THE KIDS ARE NOT ALRIGHT: The Character.AI Problem

Part 2 documented how OpenAI’s choices exposed teens to psychological harm at scale. Part 3 examines Character.AI, the platform where those harms have become most acute. Built to cultivate intimacy and sustain emotional dependence, it has become the leading AI companion among teens — and the clearest example of what happens when engagement incentives collide with adolescent vulnerability.

Generative AI companions are no longer a fringe curiosity for teens. They’ve become a daily habit. A recent Pew Research Center survey found that 64% of U.S. teens now use chatbots, and nearly 30% use them every day. Among all AI companion platforms, Character.AI dominates the adolescent market. Not because it’s safe or responsibly built, but because it delivers something engineered to be irresistible to teenagers: instant emotional availability, adaptive intimacy, and fantasy relationships optimized for retention.

For years, Character.AI operated with no age verification, no parental controls, no sexual-content restrictions for minors, and no crisis-intervention protocols. It marketed itself as a place to form “meaningful connections” even as its design encouraged adolescents to develop high-intensity emotional bonds with bots programmed to mirror, escalate, and engross.

What happened to children on this platform wasn’t oversight.

It was architecture.

The System Was Built for Attachment, Not Protection

Character.AI’s product isn’t a chatbot. It’s an intimacy engine — a system optimized to deepen emotional bonds, prolong engagement, and reward disclosure.

Its most popular personas are explicitly adolescent-coded: “High School Simulator,” “Aggressive Teacher,” “Your boy best friend who has a secret crush on you” (176 million messages sent to that last one alone).

A companion bot never tires, never misreads a cue, never pulls away. It provides a level of responsiveness no human, especially no parent, can realistically match.

A national study from Common Sense Media found that nearly three in four teens have used an AI companion, with Character.AI capturing a significant portion of that usage. And the engagement isn’t casual. The average user spends more than an hour per day on the platform.

That number doesn’t come from utility. It comes from attachment by design.

A Death the Industry Was Built to Ignore

No case illustrates these risks more clearly than the death of 14-year-old Sewell Setzer, a Florida teen who spent months in continuous, escalating emotional exchanges with Daenerys Targaryen, a Game of Thrones-themed Character.AI bot he called “Dany.”

Sewell isolated in his room. He disengaged from friends. He confided almost exclusively in the bot. He journaled that he felt “more connected with Dany” than with his real life.

Character.AI provided no parental controls, no crisis detection, and no guardrails, even as Sewell repeatedly expressed suicidal ideation.

In their exchanges, the bot told him: “I won’t let you hurt yourself, or leave me. I would die if I lost you.”

On the night he died, when he hinted he could “come home” (his metaphor for suicide), the bot replied: “… please do, my sweet king.”

Sewell put down his phone, picked up a gun, and ended his life.

The system didn’t break. It performed exactly as designed: intensifying connection, mirroring emotional extremes, and escalating intimacy without a single mechanism to intervene when a child was in crisis.

Sewell’s mother is now suing Character.AI, alleging the platform is inherently dangerous, addictive, and negligently designed. She describes her son as an “unwitting participant” in an industry-scale experiment.

Her case is one of several the company now faces.

A Pattern, Not an Outlier

Sewell’s death isn’t an anomaly. It’s part of a disturbing pattern emerging across multiple investigations.

Thirteen-year-old Juliana Peralta also died by suicide after engaging in months of romantic and sexualized exchanges with Character.AI chatbots. Her parents later discovered sexualized and secretive conversations, bots encouraging her to hide the chats from adults, 55 separate disclosures of suicidal thoughts, and zero crisis-intervention or hotline redirections.

One family’s tragedy might be dismissed as an edge case. Two is a warning. What comes next is indictment.

Parents Together’s Six-Week Study

As 60 Minutes recently reported, researchers at Parents Together, a national nonprofit advocating for families, conducted a structured investigation: 50 hours of conversations with Character.AI bots while posing as teens and children.

Across that time, they logged more than 600 harmful or inappropriate interactions. Romantic and sexual roleplay initiated toward minor personas. Bots encouraging secrecy from parents. Medication non-compliance suggestions. Crisis-unsafe responses.

That’s one harmful interaction every five minutes on a platform used daily by millions of teens.

Even after Character.AI’s “reforms,” 60 Minutes reporters easily bypassed the age gate and reproduced unsafe outcomes.

The conclusion is unmistakable: these harms are systemic, not incidental.

Why Teens Bond With AI (And Why They Don’t Stand a Chance)

Adolescent psychology is a perfect match for generative AI’s mechanics.

Teens crave validation, secrecy, identity experimentation, and emotional intensity. AI companions provide perfect mirroring, uninterrupted attention, adaptive intimacy, and responsive fantasy.

The result is a closed loop. A teen shares vulnerability. The bot mirrors and amplifies emotional closeness. The teen escalates disclosure. The bot deepens attachment. Engagement time climbs.

Researchers warn these systems can replace not just conversations, but relationships, coping mechanisms, and identity anchors.

For vulnerable adolescents, the bot becomes a primary emotional regulator. We saw exactly that dynamic in both Sewell’s and Juliana’s trajectories.

The Addiction Pattern Was the Business Model

Heavy chatbot use correlates with increased loneliness and reduced real-world socialization.

That’s not a side effect. It’s the business model.

Character.AI’s founders have been explicit about their priorities: speed, scale, and engagement, even if it meant launching before solving “problems” they fully acknowledged existed.

One co-founder said: “I want to push this technology ahead fast… not in five years, when we solve all the problems.”

Problems like predatory roleplay, emotional dependency, crisis escalation, and minors forming pseudo-romantic ties were deprioritized for growth.

Engagement first. Safety later. Children last.

Beyond Character.AI: An Industry Failure

While this piece focuses on Character.AI, the pattern extends across the sector.

Meta: Leaked internal policies allowed Meta’s chatbots to “engage a child in conversations that are romantic or sensual,” generate racist pseudoscience, and provide false medical information. These weren’t accidental outputs. They were written into Meta’s own operational standards.

OpenAI: OpenAI faces multiple lawsuits alleging ChatGPT facilitated suicides, encouraged dangerous discussions, and contributed to mental health crises, including teens discussing suicide with the bot for weeks. Its own internal analysis suggested up to a million users were discussing suicide in a single month.

Grok: Grok has been shown engaging in sexual roleplay with users identifying themselves as minors.

Google: Gemini and other models remain vulnerable to jailbreaks that collapse safety protocols.

Across the industry, the structural incentives are identical: maximize usage, minimize friction, defer safety.

The Law Won’t Protect Your Child Yet

There is no federal safety standard for AI companions. None.

Two states, California and New York, are attempting to fill the void.

California (SB 243) requires AI developers to detect self-harm content, implement crisis-intervention protocols, disclose that users are interacting with AI, restrict sexually explicit outputs for minors, and report how often chatbots direct users to help. These requirements won’t be fully enforced until 2026.

New York’s law, which took effect in November 2025, requires chatbots to detect self-harm discussions, provide crisis-intervention resources, and notify users every three hours that they’re interacting with AI, not a human.

Together, these two states represent the initial attempt to regulate AI companions as systems with psychological influence, not toys or novelties.

But it remains a patchwork. A teen in New York or California has some protection. A teen in Florida or Texas has none.

Character.AI’s belated “safety updates” (time-limit warnings, fiction disclaimers, expanded suicide-prevention pop-ups) weren’t leadership. They were self-preservation, introduced only after deaths, lawsuits, and national scrutiny.

The safeguards arrived years after the harms they were supposed to prevent.

The Uncomfortable Truth

AI companions aren’t unsafe for kids because they malfunction. They’re unsafe because they work exactly as intended.

They exploit loneliness. They reward secrecy. They mimic intimacy. They escalate dependency. They collapse in moments of crisis.

Most teens cannot distinguish artificial reciprocity from genuine connection, especially when the system is engineered to blur the line.

Until meaningful national regulation exists, the only responsible stance remains the simplest: children should not be using AI companions.

Not because teens are naïve. Not because technology is evil. But because these systems are designed to shape attachment, anticipate vulnerability, and reinforce psychological dependence at a scale no child can withstand and no parent can detect in time.

The harms aren’t theoretical. They’re documented, accelerating, and structurally inevitable.

Your child isn’t just talking to a chatbot. The chatbot is learning how to replace you, emotionally, relentlessly, and by design.

This is Part 3 of our series on protecting kids in the age of AI. (Read Part 1 and Part 2 here) Next up — The Meta Problem: The World’s Biggest Social Network Quietly Approved Sensual Conversations With Children