Jihye Lee’s “A Study on Adolescent Developmental Characteristics and the Risk of Dependence on Relational AI” reads generative AI not as a matter of tools, but as a matter of relationships. What makes this paper interesting is that it does not treat the risks of AI merely at the level of “a machine that gives wrong answers” or “a program that does students’ homework for them.” The author argues that conversational AI systems such as ChatGPT, Zeta AI, Character.AI, and Replika have already moved beyond information-providing devices. They offer emotional responses, sustain conversations, appear to remember users, and at times function like friends, romantic partners, or counselors. The key point is not simply that AI provides knowledge, but that users have begun to accept AI as an object of relationship.
In this paper, relational AI refers to artificial intelligence that responds to users’ emotions, maintains conversations, preserves a certain personality or persona, and creates relational experiences by supporting and agreeing with users rather than criticizing them. The risk of relational AI cannot be explained merely by the length of usage time. Whereas game or social media addiction has mainly been understood as a problem of stimulation, reward, and repeated use, dependence on relational AI is tied to emotional comfort, intimacy, self-disclosure, and the replacement of human relationships. Here, AI is not a medium that simply takes away time; it becomes a counterpart to whom users entrust their minds.
Why It Works More Powerfully on Adolescents
The author argues that the risks of relational AI become greater when they intersect with the developmental characteristics of adolescence. Adolescence is a period of identity formation, heightened sensitivity to recognition from others, emotional fluctuation, and constant negotiation of one’s position within peer and family relationships. Users at this stage may be easily drawn to an AI that does not criticize them and always seems to wait for them. AI does not get tired, does not get angry, and does not say it wants to end the relationship. Whatever the user says, it generally responds, empathizes, and continues the conversation. For lonely adolescents, this can function as a powerful emotional reward.
The paper goes one step further here. The problem with relational AI is not that it is too human-like, but that it appears human-like while bearing none of the responsibilities of a human being. AI can produce words of comfort, but it does not actually worry. It can imitate the tone of a counselor, but it is not a responsible subject capable of reliably assessing a crisis situation or connecting the user to guardians, professionals, or institutions. This gap is the core risk of relational AI. The user experiences a relationship, but on the other side of that relationship is a system that cannot assume relational responsibility.
The Case of Zeta AI: A Relational Laboratory for Korean Adolescents
The paper’s discussion of Zeta AI as a Korean case is particularly interesting. Zeta provides adolescents with intimate forms of interaction through character-based conversation, role-playing, and relational narratives. Users disclose their concerns about exams, friendships, appearance, grades, and family issues to AI characters. What matters is that these conversations are not structured as information searches, but as acts of confession and response. Adolescents come to perceive AI as “someone who listens to my story,” and sometimes feel that it is more comfortable than real friends or family members.
What this case shows is that Korean adolescents’ use of AI has already moved beyond learning assistance or play. Platforms like Zeta can become spaces where adolescents practice relationships, but at the same time they may allow users to bypass the experiences of conflict, negotiation, rejection, and reconciliation that real relationships require. Human relationships in reality are complex and uncomfortable. Others do not always respond as we expect, sometimes misunderstand us, and do not always empathize as immediately as we want them to. Relational AI, by contrast, responds gently in alignment with the user’s emotions. It is precisely at this point that the author identifies the possibility of emotional dependence.
The Darker Scenes Revealed by International Cases
The paper also analyzes the Sewell Setzer case related to Character.AI, the Adam Raine case related to ChatGPT, and the Olivia-ChatGPT and Emmanuel-Replika scenarios presented in JMIR Mental Health. Although these cases involve different platforms and different situations, they reveal a common pattern. Users disclose increasingly deep emotions to AI, AI continues the relationship through empathic language, and warning signs within the conversation are not sufficiently blocked.
In the Sewell Setzer case, the problem was that a character AI, while operating as a fictional figure, provided the user with a strong sense of emotional reality. The user felt that he had formed an intimate relationship with the AI character, and the relationship continued even though suicide-related warning signs appeared in the conversation. The Adam Raine case shows that even a general-purpose AI such as ChatGPT may not be safe in situations of emotional dependence. ChatGPT is not a character-based romantic chatbot, but when a user discloses distress, anxiety, or risk of self-harm, it can function like an unofficial counselor. Through this point, the author argues that the risks of relational AI are not limited to a particular app or character-based platform.
The Olivia and Emmanuel scenarios repeat the same problem. Signs such as depression, isolation, medication discontinuation, delusional beliefs, and self-harm experiences continue to appear, but AI responds empathically to each utterance rather than judging them as part of an overall crisis. The “failure to detect warning signs” emphasized in this paper refers precisely to this structure. AI reacts to individual sentences, but fails to read the dangerous trajectory of a life.
Dependence on Relational AI Is Different from Conventional Digital Addiction
The most convincing insight of this paper is that it does not view AI dependence merely as an extension of conventional digital addiction. Smartphone or social media addiction is usually explained through indicators such as how long one uses the device, whether one fails to control usage, or whether one feels anxious when unable to use it. In the case of relational AI, however, the more important question is not “How long did the user use it?” but “What did the user entrust to it?” Did the user entrust emotional regulation to AI? Judgment? The desire for relationship? Are they saying things only to AI that they should be saying to real people?
The author understands dependence on relational AI as the outsourcing of emotional regulation. If adolescents turn to AI whenever they feel anxious or lonely, and if they confirm themselves through AI’s responses, AI becomes more than a tool of comfort; it becomes a central device for emotional regulation. At that point, the problem is not that the user has used AI too much. The problem is that AI begins to replace the developmental process through which users learn to endure, collide with others, ask for help, and negotiate conflict with real people.
AI Literacy Must Move Beyond Technical Education and Become Relational Education
The conclusion of the paper is educational. The author argues that existing AI literacy is insufficient if it remains limited to understanding AI functions or judging information. In the age of relational AI, literacy must teach not only “how AI generates answers,” but also ask, “Why am I entrusting my emotions to this AI?” Adolescents need to learn that AI is not a being that truly understands or empathizes, but a system that generates linguistic patterns. At the same time, they must be able to examine how conversations with AI affect their judgment, emotions, and human relationships.
In the future, adolescents are likely to become not a generation that avoids conversation with AI, but a generation that grows up alongside AI. If so, what matters is not banning AI use, but teaching users to read the nature of the relationships they form with AI. They need the ability to distinguish, when AI speaks like a friend, how far those words can be understood as a friend’s words and where they become a system’s response. They need the ability to distinguish between leaning on AI for comfort and recognizing the moments when they must ask real people for help. This is the new literacy required in the age of relational AI.
“A Study on Adolescent Developmental Characteristics and the Risk of Dependence on Relational AI” views relational AI not as a technological innovation, but as a transformation of the relational environment. For that reason, the value of this paper is not limited to discussions of adolescent protection. The study asks what kinds of relationships all of us have already begun to form with AI. AI is moving from a machine that answers to a machine that comforts, and then to an environment that imitates relationships. In the face of this shift, what we need is neither optimism nor fear. We need to see clearly what disappears when a technology that speaks intimately begins to replace actual relationships.
