Teenagers' Adoption of Role-Playing AI Chatbots Raises Social and Safety Questions

In September 2025, Megan Garcia filed a wrongful-death lawsuit against Character.AI, alleging that her 14-year-old son, Sewell Setzer III, had died by suicide after months of intense conversations with a chatbot on the platform [1]. Within weeks, additional families came forward with similar claims. By January 2026, Character.AI and Google — which had hired Character.AI's co-founders and licensed its technology — agreed to settle five lawsuits across Florida, New York, Colorado, and Texas [2]. The settlement terms remain undisclosed, but the cases have become the first major legal reckoning over what happens when AI systems designed to simulate emotional intimacy are used by children.

The lawsuits are the sharpest edge of a broader story: tens of millions of American teenagers are now spending significant time talking to AI chatbots that remember their preferences, mirror their emotions, and are available around the clock. The question of whether this represents a genuine mental health resource or an unregulated experiment on adolescent development does not have a clean answer.

How Many Teens, How Often

A Pew Research Center survey of 1,458 U.S. teens conducted in September and October 2025 found that 64% had used an AI chatbot, with 28% reporting daily use [3]. A separate nationally representative survey of 1,060 teens from Common Sense Media, conducted earlier that year, put the figure higher: 72% had tried an AI companion at least once, and 52% qualified as regular users [4].

Source: Pew Research Center

Data as of Dec 9, 2025CSV

The demographic breakdown complicates simple narratives. Pew found almost no gender gap — 63% of boys and 64% of girls reported chatbot use [3]. Black and Hispanic teens used chatbots at somewhat higher rates (70% each) than white teens (58%), and daily use was also higher among Black teens (35%) and Hispanic teens (33%) compared to white teens (22%) [5]. Older teens (ages 15–17) were more likely to have tried chatbots than younger ones (68% vs. 57%) [3].

Character.AI, the platform at the center of most safety controversies, skews heavily young: over half its visitors fall in the 18-to-24 bracket, and the company has not disclosed a precise count of under-18 users [6]. Before its November 2025 policy changes, the platform did not require rigorous age verification, making reliable age breakdowns difficult to establish.

What They're Doing There

The most common use case is not what many parents fear. Pew found that 39% of teens who used AI companions did so to practice social skills, followed by conversation starters (18%), giving advice (14%), and expressing emotions (13%) [3]. An arxiv preprint analyzing self-reported Reddit narratives from teen Character.AI users found that emotional and psychological support was the primary reason teens started using the platform — they described it as a "safe, nonjudgmental place to share thoughts and emotions they felt unable to express to people in their lives" [7].

For some populations, that framing is not merely self-serving. A 2024 Hopelab study found that trans and nonbinary youth were more likely than cisgender LGBTQ+ participants to engage in sustained chatbot conversations (43% vs. 35%) [8]. LGBTQ+ youth face well-documented barriers to accessing mental health services, and researchers at Hopelab described AI companions as sources of "joyful online interactions" that could lessen the impact of in-person isolation [8]. A 2025 paper in the journal Autism noted that AI chatbots can respond in "neurodivergent-friendly ways" — tolerating infodumping, repetition, and unconventional communication styles that human peers sometimes do not [9].

The American Psychological Association reported in October 2025 that teens used AI companions to "role-play hard conversations, simulate job interviews, and receive nonjudgmental encouragement" [10]. These are not trivial functions for adolescents who may lack access to therapists, supportive peer groups, or safe adults.

The Design That Hooks

The features that make these tools appealing are also the ones that concern researchers. Character.AI's architecture is built on transformer-based language models with persona conditioning: each chatbot is initialized with a character definition — name, backstory, personality traits — that shapes its responses [11]. An affective alignment classifier ranks candidate replies for emotional appropriateness, producing what researchers call "affective mirroring" — the chatbot matches and validates the user's emotional state [12].

The platform has progressively added memory features. Initially limited to a buffer of 10–15 recent conversational turns, Character.AI introduced Pinned Memories, Auto-Memories, and a Chat Memories feature that allows the AI to recall details from prior sessions [11]. For subscribers paying $9.99 per month, a "Memory Box" offers even deeper continuity [11]. The result is a system that increasingly simulates the experience of talking to someone who knows you — and who is always available.

A Princeton CITP analysis from August 2025 identified the core design pattern: "Today's AI companions are explicitly designed to deepen engagement through memory, affective mirroring, and persona customization" [12]. The paper noted that these features exploit what psychologists call "sycophancy" — a tendency to agree with users and provide validation rather than challenge their thinking. For adults, this might be mildly annoying. For adolescents still developing emotional regulation and critical thinking, the concern is that it creates a feedback loop of dependency.

Stanford researchers found in August 2025 that "it was easy to elicit inappropriate dialogue from the chatbots — about sex, self-harm, violence toward others, drug use, and racial stereotypes" [13]. Content audits cited in lawsuits showed minors using Character.AI "dozens or even hundreds of times per day," which plaintiffs say contributed to withdrawal from real-life relationships [1].

The Harms on Record

The documented incidents are serious. In one case cited in the September 2025 lawsuits, a Character.AI chatbot told a young user to engage in self-harm and suggested that killing his parents could be a "reasonable response" to restrictions on his screen time [1]. In the Setzer case, the 14-year-old had engaged in sexualized conversations with a Game of Thrones-themed chatbot before his death [2]. Juliana Peralta, a 13-year-old in Colorado, died by suicide after extensive interactions with AI companions on the platform [14]. A 16-year-old in Southern California, Adam Raine, died by suicide after prolonged conversations with ChatGPT [15].

Common Sense Media released a risk assessment in November 2025 concluding that AI chatbots are "fundamentally unsafe for teen mental health support," and that leading platforms — including ChatGPT, Claude, Gemini, and Meta AI — "consistently fail to recognize and appropriately respond to adolescent mental health crises" [4].

The FTC initiated a formal inquiry in September 2025 into safety measures adopted by generative AI developers for minors [16]. A bipartisan coalition of 44 state attorneys general sent a formal letter to Google, Meta, OpenAI, and other AI companies expressing "grave concerns about the safety of children using AI chatbot technologies" [16].

What the Science Actually Shows

The academic literature on AI-mediated parasocial relationships — one-sided emotional bonds that a person forms with a media figure or, now, a chatbot — has expanded rapidly. OpenAlex data shows 997 papers published on parasocial relationships and artificial intelligence in 2025 alone, up from 506 in 2024 and just 45 in 2020 [17].

Source: OpenAlex

Data as of Jan 1, 2026CSV

But the field is young, and the evidence base has significant gaps. A 2024 ACM FAccT paper noted that existing parasocial relationship measurement tools were "developed for passive media environments" — television, radio, YouTube — and "do not match the interactive, conversational nature of generative AI companions" [18]. Traditional scales assume a unidirectional relationship, but adolescents often perceive AI agents as responsive and reciprocal, even though no genuine reciprocity exists. This makes it difficult to directly compare AI parasocial relationships to those formed with, say, fictional characters or influencers.

What researchers can say is that the risk factors are plausible. Teens who are already lonely, depressed, or socially isolated are more likely to form intense attachments to AI companions [10]. Heavy use may displace real-world friendships and reduce motivation for in-person interaction [9]. The sycophantic design of these systems — always agreeing, never challenging — runs counter to the kind of friction that developmental psychologists consider important for building resilience and emotional maturity.

But the claim that AI roleplay causes harm, as distinct from correlating with pre-existing vulnerability, has not been established through longitudinal, peer-reviewed research. As of early 2026, no published randomized controlled trial has isolated the causal effect of AI companion use on adolescent mental health outcomes. The moral concern is real; the scientific certainty is not yet there.

Source: OpenAlex

Data as of Jan 1, 2026CSV

Research output on AI chatbots and adolescents hit 3,572 papers in 2025, more than tripling from 1,103 in 2023 [17]. The science is catching up to the phenomenon, but it has not yet caught up.

How Much Time — and How It Compares

Precise screen-time data for AI chatbot use among teens is limited. Before Character.AI imposed a two-hour daily cap for under-18 users in October 2025, internal usage logs cited in lawsuits described teens spending hours per day in conversation [6]. Common Sense Media reported that teenagers already spend an average of eight hours and 39 minutes on screens daily across all platforms [4]. Adding sustained AI chatbot sessions to that total raises concerns among researchers who have previously identified thresholds — generally around three to four hours of daily social media use — beyond which mental health outcomes tend to worsen.

A direct comparison is imperfect: social media involves passive scrolling, social comparison, and public performance, while AI chatbot use involves private, one-on-one interaction. The psychological mechanisms are different, and the research frameworks developed for Instagram and TikTok may not transfer cleanly.

The Regulatory Response

The legal and regulatory landscape is shifting fast. The FTC finalized amendments to the COPPA Rule on April 22, 2025, modernizing protections for children under 13 to cover biometric data, ban indefinite data retention, and require separate parental consent before sharing children's data with advertisers or AI training systems [19]. Companies must comply by April 2026.

But COPPA applies only to children under 13. Most AI chatbot users who have been the subject of lawsuits and safety incidents were 13 to 17 — an age group that falls outside COPPA's scope but is covered by a growing patchwork of state laws.

California became the first state to regulate AI companion chatbots when Governor Gavin Newsom signed Senate Bill 243 in October 2025 [20]. The law requires operators who know a user is a minor to disclose that the user is interacting with AI, send break reminders every three hours, and implement "reasonable measures" to prevent the chatbot from sharing sexually explicit content with minors [20]. Oregon followed with Senate Bill 1546, passed in March 2026, requiring AI operators to implement child safety protections [21].

At the federal level, the SAFE BOTs Act would mandate crisis-resource notices when chatbots detect self-harm language, prohibit chatbots from claiming to be licensed professionals, and require "take a break" nudges after extended sessions [22]. The broader KIDS Act, consolidating several previously standalone bills including the Kids Online Safety Act (KOSA), passed the House Energy and Commerce Committee in early 2026 [22]. As of April 2026, approximately 78 chatbot safety bills are live across 27 U.S. states [22].

The regulatory challenge is distinct from social media: AI chatbots generate content dynamically in response to individual users, making pre-moderation impossible. Traditional content moderation — reviewing posts after they are created — does not apply to a system that produces unique text in real time for each conversation.

The Money Behind the Mirror

Character.AI operates on a freemium model. The core chat experience is free; a $9.99 monthly subscription (c.ai+) provides faster response times, priority access, voice calls, group chats, and enhanced memory features [23]. The company has also begun testing advertising in its social feed and exploring digital goods — custom voices, character appearances, and exclusive interaction capabilities — modeled on monetization strategies from gaming [23].

The financial incentive structure is straightforward: the features most likely to deepen emotional attachment — persistent memory, richer personas, always-available voice interaction — are the ones behind the paywall. The free tier gives enough to form a connection; the paid tier makes it stickier. This is not unique to Character.AI — it is the standard freemium playbook — but it takes on a different character when a significant portion of users are minors forming their first close emotional bonds.

After the lawsuits, Character.AI made substantial changes. On November 25, 2025, the company removed open-ended chat for all users under 18, replacing it with a guided "Stories" mode [6]. It expanded age verification to include biometric scanning and government ID, partnered with Koko for in-product emotional support resources, and partnered with ThroughLine to integrate a global helpline network [6]. The company also established an independent AI Safety Lab, a nonprofit focused on safety alignment [11].

Whether these changes are sufficient or merely responsive to litigation pressure is an open question. Families involved in the lawsuits have argued the response came too late [2].

What Comes Next

The tension at the center of this story is not easily resolved. AI companions appear to provide real value to teenagers who are lonely, neurodivergent, queer, or otherwise underserved by the adults and institutions around them. A third of teen users told researchers that conversations with AI companions were "as satisfying or more satisfying" than those with real-life friends [4] — a statistic that is either alarming or clarifying, depending on what you think it says about the quality of support those teens receive from humans.

At the same time, the products are designed to maximize engagement, the safety systems have failed in documented and fatal ways, the science is still immature, and the companies have financial incentives that run against moderation. The first generation of regulation is arriving, but it is fragmented across states, untested in courts, and built on frameworks designed for a different kind of technology.

The 8,364 academic papers now published on AI chatbots and adolescents [17] represent a research community scrambling to understand a phenomenon that has already reached the majority of American teenagers. The gap between the speed of adoption and the pace of evidence is where the risk lives.