AI Companion App Safety in 2026: What Protects You and What Doesn’t

Every AI companion and roleplay platform now has a safety page, and most of them look the same: an age gate, a mention of moderation, a line about protecting the community. What those pages rarely say is which measures actually keep a 14-year-old away from adult content and which exist mainly so a company can point to them later.

This isn’t a ranked list of apps. It’s a way to read the safety claims of any platform: Character.AI, Janitor AI, Replika, Chai, or whatever launches next. The test is whether a control changes behavior or mostly provides legal cover. That matters more than the specific app, because the same pattern repeats across the whole category.

For parents, one point comes first. The American Academy of Child and Adolescent Psychiatry calls AI companions unsafe for children and tells parents to strongly consider restricting or removing access. AI companion app safety for adults looks different. The main risks shift from who gets in to what happens to your conversations and how hard the product works to keep you attached. The sections below cover both.

What Counts as Real Age Verification?

Real age verification checks something the user can’t simply type: a government ID, a facial age estimate, or account signals that flag a likely minor for a closer look. A birthdate field the user fills in verifies nothing. It’s a formality, not a gate.

Self-reported birthdates were the industry standard for years, and many platforms still rely on them. A child willing to lie about their age won’t be stopped by a dropdown menu, and regulators have documented this failure directly. Italy’s data protection authority found that users could get past Replika’s age gate by entering an adult age at signup and editing their profile afterward. That finding was part of the case behind Replika’s €5 million GDPR fine.

In July 2026, the same regulator fined Character.AI’s parent company €158,000. It also ordered the company to make its age checks work properly and to stop minors with frozen accounts from simply registering again.

The category is moving toward checks that don’t depend on self-report:

• Signals first, verification second. Character.AI runs age assurance on every account using existing account signals. Accounts that may belong to a minor go to Persona, a third-party provider, for a selfie age estimate. A government ID is requested only if the selfie result is inconclusive.
• Region-triggered checks. Janitor AI asks for verification only from users connecting from countries whose laws require it: Australia and Brazil since April 24, 2026, and the UK since June 15, 2026. Everyone else uses the platform with no check at all. Brazil’s trigger is the Digital ECA (Lei nº 15.211/2025), in force since March 17, 2026.
• Vendor named, coverage unclear. Some apps name a verification provider without saying who actually gets checked. Chai’s age-check claims cite Persona but don’t confirm that every user goes through the same process.

None of these methods is foolproof. Facial age estimation has a well-documented failure mode: it misclassifies adults as minors and locks out legitimate accounts. Both Character.AI’s and Janitor AI’s 2026 rollouts drew waves of complaints from adults who were wrongly flagged or stuck in verification loops.

The honest test for any age-verification claim is simple: does it check something the user can’t just type in? If not, treat an “18+” label as a disclaimer, not a barrier.

Does Content Moderation Stop Explicit Content, or Just Slow It Down?

On platforms built around user-created characters and open-ended chat, moderation is almost always reactive. It catches reported content after the fact, not before. Real-time filtering that blocks explicit material during a live conversation is rare. Where it exists, it usually covers only specific regions or user tiers.

This structural issue sets the category apart from a typical content platform. On Janitor AI, Chub AI, and similar services, users write the characters and shape how they behave. No editorial team pre-approves what a character will say to a given prompt, because the model generates each reply live. That’s the main draw for many users, and it’s also why moderation here means something different than it does on a video platform.

In practice, moderation looks like this:

• Report-and-remove. A user flags a character or conversation, and a human or automated moderator reviews it afterward. That does nothing for the person who already saw it.
• Word filters. Character.AI lets users mute specific words, but a word filter is easily defeated by rephrasing and doesn’t cover broader topics.
• Tier-based filtering. Character.AI says its under-18 model runs more conservative classifiers than its adult model. This is one of the few published examples of filtering applied before a reply is generated.
• Regional restrictions. Laws such as the UK Online Safety Act push platforms to restrict adult content by location.
• Session-structure limits for minors. This is a different and generally stronger lever, covered below.

If a safety page mentions “community guidelines” and “reporting tools” but never describes filtering before generation, assume the moderation is reactive. That isn’t necessarily dishonest, since filtering generated text in real time is a genuinely hard problem. But it means the safety burden sits with the user, or with the parent supervising them, not the platform.

What Happens When a User Talks About Self-Harm?

Regulators now treat crisis response as mandatory, yet most safety pages describe it least clearly.

New York’s AI Companion Models Law, in force since November 5, 2025, requires two things. Operators must make reasonable efforts to detect and respond to suicidal ideation or self-harm. They must also remind users, at the start of a conversation and every three hours after, that they’re not talking to a human. California’s SB 243, effective January 1, 2026, adds a crisis-referral protocol and break reminders for minors at least every three hours.

Some platforms spell out how this works. Chai’s real-time message scanning is described as checking every message and refusing or replacing a reply when it detects self-harm content. A safety page that never mentions crisis response leaves the most important question unanswered.

Are Session Limits and Under-18 Modes Different From Age Verification?

Yes, and most coverage misses the distinction. Age verification decides who gets in. Session-structure limits decide what users can do once inside. For any platform with minor users, the second is arguably the stronger safeguard, because it keeps working when the gate fails.

Character.AI shows the split most clearly. Beyond adding age assurance, it changed the product for anyone identified as under 18:

• Open-ended chat ended for under-18 accounts on November 25, 2025, replaced by creation tools such as stories, videos, and streams.
• During the transition, minors were limited to two hours of chat a day.
• The company had already added Parental Insights reports and a notification after an hour of use.
• It also announced an independent nonprofit AI Safety Lab.

Those changes came under heavy legal pressure. In May 2025, a federal judge refused to dismiss the Florida lawsuit from Megan Garcia, mother of 14-year-old Sewell Setzer III, on First Amendment grounds. In January 2026, Google and Character.AI agreed to settle five family lawsuits filed in Florida, Colorado, New York, and Texas. Garcia had called the under-18 ban “about three years too late.”

The same month, Kentucky became the first state to sue a chatbot company, citing its consumer-protection and data-protection laws. At the federal level, the FTC ordered seven companies, including Character Technologies, to explain how they enforce age restrictions and handle conversation data. On October 12, 2026, Character.AI filed for Chapter 11 bankruptcy and began a court-supervised sale. The platform stays online during that process.

Explicit content is only part of the risk. Kentucky’s complaint centers on design that encourages emotional attachment and blurs simulated and real relationships, a pattern tied to how companion apps monetize emotional attachment. An age gate does nothing about that. A product built differently for minors at least limits it.

The practical takeaway: if a platform has only an age check and no separate under-18 tier, a minor who gets past the check has exactly the same access as an adult.

AI Companion App Safety and Privacy: What Happens to Your Data?

Verifying your age and chatting with a companion create two separate data trails, and platforms tend to explain only the one that flatters them.

On verification, the trend is to keep the platform out of the loop. Janitor AI’s checks run through k-ID, which doesn’t keep documents or facial data afterward. Janitor receives only a yes-or-no answer on whether the user is over 18. Character.AI says it gets only the result from Persona, never the selfie or ID itself. Sensitive documents are a liability if breached, so few platforms want to hold them.

Conversations are a different story. Hosted platforms typically retain chat content whether or not you ever verified your age. Texas Attorney General Ken Paxton’s August 2025 investigation alleges that chatbots on Meta AI Studio and Character.AI claim confidentiality while their terms of service allow interactions to be logged, tracked, and used for targeted advertising. Stored chats can also become an asset in a sale. What happens to them depends on what the buyer acquires, a live question for Character.AI users right now.

That gap between “we protect your ID” and “we keep your conversations” is worth checking in any privacy policy. A safety page that leads with ID-deletion language can be accurate while saying nothing about the chats themselves.

Locally run frontends such as SillyTavern change this picture only partly. SillyTavern stores chats on your own machine, so no companion platform keeps your history. Whether anyone else sees your messages depends on the model behind it. A fully local model keeps everything on your device. A hosted API key sends every message to that provider under its own retention rules, the same trade-off that applies when connecting an external API key to Janitor AI.

Local setups also give up the moderation, parental controls, and under-18 tiers that regulation now pushes hosted platforms toward. Neither structure is safer across the board; they trade off different risks.

AI Companion App Safety Guidelines for 2026: Five Questions to Ask

Whichever app you’re checking, five questions cut through most of the marketing language:

1. Does the age check verify anything, or just record what the user typed? Also check who gets checked: everyone, flagged accounts, or only users in certain countries.
2. Is there a separate product for minors, or just a locked door? A different under-18 mode keeps working when the gate fails.
3. Is moderation applied before generation, or only after reports? Report-based moderation protects the next person, not the one in the conversation now.
4. What happens when a user expresses self-harm? Look for a described detection and referral process, not a hotline link in the footer.
5. What does the privacy policy say about message content? If it covers only ID handling, that’s a gap, not an answer.

None of these requires special expertise. A platform’s own safety and privacy pages answer them within a few minutes, and the answers reveal far more than any “18+” badge on the landing page.

FAQs

Q. What does AI companion app safety look like for adults?

Age gates and under-18 modes don’t apply to adults, so other safeguards matter more:

• How the privacy policy treats message content.
• Whether the platform describes a real crisis-response process.
• Whether the design rewards constant engagement, which can turn casual use into dependence.

Q. Is age verification on AI companion apps mandatory now?

It depends on the platform and where you live. Brazil’s Digital ECA requires reliable age assurance, and Australia’s and the UK’s online safety rules have pushed adult-content services toward verification. Some platforms check every account in some form. Others check only users in the countries where the law requires it.

Q. Can a determined teenager get around age verification?

Often, yes. Italy’s regulator documented minors bypassing Replika’s age gate and blocked users returning to Character.AI with new accounts. Facial age estimation raises the bar but isn’t airtight.

Q. Did regulation or voluntary choice drive the 2025–2026 changes?

Mostly regulation and litigation. New York’s companion law took effect in November 2025, followed by California’s SB 243 in January 2026 and Brazil’s Digital ECA in March 2026. The lawsuits against Character.AI came before most of this legislation and helped shape it.

Related: Perchance Content Warning Fix (2026): Why It Happens & How to Remove It