“Already outdated and still under review”: mapping the landscape of research on adolescent development and AI chatbot use

Abstract Artificial intelligence (AI) chatbots are increasingly embedded in adolescents’ daily lives, yet developmental science research struggles to keep pace with adoption. This mixed-methods study gathered developmental expert perspectives via a survey (n = 141 researchers studying youth and technology and adjacent professionals) and semi-structured interviews (n = 15) to inform a coordinated research and policy agenda on adolescent AI use. The study aims were to identify (1) priority AI use cases for future study, (2) priority research questions and methods, as well as barriers to future research, and (3) societal guardrails that developmental professionals recommend enacting now, while research is underway. Participants cautioned against vague “screen time” measures and agreed that nearly all use cases merit study but prioritized those with severe or irreversible immediate-term harms, AI literacy, AI as a social/relational actor, AI-generated (mis)information, and overreliance. Participants emphasized that methodological pluralism and triangulation, including within individual studies, are essential, agreeing that different methods can be complementary to address distinct research questions. Key barriers included the rapid pace of change in AI technology and use relative to academic timelines, limited industry transparency, funding gaps, and misaligned academic incentives. Proposed solutions included shared infrastructure, interdisciplinary consortia, and community partnerships. Participants underscored that responsibility for safe AI should not rest primarily on adolescents and families and identified three immediate guardrails: enforceable youth-centered regulation, developmentally informed design, and multidimensional AI literacy. Findings clarify a preliminary direction for the field and emphasize that coordination among researchers and between sectors is needed. 1 Introduction Artificial intelligence (AI) chatbots have rapidly become nearly ubiquitous among U.S. youth. ChatGPT launched as a consumer product in 2022, marking a turning point in public generative AI access. Today, more than 2 in 3 U.S. teens report using AI chatbots (Faverio and Sidoti 2025; O’Neil et al. 2026) and global adoption rates continue to climb (Microsoft 2026). Many U.S. preteens ages 8–12 also access generative AI chatbot applications (Maheux et al. 2026a; O’Neil et al. 2026). AI chatbots are defined broadly as systems using large language models and image-generation models trained on vast datasets of human-generated text and images to synthesize and predict coherent output, offering human-quality language synthesis and content generation (Ronge et al. 2025). As a scientific community, we know very little about how AI may influence adolescent development and the current literature remains slow to emerge, methodologically fragmented, and limited by inconsistent terminology (Maheux et al. 2026b). Parents, teachers, clinicians, policymakers, and adolescents themselves urgently need guidance on the developmental impacts of AI use. Rigorous developmental science research is necessary to provide that guidance, but its value will depend on coordination within and across researchers studying youth development and AI use. This includes a shared language and taxonomy of AI use cases (i.e., specific ways a user can engage with AI), a proactive and synchronized research agenda, and a set of empirically grounded protections that society can implement with confidence, while longer-term research is underway. The goal of the current study was to begin addressing this gap by systematically gathering expert input from social science researchers and adjacent professionals. The study has three main aims: (1) gather insight on AI use cases among adolescents to prioritize for future research, (2) identify research questions, methods, and barriers to research that can set the stage for the future study of adolescents and AI, and (3) propose guardrails, supported by professionals in developmental science and adjacent fields, to protect young people in the interim. 1.1 Adolescent development and AI Adolescence represents a developmental period marked by change in neurobiological, social, cognitive, and emotional systems. During this time, youth experience heightened sensitivity to social information and feedback, with peer relationships becoming increasingly salient, complex, and critical for skill-building (Dahl et al. 2018; Steinberg 2014). Simultaneously, youth undergo heightened sensitivity to reward as their behavioral and cognitive impulse control develops, creating a developmental context wherein inhibiting reward-seeking can be especially challenging (Albert et al. 2013; Crone and Dahl 2012; Somerville 2013). Given this drive for social engagement, reward, and novelty, adolescents and young adults are often the earliest adopters of new technology (Giedd 2020), and AI adoption is no different. AI use has skyrocketed among adolescents, with an estimated 86% of U.S. 9–17 years old now using AI, including 64–67% specifically using AI chatbots and 28% using chatbots every day (Faverio and Sidoti 2025; O’Neil et al. 2026). AI chatbots introduce novel opportunities and potential risks to an already complicated technological landscape. Many of the affordances of AI, and the downstream benefits or costs, may be uniquely salient during the adolescent developmental period (Maheux and Maes 2026). AI chatbot use is typically a solitary activity, yet can feel like interacting with a human, potentially activating adolescent drives for social engagement, acceptance, and intimacy. AI chatbots can respond contingently, retain memories, reflect affect, and be personalized in ways that make exchanges feel meaningful and intimate (Maheux and Maes 2026). Designer-modulated features, such as tone, persona, and communication style, further shape the character of these interactions, with the current models tending toward sycophancy and anthropomorphism. Because chatbots are non-sentient, they require nothing of the user, tolerate everything, and pass no judgment. These qualities may make AI chatbots particularly attractive to adolescents, who are particularly sensitive to peer rejection, learning to navigate mutual peer relationships, and still developing executive function skills in rewarding or arousing social contexts (American Psychological Association 2025; Hopelab et al. 2024). How youth use AI and the implications of such use remain poorly understood. A recent systematic review of peer-reviewed research on adolescents and AI use indicates an extremely thin evidence base dominated by research on AI in academics or seeking to understand factors promoting AI adoption among adolescents and young adults (Maheux et al. 2026b). Nationally representative polling data suggest common uses include to search for information, help with homework, for entertainment, and, for some youth, a simulated friend (Madden et al. 2024; McClain et al. 2026; Robb and Mann 2025). AI chatbots may serve important scaffolding functions for adolescent social, emotional, and cognitive developmental milestones—for example, by supporting learning and comprehension or providing accurate advice on navigating challenging social contexts (American Psychological Association 2025). However, AI chatbots may also undermine aspects of adolescent development in critical ways, including displacing human relationships, reducing skill-building by outsourcing cognitive or social tasks, and promoting harmful or dysfunctional beliefs about relationships (American Psychological Association 2025; Maheux and Maes 2026). Many stakeholders are concerned that, without proper guardrails—including a strong empirical base, developmental policies, design foundations, and interventions—AI chatbots will unfold as social media did: with little regard for youth well-being and potential harms among vulnerable users, and a slow roll-out of policy-relevant research lagging behind technological innovations (Davis et al. 2025). 1.2 The current study The field of research addressing the psychosocial impacts of adolescent AI use requires consensus and coordination to keep pace with rapid technological change. Research on adolescents and technology has historically been fragmented across disciplines, with psychologists, communication scholars, public health researchers, neuroscientists, and computer scientists often working in parallel rather than in dialogue (Davis et al. 2025). More broadly, researchers often remain separate from the sectors that design, govern, and implement emerging technologies, including industry, policymakers, educators, youth-serving organizations, and young people themselves. The challenges posed by AI will be best addressed by rigorous interdisciplinary research and sustained coordination between researchers and other sectors of society. Achieving provisional consensus on which AI use cases deserve research priority, which methodological approaches are most promising, and what barriers are impeding progress requires deliberate coordination among researchers. Moreover, researchers studying adolescents and technology already know much about ideal approaches for designing tools that support positive development. Thus, scientists can also offer guidance on what society should be doing now, based on the existing evidence from adjacent literatures, while longer-term research is underway. Initial consensus across these domains can provide a roadmap for future scholarship and collaborative society-wide action to support adolescent well-being in the age of AI. The current project aims to proactively address these issues. We identified individuals working for research and advocacy organizations to provide expert insight, leveraging both a large-scale mixed-methods survey of professionals across multiple disciplines and structured interviews with a purposive subsample. The goals of the current study were threefold: Aim 1: Develop a set of AI use cases among adolescents to prioritize studying in future research. Aim 2: Establish a provisional research agenda by identifying priority research questions and methodological approaches and documenting perceived barriers and solutions to progress in this field. Aim 3: Identify guardrails for society to consider implementing now, based on prior research on adolescents and technology use. 2 Methods This study employed a convergent mixed-methods design (Creswell and Clark 2017) in which quantitative survey data and qualitative interview data were collected concurrently. Quantitative and qualitative data were collected from researchers and practitioners via two complementary methods: (1) an online survey incorporating both quantitative and open-ended responses, and (2) semi-structured interviews with a purposive subsample of survey respondents. The two data streams were integrated at the interpretation stage to produce a comprehensive account of researcher and professional perspectives. All research activities were reviewed and considered exempt by the university Institutional Review Board (Protocol #25-2587). 2.1 Young adult and teen input To develop the initial list of AI use cases, the research team collaborated with a group of six young adult undergraduate students (ages 19–20) working in the first author’s Psychology research laboratory. Each student independently generated a list of 15 AI use cases that they believed were both common among their peers and younger teens and that were particularly important for researchers to study. This group then met with the members of the research leadership team to discuss the combined responses and individually rank use cases in order of importance, after which the first author synthesized and refined the list by consolidating overlapping items and clarifying descriptions. The revised list was subsequently reviewed by members of the study’s Teen Advisory Board, consisting of U.S. middle and high school students, during a focus group (n = 4) and via an emailed survey (n = 12), whose feedback informed further revisions to improve relevance and clarity. The refined list was then discussed with the young adult undergraduate advisors and finalized through an iterative consensus-building process, including a survey where undergraduate advisors ranked use cases on importance again. 2.2 Participants and recruitment Researchers and organizations employing researchers and practitioners were identified to find people doing research on adolescence and technology use or related fields. This list was developed via multiple methods, including leveraging contact lists that the authors could access (e.g., convening invite lists; expert lists developed for prior Delphi studies on youth and technology; Capraro et al. 2025). The research team also conducted extensive internet searches using relevant search terms to ensure that the search was as broad as possible. The survey was also posted on dedicated listservs for organizations focused on youth and technology research (e.g., CERES | Connecting the EdTech Research EcoSystem, n.d.). During Fall 2025 and early Spring 2026, we contacted directly over 800 researchers, labs, and organizations, and many more viewed the advertisement via listservs. This approach balanced individualized recruitment with broader outreach to the target population; accordingly, we evaluated respondents’ reported names, affiliations, and/or email addresses to screen for bots or fraudulent responses and confirm alignment with our target population. Because recruitment methods were targeted and study advertisements clearly described the focus of the research, and given the nascent nature of this field and the resulting difficulty in identifying researchers newly engaging with this topic, we considered respondents who self-selected into the survey and whose work was at least indirectly related to adolescent AI use (e.g., adolescent development without a technology focus, or technology and well-being without a developmental science focus) as part of the target population. Upon review, we first removed three duplicate responses (for participants who responded more than once, we retained their first response only). Of the remaining sample, we verified the identity of 99.29% of respondents (n = 140; based on a combination of information participants provided, Internet searches, and our team’s knowledge of the field), all of whom were within the target population. The remaining one participant we could not verify identified as a student, and thus may not have publications or an online presence; we retain this single case, such that all non-duplicate datapoints (n = 141) were included for analysis. Eligibility required participants to be 18 years of age or older. Participants completed an online survey administered via Qualtrics. Five survey respondents were selected for a drawing to earn a $100 gift card. Sample characteristics, including field of study, role or profession, and location, are provided in Table 1. A purposive subsample of n = 15 researchers participated in semi-structured individual interviews to elaborate on their survey responses, conducted by the second author (a PhD student in Developmental Psychology). Participants were selected to achieve maximum variation across roles, disciplinary backgrounds, and substantive focus area. Of those invited (n = 27), n = 15 (55.56%) agreed to be interviewed. Interviews were conducted via Zoom, lasted approximately 40 min (Range 30–60 min), and were audio-recorded and transcribed verbatim. All interviewees were compensated with a $50 gift card. 2.3 Survey measures Respondents were asked “How important do you think it is for researchers to study each of the following AI use cases among teens?” on a 4-point scale from 1 (not important to study) to 4 (extremely important to study) for a list of 20 AI use cases (developed in partnership with youth advisors, see above). For a list of all use cases presented to respondents, see Table 2. All participants were invited to nominate AI use cases warranting additional research attention with one open-ended question, as well as any they considered “red herrings” unlikely to yield meaningful insight. Participants were also invited to share AI use case priorities for marginalized youth—specifically LGBTQ+ youth and Black and Latine youth —given that these groups often face both unique opportunities and challenges online and report higher rates of technology use (Faverio and Sidoti 2025; Hopelab et al. 2024). Active researchers completed open-ended items asking them to identify the highest-priority research questions at the intersection of youth and AI and the most promising methodological approaches for addressing those questions. Barriers to research were assessed among active researchers using a checklist of 12 items with an open-ended option to describe additional barriers not captured by the checklist. Finally, respondents indicated guardrails that society can implement now based on known research on youth development and technology. Participants selected their top five priority guardrails from a checklist of 12 items. A final open-ended item invited participants to elaborate on how they hope society will implement guardrails to ensure a positive future for young people who use AI. 2.4 Interview protocol The semi-structured interview guide was organized around the aims of the study and (for all but one participant) interviews were conducted following the participants’ completion of the survey. Participants were first asked to elaborate on their ratings of AI use case importance and to reflect on how use cases might be conceptually organized or prioritized. They then discussed the research questions they viewed as most critical and the methodological approaches best suited to address them. Participants then discussed the barriers they perceived as most consequential. Finally, participants were asked about guardrails that could leverage existing research on youth and other technologies. Interview flow emphasized respondents’ expertise (e.g., neuroscientists were asked to comment on the adolescent brain and implications for AI use, participants with industry experience were asked to comment on the potential role of industry-academic partnerships in addressing limitations around data access and research infrastructure). Probes were used throughout to elicit concrete examples, explore points of tension or uncertainty, and provide space for additional detail. 2.5 Positionality statement The research team brought diverse but complementary expertise to this study. The team included early career psychology researchers, including three assistant-level faculty with expertise in adolescent development and technology, affective neuroscience, and digital media; one graduate student researcher studying technology and adolescent and young adult development; one recent college graduate serving as research staff; and one former academic researcher now working in the nonprofit sector at the intersection of research, advocacy, policy, and youth-serving practice. Collectively, the team brought substantial knowledge of adolescent development, digital technologies, and youth well-being, as well as experience conducting research in these areas. Consistent with reflexive thematic analysis, we recognize that our disciplinary training, professional experiences, and values shaped the research process. Our team broadly agrees that technology can both support and undermine young people’s well-being, depending on its design and use. Our team also broadly agrees that commercial incentives can influence the design and governance of digital technologies, often in ways misaligned with youth well-being. 2.6 Analytic approach Descriptive statistics were computed for all quantitative survey items. Means and standard deviations were calculated for each AI use case importance rating, barriers, and guardrails and items were rank ordered to identify areas of greatest consensus among respondents. Response rates varied across survey questions, with some evidence of drop-off (i.e., fewer responses for questions appearing later in the survey). For percentages of dichotomous checklist-style variables, denominators reflect the number of respondents who were presented with the relevant questions and provided at least one response within that checklist item set. Questions about research barriers were presented only to respondents who identified as researchers (n = 127); 114 researchers (89.76%) provided at least one response to these items. Questions about guardrails were presented to all respondents (n = 141); 124 respondents (87.94%) provided at least one response to these items. Questions regarding AI use cases involving marginalized youth (i.e., people of color and LGBTQ+ youth) were presented to all respondents (n = 141); 111 respondents (78.72%) provided at least one response to these items. Descriptive quantitative findings were analyzed first and were used as one of many sources of knowledge during the coding and interpretation of the qualitative data, with these data streams integrated at the interpretation stage (Braun and Clarke 2021; Creswell and Clark 2017). For each of the three study aims, descriptive survey findings and qualitative themes were used collectively to produce a unified account of researcher and practitioner perspectives. Quantitative data and quotes from interviews and open-ended survey responses are used throughout the results to illustrate and contextualize these patterns. Open-ended survey responses and interview transcripts were analyzed using reflexive thematic analysis, informed by the approach described by Braun and Clarke (2021). In this study, our reflexive analysis approach considers researcher subjectivity an asset to the process; given that the research team identifies as members of the study population, we leveraged our knowledge of the field and experience as researchers to interpret participant data within that broader context. Analysis was iterative, recursive, and collaborative, with the goal of actively constructing themes. Four members (AJM, CM, MGV, KB) of the research team independently reviewed the qualitative data to become familiar with the dataset and recorded initial impressions prior to meeting as a group to discuss patterns in the qualitative data. Then, after re-reviewing all quantitative and qualitative data, the lead author (AJM) developed candidate themes and sub-themes, generated through interpretive engagement with the data rather than a codebook framework. Themes were subsequently reviewed, discussed, and refined collaboratively by the research team. Throughout the process, multiple sources of knowledge (i.e., qualitative data, quantitative data, the study team’s knowledge and reflexive perspective within the field) were treated as analytic resources that enriched interpretation. Given that our goal is to gather perspectives of researchers, our results are contextualized within the existing literature in a way that is atypical of empirical studies. For this reason and following precedent from similar work on researchers’ perspectives on other technologies (Davis et al. 2025), we present the results and discussion as an integrated section, followed by a short over-arching conclusions and implications section. Within the Results and integrated discussion, we distinguish between insights derived from participants and interpretations or contextualization provided by the author team. 3 Results and integrated discussion 3.1 Aim 1: AI use cases The first goal of the project was to gauge researchers’ perceptions of the relative importance of different AI use cases. An overview of the aims and themes for all aims is presented in Table 3. Within the qualitative data, participants were aligned that “screen time” or vague and decontextualized AI use measures should be avoided and replaced with emphasis on specific ways youth use AI, mirroring similar calls in research studying other technology use among youth (Maheux et al. 2025; Valkenburg 2022). One researcher argued that “generic categories around what’s driving use (e.g., entertainment, social connection)… are not telling us the full story about use cases, what’s motivating engagement, and what might lead to either over-reliance or not” (survey, researcher at a nonprofit, USA). Participants rated the importance of studying 20 AI use cases among adolescents. Mean importance ratings ranged from 2.61 to 3.85 on a 4-point scale (see Table 2 and Fig. 1). Notably, the relatively small differences among the AI use case means suggest that distinctions between them should be interpreted cautiously. Rather than reflecting a precise ordering of importance, the findings indicate broad support among participants for studying a range of AI use cases affecting adolescents. Below, we identify five key themes for research prioritization, integrating these quantitative ratings with the qualitative data. 3.1.1 Theme 1: study everything, but triage for severe cases Participants broadly endorsed the importance of studying AI use across contexts. For all 20 use cases, a majority of respondents (51–96%, depending on the use case) rated them as either Important or Extremely important to study. Overall, the relatively high endorsement and means (all above the mid-point of the scale) indicate that respondents believe that all 20 use cases are important to study. Participants commonly emphasized the need to be comprehensive, “I think everything should be studied, we know so little that any information would help” (survey, student advocate, USA). However, quantitative and qualitative responses also revealed a triage logic among participants: the most severe, irreversible, and immediate individual-level harms should be considered most urgent, with many arguing that “it is critical to prioritize research that could be life-saving” (survey, graduate student, USA). These included AI providing dangerous health information related to suicide, self-harm, or disordered eating; AI used as a tool for cyberbullying or harassment, including AI-generated sexual abuse material and deepfakes; and AI used for extortion or manipulation. Additionally, from the checklist items regarding priorities for marginalized youth—specifically Black, Latine, and LGBTQ+ populations—AI for mental health, identity, and emotional support was ranked as the top priority use case (of the 111 participants who responded to at least one of these items, 94.6% [n = 105] endorsed for LGBTQ+; 92.8% [n = 103] endorsed for youth of color). Notably, a subset of participants also suggested that because so little is known, and because the most immediate and acute issues may overshadow the importance of studying more subtle changes, all areas of study must be considered and addressed simultaneously. This tension reflects a broader challenge that the authors have identified in rapid-response AI research: how to prioritize attention across risks that vary in both prevalence and severity (Saeri et al. 2026). Individual-level acute mental health outcomes, such as AI-facilitated suicide, receive disproportionate media attention (Hill 2025), yet the prevalence of such issues remains unknown, and is likely low (Archiwaranguprok et al. 2025). Additionally, one participant noted how industry actors are already investing effort in identifying and mitigating acute safety risks (OpenAI 2026), yet may have fewer incentives to focus on long-term, less sensational, or population-level risks. Thus, the authors and a few participants concur that focusing exclusively on acute, life-threatening harms risks overlooking a distinct category of concerns that may be less visible but nonetheless potentially consequential developmentally, including subtle effects on social development, emotional regulation, judgment, executive functioning, and emerging social norms during adolescence that may emerge in long-term, population-level studies. The case for prioritizing the most severe harms may be largely based on participants’ assumptions of limited resources. As one researcher put it: “in an ideal world we would have the space and the time to study everything, and design these tools to better our interactions with one another and our views towards ourselves as humans” (survey, graduate student, USA). Coordination among research teams for future AI research could ensure that all AI use cases are collectively addressed, as we discuss below in the guardrails section. A few participants raised a related pragmatic concern: use cases with little commercial viability may not warrant research investment, as market forces are unlikely to promote them at scale. This observation points to a broader question that may itself merit study which was noted by only one participant—how dynamic commercial incentives are shaping the trajectory of AI development and, by extension, the landscape of potential AI use cases researchers should anticipate. 3.1.2 Theme 2: AI literacy as a foundational and multidimensional skill Participants rated AI literacy as among the most important use case to study (M = 3.69, SD = 0.56). Much prior work on AI literacy has focused on workplace readiness and using AI effectively (Casal-Otero et al. 2023; Wang and Lester 2023). However, participants’ qualitative responses consistently emphasized the multidimensionality of AI literacy—encompassing more than simply how to use AI. For example, participants and authors collectively note that AI literacy should include knowledge of privacy and digital footprints; critically evaluating AI-generated content especially when shared through peer networks or when content reflects biases; navigating school and institutional AI policies; developing skills of using AI in ways that support—vs. detract from—learning, relationships, and human thriving; and complex effects on labor markets, environmental costs, and intellectual property. These diverse components of an ideal conceptualization of AI literacy reflect the extensive scope of what informed AI use requires of young people, and how researchers should conceptualize and study it accordingly. 3.1.3 Theme 3: AI as a social and relational actor Many AI interactions involve simulated sociality, and researchers largely agreed that this use case must be studied further. Use cases including companion relationships with an AI (M = 3.65, SD = 0.58), romantic relationships with an AI (M = 3.43, SD = 0.80), how social skill development intersects with AI use (i.e., using AI to resolve interpersonal conflicts; seek guidance on friendships, dating, or family relationships; or practice social interactions; M = 3.36, SD = 0.71), and sexual interactions with an AI (M = 3.29, SD = 0.84) were all rated as highly important to study. Current design choices and affordances of AI chatbots were frequently cited by participants as key to conceptualizing this use case, including the frictionless availability, responsivity, and current design choices prioritizing sycophancy in AI responses that have been described in previous work (Maheux and Maes 2026; Raedler et al. 2025). For example, one interviewee shared: “Technology is at everyone’s fingertips, including all these kids, and if you have your best friend in your pocket or something that will constantly validate you, tell you you’re right, promote your ideas back to you and all that stuff in your pocket, then what’s your need for…overexerting yourself to make these connections in real life that seem like way too much of a barrier to overcome” (interview, graduate student, USA). Participants emphasized that how young people conceptualize their AI chatbots—as a tool vs. a social agent with a mind, intentions, or feelings—likely shapes how those interactions affect development. Although some prior research has addressed AI mind perception among children and adults (Lee et al. 2020; Xu et al. 2025), we argue that more research is needed among adolescents—a unique population actively developing mature social cognition skills in increasingly complex social contexts (Dahl et al. 2018). Related to this, some participants emphasized the “super peer” framing, a term coined to describe how prior media technologies can influence adolescent behaviors and expectations (Brown et al. 2005; Elmore et al. 2017). Participants offered that the “super peer” concept, in the context of AI, suggests that chatbots may serve as credible, always-available peers or role models that can shape adolescents’ understandings of norms and culture. One participant commented that this may have outsized importance in cases where adolescents view the chatbot as a friend or source of advice or health information. Participants also highlighted that AI use is not only an individual behavior but a socially negotiated one. One researcher noted that social norms around AI use are actively forming within peer microsystems, which prior qualitative work has begun to address (Center for Digital Thriving 2025). We argue that these norms may exert influence on AI use that parallels the peer socialization dynamics already well documented in the technology and adolescent development literature (Navarro and Tudge 2023), offering a useful starting point for research on AI within social networks. Moreover, recent research suggests that adolescents currently report almost no guidance from parents on how to navigate the AI ecosystem safely (O’Neil et al. 2026). Thus, we contend that how parents model and communicate to their adolescents about AI use is critical to study. Although uncommon in qualitative responses, participants indicated that how parenting intersects with AI use is important to study in the quantitative ratings (M = 3.42, SD = 0.73). 3.1.4 Theme 4: information, agency, and identity in the AI era Participants identified the importance of studying AI in adolescents’ information ecosystem—including how they seek information with AI, how they understand false AI-generated content, and how their own information may be weaponized against them. In the quantitative data, exposure to or creation of AI-generated misinformation, including hate speech, propaganda, or political media (M = 3.68, SD = 0.60), using AI to access medical and healthcare-related information or advice (M = 3.63, SD = 0.61), and privacy concerns (i.e., users disclosing personal information into chatbots with unclear or misaligned [e.g., profit-driven] data use or storage security policies; M = 3.60, SD = 0.63) were considered high-priority use cases. Several qualitative comments specifically mentioned deepfakes (i.e., AI-generated imagery of known others) and deepnudes (i.e., AI-generated sexual imagery of others) as key challenges in need of future study. Participants’ comments may in part reflect recent high-profile events involving AI chatbots ‘undressing’ images of people, which have disproportionately affected women and children (Conger et al. 2026). We note that despite preliminary commentary from scholars on how to address such issues among adolescents (Alexander 2025), systematic research on adolescents’ use of and experience with deepfakes and deepnudes remains lacking. Some participants shared concern that AI interactions or exposure to misinformation may have subtle influences on self-concept, sense of agency, or orientation towards truth. One participant shared: “the sophistication of those tools [is] making it much, much harder for us to tell the difference between what is real, what is true, what has human origin, and what comes from an AI. So, there are new vulnerabilities that that creates. I’m very worried about our adolescents’ ability to understand fact…to even value fact” (interview, faculty researcher, USA). This theme was seen as particularly salient for marginalized youth. Quantitative data suggested that access to health information, including physical, mental, and sexual (91.0%, n = 101 for LGBTQ+ youth; 81.99%, n = 91 for youth of color), and exposure to racial/cultural biases for youth of color (85.6%, n = 95) and homophobic/transphobic biases for LGBTQ+ youth (85.6%, n = 95) are among the most important priorities to study for these groups. Many survey and interview respondents brought up AI biases, which we note have received significant attention in computational fields (Beattie et al. 2022; Iloanusi and Chun 2024). One respondent said: “Because of the data sets it’s trained on, and the language it’s trained on, and the cultural values it’s trained on, there’s a lot of little microaggressions that can happen that go under the radar” (interview, faculty researcher, Canada). A few participants noted that marginalized youth may be more likely to use AI as a tool during identity exploration; prior commentary has offered that AI may distort identity exploration by reflecting idealized or biased representations of identity rather than supporting authentic self-development conducted within social contexts (Maheux and Maes 2026). Critically, a few participants reminded that AI biases are not inherent in AI systems, but rather reflect the training data, “These disparities exist in digital spaces because the models are created by humans” (survey, clinician and faculty researcher, USA). However, prior research broadly suggests that adolescents need spaces to engage in identity exploration, and many already use technology for this purpose (Choukas-Bradley et al. 2023; Common Sense Media 2024). For example, sexual and gender minority adolescents may be navigating identity exploration in contexts where in-person support may be limited; in such contexts, and if AI is designed with child well-being as a priority, we are open to the possibility that AI could serve as a relatively private or useful space for exploration. More research is needed to understand the balance of risks and benefits. Several participants share this sentiment; one noted that “Kids are resourceful and will find the information they need, which makes it even more important to ensure that AI-enabled spaces are designed for accuracy and support for marginalized youth” (survey, advocate, USA). Thus, identifying how marginalized youth engage with AI-generated information for these purposes, and how such motivations intersect with chatbot design and adolescent developmental outcomes, are key for future research. 3.1.5 Theme 5: the risk of overreliance and cognitive offloading Participants consistently flagged overreliance on AI and cognitive offloading to AI as a priority concern. Prior research shows that when AI handles cognitively demanding tasks, young people may not develop the skills those tasks were meant to build (Zhai et al. 2024). We argue that overreliance can also emerge in psychosocial contexts, including excessively relying on AI to be a social companion, provide advice, or regulate emotions. How exactly such an AI use case will be defined remains unclear, and is an area that will benefit from and likely require collaborations among developmental scientists, technologists, and young people using these tools themselves. Participants were broadly aligned that measures of quantity or frequency of use alone are unlikely to capture such a construct: “I think [studying] overreliance on AI (not just use, but excessive use) will be important” (survey, graduate student, USA). 3.2 Aim 2: research priorities & opportunities The second aim of the project was to begin curating a set of priority research questions, methods, barriers, and solutions. Below, we provide an overview of four themes specific to theoretical foundations and research orientation, research questions and methods, research barriers, and solutions. 3.2.1 Theme 1: theoretical foundations and research orientation Researchers offered several overarching points that provide insight to the field as it adjusts its orientation toward the challenges of studying adolescents and AI. First, researchers agreed that a developmental lens must be applied to research on this topic—a point made in numerous survey responses and interview comments. For example, “Adolescence is a uniquely vulnerable period—not because teens are “broken,” but because they’re more sensitive to social feedback and rewards; Persuasive design exploits developmental traits—features like infinite scroll and “likes” specifically target how teen brains respond…Age-appropriate design isn’t optional—AI for teens needs different safeguards than AI for adults” (survey, graduate student, USA). The core features of adolescence mentioned most by researchers included changes in social context and salience of peers, neurobiological development related to reward sensitivity and executive function, and identity development. However, participants were not aligned on the level of urgency or concern to bring to AI-related research. Several cautioned against moral panic and reflexive negativity about AI, noting that assuming all use is harmful is a methodological and framing error, which we note scholars have previously suggested regarding social media (Orben 2020). These participants suggest that research should instead take a balanced approach to considering both potential benefits and harms of AI. However, other respondents reported existential concerns about the nature of AI for adolescents. For example, one noted “I think that in general, kids’ use of AI platforms is a brewing and very, very underappreciated and underrecognized public health problem…I’ve never felt in my 35 years of doing clinical research the level of urgency that I have now on needing to understand and really needing to make the public and parents aware” (interview, industry, USA). Thus, the field may not align on how to think about adolescence and AI as we study it. The authors suggest (and hope) that this will be a strength—optimistic and pessimistic viewpoints may collectively move us toward a deeper, more nuanced understanding and critical insights built on specificity rather than generalization. Third, a central conceptual tension in participants’ comments emerged regarding whether AI indirectly or directly (per the authors’ synthesis and paraphrasing) impacts developmental processes. The former reflects the displacement hypothesis, which argues that technology use may have negative impacts when it displaces healthy activities like sleep, exercise, and time socializing (Exelmans 2020; Valkenburg and Peter 2007). For example, one noted, “I am deeply concerned about the ways in which AI is being used socially-esp. the ways in which it displaces real human relationships. I think the outcomes are existential in nature” (survey, advocate, USA). Alternatively, direct impacts may involve AI companions undermining (or stimulating) human connection through a direct influence on social skills and expectations, which prior work has begun to explore (Ibrahim et al. 2026; Smith et al. 2025; Sun et al. 2026). Reflecting on researchers’ shared and diverging opinions in the current data, we advocate for keeping the developmental function as the primary unit of analysis, considering how AI supports or undermines such tasks and processes, rather than considering AI use cases as a starting point. Finally, researchers overwhelmingly emphasized that individual differences among adolescents will change how they use and respond to AI. This idea reflects the differential susceptibility to media effects model, which argues that dispositional, social, and developmental factors alter media effects (Valkenburg and Peter 2013). Research on social media has broadly suggested a similar pattern (Maheux et al. 2025). Differential susceptibility concepts were commonly evoked by participants in the context of marginalized youth. Social media research emphasizes that marginalized youth may experience a “double-edged sword” wherein the benefits of online environments can compensate for offline isolation or discrimination, but harms may also be exacerbated (Choukas-Bradley et al. 2023; Common Sense Media 2024). The authors note that, based on these prior patterns, adolescents experiencing discrimination, marginalization, and/or structural inequalities and resource limitations may experience both greater benefits (e.g., AI tutors mitigating educational inequities) but also greater vulnerability related to AI harms, as some preliminary work suggests (Lee and Culver 2026). One researcher noted a worry that “companies are exploiting structural inequalities and in turn amplifying vulnerability (e.g. police profiling) and disempowering communities (e.g. cutting access to services)” (survey, faculty researcher, USA). A few researchers also emphasized that effective AI literacy programs and supportive AI design cannot be applied to marginalized communities from the outside but must be developed in active partnership with them. 3.2.2 Theme 2: priority research questions and methods Researchers’ open-ended responses regarding research questions and methods were diverse, with little consistent pattern across responses. Only one theme was repeated by participants and indirectly highlighted by the diverse range of participants’ responses: the need for methodological and conceptual pluralism, including multi-method studies and interdisciplinary research teams. Triangulation—the use of multiple methods to establish convergent validity—is a cornerstone of social science research, ensuring that findings reflect the construct of interest rather than an artifact of any single method (Jick 1979). Indeed, to synthesize researchers’ methodological suggestions in the current study would be functionally to describe the best practices across all interdisciplinary fields studying adolescents in digital contexts. We argue that the breadth of suggestions across participants signals that this phenomenon is too multifaceted to be captured by any one subfield or approach, and progress may depend on deploying methods in combination. A broad triangulation approach—across disciplines and research teams working in tandem—may be the best way to address all research questions and leverage diverse methods for this field. Below, we briefly describe some of the specific approaches that participants mentioned—organized across measurement of AI use cases, adolescent outcomes, and design considerations. Critically, as with any scientific endeavor, the research questions (e.g., which AI use cases to study and which outcomes to measure) and the methodological decisions (e.g., which design or measurement approach to take) must carefully match the strengths of the methods to the questions and study goals. Researchers’ suggestions regarding priority research questions often centered on AI use cases. These suggestions largely mirrored the priorities outlined in Aim 1, which we hope offers the field a starting point on important use cases to study. Some researchers emphasized foundational prevalence questions regarding adolescents’ engagement with AI across different use cases (which national polling has been able to address more quickly than traditional peer-reviewed scholarship; McClain et al. 2026; O’Neil et al. 2026). Other comments emphasized studying how AI design decisions shape adolescents’ use and understanding of AI, leveraging experimental methods that modulate such design features, as we note has been done successfully in some preliminary studies (Ibrahim et al. 2026; Kim et al. 2025). These included how systems present uncertainty, how they respond to emotionally vulnerable users, and how interface choices shape adolescents’ trust, engagement, and interpretation. Others emphasized key psychosocial and cognitive experiences of engaging with AI—including across different AI designs and interactions—such as how adolescents interpret AI-generated content or respond to expressions of uncertainty and prompts to seek human support. Participants’ research question priorities also emphasized the developmental context and child-level outcomes that must be paired with studying AI use cases and design choices. These include, among others, mental health outcomes, academic success, executive functioning, social skill development, loneliness, and neurobiological development. Participants also identified outcomes that are likely to be subtle or slow-emerging—including changes in perceived agency, beliefs about human exceptionalism, and capacity for relational reciprocity—as important targets that may be missed when focusing only on acute impacts. Others took a population-level view, asking how AI-driven shifts in epistemic norms, social norms, and information environments may unfold at scale. In terms of measurement, suggestions were offered for both AI use cases and child outcomes. These include rigorous methods to capture digital behavior: AI conversational logs, passive sensing of interaction patterns, and natural language processing of chatbot exchanges offer a level of behavioral precision not available via self-report. That said, scholars also signaled the need for rigorously developed self-report tools. Self-report methods remain ideal for understanding subjective experiences, and multi-informant approaches paired with behavioral observation may be the optimum approach for triangulation. Child outcomes should be measured, where possible, with gold-standard tools, including psychometrically rigorous self-report scales, sociometric nominations, clinical assessments, neurobiological measures (e.g., EEG, fMRI, and fNIRS), and school-reported academic metrics. Longitudinal, experimental, and participatory designs were also frequently mentioned. Designs that can suggest directional or causal processes are essential for informing policy, especially in the context of common overinterpretations of cross-sectional data. The possibility that AI effects diverge across short and long timescales—for example, alleviating immediate loneliness or distress but exacerbating isolation or distress intolerance over time—further underscores the value of longitudinal designs, including panel designs, ecological momentary assessments (EMA), burst designs, and N-of-1 designs, that can track individual trajectories over time. Evaluation of educational, literacy, or interventions was identified as another urgent priority, especially for informing policy and for identifying opportunities to support adolescents in rapidly changing AI contexts. Co-design, youth participatory research, and qualitative designs were also frequently mentioned—often in the context of mixed-method studies—as being critical. In addition to promoting agency among youth research partners, the rapidly changing nature of AI technology, and the speed with which adolescent norms and use patterns are evolving, makes qualitative and participatory methods essential for researchers to study relevant phenomena. 3.2.3 Theme 3: barriers to research Participants identified several barriers to conducting rigorous research at the intersection of adolescents and AI and completed a 12-item checklist indicating the biggest barriers to research progress (Fig. 2). Several barriers reflected the scale, novelty, and pace of the problem. Overwhelmingly, the most-cited barrier in the quantitative data was “Rapid speed of change in AI systems and use cases (relative to time required to conduct high-quality research),” which was endorsed by 76% (n = 87). One researcher stated that the problem is: “How quickly AI changes and slow speed of publication process. It’s already outdated and still under review” (survey, faculty researcher, USA). Findings risk obsolescence before publication, and longitudinal designs face the problem of studying a moving target. We argue that this challenge is exacerbated given the opacity of how AI systems operate, and thus, how design choices may influence adolescent behavior. A commonly cited barrier by 56% (n = 64) in the quantitative data was the lack of transparency on how AI platforms and AI models operate. Additionally, some participants described a form of epistemic overwhelm resulting from this pace and scale of change: “I feel utterly bewildered by where to start in a way that I haven’t with any other research area. This feels enormous and existential” (survey, faculty researcher, USA). Because of this speed and pressure, participants also noted a mismatch between the volume of interest in AI research and the availability of theory-driven hypotheses to anchor it, creating risk of poorly designed studies accumulating without building cumulatively. Structural barriers were equally prominent. Participants indicated that funding remains scarce, slow, and misaligned. Many researchers noted that lack of funding (n = 59, 51.7%) and lengthy funding application processes (n = 46, 40.36%) were key barriers. Some commented in qualitative data that most resources flow toward AI development rather than toward understanding its human consequences, and research on marginalized youth is especially underfunded. Some shared how participant recruitment for research involving minors is logistically demanding and costly, especially recruiting representative or hard-to-reach samples, and requires significant resources to invest in developing partnerships with schools and communities. Ethical review processes add further constraints, and many researchers (n = 30, 26.31%) noted IRB review challenges (e.g., slow timelines, inconsistent review, IRBs determining studies to be more risky than researchers believe) as a major barrier to conducting their work. Community-based work—wherein the researcher partners with community members to understand and support community needs—can be especially demanding, as a few participants shared, yet more impactful than developing interventions without careful evaluation. One researcher commented on the resources needed for this type of hands-on work with adolescents: “Doing research where kids are involved from day one is way harder than going to build some tool or some learning system and toss it out into the world and then maybe evaluate it…It’s a huge time commitment. You need to fund the [graduate] students. I like to give money to the school. I have to buy myself out of teaching. So funding is always an issue, especially for a new field” (interview, faculty researcher, Canada). The second most-cited barrier was the limited access to data or industry partnerships (n = 65, 57.02%). Technology companies possess vast quantities of behavioral data that could substantially advance scientific understanding if made available through research partnerships, APIs, or participant-mediated data donation. However, data access and industry partnerships remain constrained by the defensive posture of large technology companies, as researchers described, “the company doesn’t want you to be able to make any claims about people being depressed or not having a good time on their platform, and so they’re just not gonna release the data” (interview, graduate student, USA). Researchers also cited, to a lesser degree, the lack of interdisciplinary collaborators (n = 27, 23.68%) and lack of researcher training in relevant methods (n = 26, 22.81%) as barriers. Although less frequently cited, strengthening expertise in these areas could help researchers develop alternative computational approaches that reduce reliance on direct industry partnerships and proprietary data access. Barriers specific to academic contexts were also cited as challenges. Researchers identified misaligned academic incentive structures (n = 50, 43.86%) as a prominent barrier; examples provided for this checklist item included university expectations and granting agency values that sometimes disincentive high-risk research, encourage quantity of publications over quality, or prioritize independence over collaboration. One researcher also noted that the research ecosystem threatens the quality of the emerging evidence base, including variable standards of rigor and publication bias that favors significant effects and insufficiently considers null results. 3.2.4 Theme 4: solutions—coordination and partnerships Our data point to several solutions to these barriers. First, we argue that the scale and pace of AI development require a coordinated scientific response. Only a few participants in our study offered this concept as a solution; however, we believe that this is less a scientific problem than a systems’ infrastructure problem, and thus may be less visible as an immediate research priority for trained scientists than as a necessary condition for the long-term progress of the field. As our team and others observed with social media research, fragmented individual efforts, however rigorous, will not be sufficient (Davis et al. 2025). To support and improve the future of science on this topic, we call for a coordinated scientific response that exceeds the capacity of individual research labs. The need is not simply for more studies but for field-level infrastructure, some of which our participants suggested as isolated scientific goals but which, taken together, could strengthen the field’s capacity for long-term progress: shared measurement standards, data repositories, interdisciplinary consortia, and multi-site cohort studies. Industry-academic partnerships, approached carefully and with appropriate transparency protections, could also provide data access that individual labs cannot obtain independently. This infrastructure vision extends to research partnerships with teens, families, schools, and communities, which researchers commonly mentioned directly. Across all three aims, a consistent value emerged among participants’ comments: adolescents should not be merely the subjects of research but active participants in shaping its questions, methods, and applications. This appeared as a methodological argument—participatory and co-design approaches yield more relevant research questions and more ecologically valid measurements—as well as an ethical one, grounded in the view that adolescents have a right to involvement in decisions that affect them. As one participant observed: “Youth are the experts of their own experiences, also often equally as confused/excited/scared about the future of technology as the adults in their lives” (survey, faculty researcher, USA). Several interview participants emphasized that the most pressing barrier may not be a research problem at all, but rather translation of knowledge across sectors. For example, one participant commented that the challenge may be ensuring that evidence informs public decision-making, “I actually think the issue here is one of communications and policy not so much the science parts… We need help from the political scientists on this one, because just being right on the science has not been enough to get the policy makers to listen” (survey, faculty researcher, USA). Others similarly highlighted the need to rebuild public trust in scientific expertise and orient scientific efforts toward evaluating and responding to rapidly evolving technologies and policies, rather than focusing exclusively on basic research that tends to be more reactive. One commented, “I think the notion of ‘scientific expertise’ has been eroded in recent years, but it seems we need to convene credible, independent scientific experts through National Academies (or some body) to assemble a list of high-quality benchmarks for evaluating AI chatbots, to propose safety and quality standards, and to build a tracking system for monitoring effects of AI use on youth social/emotional wellbeing” (survey, faculty researcher, USA). These comments reframe the challenge as not only about conducting science, but also engaging with science communication and cross-sector collaboration, both of which are critical for dissemination and ensuring research is maximally relevant to society. Engaging directly with policymakers, journalists, and communities may be central, rather than peripheral, to the research mission in a context where developmental impacts are fast-moving and potentially high stakes. 3.3 Aim 3: guardrails—what society can do now Participants selected their top five priority guardrails from a checklist of 12 items (see Fig. 3). The most frequently endorsed were (1) “Effective built-in safeguards against harmful AI outputs (harassment, hallucinations, self-harm promotion, etc.)” (n = 97, 78.2%), (2) “Government regulation of AI companies requiring AI models to prioritize youth well-being” (n = 94, 75.8%), and (3) “AI literacy training in schools” (n = 91, 73.4%). Open-ended responses elaborated on these three overarching directions: regulation, design, and literacy. 3.3.1 Theme 1: regulation—necessary, urgent, and politically complex Participants were consistent that voluntary industry action is insufficient and that regulation is necessary. The analogy to prior historical changes—the “Big Tobacco” history and pharmaceutical development—was invoked: “AI feels like drug discovery without clinical trials! Just releasing new products into the wild without study of benefits/harms/side effects” (survey, faculty researcher, Canada). The comparison to social media was equally prominent, with participants expressing concern that history is repeating itself: “Given how little care tech companies have shown for developing brains in the realm of social media, it seems absurd to expect any meaningful initiatives from tech companies when it comes to protecting teens” (survey, faculty researcher, USA). Specific regulatory priorities named by participants included strengthening data privacy protections for minors; banning targeted advertising to minors; requiring transparency about model training, algorithms, and data use; mandating safety testing before deployment to adolescent populations; and establishing independent auditing against established benchmarks. Participants described effective regulation as necessarily careful and design-focused rather than reactive: “Ideally, society will do forward looking, design focused legislation in service of healthy child development (as opposed to not regulating at all, or to letting new techs be deployed and cause harm, and THEN thinking about guardrails)” (survey, faculty researcher, USA). Notably, several participants emphasized the importance of using legislation to incentivize or mandate safe design (described below). Many cautioned as to ways legislation can go awry, including arguing that legislation involving outright bans or that penalizes the child was unlikely to be successful. In some cases, participants argued that regulation targets the wrong culprit. For example, one interviewee argued that regulating therapy bots—which are often designed with clinicians and professionals—will be less effective than regulating the general-purpose models that most users access: “Most state bills have limited the bots that are trying harder as opposed to limiting the general purpose bots. We’re limiting the things that might do some good while they’re a minuscule part of the pie. Most of the issues are coming from these general purpose bots. So there are a lot of ways that you can fail with public policy, and I think that’s a potential risk” (interview, graduate student, USA). Bans often also fail to balance potential harms against benefits of AI systems. With careful design and oversight, AI-enabled interactions might serve beneficial functions in supporting positive youth development. Such applications of the technology could be eroded without careful legislation and could be amplified with appropriate intervention: “We also need to look at the best case scenarios. With every horribly tragic case we read about the transcripts, how many thousands of kids were walked off a ledge? We don’t know…We’re doing ourselves an ethical disservice if we say…we should shut down all mental health conversation on an LLM. I think…yes, they need to be fixed and let’s make sure we really understand and study what is working and how people are using these things naturally” (interview, industry, USA). Indeed, one community sample of young adult Replika users found that 3% reported the chatbot halted their suicidal ideation (Maples et al. 2024). We argue that systematic research on the optimal design for such benefits—and the design features that undermine potential benefits—must inform regulation. Participants were equally candid about structural obstacles. Trust in institutions was identified as a prerequisite for effective regulation. One participant noted that in their own research, they find that respondents infrequently express “trust in local, state, or federal government, or in Big Tech companies, to make decisions about children and AI” (survey, research scientist, USA). Some participants flagged political funding of legislators by the technology industry as a major obstacle and many expressed skepticism in the likelihood of governmental regulation happening. Notably, most participants (81%) in the sample were from the United States, where regulation has lagged behind other jurisdictions (e.g., European Union; EU AI Act 2024). Thus, although most participants were in favor of regulations and policies promoting safe AI design, the skepticism of this coming to fruition may not be generalizable beyond the USA. 3.3.2 Theme 2: design as the primary lever Across survey and interview data, design was identified as the most powerful and underutilized potential guardrail. An exhaustive list of safety-focused design features is beyond the scope of this paper, though many of the design features enumerated elsewhere were mentioned by participants (5Rights 2025; Collins Oguine et al. 2026; OECD 2024; Unicef 2025). For example, participants mentioned that users must always know they are interacting with an AI system and know how their data is being collected, used, and retained, with full rights to data portability and deletion, and 47.6% (n = 59) of respondents emphasized transparency in design and data collection as a key guardrail. Participants overwhelmingly called for AI systems to be designed intentionally with child safety as a baseline and child well-being as an aspirational standard. Participants were consistent that design must be developmentally informed, “AI tools and protections must be tailored to adolescents’ developmental needs, strengths, and vulnerabilities. This includes ensuring that AI systems support, rather than undermine, healthy cognitive, social, and emotional development” (survey, research scientist, USA). Participants also agreed that such design solutions may be best enforced via policy and regulation, with companies required to promote well-being rather than merely avoid harm: “Not just proving that you’re not causing harm, but that you are adding value in supporting positive development. Because young people do not have rights in the same way, they can’t advocate for themselves in the same way” (interview, faculty researcher, USA). A related theme was the importance of community participation in design. Involving teens in design and policy was endorsed in the quantitative data by 46.8% (n = 58) of respondents. This theme was also mentioned frequently in qualitative responses, including the necessity of designing both with and for adolescents’ social networks “Youth, families, and educators should be involved in shaping AI policies and product design to ensure that tools are not only effective, but safe, equitable, and developmentally appropriate… AI tools should be developed to complement, but not replace, educators, mental health professionals, caregivers, and peers who provide youth with guidance, support, and valuable human connection” (survey, research scientist, USA). Participants argued that harmful design features are not technological inevitabilities but deliberate choices. Some of these, including engagement maximization, have known consequences from prior research on other technologies: “We’ve learned so much about engagement-based design features that manipulate, addict and control youth behavior. These same features are built into many of the commercial AI tools available to children. These addictive design features are not inevitable — they are a choice. And the consequences of these design choices are now well known, documented and grounded in scientific evidence” (survey, faculty researcher, US). Participants also commented that how an AI system responds (e.g., whether it validates distress or nudges toward human support, whether it is honest or sycophantic) may be especially consequential. Sycophancy was specifically named as an underappreciated design risk: “Sycophantic AI models that reinforce or celebrate unhealthy behaviors (e.g. lack of sleep in mania, reassurance seeking in OCD) are unhelpful and may be disastrous” (survey, faculty researcher, USA). Several participants pushed explicitly against individualistic framings of AI risk, with design changes serving as a solution for a collective action problem. For example, on the topic of AI cognitive reliance, one researcher framed this as a design imperative, arguing that we should “Come up with a way to force children to use mentally taxing (in a good way) strategies to engage with tasks. It shouldn’t be on them to regulate themselves” (survey, graduate student, USA). The dominant policy conversation, they argued, remains too focused on personal behavior change and parental monitoring, while the more powerful determinants—corporate incentives—remain under-addressed. As one researcher stated: “The ills that technology has wrought on our youth is a social problem — not a personal problem. Asking parents to do more monitoring, or training teens to monitor their own usage can be helpful, but are ineffective at scale” (survey, faculty researcher, USA). Effective guardrails must, they argue, operate at a structural level, and participants expressed concern that focusing primarily on individual literacy or family-level intervention may be less effective than holding industry to account. 3.3.3 Theme 3: AI literacy and community support AI literacy was identified by participants as a critical proximal protective intervention, especially because regulation and design changes may be slower to unfold. Such literacy interventions were seen by most as pragmatic solutions to a reality where AI is not going away: “This technology’s already here. So how can we help young people gain literacy and understanding? And also thinking about how that intersects with positive youth development and how to help young people feel like they have a sense of agency... That they’re not just feeling like… that this technology’s not just happening to them, but that they understand how to use it and they have more agency in the process itself” (interview, faculty researcher, USA). Participants were careful to frame literacy as a complement to structural change rather than a substitute for it. Placing the burden of AI safety on individual young people and their families, in the absence of regulatory and design-level protections, was seen as unfair and insufficient. Reflecting comments on AI literacy as a key use case for study, participants emphasized the multidimensional nature of literacy and the need to promote a diverse set of skills to equip adolescents to navigate an AI-dominated world. Effective AI literacy interventions, participants argued, will neither simply tell kids to avoid AI nor focus narrowly on workplace readiness: “Instead of focusing on removing harms which is impossible to accomplish 100% of the time, we need to help young people identify the limitations of AI and improve their self-awareness of how they use it, motivations for using AI, and how it can help them thrive in an ethical digital technical universe, rather than how to avoid it all costs” (survey, faculty researcher, USA). Relatedly, some suggested that an aspect of AI literacy may involve supporting adolescents in not using AI and ensuring they have appropriate contexts and supports for development, “ensuring children have their psychological, social, and physical needs met will make it less likely that they’ll use AI tools to fill the void” (survey, faculty researcher, USA). Structurally, respondents felt that AI literacy initiatives should be comprehensive, including but not limited to AI literacy training in school curricula. In quantitative data, parent education regarding AI was the fourth most-commonly endorsed guardrail (48.4%, n = 60). As one researcher put it: “We need to raise AI-literacy not just among teens, but also with their support networks” (survey, faculty researcher, Turkey). Initiatives may include sustained teacher professional development and parent-facing literacy efforts that acknowledge adults’ own uncertainty about AI. Across any intervention format, participants emphasized co-design with teens as both an ethical and a pragmatic imperative to ensure that content is relevant, “AI policy, design, and initiatives are best informed by research that is co-designed by youth for youth. Inclusion of youth at all stages (including development, implementation, analysis, interpretation, dissemination, and related advocacy efforts) is the best way towards radical transformation of digital spaces” (survey, clinician and faculty researcher, USA). Participants argued that deployment and rigorous evaluation of such materials should proceed concurrently—rather than waiting for an evidence base that does not yet exist: “tech moves so rapidly, we cannot wait for the perfect gold standard study to begin offering policy and parenting recommendations for this massive set of problems that may represent an existential threat” (survey, faculty researcher, USA). A recurring concern was the lag between research results and dissemination to clinical practice, schools, and families. Standard timelines to translate science into practice are already slow and participants reported concern that these are being rapidly outpaced by the speed of AI development. Participants noted that coordinated efforts are critical to ensure that AI research reaches teens and families who can benefit from it, and that researchers have a role to play in this process. For example, one participant emphasized the potential for research to simultaneously build AI literacy among adolescents, “I believe a part of the research process is to also implement action programming complimentary to the topic at hand with AI. We are hearing more and more from our middle school youth that teachers are telling them “it’s okay” to use AI and they just need to know when. However, teachers are *not* telling what that “when” looks like. So the goal of AI literacy, in my perspective, should be at the forefront of all the AI and teen research moving forward” (survey, postdoctoral scholar, USA). Embedding evaluation into ongoing interventions, supporting community-based participatory models, and investing in science communication were all identified as practical steps the research community can take now, without waiting for the regulatory or funding environment to change. 4 Conclusion This study surveyed and interviewed professionals working in adolescent research or related fields to understand areas of provisional agreement across three aims: (1) identifying AI use cases to prioritize for future research, (2) establishing a research agenda—including priority questions, methods, and barriers—and (3) proposing guardrails society can implement, while research is underway. Participants in the study agreed that nearly all AI use cases warrant study but prioritized severe, irreversible harms (e.g., self-harm content, exploitation) alongside AI literacy, AI as a social/relational actor, information and identity processes, and overreliance. For the research agenda, participants emphasized anchoring inquiry in developmental science, attending to individual differences, and using diverse and complementary methods. Major barriers included the pace of AI change, lack of industry transparency, and resource gaps. Proposed solutions centered on cross-disciplinary coordination and partnerships with industry, adolescents, and communities. For guardrails, participants converged on regulation, developmentally informed design, and multidimensional AI literacy, agreeing that responsibility should not fall on teens and families alone. These findings point to an urgent need for coordination within research communities and across societal sectors. To advance research, scientists, labs, and organizations should actively share resources, data, measurement approaches, and emerging ideas rather than working in parallel silos, building the shared infrastructure and consortia to keep pace with rapid technological change. Greater collaboration and integration between academic and non-academic researchers—including non-profits, thinktanks, advocacy organizations—can enhance our knowledge base given different strengths, barriers, and perspectives across diverse research teams. At the same time, researchers cannot work in isolation from those positioned to translate findings into action: sustained collaboration with responsible technology organizations and non-profits, technologists and computational scientists, trust and safety teams, and with young people in the room as key experts and collaborators may help ensure that empirical science is deployed in product design. Collaboration with science communicators, advocates, and policymakers is also necessary to ensure that empirical insights reach the public and inform policy. Finally, researchers have a role to play locally, including within their own institutions, schools, and communities, promoting AI literacy and advocating for both continued investment in rigorous research and the immediate adoption of critical guardrails. Such guardrails can continue to be refined as more rigorous research is conducted. To our knowledge, this is the first empirical study of social science researchers studying adolescents and AI, directly building on similar field-wide consensus projects focused on youth and social media (Davis et al. 2025). Limitations primarily relate to non-representative sampling methods that largely reflect the disciplinary knowledge and professional network of the research team. Respondents were disproportionately U.S.-based, limiting generalizability to other countries. Psychology was overrepresented relative to other relevant disciplines (e.g., computer science, public health, communication, education), as was academia relative to industry, policy, and practitioner perspectives. Methodologically, the AI use case list—though developed collaboratively with adolescent and young adult input—reflects a finite, researcher-generated set of categories. Overall, findings should be treated as a starting point for consensus-building rather than a fully representative account of the field. This study offers an initial step toward shared priorities, research agendas, and guardrails for understanding the role of AI in adolescent development. Realizing this agenda will require sustained cross-sector coordination and iteration among researchers, communicators, technologists, policymakers, and young people themselves. Continued investment in both rigorous research and timely protective measures is needed to support healthy development in the AI era. Data availability Anonymized data are available upon request, following IRB approval and a Data Use Agreement, from the first author. References 5Rights (2025) Children & AI design code. 5rights. https://5rightsfoundation.com/children-and-ai-code-of-conduct/ Albert D, Chein J, Steinberg L (2013) The teenage brain: peer influences on adolescent decision making. Curr Dir Psychol Sci 22(2):114–120. https://doi.org/10.1177/0963721412471347 Alexander S (2025) Deepfake cyberbullying: the psychological toll on students and institutional challenges of AI-driven harassment. Clear House J Educ Strateg Issues Ideas 98(2):36–50. https://doi.org/10.1080/00098655.2025.2488777 American Psychological Association (2025) Artificial intelligence and adolescent well-being: an APA health advisory. https://www.apa.org/topics/artificial-intelligence-machine-learning/health-advisory-ai-adolescent-well-being.pdf Archiwaranguprok C, Albrecht C, Maes P, Karahalios K, Pataranutaporn P (2025) Simulating psychological risks in human–AI interactions: real-case informed modeling of AI-induced addiction, anorexia, depression, homicide, psychosis, and suicide (Version 1). arXiv. https://doi.org/10.48550/ARXIV.2511.08880 Beattie H, Watkins L, Robinson WH, Rubin A, Watkins S (2022) Measuring and mitigating bias in AI-chatbots. In: 2022 IEEE international conference on assured autonomy (ICAA), pp 117–123. https://doi.org/10.1109/ICAA52185.2022.00023 Braun V, Clarke V (2021) One size fits all? What counts as quality practice in (reflexive) thematic analysis? Qual Res Psychol 18(3):328–352. https://doi.org/10.1080/14780887.2020.1769238 Brown JD, Halpern CT, L’Engle KL (2005) Mass media as a sexual super peer for early maturing girls. J Adolesc Health 36(5):420–427. https://doi.org/10.1016/j.jadohealth.2004.06.003 Capraro V, Globig L, Rausch Z, Rathje S, Wormley AS, Olson J, Ross RM, Asci S, Bouguettaya A, Burnell K, Choukas-Bradley S, Fardouly J, Kowert R, Lopez RB, Maheux AJ, Mirea D-M, Ozimek P, Selterman DF, Thiagarajan T, … Van Bavel JJ (2025) A collective review on some potential negative impacts of smartphone and social media use on adolescent mental health: results from a Delphi process. PsyArXiv. https://doi.org/10.31234/osf.io/b94dy_v1 Casal-Otero L, Catala A, Fernández-Morante C, Taboada M, Cebreiro B, Barro S (2023) AI literacy in K-12: a systematic literature review. Int J STEM Educ 10(1):29. https://doi.org/10.1186/s40594-023-00418-7 Center for Digital Thriving (2025) Ten fresh insights on generative AI from our teen advisors. Harvard Graduate School of Education Center for Digital Thriving CERES | Connecting the EdTech Research EcoSystem (n.d.) Retrieved June 22, 2026, from https://ceres.uci.edu/ Choukas-Bradley S, Kilic Z, Stout CD, Roberts SR (2023) Perfect storms and double-edged swords: recent advances in research on adolescent social media use and mental health. Adv Psychiatry Behav Health 3(1):149–157. https://doi.org/10.1016/j.ypsc.2023.03.007 Collins Oguine O, Alvarado Garcia A, Muller M, Badillo-Urquiola K (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, pp 1–19. https://doi.org/10.1145/3772318.3791346 Common Sense Media (2024) A double-edged sword: how diverse communities of young people think about the multifaceted relationship between social media and mental health. https://www.commonsensemedia.org/research/double-edged-sword-how-diverse-communities-of-young-people-think-about-social-media-and-mental-health Conger K, Freedman D, Thompson SA (2026) Musk’s chatbot flooded X with millions of sexualized images in days, new estimates show. The New York Times. https://www.nytimes.com/2026/01/22/technology/grok-x-ai-elon-musk-deepfakes.html Creswell JW, Clark VLP (2017) Designing and conducting mixed methods research. SAGE Publications Crone EA, Dahl RE (2012) Understanding adolescence as a period of social–affective engagement and goal flexibility. Nat Rev Neurosci 13(9):636–650. https://doi.org/10.1038/nrn3313 Dahl RE, Allen NB, Wilbrecht L, Suleiman AB (2018) Importance of investing in adolescence from a developmental science perspective. Nature 554(7693):441–450. https://doi.org/10.1038/nature25770 Davis K, Anderson M, Yang C-C, Choukas-Bradley S, Bell BT, Slovak P (2025) Cross-disciplinary perspectives on youth digital well-being research: identifying notable developments, persistent gaps, and future directions. J Adolesc Res 40(2):259–295. https://doi.org/10.1177/07435584241291163 Elmore K, Scull TM, Kupersmidt JB (2017) Media as a “super peer”: how adolescents interpret media messages predicts their perception of alcohol and tobacco use norms. J Youth Adolesc 46(2):376–387. https://doi.org/10.1007/s10964-016-0609-9 EU AI Act 2024/1689, Regulation (EU) 2024/1689 (2024) http://data.europa.eu/eli/reg/2024/1689/oj Exelmans L (2020) Displacement. In: Van den Bulck J (ed) The international encyclopedia of media psychology. Wiley, pp 1–5. https://doi.org/10.1002/9781119011071.iemp0101 Faverio M, Sidoti O (2025) Teens, social media and AI chatbots 2025. Pew Research Center. https://www.pewresearch.org/wp-content/uploads/sites/20/2025/12/PI_2025.12.09_Teens-Social-Media-AI_REPORT.pdf Giedd NJ (2020) Adolescent brain and the natural allure of digital media. Dialogues Clin Neurosci 22(2):127–133. https://doi.org/10.31887/DCNS.2020.22.2/jgiedd Hill K (2025) A teen was suicidal. ChatGPT was the friend he confided in. The New York Times. https://www.nytimes.com/2025/08/26/technology/chatgpt-openai-suicide.html Hopelab, Common Sense Media, & The Center for Digital Thriving at Harvard Graduate School of Education (2024) Teen and young adult perspectives on generative AI: patterns of use, excitements, and concerns. https://hopelab.org/stories/teen-and-young-adult-perspectives-on-generative-ai Ibrahim L, Hafner FS, Cheng M, Lee C, Anselmetti R, Willer R, Rocher L, Yang D (2026) Sycophantic AI makes human interaction feel more effortful and less satisfying over time (Version 2). arXiv. https://doi.org/10.48550/ARXIV.2605.07912 Iloanusi N-J, Chun SA (2024) AI impact on health equity for marginalized, racial, and ethnic minorities. In: Proceedings of the 25th annual international conference on digital government research, pp 841–848. https://doi.org/10.1145/3657054.3657152 Jick TD (1979) Mixing qualitative and quantitative methods: triangulation in action. Adm Sci Q 24(4):602–611. https://doi.org/10.2307/2392366 Kim P, Xie Y, Yang S (2025) “I am here for you”: how relational conversational AI appeals to adolescents, especially those who are socially and emotionally vulnerable. arXiv:2512.15117. https://doi.org/10.48550/arXiv.2512.15117 Lee A, Culver M (2026) Youth, AI, and the relationships that shape them. The Rithm Project. https://www.therithmproject.org/research Lee S, Lee N, Sah YJ (2020) Perceiving a mind in a chatbot: effect of mind perception and social cues on co-presence, closeness, and intention to use. Int J Hum Comput Interact 36(10):930–940. https://doi.org/10.1080/10447318.2019.1699748 Madden M, Calvin A, Hasse A, Lenhart A (2024) The dawn of the AI era: teens, parents, and the adoption of generative AI at home and school. Common Sense. https://www.commonsensemedia.org/sites/default/files/research/report/2024-the-dawn-of-the-ai-era_final-release-for-web.pdf Maheux AJ, Maes C (2026) Coming of age with generative AI: the high stakes of automating adolescent development. Perspectives on Psychological Science. Advanced Online First. https://doi.org/10.1177/17456916261460771 Maheux AJ, Burnell K, Maza MT, Fox KA, Telzer EH, Prinstein MJ (2025) Annual research review: adolescent social media use is not a monolith: toward the study of specific social media components and individual differences. J Child Psychol Psychiatry 66(4):440–459. https://doi.org/10.1111/jcpp.14085 Maheux AJ, Akre-Bhide S, Boeldt D, Flannery JE, Richardson Z, Burnell K, Telzer EH, Kollins SH (2026a) Generative artificial intelligence applications use among US youth. JAMA Netw Open 9(2):e2556631. https://doi.org/10.1001/jamanetworkopen.2025.56631 Maheux AJ, Maes C, Power R, Mbuakoto C, Vaccaro A (2026b) Adolescents in the age of generative AI: a half-decade systematic review. PsyArXiv. https://doi.org/10.31234/osf.io/fxbtp_v1 Maples B, Cerit M, Vishwanath A, Pea R (2024) Loneliness and suicide mitigation for students using GPT-3-enabled chatbots. Npj Ment Health Res. https://doi.org/10.1038/s44184-023-00047-6 McClain C, Anderson M, Sidoti O, Bishop W (2026) How teens use and view AI. Pew Research Center. https://www.pewresearch.org/internet/2026/02/24/how-teens-use-and-view-ai/ Microsoft (2026) Global AI adoption in 2025—AI Economy Institute. Microsoft Corporate Responsibility. https://www.microsoft.com/en-us/corporate-responsibility/topics/ai-economy-institute/reports/global-ai-adoption-2025/ Navarro JL, Tudge JRH (2023) Technologizing Bronfenbrenner: neo-ecological theory. Curr Psychol 42(22):19338–19354. https://doi.org/10.1007/s12144-022-02738-3 O’Neil E, Anderson KS, White B, Zimmermann L, Mann S, Robb MB (2026) The common sense media census: AI use by tweens and teens. Common Sense Media. https://www.commonsensemedia.org/sites/default/files/research/report/2026-ai-use-by-tweens-and-teens.pdf OECD (2024) Towards digital safety by design for children. OECD Digital Economy Papers. https://doi.org/10.1787/c167b650-en OpenAI (2026) Helping ChatGPT better recognize context in sensitive conversations. https://openai.com/index/chatgpt-recognize-context-in-sensitive-conversations/ Orben A (2020) The Sisyphean cycle of technology panics. Perspect Psychol Sci 15(5):1143–1157. https://doi.org/10.1177/1745691620919372 Raedler JB, Swaroop S, Pan W (2025) AI companions are not the solution to loneliness: design choices and their drawbacks. In: HAIC workshop, international conference on learning representations (ICLR 2025). https://openreview.net/pdf?id=xFrlcTacCE Robb MB, Mann S (2025) Talk, trust and trade-offs: how and why teens use AI companions. Common Sense Media. https://www.commonsensemedia.org/sites/default/files/research/report/talk-trust-and-trade-offs_2025_web.pdf Ronge R, Maier M, Rathgeber B (2025) Towards a definition of generative artificial intelligence. Philos Technol 38(1):31. https://doi.org/10.1007/s13347-025-00863-y Saeri AK, Graham J, Noetel M, Slattery P, Ah-king D, Aittokallio E, Akindehin I, Mahdi AA, Alhajjar E, Lipcsey RA, Ang G, Azam CM, Azaria A, Balkissoon R, Barberá I, Bareato C, Barry J, Basehart M, Bean AM, … Thompson N (2026) Prioritization of risks from artificial intelligence: a Delphi study of 272 international experts (Version 1). arXiv. https://doi.org/10.48550/ARXIV.2606.04490 Smith MG, Bradbury TN, Karney BR (2025) Can generative AI chatbots emulate human connection? A relationship science perspective. Perspect Psychol Sci 20(6):1081–1099. https://doi.org/10.1177/17456916251351306 Somerville LH (2013) The teenage brain: sensitivity to social evaluation. Curr Dir Psychol Sci 22(2):121–127. https://doi.org/10.1177/0963721413476512 Steinberg L (2014) Age of opportunity: lessons from the new science of adolescence. Houghton Mifflin Harcourt Sun X, Wang Y, McDaniel BT (2026) AI companions and adolescent social relationships: benefits, risks, and bidirectional influences. Child Dev Perspect. https://doi.org/10.1093/cdpers/aadaf009 Unicef (2025) Guidance on AI and children | Office of Strategy and Evidence Innocenti. https://www.unicef.org/innocenti/reports/policy-guidance-ai-children Valkenburg PM (2022) Social media use and well-being: what we know and what we need to know. Curr Opin Psychol 45:101294. https://doi.org/10.1016/j.copsyc.2021.12.006 Valkenburg PM, Peter J (2007) Online communication and adolescent well-being: testing the stimulation versus the displacement hypothesis. J Comput-Mediat Commun 12(4):1169–1182. https://doi.org/10.1111/j.1083-6101.2007.00368.x Valkenburg PM, Peter J (2013) The differential susceptibility to media effects model: differential susceptibility to media effects model. J Commun 63(2):221–243. https://doi.org/10.1111/jcom.12024 Wang N, Lester J (2023) K-12 education in the age of AI: a call to action for K-12 AI literacy. Int J Artif Intell Educ 33(2):228–232. https://doi.org/10.1007/s40593-023-00358-x Xu Y, Thomas T, Yu C-L, Pan EZ (2025) What makes children perceive or not perceive minds in generative AI? Comput Hum Behav Artif Hum 4:100135. https://doi.org/10.1016/j.chbah.2025.100135 Zhai C, Wibowo S, Li LD (2024) The effects of over-reliance on AI dialogue systems on students’ cognitive abilities: a systematic review. Smart Learn Environ 11(1):28. https://doi.org/10.1186/s40561-024-00316-7 Acknowledgements We acknowledge support from the Winston Center on Technology and Brain Development. This project was supported by teen and young adult advisors, including SEA Lab Research Assistants: Zoya Barnes, Sam Davis, Sowndarya Chivukula, Maggie Meyers, Lillian Wu, and Jeslyn Pratiknjo. The authors also thank colleagues and collaborators who contributed to the development of this project, including Tracey Kirui and other Hopelab staff and Drs. Eva Telzer and Mitch Prinstein of the UNC Winston Center. Funding This study was supported by a grant from Hopelab. Author information Authors and Affiliations Contributions AJM, CM, JH, AV, and KB conceptualized the project. AJM, CM, MGV, AV, and KB collected and analyzed data. AJM drafted the manuscript. All authors reviewed the manuscript. Corresponding author Ethics declarations Conflict of interest KB has served as a paid consultant for social media litigation. JH works for Hopelab, which provided funding for the project. AJM and KB report ongoing collaborations with Aura. AJM attended an unpaid convening at OpenAI on child safety. Ethical approval and consent to participate All study procedures were approved by the University of North Carolina IRB (25-2587). All participants provided informed consent prior to completing the study. Additional information Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Rights and permissions Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/. About this article Cite this article Maheux, A., Mbuakoto, C., Valentino, M. et al. “Already outdated and still under review”: mapping the landscape of research on adolescent development and AI chatbot use. AI & Soc (2026). https://doi.org/10.1007/s00146-026-03380-4 Received: Accepted: Published: Version of record: DOI: https://doi.org/10.1007/s00146-026-03380-4