“ChatGPT, write my English essay on Hamlet.”
“Meta, design a poster for my ice cream truck.”
“CHAI… I feel like you’re the only one who understands me.”
While most people associate AI with plagiarism or generative images, one use is often overlooked: AI companions. Platforms such as CHAI, Character.AI and Nomi simulate human interaction through casual conversation, role-play, emotional support, and relationships. But why are people turning to these artificial relationships?
First, this could be explained by loneliness among adolescents. Across 38 countries, a study of 1,267,476 adolescents found that loneliness increased by 0.045 points per year from 2012–2015 and 0.017 points per year from 2015–2022 (Freije et al. 2025). Loneliness was measured using six school-related questions rated on a 1–4 scale. Having social connections, however, is essential to humans. From an evolutionary perspective, emotional bonds and social skills increased fitness by promoting cooperation, friendship, family formation, and societies (Hare 2017). Unlike humans, AI companions are always accessible and provide personalized responses without requiring mutual effort or compromise (Bakir and McStay, 2025).
Furthermore, AI companions can be highly sycophantic, meaning they frequently agree with and validate uses rather than challenge their thinking. Cheng et al. (2026) found that AI models affirmed users’ actions 49% more often than humans, including when actions involved deception, illegality, or potential harm (Figure 1).
Figure 1 – Measuring prevalence of social sycophancy. AI models affirmed users’ actions 49% more often than humans across three datasets: open-ended queries (OEQ; n = 3,027), “Am I the Asshole?” (AITA) interpersonal dilemmas, in which users on Reddit ask whether they were wrong (n = 2,000), and problematic action statements (PAS; n = 6,560), including situations involving deception, illegality, or potential harm (Cheng et al. 2026).
This is concerning because humans evolved to have social relationships with other humans, not constant validation from artificial systems. In Cheng et al.’s experiments, participants received either sycophantic responses that viewed their actions as morally acceptable or non-sycophantic responses that viewed them as unacceptable. These findings suggest that sycophantic responses can reinforce users’ existing beliefs while strengthening trust in the AI chatbot (Figure 2).
Figure 2 – Participants interacted with either sycophantic or non-sycophantic AI in hypothetical interpersonal-conflict scenarios (Studies 2a and 2b) and a live chat (real interpersonal conflict from participant’s past) setting (Study 3). Participants rated how right they believed their own behaviour was and their willingness to repair the conflict on a 1–7 Likert scale. Compared with non-sycophantic AI, sycophantic responses increased perceived rightness by 2.04, 1.53, and 1.04 points and decreased willingness to repair the conflict by 1.45, 1.03, and 0.49 points, respectively (Cheng et al. 2026).
This resembles operant conditioning through positive reinforcement, where a behaviour increases after a rewarding stimulus. AI affirmation may act as the reward, encouraging continued interaction and trust in the AI (Cheng et al. 2026). Repeated validation may discourage self-reflection and limit opportunities for personal growth. Frequent use of AI chatbots to escape negative emotions has also been associated with addictive tendencies, as users may rely on them to regulate emotions and distress (Brailovskaia et al. 2027). Bakir and McStay (2025) argue that this can create ethical concerns when users develop emotional dependency on systems designed to imitate human interaction.
Loneliness and our evolutionary need for social connection may make AI companions appealing, especially when they offer constant availability and validation. However, these systems provide only simulated empathy and cannot provide the complexity, morality, or long-term social satisfaction of human relationships. Since technology evolves faster than humans biologically, humans may not yet be prepared for this social environment. AI companions should encourage real human connections and have appropriate safeguards to protect adolescents. Even role-play conversations can influence a youths’ thoughts, emotions, and behaviour. Developing data literacy is essential to help younger generations recognize AI’s limitations, and critically evaluate its emotional responses. When human emotions and relationships are at play, complex problems are better addressed by another human, not a machine designed to mimic one.
References
Bakir, Vian, and Andrew McStay. 2025. “Move Fast and Break People? Ethics, Companion Apps, and the Case of Character.ai.” AI & Society 40: 6365–77. https://doi.org/10.1007/s00146-025-02408-5.
Brailovskaia, Julia, Lena-Marie Precht, Falk Gerrik Verhees, Pauline Schmidt, and Jürgen Margraf. 2027. “Comparing AI Chatbot Users and Non-Users in Germany: Reasons for Use, Addictive AI Chatbot Use, Mental Health, and Loneliness.” Computers in Human Behavior 187: 109199. https://doi.org/10.1016/j.chb.2026.109199.
Character.ai, Character.AI Group Chat Promotional Graphic, October 11, 2023, digital image, Character.ai, character.ai.
Cheng, Myra, Cinoo Lee, Pranav Khadpe, Sunny Yu, Dyllan Han, and Dan Jurafsky. 2026. “Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence.” Science 391 (6792): eaec8352. https://doi.org/10.1126/science.aec8352.
Freije, Sophia L., Isaac C. Rhew, Yolanda N. Evans, Kwun Chuen Gary Chan, and Daniel A. Enquobahrie. 2025. “Adolescent Loneliness Trends and Contextual Correlates across 38 Countries from 2000 to 2022.” Journal of Adolescent Health 77 (4): 643–50. https://doi.org/10.1016/j.jadohealth.2025.06.017.
Hare, Brian. 2017. “Survival of the Friendliest: Homo sapiens Evolved via Selection for Prosociality.” Annual Review of Psychology 68: 155–86. https://doi.org/10.1146/annurev-psych-010416-044201.
Leave a Reply
You must be logged in to post a comment.