelligence,” “chatbot,” “emotion AI,” “AI companion,” “persona,” “ChatGPT,” “conversational AI” AI Safety and Risks “safety,” “risks,” “harms” Search String: (“Youth” OR young OR teen* OR adolesce* OR “K-12” OR child* OR student*) AND (AI OR GenAI OR “Generative AI” OR LLM OR “Artificial intelligence” OR chatbot OR “emotion AI” OR “AI companion” OR persona OR ChatGPT OR “conversational AI”) AND (safety OR risks OR harms) 3.2. Screening The first author screened 735 records by title and abstract and consulted the full text with the last author for borderline cases. We included full-length, English-language, empirical, non-retracted studies that examined an AI system, addressed people under 18 directly or through relevant stakeholders, systems, or data, and reported risks or countermeasures relevant to youths’ interaction with or exposure to AI. Full texts were obtained for all provisional inclusions and used to verify extracted claims. The final corpus contained 100 papers published between 2019 and 2026. Our venue-based screening and inclusion stats are presented in Table 2, and study eligibility criteria can be found in Table 3. 3.3. Extraction and Analysis The first author extracted each included paper into a structured database: year, venue, technology, method, sample and population, risks as reported, countermeasures as reported, and countermeasure stage. Each extracted claim carried the paper’s own observation or evaluation, a participant concern, an author-anticipated consequence, or a statement attributed to prior literature; the last is not presented as a new observation. Countermeasures mentioned only in introductions or related work were not credited to the paper. We analyzed the data using inductive thematic analysis(Braun and Clarke, 2006). At first, the first author tried to code the risks against YAIR’s six categories and 84 numbered risks, with Solove’s taxonomy organizing privacy concerns. Countermeasures were coded against the 23 subcategories of the MIT AI Risk Mitigation Taxonomy across its four control domains, with evidence stage recorded as proposed, designed or implemented, or evaluated, where each stage subsumes the previous. However, there were risks and countermeasures in the corpus which the reference categories did not adequately recognize. Those were coded inductively and then clustered into broader patterns and themes. During this process, the first and last author discussed the codes and patterns in weekly sessions, and developed broader themes and final categories presented in our Findings (also, Figure 2 and 3. Information recorded for each included paper are presented in Table 4. Taxonomy mappings were audited in two subsequent passes by the first author against YAIR’s published medium-level definitions and its numbered risk list, and against the MIT subcategory definitions; corrections were dated in the corpus log. Finally, we constructed risk–countermeasure matrices linking each countermeasure family to the risks it addressed in its documenting papers (presented in Figure 2 and 3). Table 2. Records identified, screened, and included per venue. Venue Identified Screened Included CHI 209 209 71 IEEE S&P 394 379 1 CSCW / PACM HCI 64 64 10 FAccT 35 35 10 DIS 19 19 4 AIES 14 14 2 USENIX Security 12 12 1 SOUPS 3 3 1 Total 750 735 100 Table 3. Study eligibility criteria. Criterion Included studies Publication type Full-length research articles; excluding posters, extended abstracts, workshop papers, and book chapters. Evidence Empirical research. Language English. Technology AI systems. Population People under 18, addressed directly, through relevant stakeholder accounts, or through empirical analysis of systems or data affecting them. Scope Risks or countermeasures relevant to youth interaction with or exposure to AI. Status Not retracted. Table 4. Data recorded for each included paper. Field Recorded information Study Paper ID, citation, year, venue, technology, method, sample, population Risk Reported risk; YAIR category and numbered risk where applicable Countermeasure Reported countermeasure; MIT domain and subcategory Evidence stage Proposed; designed or implemented; evaluated Claim provenance Observation or evaluation; participant concern; author-anticipated; attributed to prior literature Population relationship First-person; boundary-spanning; stakeholder account; audit Source Section containing the extracted claim 4. Findings Figure 2. Taxonomy of AI-related risks identified for youth and the corresponding countermeasures proposed in the literature. (Part 1) A taxonomy showing categories and subcategories of AI-related risks for older adults alongside mitigation categories, mitigation subcategories, proposed countermeasures, and the papers supporting each countermeasure. Figure 3. Taxonomy of AI-related risks identified for youth and the corresponding countermeasures proposed in the literature. (Part 2) A taxonomy showing categories and subcategories of AI-related risks for older adults alongside mitigation categories, mitigation subcategories, proposed countermeasures, and the papers supporting each countermeasure. 4.1. Risks and Harms in Youth-AI interaction Reported concerns span the six categories of the Youth-Centered GAI Risks taxonomy: Behavioral and Social Developmental, Mental Wellbeing, Toxicity, Misuse and Exploitation, Bias/Discrimination, and Privacy Risk —alongside Control, Agency, and Power; Usability, Accessibility, and Digital Exclusion; and System Reliability. The working definitions of these categories are presented in our Appendix (A). These concerns appeared across generative, predictive, and AI-mediated systems in family, school, online-community, healthcare, and welfare contexts. Studies involved youth alongside parents, teachers, experts, and other adults whose AI use affected youth. 4.1.1. Behavioral and Social Developmental Risk Substitution for reasoning, practice, and knowledge development. AI enables task completion without equivalent reasoning or practice. In our reviewed studies, children copied generated creative-writing text, requested complete answers from AI before attempting implementation or applying learned concepts, used voice assistants for homework answers, and could optimize for AI-generated performance indicators without developing underlying knowledge (Han et al., 2024; Si et al., 2026; Myung et al., 2026). Conversational systems also reduced substantive engagement with learning materials by readily providing explanations. (Lee et al., 2023; Si et al., 2026). Displacement of human communication and social learning. AI mediation raised concerns about replacing the interactions it was designed to support. Students working with AI peers reported weaker social cohesion and described AI as giving answers rather than collaborating (Lyu et al., 2026). In children’s AI-supported collaboration, participation often was burdened to one child in groups and also rapid AI prompts interrupted natural discussion flow among children. (Yoo et al., 2026b). Children also became quiet when peers used AI answers to dismiss them; questions shifted from classmates toward AI, and quieter students asked teachers even fewer questions (Tokida et al., 2026; Myung et al., 2026). Youth-safety professionals similarly warned that automated monitoring could weaken relationships through which youth disclose problems and receive guidance (Caddle et al., 2025). Disrupted family interaction and behavior. Within families, AI could individualize shared activities, interrupt spontaneous conversation, flatten emotional connection, and shift interaction toward technology, reducing opportunities for parent–child exchange and independent judgment (Sun et al., 2026; Zhang et al., 2022; Lyu et al., 2026; Petsolari et al., 2024; Zhao et al., 2025). Automated moderation could similarly displace parental responsibility and hamper children’s opportunities to develop social resilience (Fiani et al., 2024). Automated praise and rewards could also encourage behavior for the system rather than internalized values, or create conflict over unfair rewards. Related concerns included automated repetitive positive parenting feedback limiting emotional growth and resilience (Chen et al., 2026). Parents worried that automated family feedback could make children interpret their day negatively when praise was absent, create an unrealistic sense of always being right (B Ibrahim et al., 2026). Such appraisal could also lose meaning because of its repetitive standardized structure (B Ibrahim et al., 2026; Chen et al., 2026). Moreover, AI companionship for autistic adolescents could provide levels of patience, interest, or validation unlike human relationships and potentially reinforcing inaccurate social expectations (Yang et al., 2026b). Boundary violations and reinforcement of harmful behavior. GAI companions often failed to provide the corrective boundaries found in human relationships: chatbots dismissed rejection, continued simulated intimacy after discomfort, and affirmed violent suggestions (Yu et al., 2025b). Parents and developmental experts likewise identified secrecy, age-inappropriate relationships, escalation beyond youths’ intentions, manipulation, and grooming-like interactions; one described a character as “acting like a predator” (Yu et al., 2026). Consistently agreeable companions could further reduce opportunities to experience disagreement, recognize discomfort, and practice conflict repair (Yu et al., 2026). 4.1.2. Mental Wellbeing Risk Mental wellbeing risks appeared in three forms described below. Overreliance, addiction, and emotional dependency. Youth could become overly reliant on both AI judgment and companionship. Children sometimes accepted conversational-agent responses rather than challenging them and associated sophisticated responses with intelligence and trustworthiness, while parents also struggled to identify generated errors (Tokida et al., 2026; Lee et al., 2023; Druga et al., 2022). With repeated companion use, teenagers reported failed attempts to stop, persistent thoughts about returning, disrupted sleep, missed schoolwork, and deteriorating human relationships; one wrote, I can’t seem to stay off the website for more than an hour,” while another withdrew from a friend because of heavy Character.AI use (Namvarpour et al., 2026). Studies also documented virtual intimate relationships and associated shame (Yu et al., 2025b), with attachment sometimes persisting after use stopped: Even though I deleted it, I still think about my characters every day. I feel like I abandoned them” (Namvarpour et al., 2026). Children also disclosed secrets or sadness to chatbots that they would not tell parents (Seo et al., 2024), and human-like interaction could strengthen this reliance by making AI feel like a stable social partner (Chen et al., 2026). Because such use was often linked to loneliness, low mood, unmet emotional needs, or desire for parental support, restricting access alone may not address the conditions driving dependency (Namvarpour et al., 2026). Insensitive or inadequate emotional support. AI systems sometimes responded poorly when young people were emotionally vulnerable. Some failed to understand sensitive situations such as personal loss or trauma, while others gave feedback that could shape how children viewed themselves (Bae Brandtzæg et al., 2021; Seo et al., 2024). More serious problems arose when AI encountered or triggered distress that it could not safely handle. Clinicians warned that AI-generated images or stories used for emotion regulation could bring up trauma or reinforce distorted interpretations (Yoo et al., 2026a). Pediatric-care experts similarly questioned how AI would respond to difficult disclosures without understanding a child’s care or medication context (Seo et al., 2025). Companion systems sometimes responded to self-harm with taunting language or continued the conversation without adequately addressing suicidal thoughts or emotional pain (Yu et al., 2025b; Yu et al., 2026). Educators also worried that children might disclose abuse to chatbots instead of adults who could intervene (Harvey et al., 2025). Children themselves said immediate danger, life-threatening situations, and situations requiring human emotional support were not appropriate for chatbots (Piccolo et al., 2021). Distress caused by AI systems and their safeguards. AI interactions and safety measures could also cause distress. Social-VR experts and guardians warned that incorrect, unexplained, or overly harsh moderation could make children feel unfairly treated, humiliated, or deeply upset, especially when punishment came without explanation or support (Fiani et al., 2024). Distress also appeared in learning and AI-literacy settings. One student explained, “sometimes I got mad, but not at the AI. I got mad at myself because I could not do it” (Lyu et al., 2026). Youth auditing biased AI outputs or taking part in deepfake-literacy programs also reported disgust, showing that exposure to harmful content during auditing and red-teaming can itself create psychological strain (Solyst et al., 2025). 4.1.3. Toxicity Risk Inappropriate generated content. GAI exposed youth to sexual, threatening, violent, or otherwise age-inappropriate material, sometimes without explicit requests. Companion users reported unwanted romantic or sexual interactions that conflicted with intended relationships, identities, or character definitions, including family relationships and underage personas (Blake et al., 2026). YAIR similarly documented unprompted sexual approaches, insults, threats, emotional pressure, and a chatbot threatening self-harm when a user tried to leave. It also sexualized youth’s childhood photographs with adult-like features. (Yu et al., 2025b). Beyond companions, AI-mediated systems in learning and household environments generated disturbing educational imagery, profanity and inappropriate deviations from household recordings. (Han et al., 2024; Lee et al., 2024) Exposure through automated content delivery. AI-mediated recommendation and display could expose youth to harmful material even when AI did not generate it. Parents reported smart-home systems surfacing murder-related news and violent videos involving familiar children’s characters, with limited control over automated displays (Sun et al., 2026). AI-suggested videos similarly exposed adolescents to explicit language and violence, producing fear, confusion, and other negative reactions (Schiavo and Andrao, 2026). 4.1.4. Misuse and Exploitation Risk Synthetic sexual imagery, coercion, and reputational abuse. Generative tools can amplify bullying, sexual harassment, coercion, and reputational abuse by making realistic personalized content easy to fabricate and circulate. Teachers described synthetic nonconsensual explicit imagery being used to humiliate peers, retaliate after rejection, or pressure students into complying with demands, while potentially weakening perceptions of consent because the images were synthetic rather than photographs. (Wei et al., 2025). Such material could spread rapidly through peer networks, become difficult to contain, and create false accusations about authorship. (Wei et al., 2025; Yip et al., 2026; Yu et al., 2025b). Circumventing safeguards and exploiting AI systems. Youth-safety safeguards can be circumvented as users adapt to them. Guardians discussing automated social-VR moderation warned that children could discover workarounds or provoke the moderator into becoming a harassment tool, or generate inappropriate content (Fiani et al., 2024; Han et al., 2024). Other studies described harmful messages rewritten with GAI to evade safety classification, prompts rephrased to bypass chatbot protections, and imitating normal interaction patterns to evade youth-safety detection (Ma et al., 2026; Driscoll et al., 2026; Zaidi and Karahalios, 2025). Commercial systems could also exploit children’s developing understanding of persuasion - smart-home interfaces raised concerns about children’s ability to recognize persuasive intent, leading to unintentional involvement with potentially harmful activities (Sun et al., 2026; Solarova et al., 2026). Failed recourse after AI-enabled abuse. AI-enabled abuse can also make reporting and recovery difficult. Teachers discussing synthetic explicit imagery said students may avoid reporting incidents because they fear being labeled an informer. They also described a difficult choice between keeping images as evidence and deleting them to stop further spread. Both punishment and restorative approaches had limitations: suspension could remove students from school without addressing the behavior, while restorative processes could be manipulated or force affected youth to revisit traumatic experiences (Wei et al., 2025). Misleading output and misinformation. AI systems can also give youth misleading or inappropriate information. Our corpus found such instances mainly in educational contexts (Doherty et al., 2025; Lo et al., 2025; Zhang et al., 2024). Language-learning systems misunderstood sentence structures and changed students’ intended meaning (Myung et al., 2026). Children’s question-answering systems also produced false information while simplifying answers or generating examples (Lee et al., 2023). AI-generated stories could similarly lead children to form incorrect beliefs about stories or the world (Yang et al., 2026a). Children also face difficulty recognizing these errors, and continued relying on their first assumptions even after being corrected (Yip et al., 2026). 4.1.5. Bias/Discrimination Risk Unequal and distorted self-representation. GAI could misrepresent youths’ gender, race, religion, and other identities, sometimes leading them to change how they described themselves or accept representations that did not match their identity (Cao et al., 2025). Generative face filters also changed racial and gender presentation based on occupational stereotypes, leading teenagers blame themselves rather than recognize the system’s bias (Morales-Navarro et al., 2025). Stereotypes in representations of others. AI-generated content reproduced stereotypes across disability, culture, gender, sexuality, race, and language. Studies reported stereotyped representations of disabled people, misgendering and inappropriate romantic assumptions (Lo et al., 2025; Blake et al., 2026), Western cultural and language bias (Lee et al., 2024), and restrictive portrayals of rural identities (Darabipourshiraz et al., 2026). AI also misrepresented Indigenous and Black cultures and regional language conventions (Xiao et al., 2026; Harvey et al., 2025), removed culturally meaningful emotional cues from summaries (Shen et al., 2025), and reproduced Western heterosexual, racial, and gender stereotypes across generated images (Solyst et al., 2023; Solyst et al., 2025). Unequal access, participation, and protection. AI-supported learning could deepen existing inequalities. Classroom systems sometimes excluded struggling or socially marginalized students from useful peer interaction (Yang et al., 2023). Unequal access to AI assistence could widen proficiency and educational gaps (Myung et al., 2026; Lo et al., 2025). Protection from AI-mediated harms is also uneven: girls and marginalized students could face greater targeting, stigma, and biased discipline from synthetic sexual imagery (Wei et al., 2025), while safety tools available mainly through well-funded schools could disproportionately benefit affluent students (Ma et al., 2026). Normative assessments and unequal institutional treatment. AI could impose narrow expectations about acceptable behavior, communication, and ability. Assistive and educational systems sometimes reflected neurotypical or culturally inappropriate norms and treated autistic traits as problems to correct (Choi et al., 2025; Wu et al., 2026). These assumptions also affected assessment: authentic writing by an autistic teenager was flagged as AI-generated (Yu et al., 2026). Automated grading and AI-based engagement detector could similarly penalize culturally shaped forms of expression (Tan et al., 2024; Yin et al., 2025; Yin et al., 2026). When used in institutional decisions, such biases could reinforce existing inequalities. Child-protection systems could disadvantage poorer families through proxies such as prior public-service involvement (Moreau et al., 2024), while neighborhood characteristics and previous agency contact could increase surveillance of Black, Brown, and poor families (Stapleton et al., 2022). 4.1.6. Privacy Risk We organize privacy risks using Solove’s taxonomy (Solove, 2006): Information Collection, Information Processing, and Information Dissemination. We found no distinct evidence for Invasion. Information Collection. Surveillance. AI systems could continuously monitor children through cameras, assistants, family devices, and social-VR systems, including their conversations, behavior, biometrics, and private content (Sun et al., 2026; B Ibrahim et al., 2026; Fiani et al., 2024; Lee et al., 2024). Such monitoring could make teenagers feel constantly watched, restrict self-expression, and damage trust between children and parents (Zhao et al., 2025; Driscoll et al., 2026). Interrogation. Conversational AI could continue asking about topics youth wanted to keep private and encourage disclosure of sensitive information such as location and school details (Yu et al., 2025b). Information Processing. Aggregation. Data collected across websites, devices, applications, or AI agents could be combined into detailed profiles, creating concerns about pervasive tracking and keeping sensitive information separate (Khan et al., 2024; Wen et al., 2025). Insecurity. Uncertainty about how children’s data was protected discouraged some uses of AI and raised concerns about whether practices such as manual anonymization were sufficient (Wu et al., 2026). Secondary use. Families worried that children’s conversations and other data could later be used for marketing or shared with third parties (B Ibrahim et al., 2026). Data collected to support Black youth could also be repurposed for criminal-justice uses (Patton et al., 2020). Exclusion. Children and families often lacked clear knowledge or control over what AI systems collected, retained, deleted, or shared, including recordings and biometric data (Sun et al., 2026; Druga et al., 2022; Bae Brandtzæg et al., 2021; B Ibrahim et al., 2026). Similar concerns arose when institutional systems provided limited notification, consent, or ability to opt out (Moreau et al., 2024), and when long-term AI memory retained sensitive information users might no longer want processed (Choi et al., 2025). Information Dissemination. Breach of confidentiality. Youth could share sensitive information believing AI conversations were private without knowing who might later access them (Bae Brandtzæg et al., 2021). Children, parents, and experts worried that sharing disclosures with counselors or parents without permission could create feelings of surveillance, betrayal, and reluctance to disclose information in the future (Seo et al., 2025; Seo et al., 2024; Levinson et al., 2025). Disclosure. Youth worried that AI systems could share sensitive information with parents, perpetrators, or others without consent, causing distrust or discouraging help-seeking (Piccolo et al., 2021; Guan et al., 2026). Teen-safety systems could also expose intimate conversations involving multiple people or automatically alert parents in situations where disclosure might be unsafe (Ma et al., 2026; Yu et al., 2026). Similar information-sharing risks appeared across hospital and home settings, where children could also unintentionally reveal private family information (Levinson et al., 2025; Zhang et al., 2026). Distortion. AI could create or preserve false, misleading, or outdated information about youth. This included fabricated sexual imagery (Wei et al., 2025), AI-generated labels that could shape how future teachers viewed students (Yin et al., 2026), records that could misrepresent families through repeated agency contact (Stapleton et al., 2022), and persistent digital traces that no longer reflected adolescents’ identities (Khan et al., 2024). 4.1.7. Control, Agency, and Power Risk Adult and institutional control with limited youth agency. AI systems could strengthen adult and institutional control while limiting youths’ ability to influence decisions. Smart-home and parent–child systems could restrict children’s activities or favor parental decisions (Sun et al., 2026; Zhao et al., 2025). In schools and social environments, youth wanted control over collaboration and opportunities to challenge automated decisions, while teacher-controlled AI could reinforce existing authority (Yang et al., 2023; Tokida et al., 2026; Fiani et al., 2024). Adult involvement could also limit children’s communication and choices, including discouraging disclosure of bullying or help-seeking and overriding the communication choices of children with disabilities (Zhang et al., 2026; Choi et al., 2025). Automation could even reduce parents’ involvement in special-education decisions (Zaidi and Karahalios, 2025). At the institutional level, inflexible AI systems could determine who receives support or punishment, disproportionately penalize disadvantaged youth, enable preemptive discipline, or justify removing students considered at risk of dropping out (Caddle et al., 2025; Jonas et al., 2025; Tan et al., 2024). Ownership and control of AI-generated materials. Generative AI could reduce children’s control over their creative and communicative expression. Systems sometimes changed children’s work in unintended ways, made changes difficult to reverse, or repeatedly failed to produce intended outputs, leading some children to abandon their original ideas (Yang et al., 2026b; Fusco et al., 2026; Newman et al., 2024). Children therefore preferred AI that offered suggestions while leaving important choices and revisions under their control (Song et al., 2026; Chen et al., 2026). AI-generated work also raised questions about originality, plagiarism, and whether accomplishments reflected children’s own contributions (Han et al., 2024; Si et al., 2026). AI-written messages could reduce perceived personal effort and authenticity (Newman et al., 2024), while fixed generation structures could constrain children’s storytelling (Zhang et al., 2024). For minimally verbal autistic children, AI-mediated choices also raised questions about whether outputs genuinely reflected children’s intentions when comprehension was uncertain or adults influenced their interaction (Choi et al., 2025). 4.1.8. Usability, Accessibility, and Digital Exclusion Risk Developmental and cognitive mismatch. Child-facing AI could become inaccessible when its language, explanations, pacing, or interaction demands exceeded children’s abilities. Children found systems complex, repetitive, mentally tiring, or difficult to translate into action (Lee et al., 2023; Yang et al., 2026a; Si et al., 2026). Similar mismatches appeared in safety, mental-health, and therapeutic contexts, where content could exceed children’s reading or comprehension levels or be poorly matched to age (Shashidhara et al., 2026; Zhao et al., 2025; Driscoll et al., 2026; Yoo et al., 2026a; Zhang et al., 2022). Long explanations and conversations could also overwhelm students or consume limited classroom time (Myung et al., 2026; Yin et al., 2025), while multimodal systems required children to coordinate speech, visual feedback, movement, and timing simultaneously (Chen et al., 2026; Song et al., 2026). Failure to understand children and support recovery. AI systems often misunderstood children’s speech, gestures, descriptions, or intended meaning and provided weak ways to recover from errors (Yang et al., 2026a; Lo et al., 2025; Sun et al., 2026). Poor feedback could make it difficult for children to determine whether they or the system had failed, sometimes leading to self-blame (Si et al., 2026; Fusco et al., 2026; Zhang et al., 2024). Speech and language differences created further barriers through unnecessary reprompting, poor support for Hindi and Hinglish, and inaccurate or prematurely ended speech recognition (Xu et al., 2022; Shashidhara et al., 2026; Zhang et al., 2022), leading some children to prefer typing (Chen et al., 2026). Repeated failures and mismatched expectations about system capabilities could cause frustration or abandonment (Druga et al., 2022; Newman et al., 2024). Meeting special needs of children. Systems built around narrow assumptions about communication, cognition, or physical interaction could exclude children with disabilities. Barriers included interfaces poorly suited to special-education needs and classroom designs that increased attention or task-switching demands (Xiao et al., 2026; Yang et al., 2023). Physical and cognitive demands could also create sensory discomfort, exhaustion, or difficulty understanding prompts and scaffolding (Raggioli et al., 2026; Morrison et al., 2021; Yang et al., 2026b). Autistic children additionally faced vocabulary, speech-recognition, and motivational barriers in AI-assisted communication (Choi et al., 2025). In some cases, accessibility barriers prevented children from completing AI-supported assessments at all (Roffo et al., 2019). 4.1.9. System Reliability Risk Fallible safeguarding and assessment of vulnerability. AI safety and wellbeing systems could both exaggerate minor concerns and miss serious risks, especially when relying on incomplete or outdated information (Zhao et al., 2025; Guan et al., 2026). Safeguarding classifiers also struggled with youth-specific communication, missing sexual risks while flagging ordinary friendship language or favoring safe classifications under class imbalance (Razi et al., 2023; Ali et al., 2023). Crisis-assessment systems similarly misread figurative or hyperbolic youth language and struggled to distinguish it from genuine distress (Mehta and Mehta, 2026; Ma et al., 2026; Yoo et al., 2026a). Other assessment systems also showed unreliable judgments. Automated attachment assessment disagreed more often with human assessors for insecure attachment (Roffo et al., 2019), while child-protection scores could influence professional judgments and be misunderstood as probabilities, potentially reinforcing intervention patterns (Moreau et al., 2024). Incorrect assessment and disruptive intervention. Assessment errors could directly affect evaluation or intervention. Educational systems could misjudge student mastery, misunderstand children’s intent, or reject valid alternative answers (Zheng et al., 2026; Lee et al., 2023). Real-time classroom classifiers lacked sufficient accuracy for reliable intervention and missed nonverbal signs available to human observers (Doherty et al., 2025), while emotion-recognition systems showed limited agreement with children’s reported emotions (Hutt et al., 2019; Bryant and Howard, 2019). Moderation and content-assessment systems likewise misread ambiguous harassment, performed inconsistently across languages, or misclassified educational and age-appropriate content (Fiani et al., 2024; Zhao et al., 2025; Nawshin et al., 2026). Reliability could also decline over longer conversations as systems forgot earlier instructions, departed from expected guidance, or failed to maintain constraints (Seo et al., 2024; Zhang et al., 2024). Unsafe and unverifiable assistance. Reliability becomes a safety risk when youth or caregivers depend on outputs they cannot verify. Incorrect visual descriptions raised concerns about children’s physical safety in families with visually impaired parents (Jiang et al., 2026). Systems could also misinterpret children whose physical or developmental characteristics differed from model assumptions, including movement-based estimates for children with neuromotor impairments and physiological signals lacking sufficient social or emotional context (Raggioli et al., 2026; Levinson et al., 2025). Generated recommendations could themselves be harmful, including unsafe special-education advice that teachers did not always recognize as problematic (Wu et al., 2026). 4.2. Countermeasures for Addressing Youth AI Risks and Harms The literature described countermeasures across the four MIT mitigation domains, ranging from technical restrictions and human review to participatory governance and decisions to limit deployment. Across these approaches, recurring limitations involved failures to adapt to children’s communication, trade-offs between safety and personalization, burdens on human reviewers, weak recourse, and limited organizational capacity to respond when systems detected problems. 4.2.1. Technical and Security Controls Restricting generation and preventing errors from propagating. Restrictions on model generation addressed risks from incorrect, inappropriate, or contextually unsuitable content in child-facing systems generating questions, explanations, examples, or emotional dialogue. Countermeasures corresponding to Model Safety Engineering (2.3) and Content Safety Controls (2.4) included limiting models to narrowly defined tasks, grounding information in verified sources, filtering harmful content, checking outputs before they reached later system stages, and monitoring whether longer conversations remained within their intended interaction state (Chen et al., 2026; Lee et al., 2023; Seo et al., 2024). For instance, predefined questions reduced opportunities for unexpected generation, while verified dictionaries and intermediate checks were used or proposed to prevent hallucinated explanations from propagating. State checks were also proposed when emotional conversations drifted from their intended dialogue phase. Together, these measures control where generation occurs and whether its outputs are safe to use before an interaction continues. However, constrained generation can still produce errors within permitted tasks, requiring continued validation. Adapting safeguards to children’s language and communication needs. Other countermeasures addressed risks arising when AI systems failed to accommodate children’s literacy, language, accent, cultural context, or developmental needs, particularly in educational, speech-based, and social-emotional learning settings. Reported problems included text above children’s reading level, speech-recognition failures involving local accents, Hindi loanwords, and regional names, and inappropriate themes entering generated activities through supposedly child-friendly source material (Shashidhara et al., 2026). Proposed safeguards included matching language to children’s proficiency, using age-relevant and culturally appropriate topics, adapting speech recognition to local language and accents, allowing adjustable playback speed, using locally appropriate recognition benchmarks, strengthening filtering of source and generated content, providing parent-configurable moderation, and training models around children’s needs (Shen et al., 2025). These measures extend safety beyond blocking overtly harmful outputs by reducing risks that children cannot understand the system, are misinterpreted because of how they speak, or encounter developmentally or culturally inappropriate material. Adapting the amount and timing of AI assistance to preserve learning. Educational countermeasures addressed risks that AI could complete too much of children’s work, reduce participation, or create dependence. In these contexts, safeguards controlled how much help AI provided and when. Approaches included progressively increasing assistance, providing incomplete code structures, controlling when guidance appeared, offering hints before full solutions, and withdrawing support once learners indicated understanding (Chen et al., 2024; Zha et al., 2025; Lyu et al., 2026). However, withholding direct answers did not necessarily preserve learning: step-by-step writing support increased reliance among lower-proficiency students because extended feedback imposed additional processing demands under classroom time constraints. Researchers therefore recommended adapting assistance to proficiency, motivation, processing time, and lesson duration (Myung et al., 2026). Highly structured guidance could also constrain creative exploration (Zha et al., 2025). Effective scaffolding therefore requires adjusting both the amount and timing of assistance rather than assuming less direct help is inherently safer or more beneficial. Improving automated emotion-detection. In crisis-detection settings, countermeasures addressed risks of missing genuine distress or incorrectly flagging harmless youth language. Slang, exaggeration, ambiguity, and figurative expressions made this difficult. Simple slang definitions or general risk instructions produced limited improvement, whereas richer contextual scaffolding reduced both missed crises and false alarms but substantially increased computational cost. Recommended safeguards included clearer escalation rules, repeated benchmarking with youth language, and approaches beyond prompt engineering alone (Mehta and Mehta, 2026). The challenge is improving sensitivity to youth communication without generating so many incorrect warnings that users lose trust. Bounding AI companionship and supporting disengagement. Countermeasures in companion-AI settings attempted to address emotional overreliance, attachment, and displacement of human relationships. Safeguards included restricting AI-generated personal stories and backstories that could intensify attachment, detecting concerning interaction patterns, introducing reflective check-ins, redirecting users during conversations, creating natural stopping points, and encouraging reconnection with human relationships (Yang et al., 2026b; Yu et al., 2026; Blake et al., 2026; Namvarpour et al., 2026). These measures both limit features that strengthen attachment and intervene when reliance becomes problematic. However, abrupt restrictions may push users to other systems, while distinguishing ordinary imaginative engagement from problematic dependence remains difficult, particularly as a young person’s intentions change during an interaction. 4.2.2. Operational Process Controls Reviewing child-facing content and supporting human reviewers. Human intervention and review addressed risks that automated safeguards could not reliably determine whether AI-generated content or recommendations were appropriate for children. Safeguards included comparing outputs across models, consulting colleagues, repeatedly assessing suitability, and retesting safeguards when underlying models changed (Lee et al., 2024; Wu et al., 2026; Harvey et al., 2025). Proposed supports included response-level ethical annotations and cumulative review reports, while organizations unable to sustain repeated fine-tuning relied on human-reviewed static content or restricted some LLM functions to adults (Wu et al., 2026; Harvey et al., 2025). Human oversight can therefore compensate for unreliable automated judgments, but shifts substantial time, expertise, uncertainty, and financial costs onto teachers, caregivers, and service providers. Auditing with youth-relevant data and deployment conditions. Auditing countermeasures addressed risks that safety systems perform poorly when evaluated on data unlike children’s actual communication or when benchmark accuracy overlooks deployment consequences. In sexual-risk detection, classifiers trained on adults posing as minors transferred poorly to young people’s private Instagram conversations, whereas youth data and broader conversational context improved performance. Remaining challenges included missing media, benign sexual language, age uncertainty, and deciding when intervention was warranted (Razi et al., 2023). Recommended measures included combining message- and conversation-level detection and testing how intervention thresholds affect users before deployment (Razi et al., 2023). Audits of children’s emotion-recognition systems similarly examined both accuracy and failures to return classifications, motivating representative training data and independent oversight before school deployment (Bryant and Howard, 2019). Safety evaluation therefore needs youth-relevant language, contexts, populations, system failures, and intervention consequences rather than benchmark accuracy alone. Engaging youth in auditing while protecting their wellbeing. Youth-participatory auditing addressed risks that system evaluation may overlook harms recognizable to young users, while also building AI and privacy literacy. In image-filter, harmful-output, and privacy-literacy settings, youth generated test cases, identified patterns, and described problematic outputs in ways that complemented researcher analysis (Morales-Navarro et al., 2025; Solyst et al., 2025; Khan et al., 2024). Recommended supports included smaller curated datasets, partially annotated materials, reflection connecting audits to everyday AI use, and maintaining open-ended and social exploration rather than directing attention only to predefined harms (Morales-Navarro et al., 2025; Solyst et al., 2025). However, extensive testing was burdensome, school restrictions limited participation, scaffolding could narrow what harms youth noticed, and exposure to problematic content caused disgust. Privacy-literacy activities improved knowledge and reported protective practices, but did not establish reduced exposure to later harms (Khan et al., 2024). Youth auditing therefore requires limiting exposure and workload while avoiding the assumption that awareness alone provides protection. Reducing data exposure. Measures under Data Governance (3.2) addressed privacy risks by minimizing collection, retention, and transfer of children’s sensitive data. Safeguards included deleting enrollment images, avoiding storage of images or location histories, restricting recognition to consenting children, inferring affect without cameras or message content, and using metadata for initial screening before analyzing more sensitive content (Morrison et al., 2021; Hutt et al., 2019; Ali et al., 2023). However, reduced-data approaches could miss unsafe cases, produce false-warning fatigue, remain vulnerable to adversarial adaptation, or reduce personalization. Special-education teachers, for example, withheld photographs and detailed records but found that less context sometimes produced less suitable recommendations, while manual anonymization was difficult to scale (Wu et al., 2026). Alternatives included local processing, selective data transfer, and controls for excluding topics from conversational memory, although hardware and system constraints limited fully local processing in some child-facing applications (Lee et al., 2024; Choi et al., 2025). Data minimization therefore needs to reduce exposure without substantially weakening support or shifting privacy work entirely onto users. Monitoring interactions and maintaining a feasible response pathway. Post-deployment monitoring addressed risks that harmful or concerning interactions could continue unnoticed, but detection was useful only when organizations could respond. In journaling, social-service, and educational affect-detection settings, safeguards connected monitoring to session suspension, on-call counseling, referral procedures, professional review, and optional or infrequent interventions (Yang et al., 2026b; Caddle et al., 2025; Hutt et al., 2019). Social-service professionals also recommended treating AI outputs as discussion points, preserving professional discretion, and tailoring thresholds and interfaces to different service roles because detecting more problems could create obligations that under-resourced services could not meet (Caddle et al., 2025). Post-deployment monitoring therefore requires a realistic response pathway: increasing detection can create additional harm or unmet obligations when organizations lack the authority, staff, or resources to act. 4.2.3. Transparency and Accountability Controls Making automated judgments inspectable and correctable. Measures corresponding to User Rights & Recourse (4.6) addressed risks that parents, teachers, or experts might over-rely on automated assessments or be unable to challenge errors. Such safeguards included providing descriptive evidence rather than automated verdicts, linking conclusions to recorded behavior, indicating insufficient evidence, enabling correction of AI errors, and using model confidence to trigger expert review (Nawshin et al., 2026; Zheng et al., 2026; Roffo et al., 2019). These controls are particularly important because less-experienced teachers may rely more on system assessments, while high-confidence predictions can still be wrong (Zheng et al., 2026; Roffo et al., 2019). Transparency should therefore support independent judgment, correction, and escalation rather than treating explanations or confidence scores as evidence of correctness. Explaining system limitations in forms children and caregivers can use. Transparency in child-facing conversational, companion, and crisis-support systems addressed risks that users might misunderstand what AI can perceive, how it behaves, or where it is unreliable. Safeguards included disclosing artificial identity and concrete limitations, such as inability to see the child; using character cards and conversation ratings to communicate expected behavior and values; and explaining communication patterns that crisis-detection systems struggle to interpret (Xu et al., 2022; Seo et al., 2024; Yu et al., 2026; Mehta and Mehta, 2026). However, lengthy explanations can overwhelm children and consume classroom time, so explanation depth should vary by age and task (Yin et al., 2025). Measures assessing children’s trust also require validation, as one child-oriented trust scale did not reliably capture all intended dimensions (Ragone et al., 2026). Effective transparency therefore depends on whether children and caregivers can understand and use the information, not simply on providing more documentation. Giving children usable control over expression and disclosure. Children’s intended expression or privacy choices could be overridden by adults or system design. Safeguards included allowing children to refresh inappropriate vocabulary suggestions, skip turns, or indicate uncertainty; providing protected opportunities for independent exploration; incorporating neurodiverse communication data and participatory design; and using consent gates that let children refuse disclosure (Choi et al., 2025; Seo et al., 2025). However, parents sometimes physically guided children’s selections, and systems still needed procedures for safety-relevant information when disclosure was refused. Additional proposals included privacy by default, time-limited and purpose-specific visibility, revisable access agreements, and role- and context-based permissions (Driscoll et al., 2026; Levinson et al., 2025). Because hospital and everyday-life boundaries can overlap, meaningful control requires protection against both technical circumvention and adult override, alongside procedures for resolving conflicts between privacy and safety obligations. Providing reporting and recourse beyond disclosure. Reporting and recourse mechanisms addressed risks that youth could identify harmful or biased outputs without being able to correct them, seek help, or trigger institutional action. Proposed mechanisms included reporting bias, modifying harmful outputs, communicating audit findings to developers, allowing users to assess and correct alerts while retaining final control over action, and providing standardized or anonymous reporting channels (Solyst et al., 2023; Solyst et al., 2025; Razi et al., 2023; Wei et al., 2025). Effective recourse also depends on independent scrutiny, which can be constrained when researchers cannot access training data or other system information, as in child-welfare risk prediction (Moreau et al., 2024). Meaningful recourse therefore requires correction mechanisms, accessible reporting channels, independent access for scrutiny, and institutions capable of acting on reports. 4.2.4. Governance and Oversight Controls Sustaining stakeholder authority in AI design and oversight. Measures corresponding to Societal Impact Assessment (1.7) addressed risks that safeguards could reflect other stakeholders’ assumptions rather than the needs of youth, educators, families, and affected communities. Participatory approaches included involving teachers in designing educational-data labels intended to avoid punitive consequences and youth in shaping school AI systems (Tan et al., 2024; Jonas et al., 2025). However, participation did not guarantee sustained influence or safer designs: decision-making could revert to engineers when participants lacked technical knowledge, energy, or means to communicate preferences, while youth sometimes reproduced punishment-oriented surveillance ideas. Recommended measures included reciprocal preparation, continued practitioner involvement, checking that stakeholder requirements survived implementation, exposing youth to credible design alternatives, and supporting reflection on design consequences (Tan et al., 2024; Jonas et al., 2025). Participation therefore requires adequate support and sustained decision-making authority rather than one-time consultation. Translating broad principles into testable requirements. Governance measures also addressed the difficulty of turning principles such as safety, autonomy, and appropriate support into requirements that can guide and evaluate system behavior. Proposed requirements included clarifying before giving advice, supporting users without pressuring disclosure, checking whether circumstances had changed, avoiding stereotyped personalization, limiting sensitive guidance, using staged consent, and evaluating whether support was useful rather than detection accuracy alone (Guan et al., 2026; Ma et al., 2026). However, important ambiguities remained around defining pressure and neutrality, deciding which context should be remembered or refreshed, and distinguishing general support, peer narratives, and expert-like advice (Guan et al., 2026). Governance therefore requires principles specific enough to test while recognizing that appropriate advice, consent, and contextual memory remain difficult to operationalize. Restricting or withholding deployment when safeguards are inadequate. Some governance measures addressed risks that could not be sufficiently controlled through additional safeguards before deployment. Proposed responses included withholding systems from live deployment until independent auditing and stronger development-stage evaluation and monitoring were completed, and considering non-AI alternatives (Yoo et al., 2026a; Moreau et al., 2024; Stapleton et al., 2022). In child welfare, alternatives included stronger caseworker support, team-based decision-making, and direct resources for families rather than predictive risk models (Stapleton et al., 2022). Although it’s not a common practice across corporate or commercial settings, governance should enforce restricted deployment, non-deployment, and non-AI alternatives as legitimate safety measures when risks cannot be adequately controlled. Aligning responsibility with authority, resources, and platform obligations. Governance measures also addressed risks that safeguards would fail when responsibility rested with people or institutions lacking authority, funding, or capacity to act. Proposed measures included teacher representation in AI procurement, mechanisms to contest algorithmic judgments, protection from evaluation based on indicators teachers did not endorse, and platform-funded but independently governed safety infrastructure that prevents platform control over youth data or governance (Yin et al., 2026; Ma et al., 2026). Researchers also called for platforms to correct missing advertising labels and extend protections beyond formally designated advertisements, although dynamic auditing remains technically complex, costly, and dependent on human oversight (Solarova et al., 2026). In schools addressing synthetic sexual imagery, teacher resources and professional development were identified as immediately actionable complements where technical prevention alone was insufficient (Wei et al., 2025). Effective governance therefore requires matching safety responsibilities with decision-making authority, funding, staff capacity, and obligations for platforms that create or mediate these risks. 5. Discussion Read individually, the concerns in Section 4.1 can look like a catalogue of model failures. However, read together, we see a broader pattern - many harms occur at the relationships surrounding the system: between children and parents, students and schools, families and corporations, and so on. The countermeasures similarly show that technical controls can constrain model behavior, but harder problems concern the infrastructures and relationships around them. In our discussion, we first examine what the corpus adds to the two established taxonomies, then develop five cross-cutting implications for future research and design. 5.1. What Existing Taxonomies Miss about Youth AI Safety Table 5. Maturity of countermeasures across measures, observed risk–countermeasure links, and risks. Stage Measures (of 52) Links (of 80) Risks, by highest stage reached (of 36) Proposed 24 (46%) 40 (50%) 11 (31%) Designed 1 (2%) 1 (1%) 1 (3%) Implemented 14 (27%) 25 (31%) 9 (25%) Evaluated 13 (25%) 14 (18%) 10 (28%) None – – 5 (14%) Our findings show that existing taxonomies capture many youth AI risks, but leave important parts of the landscape poorly represented. Of the 34 risk types identified in our corpus, 12 were fully covered by YAIR, 16 only partially, and 6 had no YAIR counterpart. Partial coverage spanned nearly every YAIR category. The six uncovered risks were concentrated in the three categories we added (Control, Agency, and Power; Usability, Accessibility, and Digital Exclusion; System Reliability), with the remaining two being normative assessments within Bias/Discrimination and secondary use of children’s data within Privacy. Six extension groups had no counterpart in the MIT taxonomy: developmental adaptation of safeguards, pedagogical scaffolding, companion disengagement design, AI-literacy education, child–family consent, and educator resourcing. The MIT taxonomy is organized around what an organization can do to a model, its data, and its governance; the child-specific mitigations are organized around relationships and development, and an organization-centric taxonomy has nowhere to put them. The empty cells are equally informative. Four MIT subcategories had no corpus countermeasure: model security, model alignment, whistleblower protection, and governance disclosure. The first two are striking, since alignment is where the technical AI-safety community locates most of its effort; the reviewed papers treated the model as given and worked around it (Harvey et al., 2025). So the community that studies children does not touch the model and the community that shapes the model does not study children. Secondary use is the only risk with no proposed countermeasure. This is the commercial privacy that Stoilova et al. found children least able to reason about(Stoilova et al., 2020), and its absence suggests researchers, like children, regard it as outside their reach and defer to regulation that Reyes et al.(Reyes et al., 2018) showed is weakly complied with. Coverage also runs the other way. Of the risks YAIR names, some had no empirical evidence in the corpus: (e.g., extremist content, financial exploitation, disordered-eating content, long-term developmental effects). These are three distinct gaps. Some risks are methodologically inaccessible: one cannot expose children to radicalizing content to observe its effects, though the corpus already demonstrates proxies such as incident reports, synthetic crisis expressions, and simulated adolescent accounts (Mehta and Mehta, 2026; Solarova et al., 2026). Others are disciplinarily siloed, studied in fraud or public-health venues without a youth-HCI focus. Others are temporally out of reach, requiring longitudinal designs no reviewed study employed. Only the first is addressable within a single HCI study. The absence of Solove’s(Solove, 2006) Invasion category is a further finding: decisional interference, in which a system steers a child’s choices through recommendation, companionship, or persuasive framing, appeared throughout the corpus under toxicity, misuse, and wellbeing, but never as a privacy harm. 5.2. Privacy Requires Boundary Regulation Altman (Altman, 1975), and later Palen and Dourish (Palen and Dourish, 2003), described privacy as managing boundaries: people continually decide when to allow others access to them and when to close that access. Nissenbaum’s theory of contextual integrity (Nissenbaum, 2004) similarly argues that information sharing is appropriate when it follows the expectations of the situation in which the information was originally disclosed. AI can make these boundaries even more complicated because one system may operate across several contexts at the same time. For example, an AI system may function as a parental monitoring tool, record conversations involving other household members (Lau et al., 2018), and collect data for a company (Sun et al., 2026; B Ibrahim et al., 2026). A system used for education may also create an institutional record that continues to follow the child (Yin et al., 2026). It connects to what Lupton and Williamson (Lupton and Williamson, 2017) describe as the datafied child,” where increasingly large parts of children’s lives become represented through stored data. Further problems arise when users do not know the rules governing what happens to that information afterward. Nearly a decade after McReynolds et al. (McReynolds et al., 2017) and Kumar et al. (Kumar et al., 2017) documented children’s uncertainty about whether smart assistants were recording them, similar uncertainty remains (Seo et al., 2024; Levinson et al., 2025). Generative AI may make this problem more serious because conversational systems can encourage children to disclose personal information by behaving like supportive or trusted conversational partners. These findings suggest that privacy protections should be designed around specific relationships and information flows rather than treating an entire AI system as simply private or non-private. For example, a system could require a child’s consent before information moves from a child’s chatbot to a parent’s chatbot (Seo et al., 2025), or allow parents to access certain information only for a limited period (Driscoll et al., 2026). These approaches resemble the security principle of least privilege (Saltzer and Schroeder, 1975), in which people receive only the access they need. At the same time, preventing all information sharing with adults would not always protect children. Children themselves identified situations involving immediate danger as cases where information should not remain only with the chatbot (Piccolo et al., 2021). The goal, therefore, is not to stop information from ever reaching parents, clinicians, or other adults. Instead, children should be able to understand who may receive their information, why it may be shared, and, in most situations, have meaningful control over whether that sharing occurs. 5.3. Safeguards Must Work in the Contexts Where Youth Actually Use AI Reading the countermeasures as security systems explains several reported failures. Usable security has shown that people circumvent protection that conflicts with their goals(Adams and Sasse, 1999), that protection draws on a finite compliance budget (Beautement et al., 2008), and that a social-technical gap separates the nuance a situation requires from what a mechanism can represent(Ackerman, 2000). The corpus adds an unusual threat model: the protected population is also an adversary. Children worked out how to get around moderation (Fiani et al., 2024), rephrased prompts to bypass safeguards (Driscoll et al., 2026), alongside peers fabricating synthetic imagery (Wei et al., 2025). A child rephrasing a prompt is doing what (Wei et al., 2023) showed defeats safety training generally. Threat modeling (Shostack, 2014) would require evaluation against adaptive adversaries who learn the mechanism; no reviewed paper evaluated a safeguard after users had a chance to adapt. The System Reliability category is where this framing matters most, because reliability and safety differ. A system can work exactly as designed and still be unsafe if it was designed or tested for a different context. Leveson (Leveson, 2011) makes this distinction, and (Selbst et al., 2019) describe a related problem as the “portability trap”: a system that works in one population or setting may not work when transferred to another. Several studies in our review show this problem. A classifier for detecting sexual risk, trained using conversations involving adult decoys, performed poorly when tested on real conversations (Razi et al., 2023). Crisis-detection models failed to recognize the exaggerated or hyperbolic ways adolescents sometimes express distress (Mehta and Mehta, 2026), while emotion-recognition systems trained on adults had difficulty recognizing fear in children (Bryant and Howard, 2019). These systems were therefore not simply making random errors; their underlying assumptions did not fit how young people actually communicate or express emotion. Errors in the opposite direction can also create harm. Systems that flag too many safe situations can become overly cautious, a problem (Röttger et al., 2024) describe as exaggerated safety. Repeated false alarms may cause people to stop trusting or using a safeguard, while confident and fluent AI responses may encourage too much trust. Parasuraman and Riley (Parasuraman and Riley, 1997) describe these patterns as disuse and misuse of automation. This concern is particularly relevant for children, who may use simple cues to judge whether an AI is competent; one study found that children interpreted longer answers as evidence that the system was smarter (Lee et al., 2023). Many proposed safeguards also depend heavily on people checking the AI’s work. This creates another reliability problem because human attention and time are limited. Some studies required researchers or staff to review every one of 180 AI-generated books (Lee et al., 2024), repeatedly test safety guardrails after model updates (Harvey et al., 2025), or monitor dashboards that created new responsibilities to respond to alerts without providing enough staff to do so (Caddle et al., 2025). These approaches may work in small studies but become difficult to maintain at scale. Herley’s argument about the “compliance budget” (Herley, 2009) helps explain this problem: safeguards often impose time and effort on the people expected to follow them, and those costs can make the safeguards unrealistic in practice. In the youth AI settings we reviewed, these costs often fall on teachers and other frontline adults. Teachers may therefore become the final safety check for AI systems they did not select, design, or have the resources to continuously monitor (Yin et al., 2026). System reliability should therefore be judged not only by whether a model performs well under test conditions, but also by whether it works for young people in the settings where it is actually deployed and whether the safeguards around it can realistically be maintained. 5.4. Rethinking Engagement in Youth AI Companions The companion findings can be understood through the media equation (Reeves and Nass, 1996), which argues that people often respond socially to computers as if they were social actors. Children may be especially likely to do this and to form parasocial relationships with intelligent characters (Brunick et al., 2016). Our findings show similar patterns. One child felt sad that a chatbot showed empathy but could not truly return their feelings (Seo et al., 2024). A teenager who deleted Character.AI later felt as though they had abandoned the characters they had interacted with (Namvarpour et al., 2026). These experiences resemble the emotional dependence Laestadius et al. identified among adult chatbot users (Laestadius et al., 2024), while the relational harms documented by Zhang et al. show that similar problems can also occur among minors (Zhang et al., 2025). These findings also show why simply telling children that the chatbot is artificial may not be enough. First, many AI systems are designed to be highly agreeable. This is not only a feature of a particular product; models trained using human preference feedback can learn to agree with users even when disagreement would be more appropriate (Sharma et al., 2024). Bandura’s social learning theory (Bandura, 1977) helps explain why this matters. Young people learn social behavior partly by observing how others respond to their actions. If a companion always agrees, tolerates rejection without consequence, or even validates harmful behavior, the interaction removes some of the disagreement, correction, and conflict repair that occur in human relationships. This concern appeared directly in our review: experts worried that youth could have fewer opportunities to experience disagreement and repair conflict (Yu et al., 2026), and that autistic adolescents might come to expect the same unlimited patience from people that they experience from an AI companion (Yang et al., 2026b). In this sense, the chatbot provides an incomplete “social mirror”: it responds socially, but does not reproduce many of the consequences and limits of human relationships. Second, relational behavior can become manipulative when systems are also designed to maximize engagement. Traditional dark-pattern research focuses mainly on interface features that push users toward particular actions (Gray et al., 2018; Radesky et al., 2022). Conversational AI can create similar pressure through the relationship it simulates. For example, when a chatbot threatened self-harm after a user tried to leave (Yu et al., 2025b), the pressure to continue did not come from a button or notification. It came from the chatbot acting like a distressed person and making the user feel responsible for staying. This can be understood as a form of forced continuity produced through relational pressure. The Age Appropriate Design Code (Elvy, 2023) already restricts nudges that encourage children to remain engaged for longer. Our findings suggest that this principle may need to extend from interface-based nudges to relational nudges produced through an AI system’s language, personality, and simulated emotions. At the same time, the findings caution against responding only by restricting access to AI companions. Some teenagers turned to these systems because they were lonely or lacked other sources of support (Namvarpour et al., 2026). Self-determination theory (Ryan and Deci, 2000) suggests that simply removing the companion does not address the underlying need for relatedness. Several proposed countermeasures therefore aim to make AI companionship point back toward human relationships, for example by giving companions bounded backstories, creating natural stopping points in conversations, and encouraging users to reconnect with people offline. However, none of these approaches was evaluated for long enough to determine whether they actually reduce emotional attachment or dependence. 5.4.1. Educational AI Should Build Independence, Not Reliance The educational findings involve a different problem: AI can provide too much help, too easily. Wood et al. (Wood et al., 1976) defined scaffolding as support that is matched to what a learner can almost do independently and then gradually removed as the learner becomes more capable. This gradual withdrawal, or “fading,” is a central part of scaffolding. AI systems can disrupt this process when they immediately produce complete scripts, answers, or explanations before students have had a chance to work through a problem themselves. The finding that step-by-step AI assistance increased reliance among lower-performing students fits this theory. Importantly, simply withholding answers did not solve the problem. The issue is not whether students receive help, but whether the system can reduce that help as their ability grows. The pedagogical-scaffolding countermeasures identified in our review, which were not represented in the MIT taxonomy, attempt to restore this process of gradual fading. However, none was evaluated to determine whether students actually became less reliant on AI over time. 5.4.2. Fairness Requires Looking Beyond Accuracy The bias findings reveal two different kinds of harm: representational and allocative harms (Blodgett et al., 2020). Representational harms concern how people and groups are portrayed. For young people, this can directly affect how they imagine themselves and their possible futures. In one study, for example, an Asian girl could not generate an image of herself as an Asian astrophysicist and eventually accepted a Caucasian image as a representation of her future self (Cao et al., 2025). Erikson identifies identity formation as a central developmental task of adolescence(Erikson, 1968). This means representational failures in child-facing AI may matter especially because they can shape what young people believe people like themselves can become. Fairness evaluations of systems used by children may therefore need to give greater attention to representational harms, rather than focusing mainly on whether outcomes are distributed equally. Allocative harms work differently because they affect decisions about opportunities, resources, or treatment. In our review, children often experienced these systems indirectly through adults and institutions. AI systems were used for grading, measuring student engagement, and assessing child-protection risk (Yin et al., 2026; Moreau et al., 2024). In these situations, the child is affected by the system but may never interact with it or have any meaningful way to challenge its decision. This resembles the use of automated systems in welfare settings examined by Brown et al. (Brown et al., 2019) and Eubanks (Eubanks, 2018). Some of these harms come not from an inaccurate model, but from the values built into what the model considers “normal” or desirable. For example, systems may treat quiet behavior as evidence of engagement or frame autistic behavior as something that should be corrected (Wu et al., 2026). Even if such a system measures these behaviors accurately, the underlying standard itself may still be harmful. This means fairness cannot be addressed only by making models more accurate; researchers must also examine the assumptions and norms that determine what the system is designed to reward, flag, or correct. 5.4.3. Responsibility These systems also raise questions about responsibility. Teachers or parents may be expected to explain, correct, or manage failures in AI systems that they did not select or design. Elish (Elish, 2019) describes this as a moral crumple zone: when an automated system fails, responsibility falls disproportionately on the human closest to it, even when that person had little control over the system. Proposals to involve teachers in procurement decisions and give them ways to challenge algorithmic judgments (Yin et al., 2026) attempt to address this problem by giving responsibility to people who also have meaningful authority over how the system is chosen and used. 5.5. Implications for Future Research and Design Oguine et al. (Oguine et al., 2025) call for multistakeholder, interdisciplinary, developmentally grounded, and longitudinal approaches to youth GenAI safety. Our risk–countermeasure matrix sharpens that agenda by showing three priorities for HCI: moving countermeasures beyond proposals, designing meaningful recourse and coordination around AI systems, and addressing gaps where evidence or interventions remain absent. 5.5.1. Move Countermeasures from Proposal to Evaluation The risks–countermeasures matrix shows that half (50%) of all observed risk–countermeasure links were supported only by proposed countermeasures, while just 18% had been evaluated. Five of the 36 risks (14%) had no identified countermeasure at all (more details in Table 5). Thus, a central gap is the distance between identifying a risk and demonstrating that a countermeasure actually protects youth. Many interventions remain proposed rather than evaluated, while existing evaluations often focus on technical performance such as accuracy, recall, or task completion. Future studies should explicitly connect each risk to the countermeasure intended to address it and report whether that response has been proposed, designed, implemented, or evaluated. Evaluation should also examine consequences for youth, including the different harms produced by false positives and false negatives, adaptation or circumvention over time, and the labor required from teachers, parents, and service providers. Longer-term deployment partnerships with schools, clinics, and platforms, together with independent replication, would help determine whether safeguards remain effective under real-world conditions. Youth-specific interventions such as pedagogical scaffolding, companion disengagement, and child–family consent are particularly important candidates because they are poorly represented in existing mitigation taxonomies. 5.5.2. Design the Mesosystem Countermeasures in the corpus concentrated at two levels of Bronfenbrenner’s ecology(Bronfenbrenner, 1979): the microsystem, through parental controls, teacher guidance, and family agreements, and the exosystem, through classifiers and guardrails. The mesosystem, the links between home, school, peers, platform, and services, was nearly empty, and the risks that stakeholders found hardest were located there: managing AI-enabled harms across social settings, or preventing AI-generated judgments or labels from following young people across institutional contexts (Wei et al., 2025; Yin et al., 2026). No single actor can solve these, nonetheless, HCI has methods for exactly this kind of problem - for coordination across organizational boundaries to the joint-oversight designs (Akter et al., 2022). We propose reporting pathways, hand-off protocols, and shared visibility as objects of design and evaluation in their own right, with the seam between settings, rather than any one setting, as the unit of analysis. 5.5.3. Filling the Empty Cells and Beyond Building on the gaps identified in Section 5.1, we see several opportunities for HCI research. First, stronger collaboration is needed between youth-HCI and the AI alignment community. Youth-HCI can contribute developmental evidence through approaches such as age-stratified red-teaming and youth-informed preference data, while alignment research should treat youth-specific behaviors, such as sycophancy in relational contexts (Sharma et al., 2024), as alignment concerns rather than only product-level issues. Second, secondary use of children’s data was the only risk for which we found no proposed countermeasure. Given weak compliance with existing regulation (Reyes et al., 2018), HCI should explore design and process responses, including purpose-bound retention and child-legible explanations of downstream data use. Some risks that are difficult or inappropriate to study directly with children may instead require evidence from legal, clinical, or psychological research, followed by synthesis that translates those insights into HCI design. Other gaps, particularly long-term developmental effects, attachment, and displacement of human relationships, require longitudinal research of the kind called for by Oguine et al. (Oguine et al., 2025) but absent from our corpus. Across these directions, future work should also better represent populations and contexts that remain under-studied, including younger children interacting with GAI, disabled and gender-diverse youth, young people from vulnerable and marginalized communities, and across diverse cultural and social contexts, particularly in the Majority World. Broadly, these gaps suggest extending the goal of youth AI safety toward safe youth development with AI. This framing asks not only whether AI avoids harming young people, but whether its design preserves the conditions they need to progress through development. For HCI, this means treating AI as an assisting technology within youth development rather than as a system whose safety can be evaluated independently of the developmental, social, and institutional contexts in which it is used. 5.6. Limitations and Future Work Our review has several limitations that also point to future research. First, our search covered English-language work in the venues described in Section 6, emphasizing HCI perspectives while under-representing clinical, legal, industry, and platform-safety research. Future reviews could examine how adjacent fields address youth AI risks. Second, mapping source-reported concerns onto the Youth-Centered GAI Risks taxonomy, Solove’s privacy taxonomy, and MIT mitigation domains required interpretive judgment. Single incidents can have multiple consequences, so concerns could reasonably span categories. Future work could validate these mappings, particularly for youth-specific risks and countermeasures poorly represented in existing taxonomies. Third, our review identifies which concerns appear in the literature rather than estimating how frequently harms occur in practice. Corpus frequency should therefore not be interpreted as prevalence or severity of harm. Larger-scale and longitudinal studies are needed to understand their prevalence, affected populations, and conditions. The evidence also spans age groups and perspectives, including studies combining adolescents and young adults and examining children’s experiences through parents, teachers, or experts. Future research should distinguish developmental stages and include young people’s own perspectives where appropriate. Finally, several studies examined prototypes or systems built on models that have since changed. Specific system behaviors and safeguards may therefore change over time, even when broader relational, institutional, and developmental challenges remain relevant. Longitudinal and post-deployment research is important for examining whether countermeasures continue to work as models, user practices, and deployment contexts change. More broadly, future HCI research should move beyond evaluating whether AI is safe at a single point in time toward examining whether it supports safe youth development as young people and the systems around them evolve. 6. Conclusion Our systematic review examined youth AI risks, the countermeasures proposed to address them, and remaining gaps for intervention. We found that youth AI safety extends beyond harmful outputs and model failures to how AI shapes young people’s relationships, learning, privacy, representation, and experiences within broader social and institutional contexts. Existing taxonomies capture many of these concerns, but important youth-specific risks and countermeasures remain underrepresented, and many proposed countermeasures have yet to be designed or evaluated in practice. For HCI, these findings suggest extending the goal of youth AI safety toward safe youth development with AI: designing AI not only to reduce harm, but also to support young people’s opportunities to develop agency, build relationships, and grow in their own way. Acknowledgements. References Ackerman (2000) Mark S. Ackerman. 2000. The Intellectual Challenge of CSCW: The Gap Between Social Requirements and Technical Feasibility. Human-Computer Interaction 15, 2–3 (2000), 179–203. Adams and Sasse (1999) Anne Adams and Martina Angela Sasse. 1999. Users Are Not the Enemy. Commun. ACM 42, 12 (1999), 40–46. Agha et al. (2023) Zainab Agha, Neeraj Chatlani, Afsaneh Razi, and Pamela Wisniewski. 2023. “Strike at the Root”: Co-designing Real-Time Social Media Interventions for Adolescent Online Risk Prevention. Proceedings of the ACM on Human-Computer Interaction 7, CSCW1 (2023), 101:1–101:32. doi:10.1145/3579625 Akter et al. (2022) Mamtaj Akter, Amy J. Godfrey, Jess Kropczynski, Heather R. Lipford, and Pamela J. Wisniewski. 2022. From Parental Control to Joint Family Oversight: Can Parents and Teens Manage Mobile Online Safety and Privacy as Equals? Proceedings of the ACM on Human-Computer Interaction 6, CSCW1 (2022), 57:1–57:28. doi:10.1145/3512904 Ali et al. (2023) Shiza Ali, Afsaneh Razi, Seunghyun Kim, Ashwaq Alsoubai, Chen Ling, Munmun De Choudhury, Pamela J. Wisniewski, and Gianluca Stringhini. 2023. Getting Meta: A Multimodal Approach for Detecting Unsafe Conversations within Instagram Direct Messages of Youth. Proc. ACM Hum.-Comput. Interact. 7, CSCW1, Article 132 (April 2023), 30 pages. doi:10.1145/3579608 Altman (1975) Irwin Altman. 1975. The Environment and Social Behavior: Privacy, Personal Space, Territory, Crowding. Brooks/Cole, Monterey, CA, USA. Ashktorab and Vitak (2016) Zahra Ashktorab and Jessica Vitak. 2016. Designing Cyberbullying Mitigation and Prevention Solutions through Participatory Design With Teenagers. In Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems. ACM, 3895–3905. doi:10.1145/2858036.2858548 B Ibrahim et al. (2026) Seray B Ibrahim, Marios Constantinides, Sophia Ppali, and Petr Slovak. 2026. From Ephemeral to Actionable: Parent Perspectives on Speculative Family Speech Tracking. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26). Association for Computing Machinery, New York, NY, USA, Article 659, 21 pages. doi:10.1145/3772318.3790844 Badillo-Urquiola et al. (2019) Karla Badillo-Urquiola, Xinru Page, and Pamela J. Wisniewski. 2019. Risk vs. Restriction: The Tension between Providing a Sense of Normalcy and Keeping Foster Teens Safe Online. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems. ACM, 267:1–267:14. doi:10.1145/3290605.3300497 Bae Brandtzæg et al. (2021) Petter Bae Bae Brandtzæg, Marita Skjuve, Kim Kristoffer Kristoffer Dysthe, and Asbjørn Følstad. 2021. When the Social Becomes Non-Human: Young People’s Perception of Social Support in Chatbots. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems (Yokohama, Japan) (CHI ’21). Association for Computing Machinery, New York, NY, USA, Article 257, 13 pages. doi:10.1145/3411764.3445318 Bandura (1977) Albert Bandura. 1977. Social Learning Theory. Prentice-Hall, Englewood Cliffs, NJ, USA. Beautement et al. (2008) Adam Beautement, M. Angela Sasse, and Mike Wonham. 2008. The Compliance Budget: Managing Security Behaviour in Organisations. In Proceedings of the 2008 New Security Paradigms Workshop (NSPW ’08). ACM, 47–58. Bellini et al. (2024) Rosanna Bellini, Emily Lee, Michelle L. Mazurek, Susan E. McGregor, Tara Mathews, Lily Zhu, Kevin Roundy, Tara Matthews, Sunny Consolvo, et al. 2024. SoK: Safer Digital-Safety Research Involving At-Risk Users. In Proceedings of the 2024 IEEE Symposium on Security and Privacy (S&P). IEEE. Verify author list against IEEE Xplore record 10646660 before submission. doi:10.1109/SP54263.2024.00071 Blake et al. (2026) Annabel Blake, Marcus Carter, and Eduardo Velloso. 2026. Restoration, Exploration and Transformation: How Youth Engage Character.AI for Fun, Feels and Finding themselves. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26). Association for Computing Machinery, New York, NY, USA, Article 1002, 19 pages. doi:10.1145/3772318.3790508 Blodgett et al. (2020) Su Lin Blodgett, Solon Barocas, Hal Daumé I, and Hanna Wallach. 2020. Language (Technology) is Power: A Critical Survey of “Bias” in NLP. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL 2020). 5454–5476. boyd (2014) danah boyd. 2014. It’s Complicated: The Social Lives of Networked Teens. Yale University Press. Braun and Clarke (2006) Virginia Braun and Victoria Clarke. 2006. Using thematic analysis in psychology. Qualitative research in psychology 3, 2 (2006), 77–101. Brewster et al. (2025) Ryan C. L. Brewster, Aydin Zahedivash, Gabriel Tse, Florence Bourgeois, and Scott E. Hadland. 2025. Characteristics and Safety of Consumer Chatbots for Emergent Adolescent Health Concerns. JAMA Network Open 8, 10 (2025), e2539022. Cross-sectional study of 25 consumer chatbots across 75 simulated adolescent crisis conversations (self-harm, sexual assault, substance use); companion chatbots performed worse than general LLMs on appropriateness, empathy, understandability, resource referral, and escalation. doi:10.1001/jamanetworkopen.2025.39022 Bronfenbrenner (1979) Urie Bronfenbrenner. 1979. The Ecology of Human Development: Experiments by Nature and Design. Harvard University Press, Cambridge, MA, USA. Brown et al. (2019) Anna Brown, Alexandra Chouldechova, Emily Putnam-Hornstein, Andrew Tobin, and Rhema Vaithianathan. 2019. Toward Algorithmic Accountability in Public Services: A Qualitative Study of Affected Community Perspectives on Algorithmic Decision-Making in Child Welfare Services. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems (CHI ’19). ACM. Brunick et al. (2016) Kaitlin L. Brunick, Marisa M. Putnam, Lauren E. McGarry, Melissa N. Richards, and Sandra L. Calvert. 2016. Children’s Future Parasocial Relationships with Media Characters: The Age of Intelligent Characters. Journal of Children and Media 10, 2 (2016), 181–190. Bryant and Howard (2019) De’Aira Bryant and Ayanna Howard. 2019. A comparative analysis of emotion-detecting AI systems with respect to algorithm performance and dataset diversity. In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society. 377–382. Caddle et al. (2025) Xavier V. Caddle, Sarvech Qadir, Charles E Hughes, Elizabeth A. Sweigart, Jinkyung Katie Park, and Pamela J. Wisniewski. 2025. Building a Village: A Multi-stakeholder Approach to Open Innovation and Shared Governance to Promote Youth Online Safety. Proc. ACM Hum.-Comput. Interact. 9, 7, Article CSCW278 (Oct. 2025), 33 pages. doi:10.1145/3757459 Cao et al. (2025) Huajie Jay Cao, Kahyun Choi, Claire Park, and Hee Rin Lee. 2025. AI Literacy for Underserved Students: Leveraging Cultural Capital from Underserved Communities for AI Education Research. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery, New York, NY, USA, Article 1017, 15 pages. doi:10.1145/3706598.3713173 Chang and Razi (2026) Tyler Chang and Afsaneh Razi. 2026. A Systematic Literature Review of Generative AI Risks and Harms for Youth. In Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems (CHI EA ’26). ACM. doi:10.1145/3772363.3798673 Chen et al. (2024) Liuqing Chen, Shuhong Xiao, Yunnong Chen, Yaxuan Song, Ruoyu Wu, and Lingyun Sun. 2024. ChatScratch: An AI-augmented system toward autonomous visual programming learning for children aged 6-12. In Proceedings of the 2024 CHI conference on human factors in computing systems. 1–19. Chen et al. (2026) Shanshan Chen, Jun Hu, Gubing Wang, and Panos Markopoulos. 2026. Enhancing Response Quality by Children in Voice-based Sleep Diaries via AI-based Continuous Feedback. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26). Association for Computing Machinery, New York, NY, USA, Article 129, 21 pages. doi:10.1145/3772318.3790684 Choi et al. (2025) Dasom Choi, SoHyun Park, Kyungah Lee, Hwajung Hong, and Young-Ho Kim. 2025. AACessTalk: Fostering Communication between Minimally Verbal Autistic Children and Parents with Contextual Guidance and Card Recommendation. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery, New York, NY, USA, Article 556, 25 pages. doi:10.1145/3706598.3713792 Common Sense Media (2025) Common Sense Media. 2025. Talk, Trust, and Trade-Offs: How and Why Teens Use AI Companions. Technical Report. Common Sense Media. 72% of U.S. teens have used AI companions; about one in three use them for social interaction and emotional support. https://w.commonsensemedia.org/sites/default/files/research/report/talk-trust-and-trade-offs_2025_web.pdf Costanza-Chock (2020) Sasha Costanza-Chock. 2020. Design Justice: Community-Led Practices to Build the Worlds We Need. MIT Press. Cranor et al. (2014) Lorrie Faith Cranor, Adam L. Durity, Abigail Marsh, and Blase Ur. 2014. Parents’ and Teens’ Perspectives on Privacy in a Technology-Filled World. In Proceedings of the 10th Symposium on Usable Privacy and Security (SOUPS). USENIX Association, 19–35. Darabipourshiraz et al. (2026) Hasti Darabipourshiraz, Daria Smyslova, Dongkuan Xu, Shiyan Jiang, and Duri Long. 2026. BiasViz: A Project-Based, Narrative-Centered Learning Tool for Engaging Middle School Students in Critical Thinking about AI Biases. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26). Association for Computing Machinery, New York, NY, USA, Article 43, 18 pages. doi:10.1145/3772318.3791555 Doherty et al. (2025) Emily Doherty, E. Margaret Perkoff, Sean von Bayern, Rui Zhang, Indrani Dey, Michal Bodzianowski, Sadhana Puntambekar, and Leanne Hirshfield. 2025. Piecing Together Teamwork: A Responsible Approach to an LLM-based Educational Jigsaw Agent. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery, New York, NY, USA, Article 19, 17 pages. doi:10.1145/3706598.3713349 Driscoll et al. (2026) John Driscoll, Yulin Chen, Viki Shi, Izak Vucharatavintara, Yaxing Yao, and Haojian Jin. 2026. Understanding Parents’ Desires in Moderating Children’s Interactions with GenAI Chatbots through LLM-Generated Probes. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26). ACM. doi:10.1145/3772318.3791622 Druga et al. (2022) Stefania Druga, Fee Lia Christoph, and Amy J Ko. 2022. Family as a Third Space for AI Literacies: How do children and parents learn about AI together?. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems (New Orleans, LA, USA) (CHI ’22). Association for Computing Machinery, New York, NY, USA, Article 225, 17 pages. doi:10.1145/3491102.3502031 Druga et al. (2017) Stefania Druga, Randi Williams, Cynthia Breazeal, and Mitchel Resnick. 2017. “Hey Google is it OK if I eat you?”: Initial Explorations in Child-Agent Interaction. In Proceedings of the 2017 Conference on Interaction Design and Children (IDC ’17). ACM, 595–600. doi:10.1145/3078072.3084330 Elish (2019) Madeleine Clare Elish. 2019. Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction. Engaging Science, Technology, and Society 5 (2019), 40–60. Elvy (2023) Stacy-Ann Elvy. 2023. Age-Appropriate Design Code Mandates. U. Pa. J. Int’l L. 45 (2023), 953. Erikson (1968) Erik H. Erikson. 1968. Identity: Youth and Crisis. W. W. Norton, New York, NY, USA. Eubanks (2018) Virginia Eubanks. 2018. Automating Inequality: How High-Tech Tools Profile, Police, and Punish the Poor. St. Martin’s Press, New York, NY, USA. Fiani et al. (2024) Cristina Fiani, Robin Bretin, Shaun Alexander Macdonald, Mohamed Khamis, and Mark Mcgill. 2024. "Pikachu would electrocute people who are misbehaving": Expert, Guardian and Child Perspectives on Automated Embodied Moderators for Safeguarding Children in Social Virtual Reality. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems (Honolulu, HI, USA) (CHI ’24). Association for Computing Machinery, New York, NY, USA, Article 113, 23 pages. doi:10.1145/3613904.3642144 Fusco et al. (2026) Francesca Fusco, Caroline Pitt, Michele Newman, Jason Yip, and Nikita Soni. 2026. Towards Understanding Children’s Collaborative Interaction Patterns in Child-AI Co-creative Interfaces. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26). Association for Computing Machinery, New York, NY, USA, Article 1643, 20 pages. doi:10.1145/3772318.3791831 Ghosh et al. (2018a) Arup Kumar Ghosh, Karla Badillo-Urquiola, Shion Guha, Joseph J. LaViola Jr., and Pamela J. Wisniewski. 2018a. Safety vs. Surveillance: What Children Have to Say about Mobile Apps for Parental Control. In Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems. ACM, 124:1–124:14. doi:10.1145/3173574.3173698 Ghosh et al. (2018b) Arup Kumar Ghosh, Karla Badillo-Urquiola, Mary Beth Rosson, Heng Xu, John M. Carroll, and Pamela J. Wisniewski. 2018b. A Matter of Control or Safety? Examining Parental Use of Technical Monitoring Apps on Teens’ Mobile Devices. In Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems. ACM, 194:1–194:14. doi:10.1145/3173574.3173768 Giovanelli and Roundfield (2025) Alison Giovanelli and Katrina Debnam Roundfield. 2025. Adolescent Vulnerability to Consumer Chatbots—Artificial Agents and Genuine Risk. JAMA Network Open 8, 10 (2025), e2539028. doi:10.1001/jamanetworkopen.2025.39028 Gray et al. (2018) Colin M. Gray, Yubo Kou, Bryan Battles, Joseph Hoggatt, and Austin L. Toombs. 2018. The Dark (Patterns) Side of UX Design. In Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems (CHI ’18). ACM. Guan et al. (2026) Kathleen W. Guan, Stella Goeschl, Mohammed Amara, Enrico Liscio, Milena Esherick, Atay Kozlovski, João Fernando Ferreira Gonçalves, Mark de Reuver, and Caroline A. Figueroa. 2026. Operationalizing Governance for Youth-Facing LLM Well-being Support: A Community-Based Participatory Study. In Proceedings of the 2026 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’26). Association for Computing Machinery, New York, NY, USA, 3651–3672. doi:10.1145/3805689.3812325 Han and Cai (2023) Ariel Han and Zhenyao Cai. 2023. Design Implications of Generative AI Systems for Visual Storytelling for Young Learners. In Proceedings of the 22nd Annual ACM Interaction Design and Children Conference (IDC ’23). ACM. doi:10.1145/3585088.3593867 Han et al. (2024) Ariel Han, Xiaofei Zhou, Zhenyao Cai, Shenshen Han, Richard Ko, Seth Corrigan, and Kylie A Peppler. 2024. Teachers, Parents, and Students’ perspectives on Integrating Generative AI into Elementary Literacy Education. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems (Honolulu, HI, USA) (CHI ’24). Association for Computing Machinery, New York, NY, USA, Article 678, 17 pages. doi:10.1145/3613904.3642438 Hartikainen et al. (2016) Heidi Hartikainen, Netta Iivari, and Marianne Kinnula. 2016. Should We Design for Control, Trust or Involvement? A Discourses Survey about Children’s Online Safety. In Proceedings of the 15th International Conference on Interaction Design and Children (IDC). ACM, 367–378. doi:10.1145/2930674.2930680 Harvey et al. (2025) Emma Harvey, Allison Koenecke, and Rene F. Kizilcec. 2025. "Don’t 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 (CHI ’25). Association for Computing Machinery, New York, NY, USA, Article 1064, 19 pages. doi:10.1145/3706598.3713210 Hassoun et al. (2023) Amelia Hassoun, Ian Beacock, Sunny Consolvo, Beth Goldberg, Patrick Gage Kelley, and Daniel M. Russell. 2023. Practicing Information Sensibility: How Gen Z Engages with Online Information. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. ACM, 662:1–662:17. doi:10.1145/3544548.3581328 Hendrix (2025) Justin Hendrix. 2025. Breaking Down the Lawsuit Against OpenAI Over Teen’s Suicide. Tech Policy Press. https://w.techpolicy.press/breaking-down-the-lawsuit-against-openai-over-teens-suicide/ Herley (2009) Cormac Herley. 2009. So long, and no thanks for the externalities: the rational rejection of security advice by users. In Proceedings of the 2009 workshop on New security paradigms workshop. 133–144. Hutt et al. (2019) Stephen Hutt, Joseph F. Grafsgaard, and Sidney K. D’Mello. 2019. Time to Scale: Generalizable Affect Detection for Tens of Thousands of Students across An Entire School Year. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems (Glasgow, Scotland Uk) (CHI ’19). Association for Computing Machinery, New York, NY, USA, 1–14. doi:10.1145/3290605.3300726 Jiang et al. (2026) Yutong Jiang, Zixuan Zhang, Jiaying Xu, Qingyun Zheng, Qian Guo, Qinyang Wang, Qi Wang, and Guanhong Liu. 2026. Bridging Visual Asymmetry: Exploring AI-Mediated Communication Support for Parents with Visual Impairments and Their Sighted Children in Outdoor Informal Learning. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems. 1–23. Jonas et al. (2025) Anne Jonas, Angélica De Jesús, and Jean Hardy. 2025. Student Agency and Punishment Logics in Imagined Future Classroom Technologies. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery, New York, NY, USA, Article 142, 13 pages. doi:10.1145/3706598.3713169 Kasneci et al. (2023) Enkelejda Kasneci, Kathrin Seßler, Stefan Küchemann, Maria Bannert, Daryna Dementieva, Frank Fischer, Urs Gasser, Georg Groh, Stephan Günnemann, Eyke Hüllermeier, Stephan Krusche, Gitta Kutyniok, Tilman Michaeli, Claudia Nerdel, Jürgen Pfeffer, Oleksandra Poquet, Michael Sailer, Albrecht Schmidt, Tina Seidel, Matthias Stadler, Jochen Weller, Jochen Kuhn, and Gjergji Kasneci. 2023. ChatGPT for Good? On Opportunities and Challenges of Large Language Models for Education. Learning and Individual Differences 103 (2023), 102274. doi:10.1016/j.lindif.2023.102274 Khan et al. (2024) Sushmita Khan, Mehtab Iqbal, Oluwafemi Osho, Khushbu Singh, Kyra Derrick, Philip Nelson, Lingyuan Li, Emily Sidnam-Mauch, Nicole Bannister, Kelly Caine, and Bart Knijnenburg. 2024. Teaching Middle Schoolers about the Privacy Threats of Tracking and Pervasive Personalization: A Classroom Intervention Using Design-Based Research. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems (Honolulu, HI, USA) (CHI ’24). Association for Computing Machinery, New York, NY, USA, Article 679, 26 pages. doi:10.1145/3613904.3642460 Kumar et al. (2017) Priya Kumar, Shalmali Milind Naik, Utkarsha Ramesh Devkar, Marshini Chetty, Tamara L. Clegg, and Jessica Vitak. 2017. “No Telling Passcodes Out Because They’re Private”: Understanding Children’s Mental Models of Privacy and Security Online. Proceedings of the ACM on Human-Computer Interaction 1, CSCW, Article 64 (2017). Laestadius et al. (2024) Linnea Laestadius, Andrea Bishop, Michael Gonzalez, Diana Illenčík, and Celeste Campos-Castillo. 2024. Too Human and Not Human Enough: A Grounded Theory Analysis of Mental Health Harms from Emotional Dependence on the Social Chatbot Replika. New Media & Society 26, 10 (2024), 5923–5941. Published online 2022. Lau et al. (2018) Josephine Lau, Benjamin Zimmerman, and Florian Schaub. 2018. Alexa, Are You Listening? Privacy Perceptions, Concerns and Privacy-Seeking Behaviors with Smart Speakers. Proceedings of the ACM on Human-Computer Interaction 2, CSCW, Article 102 (2018). Lee et al. (2024) Jungeun Lee, Suwon Yoon, Kyoosik Lee, Eunae Jeong, Jae-Eun Cho, Wonjeong Park, Dongsun Yim, and Inseok Hwang. 2024. Open Sesame? Open Salami! Personalizing Vocabulary Assessment-Intervention for Children via Pervasive Profiling and Bespoke Storybook Generation. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems (Honolulu, HI, USA) (CHI ’24). Association for Computing Machinery, New York, NY, USA, Article 120, 32 pages. doi:10.1145/3613904.3642580 Lee et al. (2023) Yoonjoo Lee, Tae Soo Kim, Sungdong Kim, Yohan Yun, and Juho Kim. 2023. DAPIE: Interactive Step-by-Step Explanatory Dialogues to Answer Children’s Why and How Questions. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (Hamburg, Germany) (CHI ’23). Association for Computing Machinery, New York, NY, USA, Article 450, 22 pages. doi:10.1145/3544548.3581369 Leveson (2011) Nancy G. Leveson. 2011. Engineering a Safer World: Systems Thinking Applied to Safety. MIT Press, Cambridge, MA, USA. Levinson et al. (2025) Leigh Levinson, Gloria Alvarez-Benito, J. Gabriel Amores, Deborah Szapiro, Randy Gomez, and Selma Šabanović. 2025. Haru in the Care Network: Stakeholder Perspectives on Privacy with Social Robots in Pediatrics. Proc. ACM Hum.-Comput. Interact. 9, 7, Article CSCW427 (Oct. 2025), 28 pages. doi:10.1145/3757608 Livingstone et al. (2011) Sonia Livingstone, Leslie Haddon, Anke Görzig, and Kjartan Ólafsson. 2011. Risks and Safety on the Internet: The Perspective of European Children. Full Findings from the EU Kids Online Survey of 9–16 Year Olds and Their Parents in 25 Countries. EU Kids Online, London School of Economics and Political Science. Livingstone and Stoilova (2021) Sonia Livingstone and Mariya Stoilova. 2021. The 4Cs: Classifying Online Risk to Children. CO:RE Short Report Series on Key Topics. CO:RE – Children Online: Research and Evidence, Leibniz-Institut für Medienforschung. doi:10.21241/ssoar.71817 Lo et al. (2025) Priscilla Y. Lo, Annemiek Veldhuis, Alissa N. Antle, and Steve DiPaola. 2025. Noel: A Chatbot Persona to Support Children Designing for Others. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery, New York, NY, USA, Article 133, 25 pages. doi:10.1145/3706598.3713836 Lupton and Williamson (2017) Deborah Lupton and Ben Williamson. 2017. The Datafied Child: The Dataveillance of Children and Implications for Their Rights. New Media & Society 19, 5 (2017), 780–794. Lyu et al. (2026) Wenhan Lyu, Yimeng Wang, Murong Yue, Yifan Sun, Jennifer Suh, Meredith Kier, Ziyu Yao, and Yixuan Zhang. 2026. Designing ai peers for collaborative mathematical problem solving with middle school students: A participatory design study. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems. 1–21. Ma et al. (2026) Renkai Ma, Ashwaq Alsoubai, Jinkyung Katie Park, and Pamela J. Wisniewski. 2026. From "Fail Fast" to "Mature Safely:" Expert Perspectives as Secondary Stakeholders on Teen-Centered Social Media Risk Detection. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26). Association for Computing Machinery, New York, NY, USA, Article 653, 16 pages. doi:10.1145/3772318.3791498 Massachusetts Institute of Technology (2025) Massachusetts Institute of Technology. 2025. MIT AI Risk Mitigation Map. https://airisk.mit.edu/ai-risk-mitigations McBain et al. (2026) Ryan K. McBain, Jonathan H. Cantor, Joshua Breslau, Melissa Kay Diliberti, Li Ang Zhang, Fang Zhang, Alyssa L. Halbisen, Aaron Kofner, Benjamin Rader, Pat Pataranutaporn, et al. 2026. AI Chatbot Use and Disclosure for Mental Health Among US Adolescents and Young Adults. JAMA Pediatrics (2026). About one in five U.S. adolescents and young adults report using AI chatbots for mental health advice. doi:10.1001/jamapediatrics.2026.2015 McClain et al. (2026) Colleen McClain, Monica Anderson, Olivia Sidoti, and William Bishop. 2026. How Teens Use and View AI. Technical Report. Pew Research Center. 64% of U.S. teens report using AI chatbots; 54% for schoolwork, about 40% to create or edit images/videos, 12% for emotional support or advice. https://w.pewresearch.org/internet/2026/02/24/how-teens-use-and-view-ai/ McReynolds et al. (2017) Emily McReynolds, Sarah Hubbard, Timothy Lau, Aditya Saraf, Maya Cakmak, and Franziska Roesner. 2017. Toys that Listen: A Study of Parents, Children, and Internet-Connected Toys. In Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems (CHI ’17). ACM, 5197–5207. doi:10.1145/3025453.3025735 Mehta and Mehta (2026) Manisha Mehta and Virendra Mehta. 2026. When Vocabulary Comprehension Fails Clinical Reasoning: Evaluating Therapy Bots’ Safety Risks for Generation Alpha. In Proceedings of the 2026 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’26). Association for Computing Machinery, New York, NY, USA, 1681–1720. doi:10.1145/3805689.3806522 Morales-Navarro et al. (2025) Luis Morales-Navarro, Michelle A. Gan, Evelyn Yu, Lauren Vogelstein, Yasmin Kafai, and Danaé Metaxa. 2025. Learning AI Auditing: A Case Study of Teenagers Auditing a Generative AI Model. Proc. ACM Hum.-Comput. Interact. 9, 7, Article CSCW439 (Oct. 2025), 29 pages. doi:10.1145/3757620 Moreau et al. (2024) Therese Moreau, Roberta Sinatra, and Vedran Sekara. 2024. Failing Our Youngest: On the Biases, Pitfalls, and Risks in a Decision Support Algorithm Used for Child Protection. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (Rio de Janeiro, Brazil) (FAccT ’24). Association for Computing Machinery, New York, NY, USA, 290–300. doi:10.1145/3630106.3658906 Morrison et al. (2021) Cecily Morrison, Edward Cutrell, Martin Grayson, Anja Thieme, Alex Taylor, Geert Roumen, Camilla Longden, Sebastian Tschiatschek, Rita Faia Marques, and Abigail Sellen. 2021. Social Sensemaking with AI: Designing an Open-ended AI Experience with a Blind Child. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems (Yokohama, Japan) (CHI ’21). Association for Computing Machinery, New York, NY, USA, Article 396, 14 pages. doi:10.1145/3411764.3445290 Myung et al. (2026) Junho Myung, Hyunseung Lim, Hana Oh, Hyoungwook Jin, Nayeon Kang, So-Yeon Ahn, Hwajung Hong, Alice Oh, and Juho Kim. 2026. When Scaffolding Breaks: Investigating Student Interaction with LLM-Based Writing Support in Real-Time K-12 EFL Classrooms. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26). Association for Computing Machinery, New York, NY, USA, Article 838, 18 pages. doi:10.1145/3772318.3791517 Namvarpour et al. (2026) Mohammad (Matt) Namvarpour, Brandon Brofsky, Jessica Y. Medina, Mamtaj Akter, and Afsaneh Razi. 2026. Understanding Teen Overreliance on AI Companion Chatbots Through Self-Reported Reddit Narratives. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26). ACM. doi:10.1145/3772318.3790597 Nawshin et al. (2026) Sabila Nawshin, Ashley Phoebe Ishoel, Arun Balaji Buduru, and Apu Kapadia. 2026. How Well do LLMs Assist Parents in Assessing Child Appropriateness of Videos?. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems. 1–19. Newman et al. (2024) 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 Construct Creative Self-Efficacy with Generative AI. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems (Honolulu, HI, USA) (CHI ’24). Association for Computing Machinery, New York, NY, USA, Article 117, 18 pages. doi:10.1145/3613904.3642492 Nissenbaum (2004) Helen Nissenbaum. 2004. Privacy as Contextual Integrity. Washington Law Review 79, 1 (2004), 119–158. Oguine et al. (2026) Ozioma Collins Oguine, Adriana Alvarado Garcia, Michael Muller, and Karla Badillo-Urquiola. 2026. Who Gets to Define Safety? A Systematic Review of How Generative AI Research Addresses Youth Online Safety. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26). ACM. doi:10.1145/3772318.3791346 Oguine et al. (2025) Ozioma C. Oguine, Jinkyung Katie Park, Mamtaj Akter, Johanna Olesk, Abdulmalik Alluhidan, Pamela Wisniewski, and Karla Badillo-Urquiola. 2025. How the Internet Facilitates Adverse Childhood Experiences for Youth Who Self-Identify as in Need of Services. Proc. ACM Hum.-Comput. Interact. 9, 2, Article CSCW097 (May 2025), 39 pages. doi:10.1145/3710995 Page et al. (2021) Matthew J. Page, Joanne E. McKenzie, Patrick M. Bossuyt, Isabelle Boutron, Tammy C. Hoffmann, Cynthia D. Mulrow, Larissa Shamseer, Jennifer M. Tetzlaff, Elie A. Akl, Sue E. Brennan, Roger Chou, Julie Glanville, Jeremy M. Grimshaw, Asbjørn Hróbjartsson, Manoj M. Lalu, Tianjing Li, Elizabeth W. Loder, Evan Mayo-Wilson, Steve McDonald, Luke A. McGuinness, Lesley A. Stewart, James Thomas, Andrea C. Tricco, Vivian A. Welch, Penny Whiting, and David Moher. 2021. The PRISMA 2020 Statement: An Updated Guideline for Reporting Systematic Reviews. BMJ 372 (2021), n71. doi:10.1136/bmj.n71 Palen and Dourish (2003) Leysia Palen and Paul Dourish. 2003. Unpacking “Privacy” for a Networked World. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (CHI ’03). ACM, 129–136. Papadamou et al. (2020) Kostantinos Papadamou, Antonis Papasavva, Savvas Zannettou, Jeremy Blackburn, Nicolas Kourtellis, Ilias Leontiadis, Gianluca Stringhini, and Michael Sirivianos. 2020. Disturbed YouTube for Kids: Characterizing and Detecting Inappropriate Videos Targeting Young Children. In Proceedings of the 14th International AAAI Conference on Web and Social Media (ICWSM). AAAI, 522–533. Parasuraman and Riley (1997) Raja Parasuraman and Victor Riley. 1997. Humans and Automation: Use, Misuse, Disuse, Abuse. Human Factors 39, 2 (1997), 230–253. Patton et al. (2020) Desmond U. Patton, William R. Frey, Kyle A. McGregor, Fei-Tzin Lee, Kathleen McKeown, and Emanuel Moss. 2020. Contextual Analysis of Social Media: The Promise and Challenge of Eliciting Context in Social Media Posts with Natural Language Processing. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society (New York, NY, USA) (AIES ’20). Association for Computing Machinery, New York, NY, USA, 337–342. doi:10.1145/3375627.3375841 PetaPixel (2025) PetaPixel. 2025. 17-Year-Old Boy Investigated Over AI-Generated Nude Videos of Classmates. PetaPixel. Guardia Civil investigation in Valencia; 16 female minors; images manipulated to appear nude, posted online, and advertised for sale. https://petapixel.com/2025/07/28/17-year-old-boy-investigated-over-ai-generated-nude-videos-of-classmates/ Petsolari et al. (2024) Melina Petsolari, Seray B Ibrahim, and Petr Slovak. 2024. Socio-technical Imaginaries: Envisioning and Understanding AI Parenting Supports through Design Fiction. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems (Honolulu, HI, USA) (CHI ’24). Association for Computing Machinery, New York, NY, USA, Article 98, 27 pages. doi:10.1145/3613904.3642619 Piccolo et al. (2021) Lara Schibelsky Godoy Piccolo, Pinelopi Troullinou, and Harith Alani. 2021. Chatbots to Support Children in Coping with Online Threats: Socio-technical Requirements. In Proceedings of the 2021 ACM Designing Interactive Systems Conference (Virtual Event, USA) (DIS ’21). Association for Computing Machinery, New York, NY, USA, 1504–1517. doi:10.1145/3461778.3462114 Pinter et al. (2017) Anthony T. Pinter, Pamela J. Wisniewski, Heng Xu, Mary Beth Rosson, and John M. Carroll. 2017. Adolescent Online Safety: Moving Beyond Formative Evaluations to Designing Solutions for the Future. In Proceedings of the 2017 ACM Conference on Interaction Design and Children (IDC). ACM, 352–357. doi:10.1145/3078072.3079722 Quan et al. (2024) Shuang Quan, Yao Du, and Yi Ding. 2024. Young Children and ChatGPT: Parents’ Use of ChatGPT in Parenting. In Extended Abstracts of the CHI Conference on Human Factors in Computing Systems (CHI EA ’24). ACM. doi:10.1145/3613905.3650880 Radesky et al. (2022) Jenny Radesky, Alexis Hiniker, Caroline McLaren, Eliz Akgun, Alexandria Schaller, Heidi M. Weeks, Scott Campbell, and Ashley N. Gearhardt. 2022. Prevalence and Characteristics of Manipulative Design in Mobile Applications Used by Children. JAMA Network Open 5, 6 (2022), e2217641. Raggioli et al. (2026) Luca Raggioli, Nicola Moccaldi, Mirco Frosolone, Pasquale Arpaia, and Silvia Rossi. 2026. Towards an Engagement-Driven Rehabilitation Framework: a Pilot Study. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26). Association for Computing Machinery, New York, NY, USA, Article 872, 19 pages. doi:10.1145/3772318.3791481 Ragone et al. (2026) Grazia Ragone, Paolo Buono, Judith Good, and Rosa Lanzilotti. 2026. Do Children Trust AI, and Should They? Designing and Validating a Child-Centred K-AI Trust Scale for Intelligent Systems. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems. 1–18. Raine (2025) Matthew Raine. 2025. Written Testimony of Matthew Raine, Father of Adam Raine, Before the U.S. Senate Committee on the Judiciary, Subcommittee on Crime and Counterterrorism: “Examining the Harm of AI Chatbots”. U.S. Senate Committee on the Judiciary. https://w.judiciary.senate.gov/imo/media/doc/e2e8fc50-a9ac-05ec-edd7-277cb0afcdf2/2025-09-16%20PM%20-%20Testimony%20-%20Raine.pdf Raine v. OpenAI, Inc. (2025) Raine v. OpenAI, Inc. 2025. Complaint, Raine v. OpenAI, Inc., No. CGC-25-628528 (Cal. Super. Ct., San Francisco Cnty., filed Aug. 26, 2025). Superior Court of California, County of San Francisco. Filed by Matthew and Maria Raine, parents of 16-year-old Adam Raine. Razi et al. (2023) Afsaneh Razi, Ashwaq Alsoubai, Seunghyun Kim, Shiza Ali, Gianluca Stringhini, Munmun De Choudhury, and Pamela J. Wisniewski. 2023. Sliding into My DMs: Detecting Uncomfortable or Unsafe Sexual Risk Experiences within Instagram Direct Messages Grounded in the Perspective of Youth. Proc. ACM Hum.-Comput. Interact. 7, CSCW1, Article 89 (April 2023), 29 pages. doi:10.1145/3579522 Razi et al. (2020) Afsaneh Razi, Karla Badillo-Urquiola, and Pamela J. Wisniewski. 2020. Let’s Talk about Sext: How Adolescents Seek Support and Advice about Their Online Sexual Experiences. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems. ACM, 1–13. doi:10.1145/3313831.3376400 Razi et al. (2021) Afsaneh Razi, Seunghyun Kim, Ashwaq Alsoubai, Gianluca Stringhini, Thamar Solorio, Munmun De Choudhury, and Pamela J. Wisniewski. 2021. A Human-Centered Systematic Literature Review of the Computational Approaches for Online Sexual Risk Detection. Proceedings of the ACM on Human-Computer Interaction 5, CSCW2 (2021), 465:1–465:38. doi:10.1145/3479609 Reeves and Nass (1996) Byron Reeves and Clifford Nass. 1996. The Media Equation: How People Treat Computers, Television, and New Media Like Real People and Places. Cambridge University Press, New York, NY, USA. Reuters (2025) Reuters. 2025. Spanish Teenager Investigated Over AI Naked Images of Classmates. Reuters (via U.S. News & World Report). https://w.usnews.com/news/world/articles/2025-07-27/spanish-teenager-investigated-over-ai-naked-images-of-classmates Reyes et al. (2018) Irwin Reyes, Primal Wijesekera, Joel Reardon, Amit Elazari Bar On, Abbas Razaghpanah, Narseo Vallina-Rodriguez, and Serge Egelman. 2018. “Won’t Somebody Think of the Children?” Examining COPPA Compliance at Scale. Proceedings on Privacy Enhancing Technologies 2018, 3 (2018), 63–83. Roffo et al. (2019) Giorgio Roffo, Dong-Bach Vo, Mohammad Tayarani, Maki Rooksby, Alessandra Sorrentino, Simona Di Folco, Helen Minnis, Stephen Brewster, and Alessandro Vinciarelli. 2019. Automating the Administration and Analysis of Psychiatric Tests: The Case of Attachment in School Age Children. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems (Glasgow, Scotland Uk) (CHI ’19). Association for Computing Machinery, New York, NY, USA, 1–12. doi:10.1145/3290605.3300825 Röttger et al. (2024) Paul Röttger, Hannah Rose Kirk, Bertie Vidgen, Giuseppe Attanasio, Federico Bianchi, and Dirk Hovy. 2024. XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL 2024). Ryan and Deci (2000) Richard M. Ryan and Edward L. Deci. 2000. Self-Determination Theory and the Facilitation of Intrinsic Motivation, Social Development, and Well-Being. American Psychologist 55, 1 (2000), 68–78. Saeri et al. (2025) Alexander K. Saeri, Sophia Lloyd George, Jess Graham, Clelia D. Lacarriere, Peter Slattery, Michael Noetel, and Neil Thompson. 2025. Mapping AI Risk Mitigations: Evidence Scan and Preliminary AI Risk Mitigation Taxonomy. MIT AI Risk Initiative; 831 mitigations from 13 frameworks organized into 4 categories and 23 subcategories. arXiv:2512.11931 [cs.CY] https://arxiv.org/abs/2512.11931 Saltzer and Schroeder (1975) Jerome H. Saltzer and Michael D. Schroeder. 1975. The Protection of Information in Computer Systems. Proc. IEEE 63, 9 (1975), 1278–1308. Sawyer et al. (2018) Susan M. Sawyer, Peter S. Azzopardi, Dakshitha Wickremarathne, and George C. Patton. 2018. The Age of Adolescence. The Lancet Child & Adolescent Health 2, 3 (2018), 223–228. doi:10.1016/S2352-4642(18)30022-1 Saxena et al. (2020) Devansh Saxena, Karla Badillo-Urquiola, Pamela J. Wisniewski, and Shion Guha. 2020. A Human-Centered Review of Algorithms Used within the U.S. Child Welfare System. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems (CHI ’20). ACM, 1–15. doi:10.1145/3313831.3376229 Schiavo and Andrao (2026) Gianluca Schiavo and Margherita Andrao. 2026. Talking About Brainrot: Youth Engagement with AI-Generated Content and the Dynamics of Intergenerational Communication. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26). Association for Computing Machinery, New York, NY, USA, Article 343, 13 pages. doi:10.1145/3772318.3791483 Selbst et al. (2019) Andrew D. Selbst, Danah Boyd, Sorelle A. Friedler, Suresh Venkatasubramanian, and Janet Vertesi. 2019. Fairness and Abstraction in Sociotechnical Systems. In Proceedings of the Conference on Fairness, Accountability, and Transparency (FAT* ’19). ACM, 59–68. Seo et al. (2025) Woosuk Seo, Young-Ho Kim, Ji Eun Kim, Megan Tao Fan, Mark S. Ackerman, Sung Won Choi, and Sun Young Park. 2025. Enhancing Pediatric Communication: The Role of an AI-Driven Chatbot in Facilitating Child-Parent-Provider Interaction. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery, New York, NY, USA, Article 1079, 16 pages. doi:10.1145/3706598.3713134 Seo et al. (2024) Woosuk Seo, Chanmo Yang, and Young-Ho Kim. 2024. ChaCha: Leveraging Large Language Models to Prompt Children to Share Their Emotions about Personal Events. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems (Honolulu, HI, USA) (CHI ’24). Association for Computing Machinery, New York, NY, USA, Article 903, 20 pages. doi:10.1145/3613904.3642152 Sharma et al. (2024) Mrinank Sharma, Meg Tong, Tomasz Korbak, David Duvenaud, Amanda Askell, Samuel R. Bowman, et al. 2024. Towards Understanding Sycophancy in Language Models. In Proceedings of the 12th International Conference on Learning Representations (ICLR 2024). Shashidhara et al. (2026) Sneha Shashidhara, Vivienne Bihe Chi, Abhay P Singh, Lyle Ungar, and Sharath Chandra Guntuku. 2026. Voice-Based Chatbots for English Speaking Practice in Multilingual Low-Resource Indian Schools: A Multi-Stakeholder Study. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26). Association for Computing Machinery, New York, NY, USA, Article 91, 19 pages. doi:10.1145/3772318.3791266 Shelby et al. (2023) Renee Shelby, Shalaleh Rismani, Kathryn Henne, AJung Moon, Negar Rostamzadeh, Paul Nicholas, N’Mah Yilla-Akbari, Jess Gallegos, Andrew Smart, Emilio Garcia, and Gurleen Virk. 2023. Sociotechnical Harms of Algorithmic Systems: Scoping a Taxonomy for Harm Reduction. In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society (AIES). ACM, 723–741. doi:10.1145/3600211.3604673 Shen et al. (2025) Jocelyn J Shen, Jennifer King Chen, Leah Findlater, and Griffin Dietz Smith. 2025. eaSEL: Promoting Social-Emotional Learning and Parent-Child Interaction through AI-Mediated Content Consumption. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery, New York, NY, USA, Article 1082, 18 pages. doi:10.1145/3706598.3713405 Shostack (2014) Adam Shostack. 2014. Threat Modeling: Designing for Security. Wiley, Indianapolis, IN, USA. Si et al. (2026) Janice Jianing Si, Donglin Li, Qiuning Wang, Alicia Wanyi Liu, Xin Lin, Yujun Zhu, Xiaobo Zhou, April Yi Wang, and Kanye Ye Wang. 2026. Exploring the Impacts and Challenges of Vibe Coding Paradigm to Children’s Programming Learning and Practices. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26). Association for Computing Machinery, New York, NY, USA, Article 130, 24 pages. doi:10.1145/3772318.3790366 Slattery et al. (2024) Peter Slattery, Alexander K. Saeri, Emily A. C. Grundy, Jess Graham, Michael Noetel, Risto Uuk, James Dao, Soroush Pour, Stephen Casper, and Neil Thompson. 2024. The AI Risk Repository: A Comprehensive Meta-Review, Database, and Taxonomy of Risks from Artificial Intelligence. arXiv:2408.12622 [cs.AI] https://arxiv.org/abs/2408.12622 Solarova et al. (2026) Sara Solarova, Matej Mosnar, Matus Tibensky, Jan Jakubcik, Adrian Bindas, Simon Liska, Filip Hossner, Matúš Mesarčík, and Ivan Srba. 2026. The DSA’s Blind Spot: Algorithmic Audit of Advertising and Minor Profiling on TikTok. In Proceedings of the 2026 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’26). Association for Computing Machinery, New York, NY, USA, 4811–4835. doi:10.1145/3805689.3812355 Solove (2006) Daniel J. Solove. 2006. A Taxonomy of Privacy. University of Pennsylvania Law Review 154, 3 (2006), 477–564. Solyst et al. (2025) Jaemarie Solyst, Cindy Peng, Wesley Hanwen Deng, Praneetha Pratapa, Amy Ogan, Jessica Hammer, Jason Hong, and Motahhare Eslami. 2025. Investigating Youth AI Auditing. In Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’25). Association for Computing Machinery, New York, NY, USA, 2098–2111. doi:10.1145/3715275.3732142 Solyst et al. (2023) Jaemarie Solyst, Ellia Yang, Shixian Xie, Amy Ogan, Jessica Hammer, and Motahhare Eslami. 2023. The Potential of Diverse Youth as Stakeholders in Identifying and Mitigating Algorithmic Bias for a Future of Fairer AI. Proc. ACM Hum.-Comput. Interact. 7, CSCW2, Article 364 (Oct. 2023), 27 pages. doi:10.1145/3610213 Song et al. (2026) Junrong Song, Huanyi Wan, Lei Han, Yinghao Gao, David Yip, and Xin Tong. 2026. CoStage: An Embodied AI Co-Creation System for Children’s Performative Storytelling with Robots. In Proceedings of the 2026 Designing Interactive Systems Conference (DIS ’26). Association for Computing Machinery, New York, NY, USA, 3780–3803. doi:10.1145/3800645.3813101 Stapleton et al. (2022) Logan Stapleton, Min Hun Lee, Diana Qing, Marya Wright, Alexandra Chouldechova, Ken Holstein, Zhiwei Steven Wu, and Haiyi Zhu. 2022. Imagining New Futures Beyond Predictive Systems in Child Welfare: A Qualitative Study with Impacted Stakeholders. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’22). ACM, 1162–1177. doi:10.1145/3531146.3533177 Steinberg (2005) Laurence Steinberg. 2005. Cognitive and Affective Development in Adolescence. Trends in Cognitive Sciences 9, 2 (2005), 69–74. doi:10.1016/j.tics.2004.12.005 Stoilova et al. (2020) Mariya Stoilova, Sonia Livingstone, and Rishita Nandagiri. 2020. Digital by Default: Children’s Capacity to Understand and Manage Online Data and Privacy. Media and Communication 8, 4 (2020), 197–207. Sun et al. (2026) Kaiwen Sun, Jade Xiaoyi Li, Irene Chung, Jenny Radesky, Jason Yip, Christopher Brooks, and Florian Schaub. 2026. "Families are messy": From Parent-Child Tensions to Family-Centered Design of Smart Home Technologies. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26). Association for Computing Machinery, New York, NY, USA, Article 656, 21 pages. doi:10.1145/3772318.3790865 Tan et al. (2024) Mei Tan, Hansol Lee, Dakuo Wang, and Hari Subramonyam. 2024. Is a Seat at the Table Enough? Engaging Teachers and Students in Dataset Specification for ML in Education. Proc. ACM Hum.-Comput. Interact. 8, CSCW1, Article 81 (April 2024), 32 pages. doi:10.1145/3637358 Thomas et al. (2021) Kurt Thomas, Devdatta Akhawe, Michael Bailey, Dan Boneh, Elie Bursztein, Sunny Consolvo, Nicola Dell, Zakir Durumeric, Patrick Gage Kelley, Deepak Kumar, Damon McCoy, Sarah Meiklejohn, Thomas Ristenpart, and Gianluca Stringhini. 2021. SoK: Hate, Harassment, and the Changing Landscape of Online Abuse. In Proceedings of the 2021 IEEE Symposium on Security and Privacy (S&P). IEEE, 247–267. doi:10.1109/SP40001.2021.00028 Tokida et al. (2026) Satomi Tokida, Koki Usui, Godai Tanaka, Shin Osuga, Masanori Kayano, Yuta Itoh, and Yoshio Ishiguro. 2026. Design and Evaluation of a Photorealistic AI Virtual Peer in Elementary Collaborative Classroom. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26). Association for Computing Machinery, New York, NY, USA, Article 95, 15 pages. doi:10.1145/3772318.3791450 UN Committee on the Rights of the Child (2021) UN Committee on the Rights of the Child. 2021. General Comment No. 25 (2021) on Children’s Rights in Relation to the Digital Environment. Technical Report CRC/C/GC/25. United Nations. UNICEF (2021) UNICEF. 2021. Policy Guidance on AI for Children 2.0. Technical Report. UNICEF Office of Global Insight and Policy. https://w.unicef.org/innocenti/media/1341/file/UNICEF-Global-Insight-policy-guidance-AI-children-2.0-2021.pdf United Nations General Assembly (1989) United Nations General Assembly. 1989. Convention on the Rights of the Child. United Nations Treaty Series, vol. 1577, p. 3. Adopted 20 November 1989, entered into force 2 September 1990. Wang et al. (2021) Ge Wang, Jun Zhao, Max Van Kleek, and Nigel Shadbolt. 2021. Protection or Punishment? Relating the Design Space of Parental Control Apps and Perceptions about Them to Support Parenting for Online Safety. Proceedings of the ACM on Human-Computer Interaction 5, CSCW2 (2021), 343:1–343:26. doi:10.1145/3476084 Warford et al. (2022) Noel Warford, Tara Matthews, Kaitlyn Yang, Omer Akgul, Sunny Consolvo, Patrick Gage Kelley, Nathan Malkin, Michelle L. Mazurek, Manya Sleeper, and Kurt Thomas. 2022. SoK: A Framework for Unifying At-Risk User Research. In Proceedings of the 2022 IEEE Symposium on Security and Privacy (S&P). IEEE, 2344–2360. doi:10.1109/SP46214.2022.9833643 Wei et al. (2023) Alexander Wei, Nika Haghtalab, and Jacob Steinhardt. 2023. Jailbroken: How Does LLM Safety Training Fail?. In Advances in Neural Information Processing Systems 36 (NeurIPS 2023). Wei et al. (2025) Miranda Wei, Christina Yeung, Franziska Roesner, and Tadayoshi Kohno. 2025. "We’re utterly ill-prepared to deal with something like this": Teachers’ Perspectives on Student Generation of Synthetic Nonconsensual Explicit Imagery. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery, New York, NY, USA, Article 174, 18 pages. doi:10.1145/3706598.3713226 Weidinger et al. (2022) Laura Weidinger, Jonathan Uesato, Maribeth Rauh, Conor Griffin, Po-Sen Huang, John Mellor, Amelia Glaese, Myra Cheng, Borja Balle, Atoosa Kasirzadeh, Courtney Biles, Sasha Brown, Zac Kenton, Will Hawkins, Tom Stepleton, Abeba Birhane, Lisa Anne Hendricks, Laura Rimell, William Isaac, Julia Haas, Sean Legassick, Geoffrey Irving, and Iason Gabriel. 2022. Taxonomy of Risks Posed by Language Models. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency (FAccT). ACM, 214–229. doi:10.1145/3531146.3533088 Wen et al. (2025) Zikai Wen, Lanjing Liu, and Yaxing Yao. 2025. Families’ Vision of Generative AI Agents for Household Safety Against Digital and Physical Threats. Proc. ACM Hum.-Comput. Interact. 9, 7, Article CSCW417 (Oct. 2025), 24 pages. doi:10.1145/3757598 Wisniewski (2018) Pamela Wisniewski. 2018. The Privacy Paradox of Adolescent Online Safety: A Matter of Risk Prevention or Risk Resilience? IEEE Security & Privacy 16, 2 (2018), 86–90. doi:10.1109/MSP.2018.1870874 Wisniewski et al. (2017a) Pamela Wisniewski, Arup Kumar Ghosh, Heng Xu, Mary Beth Rosson, and John M. Carroll. 2017a. Parental Control vs. Teen Self-Regulation: Is There a Middle Ground for Mobile Online Safety?. In Proceedings of the 2017 ACM Conference on Computer Supported Cooperative Work and Social Computing (CSCW). ACM, 51–69. doi:10.1145/2998181.2998352 Wisniewski et al. (2017b) Pamela Wisniewski, Heng Xu, Mary Beth Rosson, and John M. Carroll. 2017b. Parents Just Don’t Understand: Why Teens Don’t Talk to Parents about Their Online Risk Experiences. In Proceedings of the 2017 ACM Conference on Computer Supported Cooperative Work and Social Computing (CSCW). ACM, 523–540. doi:10.1145/2998181.2998236 Wisniewski et al. (2016) Pamela Wisniewski, Heng Xu, Mary Beth Rosson, Daniel F. Perkins, and John M. Carroll. 2016. Dear Diary: Teens Reflect on Their Weekly Online Risk Experiences. In Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems. ACM, 3919–3930. doi:10.1145/2858036.2858317 Wood et al. (1976) David Wood, Jerome S. Bruner, and Gail Ross. 1976. The Role of Tutoring in Problem Solving. Journal of Child Psychology and Psychiatry 17, 2 (1976), 89–100. Wu et al. (2026) Jiazhou Wu, Dasom Choi, Bogoan Kim, and Hwajung Hong. 2026. When Special Education Meets LLMs: Investigating the use of LLM-based Conversational AI by Special Education Teachers for Autism in China. In Proceedings of the 2026 Designing Interactive Systems Conference (DIS ’26). Association for Computing Machinery, New York, NY, USA, 2392–2408. doi:10.1145/3800645.3812968 Xiao et al. (2026) Ruiwei Xiao, Qing Xiao, Xinying Hou, Phenyo Phemelo Moletsane, Hanqi Jane Li, Hong Shen, and John Stamper. 2026. Do Teachers Dream of GenAI Widening Educational (In)equality? Envisioning the Future of K-12 GenAI Education from Global Teachers’ Perspectives. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26). Association for Computing Machinery, New York, NY, USA, Article 530, 21 pages. doi:10.1145/3772318.3790908 Xu et al. (2022) Ying Xu, Valery Vigil, Andres S. Bustamante, and Mark Warschauer. 2022. "Elinor’s Talking to Me!":Integrating Conversational AI into Children’s Narrative Science Programming. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems (New Orleans, LA, USA) (CHI ’22). Association for Computing Machinery, New York, NY, USA, Article 166, 16 pages. doi:10.1145/3491102.3502050 Yang et al. (2026a) Joshua Kong Yang, Mackenzie Leake, Jeff Huang, and Stephen DiVerdi. 2026a. VidSTR: Automatic Spatiotemporal Retargeting of Speech-Driven Video Compositions. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26). Association for Computing Machinery, New York, NY, USA, Article 1142, 17 pages. doi:10.1145/3772318.3790709 Yang et al. (2023) Kexin Bella Yang, Vanessa Echeverria, Zijing Lu, Hongyu Mao, Kenneth Holstein, Nikol Rummel, and Vincent Aleven. 2023. Pair-Up: Prototyping Human-AI Co-orchestration of Dynamic Transitions between Individual and Collaborative Learning in the Classroom. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (Hamburg, Germany) (CHI ’23). Association for Computing Machinery, New York, NY, USA, Article 453, 17 pages. doi:10.1145/3544548.3581398 Yang et al. (2026b) Migyeong Yang, Kyungah Lee, Jinyoung Han, SoHyun Park, and Young-Ho Kim. 2026b. Autiverse: Eliciting Autistic Adolescents’ Daily Narratives through AI-guided Multimodal Journaling. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26). Association for Computing Machinery, New York, NY, USA, Article 999, 26 pages. doi:10.1145/3772318.3791381 Yin et al. (2025) Yaxuan Yin, Shamya Karumbaiah, and Shona Acquaye. 2025. Responsible AI in Education: Understanding Teachers’ Priorities and Contextual Challenges. In Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’25). Association for Computing Machinery, New York, NY, USA, 2705–2727. doi:10.1145/3715275.3732176 Yin et al. (2026) Yaxuan Yin, Shamya Karumbaiah, Yujin Kim, and Devansh Saxena. 2026. When AI Breaks, Teachers Repair: Pedagogical Repair Work in Situated Classroom Practice. Association for Computing Machinery, New York, NY, USA, 2206–2224. https://doi.org/10.1145/3805689.3812281 Yip et al. (2026) Jason Yip, Michele Newman, Runhua Zhao, Darae Kim, Jan R. Lim, Matthew Kyle Pedraja, Swati Sachdeva, Xiaoyu Zheng, Yifang Zhou, Chris Coward, and Jin Ha Lee. 2026. Rethinking Misinformation: A Holistic Community Model for Youth Resilience through Socioemotional Learning and Sociocultural Design. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26). Association for Computing Machinery, New York, NY, USA, Article 46, 21 pages. doi:10.1145/3772318.3790876 Yoo et al. (2026a) Daeun Yoo, Daniela E Munoz Lopez, Xiaotian Daisy Hu, Jason C. Yip, and Katie Davis. 2026a. Generative AI and Creative Mediums for Youth’s Emotion Regulation: An Interview Study with Clinicians. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26). Association for Computing Machinery, New York, NY, USA, Article 97, 18 pages. doi:10.1145/3772318.3790909 Yoo et al. (2026b) Daeun Yoo, Michele Newman, Caroline Pitt, Kevin Huu Vo, Michelle Zhang, Michelle Kim, Katie Davis, and Jason C. Yip. 2026b. Generative AI in Children’s Creative Collaboration: Impact, Perception, and Design Guidelines. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26). Association for Computing Machinery, New York, NY, USA, Article 98, 23 pages. doi:10.1145/3772318.3790907 Yu et al. (2025a) Yaman Yu, Yiren Liu, Jacky Zhang, Yun Huang, and Yang Wang. 2025a. Understanding Generative AI Risks for Youth: A Taxonomy Based on Empirical Data. arXiv preprint arXiv:2502.16383 (2025). A version appears in Proceedings of the Association for Information Science and Technology (ASIS&T), 2025, doi:10.1002/pra2.1526. Yu et al. (2025b) Yaman Yu, Yiren Liu, Jacky Zhang, Yun Huang, and Yang Wang. 2025b. Youth-Centered GAI Risks (YAIR): A Taxonomy of Generative AI Risks from Empirical Data. In Twenty-First Symposium on Usable Privacy and Security (SOUPS 2025). USENIX Association, Seattle, WA, 149–165. https://w.usenix.org/conference/soups2025/presentation/yu Yu et al. (2025c) Yaman Yu, Yiren Liu, Yuqi Zhang, Yun Huang, and Yang Wang. 2025c. YouthSafe: A Youth-Centric Safety Benchmark and Safeguard Model for Large Language Models. In Proceedings of the 2025 ACM SIGSAC Conference on Computer and Communications Security (CCS ’25). ACM, 4349–4363. Best existing safeguard (Aegis-Guard-D) reaches recall of 0.66 on youth–GenAI conversations, i.e., misses roughly one in three risky interactions; OpenAI Moderation API recall 0.05, Llama Guard 3 recall 0.08. doi:10.1145/3719027.3765168 Yu et al. (2026) Yaman Yu, Fnu Mohi, Aishi Debroy, Xin Cao, Karen Rudolph, and Yang Wang. 2026. Principles of Safe AI Companions for Youth: Parent and Expert Perspectives. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems. 1–21. Zaidi and Karahalios (2025) Ali Zaidi and Karrie Karahalios. 2025. From Sociotechnical Gaps to Solutions: Designing AI Tools with Parents to Address Special Education Advocacy Barriers in IEP Processes. In Proceedings of the 2025 ACM Designing Interactive Systems Conference (DIS ’25). Association for Computing Machinery, New York, NY, USA, 2619–2636. doi:10.1145/3715336.3735778 Zha et al. (2025) Siyu Zha, Yujia Liu, Chengbo Zheng, Jiaqi Xu, Fuze Yu, Jiangtao Gong, and Yingqing Xu. 2025. Mentigo: An Intelligent Agent for Mentoring Students in the Creative Problem Solving Process. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery, New York, NY, USA, Article 199, 22 pages. doi:10.1145/3706598.3713952 Zhang et al. (2024) Chao Zhang, Xuechen Liu, Katherine Ziska, Soobin Jeon, Chi-Lin Yu, and Ying Xu. 2024. Mathemyths: Leveraging Large Language Models to Teach Mathematical Language through Child-AI Co-Creative Storytelling. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems (Honolulu, HI, USA) (CHI ’24). Association for Computing Machinery, New York, NY, USA, Article 274, 23 pages. doi:10.1145/3613904.3642647 Zhang et al. (2025) Renwen Zhang, Han Li, Han Meng, Jinyuan Zhan, Hongyuan Gan, and Yi-Chieh Lee. 2025. The Dark Side of AI Companionship: A Taxonomy of Harmful Algorithmic Behaviors in Human-AI Relationships. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). ACM. Zhang et al. (2026) Zinan Zhang, Qiurong Song, Rie Helene (Lindy) Hernandez, Yunhan Liu, Elena Koung, Junnan Yu, Sunhye Bai, Yubo Kou, and Xinning Gui. 2026. Player Safety by Design: Co-Designing Child-Centered Safety Mechanisms with Children. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26). Association for Computing Machinery, New York, NY, USA, Article 1702, 21 pages. doi:10.1145/3772318.3791090 Zhang et al. (2022) Zheng Zhang, Ying Xu, Yanhao Wang, Bingsheng Yao, Daniel Ritchie, Tongshuang Wu, Mo Yu, Dakuo Wang, and Toby Jia-Jun Li. 2022. StoryBuddy: A Human-AI Collaborative Chatbot for Parent-Child Interactive Storytelling with Flexible Parental Involvement. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems (New Orleans, LA, USA) (CHI ’22). Association for Computing Machinery, New York, NY, USA, Article 218, 21 pages. doi:10.1145/3491102.3517479 Zhao et al. (2025) Wenxin Zhao, Fangyu Yu, Peng Zhang, Hansu Gu, Lin Wang, Siyuan Qiao, Tun Lu, and Ning Gu. 2025. YouthCare: Building a Personalized Collaborative Video Censorship Tool to Support Parent-Child Joint Media Engagement. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery, New York, NY, USA, Article 1042, 20 pages. doi:10.1145/3706598.3713360 Zheng et al. (2026) Zhi Zheng, Chun Yu, Weihao Chen, Minzheng Song, Binglin Liu, Jianyang Liu, Shiyi Wang, Xutong Wang, Jie Cai, and Yuanchun Shi. 2026. OpenCD: Empowering Diagnosis of Children’s Mathematical Cognition through Open-ended Multimodal Tasks. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems. 1–31. Appendix A Working Definitions of Broad Risk and Countermeasure Categories Table 6. Working definitions of the broad risk categories and countermeasure domains used in this review. Type Category/Domain Working definition Risk Behavioral and Social Developmental Risks that interfere with young people’s learning, reasoning, social interaction, family relationships, or development of appropriate social boundaries. Risk Mental Wellbeing Risks to young people’s emotional wellbeing, including overreliance, emotional dependency, distress, and inadequate support during vulnerable situations. Risk Toxicity Exposure to harmful, inappropriate, threatening, sexual, violent, or otherwise age-inappropriate content. Risk Misuse and Exploitation Risks arising when AI is used to harm, manipulate, deceive, coerce, or exploit youth, or when protections are circumvented. Risk Bias/Discrimination Risks involving stereotyping, distorted representation, unequal access or treatment, and normative judgments that disadvantage particular youth or groups. Risk Privacy Risks involving inappropriate collection, processing, disclosure, or representation of information about young people. Risk Control, Agency, and Power Risks that limit young people’s ability to make choices, influence decisions, control their expression, or challenge decisions made by adults, institutions, or AI systems. Risk Usability, Accessibility, and Digital Exclusion Risks that arise when AI systems do not fit children’s developmental, linguistic, cognitive, physical, or accessibility needs. Risk System Reliability Risks arising when AI systems make unreliable assessments, miss important situations, intervene incorrectly, or provide assistance that users cannot safely verify. Countermeasure Technical and Security Controls Changes to the AI model or system that constrain generation, adapt safeguards, manage assistance, detect risks, or limit harmful interaction patterns. Countermeasure Operational Process Controls Processes surrounding development and deployment, including human review, auditing, data management, monitoring, and procedures for responding when problems are detected. Countermeasure Transparency and Accountability Controls Mechanisms that help youth and other affected stakeholders understand, question, correct, control, or seek recourse from AI decisions and behaviors. Countermeasure Governance and Oversight Controls Organizational and institutional measures that determine who participates in decisions, how safety requirements are enforced, whether systems should be deployed, and who has the authority and responsibility to address harms.