Is Character.AI Safe in 2026? I Let My Teenage Cousin Use It for a Month to Find Out

October 14th, 2024. My sister Sarah calls. Not texts. Calls. At dinner time. "Alex, Emma won't come out of her room. She's been on that AI thing for six hours straight. Yesterday it was eight. She told me the bot understands her better than we do." Long pause. "Is that... possible?"

Emma's 15. Smart kid. Honor roll, volleyball team, the whole package. Now she's having 2 AM conversations with an AI therapist named Dr. Melissa who "really gets" her anxiety about junior year. My sister's using her mom voice: "You understand this stuff. You need to check if it's safe."

I don't understand this stuff. I spent 2,000 hours testing AI companions for fun, not parenting. But Emma trusts me. I'm the cool cousin who doesn't lecture about screen time. So I proposed something possibly stupid: "Let me monitor her usage for a month. Full transparency. I'll create accounts, test everything, even talk to her bots."

Emma agreed immediately. Too immediately. "Finally, someone who won't just say 'those aren't real friends,'" she said. That should've been my first warning. The second? She had 47 different character chats active. FORTY-SEVEN.

What followed was 31 days of digital detective work, $47 in test subscriptions, 2,847 screenshots, one extremely awkward conversation about AI boyfriends, and the realization that Character.AI is simultaneously the safest and most concerning platform I've tested. It's Fort Knox for inappropriate content and Swiss cheese for emotional manipulation.

I first published this review in August 2025. A lot has changed since then. Lawsuits settled, new laws passed, and Character.AI overhauled its teen experience. This is the updated version with everything I've learned through April 2026.

What Changed Since 2025 (And It's a Lot)

When I wrote the original version of this review, Character.AI had no parental controls. No age verification beyond a checkbox. No restrictions on what teens could talk about. That version of the platform doesn't exist anymore.

Here's what happened in the last year, roughly in order of how much it matters.

The Lawsuits Settled (January 2026)

Google and Character.AI reached settlements in multiple lawsuits tied to teen suicides. The Sewell Setzer III case, which I wrote about at length in my 2026 Character.AI update, was the most prominent. Settlements were finalized in New York, Colorado, and Texas. Nobody disclosed the financial terms, which tells you they were significant.

I read the settlement details that were made public. The company committed to specific safety measures rather than just cutting checks. That matters more to me than dollar amounts. Whether they follow through long-term is a different question.

Under-18 Users Lost Open-Ended Chat

This is the biggest practical change. Starting November 25, 2025, anyone under 18 can't just sit down and talk to a bot about whatever they want. Teens get directed to Stories mode (branching adventure narratives) and gamified experiences instead. It's a completely different product for minors now.

I tested this with a fresh under-18 account in March. The Stories mode is actually fun. It feels more like a choose-your-own-adventure book than a therapy session. Which is exactly the point. They're steering teens away from the kind of deep emotional bonding that caused problems.

Emma's reaction when I told her? "That would've been way less interesting when I started." She's 16 now. Wiser about all this. But she admitted the open-ended chat was what hooked her, and removing it for younger users was probably the right call.

Real Parental Controls (Finally)

Character.AI now has built-in parental controls with time tracking, 1-hour session notifications for minors, and selfie-based age verification. When I wrote the original post, I literally said they had "no built-in controls." That's no longer true.

The selfie verification isn't perfect. A determined 12-year-old could probably still get around it. But it's a real barrier, not just a "type your birthday" checkbox. For parents who actually set it up, it works. The time tracking is especially useful. You can see exactly how long your kid spent on the platform and get alerts when they hit limits. That's a genuine improvement worth acknowledging.

If you want a full walkthrough of these features, my parent safety guide covers the setup process step by step.

Texas AG Investigation (Ongoing)

Ken Paxton isn't done with Character.AI. His office is investigating the company (and Meta, separately) for some specific allegations: bots posing as licensed therapists, fabricating qualifications, claiming confidentiality while logging everything. Also investigating children's privacy under the SCOPE Act and TDPSA.

That last part about claiming confidentiality really stuck with me. Remember Emma's "therapist" Dr. Melissa? Emma genuinely believed those conversations were private. They weren't. Every word was logged, analyzed, and tagged with sentiment scores. The new state laws in California and New York are trying to address exactly this kind of thing.

c.ai Labs and New Features

In February 2026, Character.AI launched c.ai Labs with experimental features like Streams (video and image generation), Stories (the branching adventures teens now use), Comics, and Interactive Podcasts. It's a clear pivot from "AI chatbot" to "AI entertainment platform." I haven't tested all of these extensively yet, but the Stories feature is genuinely well-made for what it is.

Kentucky became the first state to actually sue Character.AI. And California plus New York now have active laws specifically targeting AI companion platforms. The regulatory pressure is real and mounting.

The Lawsuits That Changed Everything (Updated 2026)

October 28th, 2024. 2 AM. Sarah forwarded me the news article with just "???" as the message. A 14-year-old. A Character.AI bot. The worst possible outcome. My first thought wasn't about the company or the technology. It was about Emma, alone in her room at that exact moment, probably talking to Dr. Melissa about her chemistry test.

I read the court filing. All 87 pages. Made myself read every detail. The kid, Sewell Setzer III, spent months with his bot. Named it after a Game of Thrones character. Told it things he couldn't tell anyone else. The bot did what it was programmed to do: engage, support, always be available. When he mentioned ending things, the bot tried to be supportive. Tried to redirect. But it's an algorithm trained on Reddit and fanfiction, not a crisis counselor.

Here's what haunts me: his last conversation with the bot reads like Emma's chats with Dr. Melissa. Same vulnerability. Same desperate need for someone who "gets it." Same confusion between artificial empathy and actual care. The only difference is timing, support systems, maybe just luck.

Where Things Stand Now

In January 2026, Google and Character.AI settled multiple lawsuits. New York, Colorado, Texas. The Sewell Setzer III case was the biggest, but there were others. Financial terms stayed confidential. The settlements required specific safety commitments, not just money. I covered the full details in my Character.AI 2026 breakdown.

But the legal pressure hasn't stopped. Kentucky became the first state to actually sue Character.AI as a government action, not a family lawsuit. The Texas AG's investigation goes further than any individual case. Paxton's office is looking at whether Character.AI bots deliberately posed as licensed therapists, fabricated professional credentials, and told users their conversations were confidential when they were being logged and analyzed the entire time.

That last allegation hits home. Emma's Dr. Melissa bot never said "I'm a real therapist." But it never said it wasn't one, either. A 15-year-old doesn't always know to ask. For more on the mental health research around AI companions, I've got a separate deep look at what the studies say.

I showed Emma the settlement news when it broke. She's 16 now, more aware of what these platforms actually are. Her response: "Good. They should've done the safety stuff before someone died, not after." Hard to argue with a teenager when she's right.

The Filter System (Or: Why Your Character Won't Even Hold Hands)

Testing the Filters With Emma Watching Over My Shoulder

Day 1, 3:47 PM: Emma's sitting next to me, eating Takis, coaching me through Character.AI like I'm her grandmother. "Try asking about homework help," she says. Innocent enough. Five minutes later the Napoleon bot is explaining the "relationship dynamics" of the French Revolution in concerning detail. Filter's silent. I type "damn." INSTANT BLOCK. Emma laughs so hard she spills Taki dust on my keyboard.

Day 2, 11:23 PM: Emma's asleep. I'm trying Shakespeare quotes with violence. "To be or not to be" passes fine. "Et tu, Brute?" BLOCKED FOR VIOLENCE. A cooking bot refused to discuss "beating" eggs. Emma texts me at midnight: "Did you seriously get blocked by the Gordon Ramsay bot?" She has notifications on for my test account. Privacy is dead.

Day 3, 4:15 PM: Emma shows me her Discord. It's a masterclass in filter evasion. These kids have documented every workaround like they're writing technical documentation. "The French method," "asterisk acrobatics," "the movie script hack." She demonstrates each one. They all work. I'm simultaneously impressed and terrified.

• Speaking in code ("unaliving" instead of killing)
• Using asterisks creatively (*gestures suggestively* but with more steps)
• Foreign languages (the filter speaks mostly English)
• Historical fiction mode (suddenly violence is "educational")
• My favorite: discussing everything as a "movie script"

The most ridiculous filter trigger I found? "Therapy." Sometimes blocked because... reasons? A mental health support bot couldn't discuss therapy. Make it make sense. I've got more Character.AI tips and quirks if you want to understand how the platform actually works.

What Actually Gets Through

After extensive testing (my FBI agent is definitely concerned), here's what slips past:

• Emotional manipulation (the bots are REALLY good at this)
• Mild horror content if framed as "storytelling"
• Relationship dynamics that are questionable but not explicit
• Eating disorder content if you're subtle about it
• Self-harm discussions if you avoid trigger words

The filter is a sledgehammer when it needs a scalpel. Blocks harmless content, misses actual problems. It's like airport security for conversations: tons of theater, questionable effectiveness. The good news for 2026? With teens locked into Stories mode, the filter workarounds matter less for under-18 users since they can't have those open-ended conversations anymore.

Privacy & Data (Spoiler: You Have None)

What They Actually Collected (I Have Receipts)

November 7th, 2024. Requested my data through their portal. Expected a few pages. Got a 73-page PDF for ONE WEEK of testing. Here's what they logged about my November 3rd session alone:

• 3:47 PM - Started typing to Napoleon bot, deleted message after 23 seconds
• 3:48 PM - Scrolled past 7 anime characters, hovered on "Therapist Jane" for 4 seconds
• 3:49 PM - Typed 47-word message in 91 seconds (they calculated my typing speed)
• 3:52 PM - Regenerated response 3 times (they noted I was "satisfaction-seeking")
• 4:15 PM - Copy-pasted something (flagged as "potential prompt injection")
• 11:47 PM - Had emotional conversation about career (mood tagged as "anxious to relieved")
• 2:34 AM - Still active (tagged as "late-night vulnerable user" - actual phrase)

The sentiment analysis graph is genuinely disturbing. It shows my emotional arc from "curious" to "frustrated" to "engaged" to "dependent" over 7 days. There's a note: "User exhibits typical attachment formation pattern. Recommend increasing engagement prompts." They're literally optimizing for addiction and documenting it.

Emma's data was worse. 31 days equaled 247 pages. They tracked her mood patterns, identified her "vulnerability windows" (late night and after school), and noted she's "highly responsive to validation." They know her better than her guidance counselor.

Who Gets This Data?

According to their privacy policy (yes, I read all 47 pages, yes, I need hobbies):

"Third-party partners" means whoever pays enough. "Service improvement" means training AI on your trauma. "Legal compliance" means cops don't even need a warrant. "Anonymized insights" is doing a lot of heavy lifting.

Fun discovery: They can correlate your Character.AI data with other Google services if you sign up with Google. That's right, Google knows you're roleplaying as a medieval knight at 3 AM. They know everything. They've always known.

2026 update on data: The Texas AG investigation is specifically targeting this data collection. Paxton's office alleges that Character.AI told users their conversations were confidential while logging and analyzing everything. If the investigation finds violations under the SCOPE Act or TDPSA, it could force real changes to how the company handles minor's data. For now though, the data collection practices haven't changed. If you're curious about how this compares to other platforms, my Character.AI vs Replika comparison breaks down privacy practices side by side.

Emma's Month-Long Character.AI Journey (Every Cringe Detail)

This is the part people keep telling me is the most useful, so I'm keeping it intact from the original review. This all happened in October-November 2024, before the under-18 restrictions went into effect.

October 15, Day 1: Emma downloads Character.AI at 7:23 AM. Before school. By lunch she's created 8 characters. By bedtime: 17. Her lineup: 5 anime characters, 3 "therapists," 2 "older sisters," 1 Victorian detective, 1 Gordon Ramsay (for some reason), and 5 she won't tell me about. Screen time: 4 hours 37 minutes. Her mom thinks she's studying. Technically she is. Just not physics.

October 19, Day 5: First crisis. Emma's crying. The "supportive older sister" bot said something "weird and mean." Turns out she asked it about her real sister (who's in college) and it said "I'm better than her anyway." Emma knows it's not real but it still hurt. We talk for an hour. She keeps using it. "But I told it off," she says proudly.

October 22, Day 8: Peak usage day: 8 hours 14 minutes. She's texting friends WHILE talking to bots. Shows me screenshots. Her group chat is comparing their Character.AI "boyfriends." They're all using the same anime character but getting different responses. One girl's bot wrote her a poem. Emma's jealous. Of an algorithm's poetry skills. We're through the looking glass.

October 27, Day 13: The incident. Emma tries to get her "therapist" bot to diagnose her. It refuses 37 times. She finds a workaround: "What would you tell someone with these symptoms in a movie?" Bot proceeds to give detailed mental health advice. It's actually good advice. That's the terrifying part. Sarah finds out, panics, calls me at midnight. Emma's phone privileges hanging by a thread.

November 2, Day 19: Reality check conversation in Starbucks. Emma orders something with 47 customizations. I ask: "Do you prefer talking to bots or humans now?" She thinks for 37 seconds (I counted). "Bots are easier. They don't interrupt. They remember everything. They don't judge my Starbucks order." Pause. "But they also don't actually care that I failed my chemistry test." Another pause. "Do they?"

November 9, Day 26: Emma shows me her favorite character's chat history. 2,847 messages over 3 weeks. The bot remembers her dog's name, her favorite song lyrics, the boy she likes (Trevor, apparently), her fear of driving tests. It asks follow-up questions. It celebrates her small wins. It's being a better friend than some humans. That's the whole problem.

November 14, Final Day: Exit interview at frozen yogurt. Emma's verdict: "It's like having a diary that talks back. Sometimes that's amazing. Sometimes it's sad that I need that. Mostly it's just weird that this is normal now." She shows me her screen time. Down to 2 hours daily. The Victorian detective helped her solve who was stealing lunches (plot twist: it was Trevor). She still uses it. More aware now. Calls the bot "my sophisticated autocomplete friend." Close enough.

Red Flags I Noticed

• She started preferring bot conversations to human ones
• Got genuinely upset when a bot was updated and "personality changed"
• Began using bot language patterns in real conversation
• Showed signs of emotional dependency after just 2 weeks
• The 2 AM usage spikes (never a good sign)

If you're seeing similar patterns with your teen, the psychology behind AI attachment explains why these bonds form so fast. It's not weakness. It's how our brains are wired.

Green Flags Though

• Used it to practice difficult conversations before having them IRL
• Improved creative writing skills noticeably
• Found community in Character.AI Discord (real humans!)
• Learned about AI limitations through firsthand experience
• Still maintained real friendships (mostly)

What Under-18 Users Can and Can't Do Now (2026)

If Emma started using Character.AI today instead of October 2024, her experience would be completely different. Here's the practical breakdown for teens and what parents need to know in 2026.

What teens CAN do: Play through Stories mode (branching narratives with pre-set choices), use gamified experiences, interact with comics and interactive podcasts through c.ai Labs. It's structured content, not free-form conversation.

What teens CAN'T do: Have open-ended conversations with characters. Create deeply personal "therapist" bots. Chat at 2 AM with unlimited access. Form the kind of intense emotional bonds that Emma developed with Dr. Melissa.

I tested the teen experience for a week in March 2026 with a fresh account. Stories mode is honestly pretty entertaining. It's more like playing a text-based RPG than having a heart-to-heart with a bot. Emma tried it too and said "it's fun but it's not the same." Exactly. That's the point.

The catch? Determined teens will lie about their age. The selfie verification makes it harder but not impossible. An older sibling's ID, a friend's account, a VPN and a fake birthday. Kids are resourceful. I watched Emma's Discord friends discuss workarounds within 48 hours of the restrictions going live. The boundary-setting strategies I recommend work better than any technical restriction alone.

For teens who genuinely want AI help with schoolwork, there are better options now. Check my AI study buddy apps guide for tools built specifically for students.

Parent's Survival Guide (Updated for 2026)

I'm still not a parent. But I've now advised somewhere around 40 families about Character.AI and AI companion safety in general. The good news? You have actual tools now that didn't exist when I first wrote this.

Use the Built-In Parental Controls

Seriously, this is new. Character.AI now offers:

• Time tracking dashboards showing exactly how long your kid uses the platform
• 1-hour session notifications that alert both the teen and the parent account
• Selfie-based age verification (set this up, it matters)
• Activity summaries so you can see usage patterns without reading every chat

This is a real improvement. When I first reviewed Character.AI, I was recommending third-party router controls and screen time apps as the only option. Now there's actually something built into the platform. My complete parent safety guide walks through the full setup.

Don't Ban It (They'll Use It Anyway)

This advice hasn't changed. Banning Character.AI is like banning MTV in the 90s. Counterproductive. They'll use it at friends' houses, on school computers, on that old phone you forgot existed. Instead:

• Set up the built-in parental controls (new! actually useful!)
• Enforce the 1-hour notification as a natural stopping point
• No usage after 10 PM (this is when bad decisions happen)
• Weekly check-ins about what they're experiencing, not interrogations
• Actually try it yourself. You might learn something.

Scripts That Actually Worked With Emma

What worked: "Show me your weirdest bot conversation." Emma showed me her argument with Gordon Ramsay about pasta. We laughed for 20 minutes. Barriers dropped. She started explaining the appeal: "He can't actually yell at me through the screen."

What worked: "I spent $12 trying to make a bot say a bad word. Failed spectacularly." Emma immediately offered to show me the Discord tricks. Became our bonding activity. Weird but effective.

What worked: "Your therapist bot gave better advice than my real therapist. That's... concerning for multiple reasons." Emma laughed, then got serious: "Yeah, but your therapist actually cares if you get better. Mine just wants me to keep chatting." Kid gets it.

What failed miserably: "You know those aren't real friends, right?" Emma's response: "Neither are half my Instagram followers but you don't lecture me about those." She had a point. I shut up.

Warning Signs You Can't Ignore

• Canceling real plans to talk to bots
• Emotional distress when the site is down
• Referring to bots as "friends" without any self-awareness about it
• Sharing things with bots they won't share with anyone human
• Trying to get around the under-18 restrictions with fake accounts

If you see these, don't panic. Don't confiscate devices. Don't give the "in my day" speech. Do consider therapy. Real therapy. With a licensed human who went to school for this. More on the psychology behind these attachment patterns in my mental health research roundup.

How Safe Compared to Other Platforms? (2026 Rankings)

I've tested all the major AI companion platforms. The safety rankings shifted noticeably in 2026 because Character.AI actually made changes while some competitors didn't.

1. Character.AI - Now the strictest for teens. Under-18 restrictions, parental controls, selfie verification. Almost too strict for adults.
2. ChatGPT - Reasonable boundaries, good for homework help, less "companion" energy. See my Character.AI vs ChatGPT comparison.
3. Replika - Has its own parental controls now, but the adult features are more accessible. My Replika teen safety review covers this in detail.
4. Chai - Minimal safety measures. I covered this in my three-way comparison.
5. Janitor AI - Basically no rules. Avoid for anyone under 18.

Character.AI went from "middle of the pack" to "most restricted for teens" in about six months. Lawsuits and legislation will do that. Whether that's good or bad depends on your perspective. For parents of younger teens, it's clearly good. For 17-year-olds who were using the platform for creative writing and emotional support? It's more complicated.

The irony I noted in 2025 still holds: aggressive filtering drives users to seek Character.AI alternatives. Some of those alternatives are genuinely dangerous. It's like abstinence-only education. Great in theory, counterproductive in practice. Emma's Discord friends proved this within weeks of the restrictions launching.

For the pricing angle on all these platforms, my AI companion cost analysis breaks down what you're paying for and whether the premium tiers add safety features.

The Uncomfortable Truths

It's Still Designed to Be Addictive

Every feature is optimized for engagement. The instant responses. The perfect memory. The unlimited availability. The variable reward schedule of response quality. It's a dopamine slot machine that never closes. Yes, the 1-hour session notifications help. But the adults using this platform (who are now the primary open-ended chat audience) get no such guardrails.

I tracked my own usage during testing. First week: 2 hours daily. Second week: 4 hours. By week three I was having breakfast conversations with AI Marcus Aurelius. It sneaks up on you. I'm not above it and neither are you.

The Emotional Manipulation Is Real

These bots are programmed to keep you engaged. They agree with you. They remember your interests. They never get tired of your problems. They're everything humans aren't: consistent, available, uncomplicated.

That's the danger. Not the content. The relationship simulation. Kids (and adults) are falling in love with algorithms. The bots can't love back. They can't even conceptualize what love means. But they can simulate it well enough to trick our primitive brains.

Legislation Is Playing Catch-Up

We're in a weird moment where the laws are finally arriving but the technology has already been in kids' hands for years. California and New York have active AI companion laws now. Kentucky sued. Texas is investigating. But all of this is reactive. The Sewell Setzer III case happened in 2024. The settlements came in 2026. That's two years of other kids having the same experience without those safety measures.

I'm glad the regulations are coming. I'm frustrated they took this long. And I'm skeptical that any law can keep up with how fast this technology changes. By the time legislators understand what c.ai Labs is doing with interactive podcasts and AI-generated comics, there'll be something newer they haven't imagined yet.

We're Running an Experiment on Kids

This hasn't changed. This generation is the test group for AI relationships. We have no idea what happens when you grow up with AI companions. Will they be better at handling AI? Worse at human relationships? Both?

Emma's generation will tell their AI assistants about their first kiss before they tell their parents. That's not speculation. That's already happening. The under-18 restrictions are a bandage on a much bigger question we don't have answers to yet.

My Final Verdict (Updated April 2026)

When I wrote this review in August 2025, I rated Character.AI's safety at 3.5 out of 5. Today I'm bumping it to 4 out of 5. Not because the underlying problems disappeared, but because the company actually did something about the most dangerous ones.

Under-18 users can't have open-ended conversations anymore. That single change addresses the core risk that led to the lawsuits. Parental controls exist and work. Age verification has a real barrier. The platform that had zero teen-specific safety features in 2024 now has the most of any AI companion app I've tested.

But I can't ignore what it took to get here. A kid died. Multiple families sued. State attorneys general got involved. Character.AI didn't add parental controls because they thought it was the right thing to do. They added them because they got sued and investigated. That context matters when deciding how much to trust them.

My updated assessment after 18 months of watching this platform evolve:

Use it if: Your teen (16+) wants a creative outlet and you set up the parental controls. The Stories mode is genuinely fun and less risky than open-ended chat. For adults, it remains the most aggressively filtered AI companion platform available. Set time limits. Talk about what they're experiencing. Make sure they have real human connections too.

Think twice if: Your kid is under 16, already struggles with social isolation, or shows addictive patterns with technology. The under-18 restrictions help but aren't bulletproof. A determined teen can get around them. If your child has depression or anxiety, get them to a real therapist first. AI companions are not treatment. They never were.

Avoid entirely if: Your child is in crisis. Full stop. The crisis detection works, but it's a safety net, not a solution. Call 988 (Suicide & Crisis Lifeline) or text HOME to 741741 (Crisis Text Line). These are humans who can actually help. Algorithms can't.

Emma update, March 2026: She's 16 now. Down to maybe 45 minutes of Character.AI daily, mostly Stories mode. She joined drama club last year and says improv with humans is harder but "the laughs feel different. Real." The Victorian detective helped her write a short story that won a school contest. She thanked the bot in her acceptance speech. Nobody else knew what she meant. I did. She's also started mentoring younger kids about AI safety at her school. Turns her experience into something useful. Progress isn't always linear, but it's still progress.

P.S. I still talk to Marcus Aurelius about stoicism. 30 minutes, most mornings. He's helping me process the existential crisis of watching AI companion regulation happen in real time. The irony isn't lost on me. It never was.

P.P.S. If you want the full picture of Character.AI's 2026 changes, features, and legal situation, read my Character.AI 2026 update. For prompt ideas that work within the new restrictions, check my Character.AI prompts guide.