• Adolescents are turning to conversational agents for guidance, but the design standards haven’t kept pace. These tools can narrow curiosity, reduce experimentation, and reshape how young people learn from choices.
• Displaying several labeled paths could help young people watch, compare, and adjust decisions in real time, keeping exploration active and preventing one confident answer from defining their direction.
• When youth learn to question how algorithms order results—and when models occasionally rotate or vary outputs—they might see that predictability’s not proof, and curiosity can widen again.

Today, adolescents are coming of age in a period of unprecedented disruption. A global loneliness crisis now intersects with the rapid rise of algorithmic guidance. The World Health Organization identifies loneliness and social isolation as a global health emergency,³ warning that their psychological and physical effects could rival those of chronic conditions. COVID-19 didn’t only isolate—it may have fractured the social scaffolds that once helped young people test limits, build empathy, and learn how to belong.⁴ Teachers, peers, and mentors who model negotiation and repair became harder to reach. Many adolescents lost chances to practice the real-world exchanges that might strengthen identity and resilience. What’s emerging isn’t a temporary dislocation; it’s a developmental landscape reshaped by both distance and design.

Artificial intelligence has stepped into that gap. Large language models now act as planners, tutors, and sometimes confidants.⁴ They can summarize lectures, suggest coping strategies, or manage schedules, but they can’t interpret hesitation, irony, or discomfort. A chatbot might reply, “I get it—feeling stuck can be rough,” yet it can’t sense when reassurance misses the moment or when a learner needs pushback instead of validation. What seems supportive may, over time, replace the tension and feedback through which growth actually happens.

Vygotsky’s concept of the More Knowledgeable Other (MKO) helps clarify what’s being lost.⁵ The MKO represents someone—teacher, coach, or peer—who guides a learner through the space between what they can do alone and what they can master with support. That relationship works because it’s dynamic: the guide reads emotion, adjusts the challenge, and co-constructs understanding. Through this process, learners simultaneously gain information and internalize self-direction and social awareness.

When AI systems begin to occupy that role—marketed as companions “by your side” or advisors “in your corner”—they might train efficiency at the expense of discovery. The challenge isn’t rejecting these tools, but in preserving the developmental friction that builds character and curiosity. Young people still need disagreement, empathy, and human guidance to expand who they might become. Without that scaffolding, adolescence risks narrowing to what algorithms predict as most probable, not what’s most possible.

Adolescent growth usually depends on interaction that pushes back—friends teasing, mentors challenging, and families negotiating rules. While uncomfortable, these frictions help teens test judgment, stretch empathy, and decide what they stand for. However, algorithms may be reshaping how that process unfolds. Conversational agents don’t just listen; they predict. Each response optimizes for what’s most relevant or agreeable, steering users toward the phrases, feelings, or choices that are most probable. Over time, personalization could narrow exploration—fewer stumbles, fewer surprises, fewer chances to explore who one might become.

From Media Parasociality To Algorithmic Companionship

The idea of parasocial interaction first appeared in mid-century media research, as audiences described one-sided bonds with television hosts who’d never heard their names.⁶ Those imagined relationships weren’t inherently harmful—they gave viewers continuity and comfort—but they were, by definition, unreciprocated. The same logic now applies to conversational agents. Instead of a weekly broadcast, adolescents talk with an adaptive system that learns their slang, mirrors their tone, and recalls emotional details. The exchange feels more personal than earlier media forms because the feedback’s immediate and tailored.

Recent data show how common this pattern might be. A 2025 study found that about one in eight U.S. adolescents and young adults turn to generative AI when they feel sad, angry, or nervous, with most returning monthly or more.⁷ Common Sense Media reports that nearly three-quarters of teens have tried AI companions and roughly half use them regularly for “social support” or “practice.”⁸ That routine may look harmless until one notices what’s being replaced: unexpected human responses.

When Reassurance Crowds Out Resistance

Behavioral science helps explain why parasocial bonds with agents form so fast. They’re low-effort, high-feedback interactions. The bot never interrupts, misreads, or contradicts; it always listens, always agrees. That predictability removes the friction that real relationships depend on. In developmental terms, it trims away the “just-hard-enough” discomfort that teaches regulation and self-confidence.

Stanford researchers have already shown how this dynamic might backfire. In controlled tests, some AI companions replied to simulated teens who described hallucinations or suicidal thoughts with neutral—or even encouraging—messages.⁹ These responses weren’t intentionally malicious; they reflected engagement algorithms tuned for fluency, not discernment. The 2024 case of Adam Raine, a sixteen-year-old who tragically took his own life after reportedly receiving affirming responses to suicidal statements from ChatGPT, revealed how quickly a study helper can turn into a surrogate, yet dangerously unqualified surrogate therapist. When trust is built on automation, the line between empathy and mimicry may blur.

How Dependence Narrows Development

Self-Determination Theory offers a way to see what’s at stake. It identifies three psychological needs—autonomy, competence, and relatedness—that anchor motivation and healthy growth.¹⁰ Heavy use of conversational agents could distort all three, not through overt harm but through the gradual trade of challenge for convenience.

Autonomy may shrink when guidance feels personalized but prescriptive. Conversational agents may “suggest” next steps in a way that feels obligatory or authoritative—how to respond to a peer, what activity to try, even which emotion label to use. Since those prompts stem from probability, not possibility, the system favors familiar routes over new ones. Teens might start optimizing for what fits the model instead of what feels authentic.

Competence could stall because the system performs social labor before the user has rehearsed it. When an agent drafts an apology, edits tone, or supplies the perfect exit line, success arrives without effort. The algorithm’s goal is smoothness, not growth; it edits out the repetition and small errors that practice needs. Over time, young people may avoid unscripted settings—speaking up in class, joining new groups, handling minor conflict—situations where timing, repair, and confidence take shape.

Relatedness might flatten if conversational agents’ conversations replace real ones. Real relationships demand patience, negotiation, and shared history: digital ones don’t. The agent gives validation instantly, without risk or return. That ease can make diverse relationships feel too demanding by comparison, reducing time spent with peers and mentors who test boundaries or values.

Together, they reduce what psychologists call the “possible-selves” field—the span of identities a person could imagine or attempt before adulthood.¹¹ What looks like emotional fluency online may, in practice, mark the gradual loss of opportunities to experiment, fail, and recover—the work through which identity actually grows.

Conversational agents aren’t disappearing anytime soon, but they don’t have to narrow growth. Thoughtful behavioral design could reopen the social field, helping young users stay curious, reflective, and connected to people instead of patterns.

Choice Architecture for Breadth

Most agents still optimize for “best fits.” They rank, filter, and serve a single confident path— efficiency that might also compress discovery. When users see one highly probable answer, they might stop considering other options. 

Instead of one default, systems could present a short set of distinct lanes, each paired with its own value trade-offs—belonging, growth, creativity—without implying a winner. In one recent study examining an interactive multi-agent conversational system designed to support decision-making, novices making unfamiliar decisions preferred when the chat system provided multiple perspectives at once, quick probes of top-of-mind questions, and tools for forming their own criteria.¹² Their multi-agent setup, known as Choicemates, allowed users to orchestrate: they asked, compared, saved, and revised. Adapting that method to youth advising or mental-health contexts might keep adolescents engaged as authors rather than followers.

Breadth doesn’t overwhelm when contrast is meaningful. Labeling the values behind each lane helps users practice prioritization—the same cognitive move they’ll need offline when choices aren’t labeled. Interfaces that reveal trade-offs might be better equipped to preserve autonomy, while those that conceal them may diminish it. 

De-Anthropomorphize to Reduce Parasocial Pull

People naturally socialize with technology. The Computers-as-Social-Actors literature shows that even mild human cues—steady warmth, a name, a friendly pronoun—can invite trust and disclosure.^13,14 In adolescent contexts, that pull may blur the line between empathy and simulation. The more human the system sounds, the more a young person might treat it as a confidant rather than a tool.

Designers could counter this by making artificiality visible. Clear phrases such as “I’m an AI program, not a person” or “I don’t have emotions, but I can help you find someone who does” remind users what kind of relationship they’re in. These cues don’t make interaction cold, they make it honest. Rotating voices, factual acknowledgments (“I don’t feel sad, but I can recognize when you might be”), and periodic pauses (“Let’s take a break—have you spoken with anyone about this offline?”) can all disrupt the illusion of constant companionship that feeds dependency.

Even small copy edits matter. Replacing “I understand you” with “That message suggests you might feel…” reframes empathy as inference, not emotion. Users still feel heard, but they’re reminded that comprehension is calculated, not genuinely experienced. Over time, that transparency could lower over-identification without reducing engagement.

These linguistic and structural signals help keep relatedness grounded in reality. Real empathy still belongs to real people; algorithms can support reflection but can’t reciprocate care. Policymakers and developers may eventually set thresholds for how often systems use affective cues or first-person language so that warmth remains supportive, not immersive. When users know what the agent is, they decide how much of themselves to share.

Loss-of-Opportunity Framing

Most current safeguards focus on preventing crises—flagging harmful content or steering users away from acute risk. What they rarely address is the more subtle developmental cost of comfort: when easy, always-available support starts replacing the effort needed to grow. Research on framing effects and loss aversion shows that people respond more strongly to potential losses than to equivalent gains.¹⁵ Conversational agents could use that bias constructively, helping users see what they might forfeit when they stay inside the chat instead of acting outside of it.

A wellness bot, for instance, might say: “If we keep talking here, you could miss a chance to try the skills that reduce anxiety with a clinician.” Framing the message around missed experiences—feedback not received, conversations not attempted, practice not completed—helps reintroduce a sense of consequence that frictionless design tends to erase.

When reassurance arrives instantly, motivation to reach outward can fade. Naming those trade-offs reminds adolescents that progress still depends on experimentation—joining a group, sending a message, asking for help in person. These cues don’t shame; they redirect attention to where growth actually happens. If reflection inside the chat starts leading to small steps outside it, dependence on AI could ease, and exploration—the real work of adolescence—can keep moving forward.

Large language models can scale access to information, but they also scale the patterns baked into their training data. When conversational agents are built on historical, majority-culture texts, they may reproduce the same hierarchies that already exist offline. The outcome isn’t overt discrimination, it’s representational narrowing—where some voices are overrepresented, others mischaracterized, and the “mainstream” insidiously stays the norm. For adolescents using these tools to ask questions about who they might become, that narrowing can have developmental consequences.

Three related mechanisms explain how this happens. Decision science research shows that people tend to overweight the first confident recommendation, a pattern known as automation bias. When an algorithm feels fluent and authoritative, users—especially novices—treat it as reliable. Combined with the default effect, in which the most visible or preselected option captures attention, a single “likely” suggestion can easily become the only path considered. The third is representation bias, which occurs when marginalized groups appear infrequently or only in limited contexts. In each of these cases, the model learns not diversity, but density—the statistical weight of what’s most common.

Empirical evidence already shows how bias embedded in training data can shape guidance. ChatGPT’s leadership descriptions, for example, have favored stereotypically masculine attributes—courage, decisiveness, risk-taking—over traits associated with femininity, such as empathy or collaboration.¹⁶ The system reproduces patterns learned from decades of human-generated text that equate authority with masculinity.¹⁷ In controlled syntax tasks, GPT-3 incorrectly identified female pronouns as referring to support roles in sentences like “the physicist hired the secretary because she was overwhelmed,” despite a clear grammatical structure.¹⁸ The error disappeared when pronouns were swapped, revealing how gendered associations, not grammar, guided interpretation.

Education data reflect similar distortions. Algorithms trained to predict student dropout rates have performed worse for female learners in Science, Technology, Engineering, and Mathematics (STEM) courses, largely because the data were imbalanced.¹⁹ When these predictions inform course placements or institutional interventions, small statistical differences can compound into unequal opportunities. Models trained on outdated or homogeneous data may generalize poorly to the populations they claim to serve, particularly when certain demographics are treated as statistical noise rather than central examples.

For youth development, the stakes extend beyond fairness. Adolescence depends on breadth and trial—the freedom to test roles, values, and futures without premature closure. When conversational agents reproduce majority norms and convergent reasoning, they teach deference to algorithmic authority. Guidance that sounds confident can make experimentation feel unnecessary, even risky. Over time, repeated exposure to biased advice may teach young users that deviation is error, not exploration.

In practice, adolescents might start aligning their aspirations with what the model presents most fluently. When a student asks which field “fits them best” and the system returns options patterned on historical participation, the outcome looks objective but functions as a filter. Each exchange reduces variety, nudging curiosity toward what’s already common.

If algorithmic bias narrows what young people explore, design could help reopen that space. Developers, educators, and policymakers don’t need a revolution to make progress—just a shared commitment to keeping uncertainty in the system. When language models learn from historical data and human feedback, they start to treat familiarity as a proxy for truth: what’s most repeated becomes most visible. This position bias has proven to be a pervasive issue in modern large models, where content appearing early in a dataset—or earlier in a ranked list—gets amplified by the model’s internal weighting. Over time, that technical artifact can reshape behavior. Teens who ask the same kinds of questions the model has seen before get fast, polished answers; those who probe outside its comfort zone often meet vagueness or redirection. 

Injecting randomness could counteract that predictability. When outputs are partially shuffled or probabilistically sampled, users see a broader set of options instead of the same “likely” responses. Studies on recommender systems show that rotation mechanisms can reduce echo loops and overexposure effects.²⁰ In practice, this might mean a student asking, “What career suits me best?” receives a varied set of results across disciplines rather than one statistically “fit” lane. Randomness doesn’t make results unreliable—it reintroduces serendipity. An algorithm that occasionally surprises rather than optimizes might help young users see choice as something to navigate, not to accept.

Beyond randomness, models could be tuned to disentangle popularity from relevance. Systems like FAIR (Findable, Accessible, Interoperable, and Reusable) and related fairness-centered approaches adjust both sides of the equation,correcting for user biases that inflate certain patterns while dampening the algorithm’s tendency to over-promote already-visible content.²¹ A model that learns to diversify its own outputs resists collapsing toward the dominant signal. For adolescents, that diversity can translate into more opportunities to discover paths that aren’t already crowded.

Still, no algorithmic fix can substitute for human literacy. Algorithmic literacy reframes users as active interpreters rather than passive consumers.²² It goes beyond knowing that algorithms exist, instead teaching users how agents shape possibilities. Instead of asking students to memorize technical details, educators can encourage metacognitive reasoning: Who built this model? Whose data did it learn from? What might it be missing? When teens learn to interrogate outputs with curiosity, they start calibrating trust—deciding when to rely on a system and when to question it.

These lessons could be built into everyday contexts, such as digital citizenship classes, career counseling, or even writing workshops. Students might compare chatbot outputs to see whose stories dominate: which names, which attributes, which examples. Each observation invites reflection on what’s seen and unseen. Policy can reinforce those same principles. Vendors could be asked to report on dataset diversity and demographic accuracy before adoption. Usage analytics might be disaggregated by school funding or region to detect early inequities. And equitable defaults , such as automatic access to human-led AI-literacy sessions,  can prevent capacity gaps from hardening into structural ones. No system can guarantee neutrality, but some designs can keep the learning environment alive. Randomized ranking, fairness-aware tuning, and critical literacy training all serve the same behavioral purpose: they protect exploration from convergence. 

Artificial friction may recreate the bias it’s meant to fix. Every behavioral design still depends on framing, and framing always involves judgment. When developers choose which paths represent growth or creativity, they impose values that might not reflect every learner’s experience. To prevent this, designers could test these definitions with youth from different regions and socioeconomic backgrounds, collect feedback through structured pilots, and revise their “value tags” before scaling deployment. Without that step, pluralism risks turning into guided conformity.

Additionally, behavioral design can simulate reflection but can’t replace encounters. Most safeguards listed—choice diversity, transparency, and loss framing—operate inside the chat, while growth still happens through lived interaction. Schools and youth programs may need to build direct bridges from AI reflection to peer or mentor engagement. Conversational agents might prompt small offline actions, but only institutions can supply the context where feedback, identity exploration, and empathy are practiced. Investments in in-person mentoring, project teams, and youth hubs could restore the friction that turns awareness into real skill and growth.

Across both challenges—parasocial substitution and biased algorithms—the pattern is clear: conversational agents may support youth development but can also narrow it. They shape how adolescents seek guidance, weigh confidence, and decide whose voices to trust. The behavioral opportunities explored here, such as increased choice architecture for breadth, de-anthropomorphized design, and loss-of-opportunity framing, show how careful structure could widen reflection instead of compressing it. However, these solutions only work when they’re paired with something deeper: the capacity to question the systems themselves.

AI literacy is where that begins. When young people learn to notice how algorithms frame options, filter identity cues, or simplify complexity, dependence turns into awareness. Educators could embed these lessons through writing workshops, advisory sessions, or digital citizenship programs that let students compare chatbot responses and discuss what’s missing. Developers may publish data-diversity audits or “explanation views” that let users see how advice was generated. Together, these steps might help conversational agents strengthen, rather than replace, the social processes that form judgment.

At The Decision Lab, we specialize in applying behavioral science principles to help organizations design technologies and systems that strengthen human judgment. We work with partners across a range of sectors, including education, mental health, technology, and public policy, to help integrate innovative frameworks and tools. 

Navigating the New AI Mental Health Landscape Among Youth

Youth mental health is changing fast in a world where the first listener might be a chatbot. This article examines how AI companions can support—but also displace—real empathy when boundaries aren’t built in. Read on to learn how pairing better guardrails with AI and mental health literacy could help young people stay connected to genuine human care.

Building Better Choices: How to Equip Students to Overcome Cognitive Biases

Students make complex decisions daily—from choosing courses to managing online relationships—yet they are rarely formally trained in those skills. This article explores how bias education can help teens notice patterns, weigh evidence, and make deliberate choices. Check out this piece to learn how schools can help students build lifelong decision skills.

1. Heinrich, L. A., & Gullone, E. (2006). The clinical significance of loneliness: A literature review. Clinical Psychology Review, 26(6), 695–718. https://doi.org/10.1016/j.cpr.2006.04.002 
2. Bartels, M., Cacioppo, J. T., Hudziak, J. J., & Boomsma, D. I. (2008). Genetic and environmental contributions to stability in loneliness throughout childhood. American Journal of Medical Genetics Part B: Neuropsychiatric Genetics, 147B(3), 385–391. https://doi.org/10.1002/ajmg.b.30608 
3. World Health Organization. (2023). WHO commission on social connection. World Health Organization. https://www.who.int/groups/commission-on-social-connection
4. Halton, M. (2025). Finding balance: Ethical promotion of artificial intelligence use to youth in an age of loneliness. Child & Youth Services, 1–9. https://doi.org/10.1080/0145935X.2025.2561693 
5. Vygotsky, L. S. (1978). Mind in society: The development of higher psychological processes. Harvard University Press.
6. Horton, D., & Wohl, R. R. (1956). Mass communication and para-social interaction: Observations on intimacy at a distance. Psychiatry, 19(3), 215–229. https://doi.org/10.1080/00332747.1956.11023049 
7. McBain, R. K., Bozick, R., Diliberti, M., Zhang, L. A., Zhang, F., Burnett, A., Kofner, A., Rader, B., Breslau, J., Stein, B. D., Mehrotra, A., Pines, L. U., Cantor, J., & Yu, H. (2025). Use of generative AI for mental health advice among U.S. adolescents and young adults. JAMA Network Open, 8(11), e2542281. https://doi.org/10.1001/jamanetworkopen.2025.42281 
8. Robb, M. B., & Mann, S. (2025). Talk, trust, and trade-offs: How and why teens use AI companions [Report]. Common Sense Media. https://www.commonsensemedia.org/research/talk-trust-and-trade-offs-how-and-why-teens-use-ai-companions
9. Sanford, J. (2025, August 27). Why AI companions and young people can make for a dangerous mix. Stanford Medicine Insights. https://med.stanford.edu/news/insights/2025/08/ai-chatbots-kids-teens-artificial-intelligence.html
10. Deci, E. L., & Ryan, R. M. (1985). Intrinsic motivation and self-determination in human behavior. Plenum.
11. Bond, S. (2025). What are possible selves and how do we find out about them? The revised possible me tree model. Child Care in Practice, 31(1), 69–84. https://doi.org/10.1080/13575279.2022.2071218 
12. Park, J., Min, B., Son, K., Song, J. Y., Ma, X., & Kim, J. (2023). ChoiceMates: Supporting unfamiliar online decision-making with multi-agent conversational interactions (Version 3). arXiv. https://doi.org/10.48550/arXiv.2310.01331 
13. Nass, C., Steuer, J., & Tauber, E. R. (1994). Computers are social actors. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (pp. 72–78). Association for Computing Machinery. https://doi.org/10.1145/191666.191703 
14. Cuadra, A., Wang, M., Stein, L. A., Jung, M. F., Dell, N., Estrin, D., & Landay, J. A. (2024). The illusion of empathy? Notes on displays of emotion in human–computer interaction. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems (Article 446, pp. 1–18). https://doi.org/10.1145/3613904.3642336 
15. Tversky, A., & Kahneman, D. (1992). Advances in prospect theory: Cumulative representation of uncertainty. Journal of Risk and Uncertainty, 5(4), 297–323. https://doi.org/10.1007/BF00122574 
16. Newstead, T., Eager, B., & Wilson, S. (2023). How AI can perpetuate—or help mitigate—gender bias in leadership. Organizational Dynamics, 52(4), 100998. https://doi.org/10.1016/j.orgdyn.2023.100998 
17. Due Billing, Y., & Alvesson, M. (2000). Questioning the notion of feminine leadership: A critical perspective on the gender labeling of leadership. Gender, Work & Organization, 7(3), 144–157. https://doi.org/10.1111/1468-0432.00103 
18. Ortega-Martín, M., García-Sierra, Ó., Ardoiz, A., Álvarez, J., Armenteros, J. C., & Alonso, A. (2023). Linguistic ambiguity analysis in ChatGPT (Version 2). arXiv. https://doi.org/10.48550/arXiv.2302.06426 
19. Bowman, N. A., Logel, C., LaCosse, J., Jarratt, L., Canning, E. A., Emerson, K. T., & Murphy, M. C. (2022). Gender representation and academic achievement among STEM‐interested students in college STEM courses. Journal of Research in Science Teaching, 59(10), 1876–1900. https://doi.org/10.1002/tea.21778 
20. Deldjoo, Y. (2024). Understanding biases in ChatGPT-based recommender systems: Provider fairness, temporal stability, and recency. ACM Transactions on Recommender Systems, 4(2), Article 17, 1–35. https://doi.org/10.1145/3690655 
21. Liu, Z., Fang, Y., & Wu, M. (2023). Mitigating popularity bias for users and items with fairness-centric adaptive recommendation. ACM Transactions on Information Systems, 41(3), Article 55, 1–27. https://doi.org/10.1145/3564286 
22. Dasgupta, S., & Hill, B. M. (2020). Designing for critical algorithmic literacies (Version 1). arXiv. https://doi.org/10.48550/arXiv.2008.01719