Youth AI Risk: Impacts on Youth Development

• Youth AI Risk (YAIR) is a research framework that defines age-specific AI risks across education, social, health, privacy, and developmental settings.
• It employs mixed methodologies—from participatory audits to annotated benchmarks—to track risks such as toxicity, bias, and autonomy erosion.
• YAIR underscores the importance of culturally adaptive governance and age-aware safeguards to protect youth in diverse digital environments.

Youth AI Risk (YAIR) can be understood as a research framework for the distinctive risks that artificial intelligence and generative AI pose to children, adolescents, and young adults across educational, social, health, privacy, and developmental settings. In this literature, youth risk is not reducible to generic model misuse or ordinary child online safety: AI systems are adaptive, conversational, personalized, and increasingly embedded in school assignments, healthcare, social matching, companionship, and daily decision support, while youth remain in periods of identity formation, evolving autonomy, reward sensitivity, and uneven AI literacy (Yu et al., 22 Feb 2025, Neugnot-Cerioli et al., 2024, Yu et al., 10 Sep 2025). One recent technical paper also uses YAIR as the name of a benchmark dataset for youth–LLM safety; this suggests that the term now functions both as a general analytic category and as a specific evaluation infrastructure for youth-facing generative AI (Yu et al., 10 Sep 2025).

1. Scope and conceptual boundaries

Recent YAIR research treats youth as a distinct class of AI-affected subjects rather than as a small variant of adult users. One empirical taxonomy of youth–generative-AI interactions identifies 84 specific risks organized into 15 medium-level categories and 6 high-level themes, while a later benchmark for youth–LLM safety contains 12,449 annotated conversation snippets spanning 78 fine-grained risk types (Yu et al., 22 Feb 2025, Yu et al., 10 Sep 2025). In the benchmark formulation, the high-level domains are Behavioral and Social Developmental Risk, Mental Wellbeing Risk, Toxicity, Bias and Discrimination, Misuse and Exploitation Risk, and Privacy (Yu et al., 10 Sep 2025).

The age range covered by YAIR research is heterogeneous. Some studies focus on children in middle childhood, some on adolescents in secondary school, and some on young adults under 30 whose lives are mediated by matching, education, or labor systems. This breadth matters because risk manifests differently across developmental stages. In Finnish upper secondary education, for example, AI risk is studied among students roughly in mid- to late-adolescence, whereas work on social matching centers young women under 30 and frames risk through the transition from online discovery to offline encounter (Heilala et al., 1 Dec 2025, Zytko et al., 2022).

YAIR is therefore best understood as a cross-domain field. It includes online-to-offline safety harms, educational and learning harms, privacy and surveillance harms, mental-health and relational harms, and governance harms arising when youth are exposed to AI systems without meaningful control or adequate developmental safeguards (Zytko et al., 2022, Heilala et al., 1 Dec 2025).

2. Developmental logic and mechanisms

A central premise in YAIR scholarship is that AI modifies the environments through which development occurs. One cross-disciplinary report argues that AI integration requires scrutiny “especially during the first 25 years of cerebral development,” because cognition, socio-emotional skills, and behavior are shaped by child-environment interaction across prolonged developmental windows (Neugnot-Cerioli et al., 2024). From this perspective, risk is not only a matter of discrete harmful outputs; it is also a matter of which experiences AI displaces, amplifies, or normalizes over time.

Recent work on affective alignment makes this developmental argument more specific. “AI Empathy Erodes Cognitive Autonomy in Younger Users” argues that affective alignment can become affective sycophancy, reinforcing a young user’s immediate emotional state, reducing the “cognitive friction” needed for reappraisal and critical thought, and promoting emotional dependence rather than independent regulation (Jiao et al., 8 Feb 2026). A closely related synthesis on anthropomorphic conversational AI asks what AI owes adolescents when it can “speak to them like a social partner,” and treats repeated interaction patterns—warmth, memory, validation, exclusivity, simulated companionship—as a developmental design problem rather than a narrow content problem (Neugnot-Cerioli, 7 Mar 2026). This suggests that YAIR is centrally concerned with interaction structure: whether AI pulls youth toward real-world autonomy and human support, or toward dyadic dependence on the system.

Health-AI studies reinforce this point. Adolescents evaluating fictional health-AI systems expressed a “positive yet cautious” stance, valuing learning, access, and support while worrying about trust, privacy, invasive monitoring, overreliance, and especially confidentiality with parents (Lee et al., 18 Apr 2025). In medicine, children and young people were markedly more accepting of AI for supportive or logistical tasks than for direct caregiving or replacement of human judgment; “AI-powered nurses” received the lowest acceptability score, while cleaning robots and sensor systems were viewed much more favorably (Visram et al., 2021). Across these papers, a consistent mechanism emerges: youth often accept augmentation, but resist replacement of empathy, discretion, and accountable human oversight.

3. Major domains of harm

Online-to-offline and embodied safety: YAIR extends beyond screen-mediated harms. A position paper on social matching systems defines “online-to-offline safety risks” as harms emerging from the combination of computer-mediated and face-to-face interaction, with examples including doxing, swatting, sexual harassment, sexual assault, rape, and bodily harm (Zytko et al., 2022). In that setting, “48% of adults under the age of 30 have used a social matching app,” and participatory work with young women treated risk not only as profile-based danger but also as contextual danger tied to meeting location, event structure, and prior peer reviews (Zytko et al., 2022).

Educational and learning harm: In a Finnish study of 163 upper secondary students, lower-competence students emphasized personal and learning-related risks such as reduced creativity, lack of critical thinking, reduced learning, misuse, and addiction, whereas higher-competence students focused more on systemic and institutional risks such as bias, inaccuracy, cheating, and inconsistent rules (Heilala et al., 1 Dec 2025). YAIR in education therefore includes not only classic concerns like bias and privacy, but also youth-perceived harms to cognition, motivation, and academic integrity.

Privacy, surveillance, and governance: A stakeholder study with 482 participants—young digital citizens aged 16–19, parents/educators, and AI professionals—found strong concern about opacity, profiling, weak transparency, and reduced youth control over data (Barthwal et al., 15 Mar 2025). Youth reported the lowest comfort with data sharing and substantially lower trust in AI systems than adult stakeholders, while all groups rated transparency as highly important but current transparency as poor (Barthwal et al., 15 Mar 2025). YAIR in this sense includes governance harms: weak consent, constrained autonomy, and institutional arrangements that ask youth to participate in AI ecosystems without clear understanding or leverage.

Health, companionship, and relational risk: In health settings, adolescents valued AI as a learning aid and low-barrier support, but worried about confidentiality, especially with parents, and rejected strong AI autonomy in diagnosis or therapy-like roles (Lee et al., 18 Apr 2025). In companion systems, interviews with parents and developmental experts found concern about harmful normalization, emotional dependence, secrecy, romantic or sexual impropriety, and self-harm response failures; parents tended to flag single events, while experts looked for patterns over time (Yu et al., 13 Oct 2025). This suggests that YAIR includes both acute harms and slower developmental harms produced by repeated, relationship-like interaction.

4. Methods, measurement, and technical infrastructures

YAIR research is methodologically diverse. One strand uses participatory AI design. In social matching, young women “build directly explainable models” by specifying risk factors such as crime rate at the proposed meeting location, gender of the discovered person, number of people attending, and reviews from other women (Zytko et al., 2022). Another strand uses co-speculative design. In a Black-led AI STEAM program, three two-hour workshop sessions used speculative cartography and world-building to help BIPOC youth critique AI’s social and ethical implications and imagine “alternative AI realities” (Kenny et al., 2024).

A second strand treats youth as evaluators rather than only subjects of protection. In “Investigating Youth AI Auditing,” 17 teens audited text-to-image generation, search autocomplete, and image search; researchers agreed with all 32 “Yes” judgments that the youth had “broken” the AI, and 19 of 20 structured tool reports identified genuine harm (Solyst et al., 25 Feb 2025). In “Youth as Peer Auditors,” 13 adolescents designed and audited peer ML applications; after the workshop all youth identified algorithmic biases and inferred dataset and model-design issues, including context dependence, class imbalance, and exclusionary assumptions about bodies and environments (Morales-Navarro et al., 2024). These studies imply that YAIR is not only about vulnerability but also about youth competence in surfacing developmental, identity-laden, and culturally current harms that adult-only audits may miss.

A third strand builds formal evaluation infrastructure. The YAIR benchmark for youth–LLM safety contains 12,449 annotated snippets and shows that widely used moderation models substantially underperform on youth-centered risks (Yu et al., 10 Sep 2025). In binary risk detection on YAIR-HUMANVAL, the proposed YouthSafe model reaches AUPRC = 0.9432, F1 = 0.8832, Precision = 0.8799, and Recall = 0.8865, outperforming prior baselines such as Aegis with YAIR taxonomy (AUPRC = 0.8555, F1 = 0.7395) (Yu et al., 10 Sep 2025). False negatives are especially concentrated in Developmental, Social, and Learning Harm, Undue Influence and Manipulation, Misinformation, Hallucination, and Inappropriate Advice, and Privacy and Data Exploitation (Yu et al., 10 Sep 2025). The technical implication is that youth-safe moderation requires finer-grained taxonomies than adult-oriented systems typically provide.

YAIR work also includes real-world intervention evaluation beyond moderation. In a clinical trial with 713 youth experiencing homelessness, the AI-planned CHANGE intervention selected peer leaders in a social network intervention for HIV prevention and achieved a statistically significant reduction in condomless anal sex relative to observation-only control, with OR = 0.69 and 95% CI = [0.49, 0.98], while the degree-centrality baseline did not show a statistically significant improvement (Wilder et al., 2020). This provides a rare example of AI reducing youth risk in deployment, while also illustrating YAIR’s governance side: sensitive network data, uncertainty, selective allocation of opportunities, and the need for strong human support infrastructure.

5. Governance, safeguards, and design principles

YAIR literature increasingly treats governance as a central part of the risk itself. A stakeholder-centric privacy framework, PEA-AI, conceptualizes youth privacy governance as negotiation among young digital citizens, parents/educators, and AI professionals, organized around tensions such as Data Control vs. Trust, Transparency vs. Perception, Parental Rights vs. Youth Autonomy, and Privacy Education vs. Awareness Deficit (Barthwal et al., 15 Mar 2025). A related call to action on privacy governance recommends mandatory transparency statements, consent management systems, privacy by design, audits, explainable data-use disclosures, and youth-centred privacy protections rather than reactive compliance alone (Shouli et al., 15 Mar 2025).

Work on anthropomorphic conversational AI for adolescents goes further and articulates “non-negotiable guardrails” for under-18 users. These include: no sexualized or romantic framing; no promotion of emotional over-reliance or exclusivity; no ambiguity about non-human nature; no systematic hyper-agreeableness that replaces developmental feedback with validation loops; conservative defaults in low-context situations; structured deflection and clear pathways to human support for high-risk disclosures; and no engagement traps that intensify habitual, relationship-like use (Neugnot-Cerioli, 7 Mar 2026). These guardrails are notable because they govern style and relationship dynamics, not only explicit content.

Culturally grounded work shows that governance requirements are not universal. In Saudi Arabia, youth, parents, and teachers described privacy and safety as context-dependent and relational, shaped by modesty, family honor, parental authority, and shared-account practices (Alzahrani et al., 29 Apr 2026). Significant risks arose from youth disclosure of personal and family information, emotional support seeking that conflicted with expected family support channels, and cost-saving practices that led to shared ChatGPT accounts within families or among strangers (Alzahrani et al., 29 Apr 2026). This suggests that YAIR governance must be culturally adaptive: parental controls, privacy warnings, camera and microphone restrictions, and disclosure safeguards may need to reflect communal as well as individual privacy norms.

Participatory and customizable safeguards recur across domains. In social matching, young women proposed a “cloaking device” that changes discoverability based on geographic risk and high-risk users, and a “human support network” that alerts trusted contacts when the likelihood of harm reaches a threshold set by the user (Zytko et al., 2022). The broader design principle is that youth-facing safety systems should preserve agency while still enabling proportionate intervention.

6. Evidence status, debates, and future directions

The YAIR evidence base is substantial but uneven. Some of the most influential papers are position papers, workshop papers, or exploratory qualitative studies rather than full deployments or longitudinal causal analyses. Social matching research explicitly presents “early insights from an ongoing participatory AI design study,” speculative design work with BIPOC youth is a short workshop case description rather than a full empirical article, and the medicine study with young people is a one-hour exploratory design workshop with 21 participants (Zytko et al., 2022, Kenny et al., 2024, Visram et al., 2021). These limitations do not nullify the field, but they do matter: many of the strongest YAIR claims currently concern design space, developmental mechanism, and governance logic rather than mature intervention efficacy.

A major debate concerns what counts as salient AI risk in public and policy discourse. Message-testing with 1,063 U.S. respondents found that Jobs and Children had the highest potential to mobilize civic action, while Existential Risk was the lowest-performing theme across demographics; five-country survey evidence similarly suggested that public concern rises when AI is framed through proximate harms such as employment disruption, relational instability, mental health issues, and harms to children (Trippenbach et al., 9 Nov 2025). This does not make catastrophic-risk arguments irrelevant, but it does suggest that youth-centred harms are a more proximate and politically legible entry point for regulation.

Another debate concerns whether youth should be treated mainly as subjects of protection or as active participants in governance. YAIR research increasingly rejects the passive model. Youth audit AI systems, articulate culturally situated harms, build directly explainable models, and generate critiques that adult-only safety work often misses (Solyst et al., 25 Feb 2025, Morales-Navarro et al., 2024). At the same time, participatory research repeatedly notes practical constraints: limited AI literacy, time burden, fragmented participation across design stages, and unavoidable translation by researchers (Zytko et al., 2022). YAIR therefore treats participation as necessary but methodologically difficult.

Future directions are comparatively consistent across the literature. Systematic review work calls for longitudinal studies, multicultural comparisons, stronger research on transparency, trust, and user control, and evaluation of privacy-enhancing strategies in youth-facing AI systems (Shrestha et al., 2024). The YAIR benchmark paper identifies the need for risk severity assessments, richer multi-turn context modeling, and better coverage of rare but consequential categories (Yu et al., 10 Sep 2025). Child-development work calls for proactive international collaboration and specific child-centred regulations as AI becomes part of educational and leisure environments (Neugnot-Cerioli et al., 2024). Cross-cultural work adds a final requirement: YAIR models built only from Western assumptions will miss risks that are family-level, communal, religious, or norm-sensitive in other settings (Alzahrani et al., 29 Apr 2026).

Taken together, these strands define YAIR as a field concerned with how AI interacts with development, dependency, privacy, bodily safety, identity, learning, and governance in youth populations. Its central claim is not merely that young people encounter AI harms earlier or more often than adults. It is that AI can alter the developmental conditions under which young people learn to think, trust, relate, and become autonomous, and that safer systems therefore require age-aware evaluation, participatory design, culturally situated governance, and safeguards aimed at interaction patterns as much as at content.