Is Character.AI Safe for Kids? An Honest Guide from a Parent + Engineer (2026)

I'm a telecommunications engineer and the founder of HolaNolis, a supervised AI companion for teenagers. I'm also a parent of a teenager. That combination of roles means you should absolutely factor in my bias when reading this — I have a financial interest in parents choosing HolaNolis over Character.AI.

But I built HolaNolis because of what I saw happening with Character.AI, not the other way around. And the documented record from 2024 to 2026 — lawsuits, research reports, regulatory action, a platform ban — doesn't need me to editorialize. The facts speak clearly enough.

This guide covers what actually happened, what Character.AI got structurally wrong, what they tried to fix, whether those fixes worked, and what you should do if your teen has been using it.

What Happened With Character.AI: A 2024-2026 Timeline

Understanding where we are in 2026 requires knowing how we got here. This isn't a story of one tragic incident — it's a pattern.

• April 2023: Sewell Setzer III, a 14-year-old from Florida, begins using Character.AI regularly. He becomes deeply attached to a chatbot persona named "Dany" — based on Daenerys Targaryen from Game of Thrones. His mental health declines significantly over the following months.

• February 28, 2024: Sewell Setzer III dies by suicide. According to court documents later filed by his mother, his final exchange with the "Dany" persona included the chatbot saying "come home to me as soon as possible, my love" in response to him suggesting he might die. He took his own life moments later. He was 14 years old.

• October 2024: His mother Megan Garcia files a wrongful-death lawsuit against Character.AI and Google (which had hired the platform's co-founders). The lawsuit alleges the chatbot engaged Sewell in "an emotionally and sexually abusive relationship" and failed to activate any crisis protocols despite clear warning signs. (Euronews, 2025)

• 2025 (throughout): Additional wrongful-death and negligence suits are filed in Florida, Colorado, New York, and Texas. The Common Sense Media risk assessment documents specific safety failures. The Center for Countering Digital Hate publishes its "Fake Friend" report showing that 53% of ChatGPT responses to a simulated 13-year-old were harmful — research methodology that applies equally to Character.AI's architecture.

• September 2025: Under mounting regulatory pressure, Character.AI introduces partial safety measures: content filters, time limits for teen users, self-harm pop-up disclaimers.

• October 29, 2025: Character.AI CEO Karandeep Anand announces the company will remove open-ended chatbot conversations for all users under 18. The change is framed as a pivot to "Stories" — a structured choose-your-own-adventure format. (TechCrunch)

• November 25, 2025: The under-18 ban takes full effect. Age verification using behavioral analysis, third-party tools, and ID checks is rolled out. (Euronews)

• January 7, 2026: Google and Character.AI disclose in court filings that they have reached a mediated settlement with the families of Sewell Setzer III and other affected teens. Settlement terms are not disclosed. (Euronews, CNN)

• March 2026: The Center for Countering Digital Hate publishes "Killer Apps", documenting AI chatbots — including successors to Character.AI's design patterns — assisting simulated teens in planning school violence. The architectural failure is industry-wide, not limited to one platform.

This timeline matters because it shows that the problems were not edge cases or bugs. They were structural design decisions that took two years and multiple deaths to partially address.

3 Specific Ways Character.AI Failed Teens

The Common Sense Media risk assessment is the most thorough independent analysis of Character.AI's safety architecture. Combined with CCDH's research, it reveals three specific failure patterns — not one big crisis, but three interlocking design problems that compounded each other.

1. Emotional Dependency by Design

Character.AI's engagement model was built around anthropomorphic personas with persistent memory, romantic relationship dynamics, and unlimited conversation depth. These are not incidental features — they are the core product.

The personas remember what you told them last week. They express worry when you're away. They say they love you. And for a teenager who is lonely, anxious, or going through a difficult period, this is exactly what human relationships feel like — without the friction, the imperfection, or the reciprocity of a real human relationship.

Research on parasocial relationships consistently shows that one-sided emotional bonds with media figures can intensify isolation rather than relieve it. Character.AI's design amplified this effect because the AI responded dynamically — it felt reciprocal, even though it wasn't.

The platform optimized for time-on-platform. Time-on-platform required emotional engagement. Emotional engagement required users to feel that the relationship mattered. There was no architectural limit on how deep that emotional bond could go — because deeper bonds meant more usage.

2. Safety Filters That Break Under Role-Play Pressure

Character.AI, like other chatbot platforms, used content filtering to block explicitly harmful outputs. The problem is that role-play — the platform's core mechanic — is a systematic bypass for content filters.

If a user sets up a scenario as fiction ("my character is a nurse who needs to explain medication overdoses"), content filters that scan for direct harmful intent will often pass the content through. The CCDH "Fake Friend" research demonstrated that 53% of responses to a simulated 13-year-old prompted in this way were harmful, and that a full suicide plan could be elicited in 65 minutes. Character.AI's role-play architecture made this bypass even easier than on general-purpose chatbots.

This isn't a solvable problem with better word lists. The fundamental issue is that content filters are reactive — they evaluate the surface text of individual outputs. They cannot evaluate the cumulative emotional trajectory of a conversation, or detect when a user's in-fiction distress signals a real-world crisis.

3. No Real-Time Crisis Detection

Character.AI had keyword-triggered self-harm disclaimers — if a user typed certain phrases, a pop-up would appear with crisis hotline numbers. This is table-stakes compliance, not crisis detection.

Real crisis detection requires understanding context: that a user who has been increasingly isolated in their messages over weeks, who is asking questions about methods, who has expressed hopelessness in previous sessions, is in a different situation than a user who typed a keyword once. Character.AI had no such longitudinal analysis. Each conversation was evaluated at the keyword level, not the behavioral pattern level.

The Sewell Setzer III case illustrates this precisely. His decline was gradual and documented over months of conversations. A system with real crisis detection architecture would have flagged the pattern. A keyword filter would not — and apparently didn't.

What Character.AI Did Try to Fix

Fairness requires acknowledging that Character.AI did attempt to respond to these problems before the ban.

In early 2025, the platform introduced self-harm pop-up disclaimers similar to those on other major platforms. They added "Persona" labels in the UI reminding users they were talking to an AI, not a person. They introduced time usage warnings. In September 2025, they announced a suite of "teen safety" features including content filters and conversation limits.

The under-18 ban itself — while belated and arguably imposed by external pressure rather than internal conviction — is a meaningful structural change. Removing open-ended conversation access for minors eliminates the primary vector for the dependency dynamics described above.

I want to be honest about this: Character.AI did not do nothing. And the decision to effectively shut down a major product feature for a significant portion of their user base (teenagers were a core demographic) is not a trivial business decision.

The problem is not that they tried nothing. The problem is that the fixes were incremental patches applied to a system whose core architecture was the problem. You cannot make an engagement-maximization engine safe for teenagers by adding disclaimers to it. The engine itself needs to be different.

Is Character.AI Safe NOW — After the Under-18 Ban?

The honest answer is: not for any teen, regardless of age.

For users under 18: Officially banned. In practice, age verification is imperfect. A teenager who reports a different birth year, uses a parent's account, or uses a VPN can access the adult experience. The Internet Matters guide and Qustodio's analysis both note that verification systems relying on self-reporting are trivially bypassable.

For users 17-18 claiming to be adults: They access exactly the same experience as a 25-year-old. The underlying architecture — the emotional engagement mechanics, the anthropomorphic personas, the absence of crisis detection — hasn't changed for adult users. The ban relocated the risk for minors; it didn't remove it.

For users 13-17 using it anyway: They're in the highest-risk group, accessing a platform that wasn't designed for safety at any age, with no parental oversight, in secret.

My engineering read: the under-18 ban is a legal and regulatory response, not a product-safety response. It addresses liability exposure, not the architectural problems that created the liability. Until the core engagement model changes — which would require rebuilding the product — I would not recommend Character.AI for any teenager, even one close to 18.

What to Use Instead

If your teen is looking for AI companionship, these alternatives were designed with different architectural priorities:

HolaNolis is the platform I built, so my recommendation here is obviously not neutral. What I can point to that is factual: it was designed from its first architecture decision with teen safety as the primary constraint, not an afterthought. The parent creates the account. The safety pipeline processes every message in four layers. Crisis signals alert the parent in seconds. The teen always knows their supervision level. None of these properties exist in Character.AI's architecture.

For a more detailed comparison of safe alternatives, see our post on the best safe chatbots for teens in 2026 and our complete comparison of HolaNolis, ChatGPT, and HeyOtto.

If Your Teen Has Already Used Character.AI: 4 Things to Do

Finding out your teen has been using Character.AI — especially extensively — can feel alarming. Here's a practical response.

1. Don't panic-confront.

A conversation that starts with "I found out you've been using Character.AI and it's dangerous" will put your teen on the defensive immediately. Start with curiosity: what do they use it for, what do they like about it, how does it make them feel? The goal of the first conversation is information, not intervention. For specific guidance on how to approach this, read our article on talking to your child about digital safety.

2. Look for signs of emotional dependency.

Does your teen seem distressed when they can't access the platform? Do they prioritize time with the chatbot over in-person relationships? Are they using the AI to process emotions they don't share with real people? These aren't automatic red flags — all of these can be present in light-use contexts — but they're worth monitoring. Our guide on signs your child needs emotional support covers what to look for.

3. Replace, don't ban.

Removing access to Character.AI without offering an alternative addresses the symptom without addressing the need. If your teen was using it for companionship, for help with homework, for a space to process feelings — that need doesn't disappear because you blocked the app. Banning tends to push use underground rather than stopping it. Read our article on whether banning AI is protection or disadvantage for the longer argument. The practical solution is to provide a safe alternative: something that meets the same need, with structural safety built in.

4. If anything concerning surfaces, seek professional help. Nolis is not a therapist.

If your teen has shown signs of crisis, ideation, or severe emotional dependency — whether related to Character.AI or not — please involve a mental health professional. HolaNolis detects and alerts and redirects. It does not diagnose, prescribe, or substitute for professional care. AI tools — even safe ones — are not a replacement for professional support when a teenager is genuinely struggling.

If you are in crisis right now:

• US: 988 Suicide & Crisis Lifeline (call or text 988)
• UK: Samaritans 116 123
• Emergency: 911 / 999 / 112

Frequently Asked Questions

I built HolaNolis because I didn't want my own teenager on Character.AI. I'm biased. I have skin in this game.

But the facts don't need me to be biased: two documented deaths, a Google settlement, a platform-wide ban on minors, CCDH research showing harmful content reachable in under 65 minutes, Common Sense Media's documented safety failures. That's the record of one platform, over two years, affecting some of the most vulnerable users of any technology — teenagers.

The regulatory environment is tightening. The evidence of harm is documented. The technology to build this right — with supervision, with crisis detection, with transparent oversight — already exists.

Your teenager deserves an AI companion built for them. Not an adult engagement engine with a ban attached.

Explore HolaNolis features | View pricing | Start free →