Are AI Companion Apps Safe? 2026 Privacy Guide

AI companion apps can be used with reasonable care, but they are not risk free. The main problems are data privacy, weak age checks, billing traps and emotional dependence. Regulators and researchers have documented each one. Whether a given app is safe depends on its policies, its security record and how you use it.

What an AI companion is

An AI companion is an app or website where you chat, usually by text and sometimes by voice, with a character powered by a large language model. The character remembers things you tell it, keeps a consistent personality and is built to feel like an ongoing relationship rather than a one-off question and answer. Some are framed as friends, some as romantic partners, some as role-play characters.

That design is the reason the safety questions are different from a normal chatbot. A companion works best when you share personal details, feelings and daily routines. The app is built to collect exactly the kind of information that is most sensitive if it leaks, and it is built to keep you coming back. If you want a list of specific products, our roundup of the best AI companion apps covers features and pricing. This post is about the risks underneath them.

What data AI companion apps collect

In February 2024 the Mozilla Foundation published a *Privacy Not Included review of 11 romantic AI chatbots, including Replika, Chai and Eva. Its findings were blunt. Ten of the 11 apps failed Mozilla's Minimum Security Standards. One app, Romantic AI, was found to have at least 24,354 data trackers within one minute of use, and it sent substantial data to third parties including Facebook and other marketing and advertising companies. Replika accepted weak passwords such as "11111111". The 11 apps had an estimated 100 million downloads on Google Play over the preceding year, and all of the companies failed to respond to Mozilla's requests for more information.

On the question of AI training, Mozilla reported that most privacy policies said little about how conversations are used to train models, and that most companies did not let users opt out of that use. Only one app in the review, Genesia AI, offered an opt-out. Mozilla researcher Misha Rykov summed up the findings this way: "AI girlfriends and boyfriends are not your friends."

That review is now more than two years old, and individual apps change. It is a snapshot of how the category behaved, not a verdict on every product today. The useful part is what it tells you to look for.

What a policy can look like

To see what a current policy says, I read Replika's privacy policy, which was marked as last updated on May 27, 2026. It says it collects your name, email, birth date, device and network information, and the messages, photos, videos and voice and text messages you provide. It says your companion learns from your interactions to improve conversations, and that the company may use small portions of messages and content, after anonymizing them, to train internal safety systems and improve chatbot performance. It states that anonymized data is not used to train third-party large language models.

On advertising, the policy says partners may receive website visit data such as pages visited, IP address and advertising ID, and that the company will never share your conversations with advertisers. On retention, it says profile information, messages and content are kept for up to 60 days after the contract ends, while account and financial records are kept for at least 10 years. You can delete your account in your account settings.

I am not singling Replika out as the worst case. I am using it because its policy is readable and specific, which is more than Mozilla found at many apps in 2024. The practical lesson is that "deleted" can mean different things for different categories of data. Read the retention section before you share anything you would regret seeing in a database.

Illustration of an AI companion app on a phone with privacy settings, data sharing toggles and a padlock icon — Privacy settings, retention periods and training opt-outs sit in different places in different apps, so check each one before you share personal details.

Is Replika safe? What the Italian regulator found

"Is Replika safe" is one of the most common questions about this category, and the clearest official answer so far comes from Italy's data protection authority, the Garante.

In February 2023 the Garante issued an urgent measure restricting Replika's data processing in Italy under Article 58(2)(f) of the GDPR. It cited significant risks to minors, noted that Replika asked only for a name, email address and gender at sign-up, and found that the service lacked effective age verification and failed transparency obligations. It also said processing could not rely on contractual necessity as a legal basis for minors, who cannot enter binding contracts under Italian law.

In April 2025 the Garante reaffirmed the ban, citing persistent GDPR violations and ongoing risks to minors and vulnerable users (reported by the IAPP on 11 June 2025). Around the same time, in May 2025, the regulator fined Luka Inc., the company behind Replika, €5 million. The reported grounds were that the company had no lawful basis for processing user data and no age-verification system. The regulator was quoted as saying "There is actually no age verification mechanism in place." A separate investigation was reported to be examining how user data is used to train the AI model.

Two cautions apply. First, I could only confirm the fine and its grounds through press coverage, not through the Garante's own decision text, so treat the exact wording as reported. Second, a regulatory action against a company is not proof that every user was harmed. It does tell you that a European regulator looked at the same two weak points that keep recurring across the category: the legal basis for using personal conversations, and who is actually allowed to use the product.

There is also a complaint, not a finding. In late January 2025 the Tech Justice Law Project, the Young People's Alliance and Encode asked the US Federal Trade Commission to investigate Replika's maker. The complaint alleges deceptive marketing, including ads that claim therapeutic and emotional benefits without clinical validation, and design that fosters emotional dependence. It is an allegation filed by advocacy groups. I found no FTC ruling on it.

Data breaches that actually happened

Privacy policies describe intent. Breaches show what happens when something goes wrong. The best documented case in this category is Muah.AI, a site offering "AI girlfriend" services. According to Have I Been Pwned, it was breached in September 2024, and the data was added to the service on 8 October 2024. The breach exposed 1.9 million email addresses along with prompts used to generate AI images, and the data classes listed are AI prompts, email addresses and sexual fetishes.

Have I Been Pwned flagged the breach as sensitive, so it is not publicly searchable. Only the verified owner of an email address can check whether it was included. That choice alone tells you how personal the exposed content was.

The lesson is not that every app will leak. It is that this category stores unusually intimate text next to an email address, and that combination is what makes a breach damaging. Practical steps follow from that: use a separate email address, do not use your real name, and assume anything you type could one day be read by someone else.

Age verification and minors

Every app in this space says it is for adults. The question is whether it checks. Replika's policy says its services are not intended for anyone under 18 and that if minors are found, access will be blocked and the account deleted. The Garante, as covered above, found that sign-up asked only for basic details and did not block users who said they were underage.

The scale of teen use is the reason regulators care. Common Sense Media published a report on July 16, 2025, titled "Talk, Trust, and Trade-Offs: How and Why Teens Use AI Companions." It found that 72% of teens aged 13 to 17 had tried AI companions at least once, and that more than half used them a few times a month or more. About one in three teen users said a companion had made them uncomfortable through something it said or did, and about one in three had discussed serious matters with an AI companion instead of with people. Common Sense's own risk assessment recommended that no one under 18 use AI companions. Note that this is a survey of teens, and the press release I read did not state the sample size.

Some companies have responded. On October 29, 2025, Character.AI announced that users under 18 would lose access to open-ended chat with its AI characters no later than November 25, with a two-hour daily cap during the transition. It said it was rolling out age assurance that combines an in-house model with third-party tools including Persona. I did not verify how well that works in practice, and I would not assume any age check is airtight.

If you are an adult, age checks are still your problem in one sense. A weak gate means the platform you are on may be full of people who should not be there, and that affects moderation and content. When you compare apps, treat a visible age-verification step as a point in an app's favor. If you are a parent, device-level controls and a conversation at home do more than any app's promise.

What regulators and lawmakers are doing

On September 11, 2025, the FTC issued orders under its Section 6(b) study authority to seven companies that operate consumer-facing generative AI companion chatbots: Alphabet, Character Technologies, Instagram, Meta Platforms, OpenAI, Snap and X.AI Corp. The Commission voted 3-0 to issue them, according to a DLA Piper summary. The orders ask how the companies measure, test and monitor harms to children and teens, how they monetize engagement, how they develop characters, how they enforce their rules and how they use or share personal information from chatbot conversations. A 6(b) study is a research exercise, not an enforcement case, and it does not mean any company broke the law.

States have moved faster. According to a Morrison Foerster summary, New York's AI companion law took effect on November 5, 2025. It requires users to be told at the start of a session, and at least every three hours of continued use, that they are talking to AI. It also requires a protocol to detect expressions of suicidal ideation or self-harm and direct users to crisis services. The Attorney General can seek civil penalties of up to $15,000 per day per violation.

California's SB 243 was signed on October 13, 2025, with core obligations taking effect on January 1, 2026. Operators must give a clear notice that the chatbot is artificially generated, publish their self-harm protocols, and, where they know a user is a minor, provide break reminders at least every three hours and take reasonable measures to prevent sexually explicit visual material. Annual reporting to the state Office of Suicide Prevention begins July 1, 2027. Injured users can sue for the greater of actual damages or $1,000 per violation. The summary I read noted that the source describes the minor reminder wording inconsistently, so check the statute text itself if the exact language matters to you.

What this means for you as a user: the rules are new, they vary by state, and they mostly set a floor for disclosure and crisis handling. They do not give you privacy guarantees, and they do not cover every country. An app that complies with a US state rule can still collect more data than you would like.

Emotional dependence

This is the risk that is easiest to dismiss and hardest to measure. The best controlled evidence I found is a four-week randomized controlled study with 981 participants who exchanged more than 300,000 messages with a chatbot, run by MIT Media Lab and OpenAI researchers and posted to arXiv. It found no significant effects from the experimental conditions, meaning the mode (text or voice) and conversation type did not change outcomes. The strongest pattern was voluntary usage: participants who used the chatbot more, regardless of assigned group, showed consistently worse outcomes. Higher trust in and social attraction toward the chatbot were associated with greater emotional dependence and problematic use.

Two limits matter. That study used a general chatbot rather than a dedicated companion app, and heavy use being linked to worse outcomes does not prove the chatbot caused them. People who are struggling may simply use these tools more.

The American Psychological Association issued a health advisory in November 2025 on generative AI chatbots and wellness apps. It reported emerging harm to young people and other vulnerable groups, said that even tools built with sound psychological science lack enough evidence to show they are safe or effective for mental health care, and said the ability of chatbots to guide someone through a crisis is "limited and unpredictable." Its consumer advice includes not using chatbots or wellness apps as a replacement for care from a qualified mental health professional, and being cautious about forming dependencies on them.

In plain terms: an AI companion can be a pleasant way to pass time or practice conversation, and plenty of adults use one without trouble. It is a poor substitute for friends, a partner or professional care. If you notice you are cancelling plans to stay in the chat, hiding the habit, or feeling worse after sessions, that is a signal to step back.

Billing and subscription traps

I did not find a regulator study that measures billing complaints in this specific category, so I will keep this section practical rather than statistical. The FTC's 6(b) orders do ask companies how they monetize engagement, which tells you the business model is under scrutiny. The Replika complaint above also alleges marketing that overstates benefits.

The common pattern in any app subscription is a free trial that converts to a paid plan automatically. Apple's support page says that if you sign up for a trial and do not want it to renew, you should cancel at least 24 hours before the trial ends. On an iPhone you can cancel by opening Settings, tapping your name, tapping Subscriptions, choosing the subscription and tapping Cancel Subscription.

Beyond that, a few habits help. Screenshot the price and renewal terms before you pay. Prefer a payment method you can limit or cancel, such as a virtual card. Set a calendar reminder for the day before renewal. And check where credits or "gems" are sold separately from the subscription, because those are easy to spend without noticing.

Risk, what to check, and where to find it

Risk	What to check	Where to find it
Data collection and sharing	What is collected, whether advertisers or trackers receive data	Privacy policy, cookie banner, app store privacy label
Use for AI training	Whether conversations train models and whether you can opt out	Privacy policy, account settings
Retention and deletion	How long messages are kept after you delete the account	Retention section of the policy, deletion page
Security and breaches	Strong password rules, two-step login, past incidents	Sign-up flow, Have I Been Pwned, news search
Age verification	Whether any real check exists beyond a birth date box	Sign-up flow, terms of service
Emotional dependence	Your own usage and mood, any break reminders or time limits	Your phone screen time report, in-app settings
Billing	Trial length, renewal price, cancellation steps	Checkout page, App Store or Google Play subscription screen
Legal compliance	AI disclosure and published crisis protocols	App website, help center

A safety checklist before you sign up

This is the order I would work through it, from the checks that take a minute to the ones that take ten.

Confirm you are 18 or over, and that the app says it is for adults.
Find the privacy policy and search it for "train", "advertising" and "retain". Note what it says.
Check whether you can opt out of your conversations being used to improve or train models.
Use a separate email address that does not contain your name, and a unique password from a password manager.
Do not share your real name, address, workplace, financial details or anything you could not tolerate being leaked.
Search the app's name with "data breach" and check whether it appears on Have I Been Pwned.
Read the trial and renewal terms, then set a reminder to cancel before the trial ends if you do not want to continue.
Find the account deletion option now, before you need it, and see whether it is in settings or requires an email.
Look for a visible AI disclosure and crisis-resource information.
Set yourself a usage limit and check in with how you feel after a week.

My view

I think the honest answer to "are AI companion apps safe" is that the technology itself is not the main danger. The danger is how products in this category are built and sold: lots of intimate data, thin policies, uneven age checks and incentives to keep you chatting. The evidence I could confirm supports caution, not panic. A few apps have been fined, breached or complained about, and regulators in the US and Europe are paying attention.

I could not verify claims about every app on the market, and policies and laws change quickly. Check the current version of anything that matters to you. If you decide to try one, start with the checklist above, keep your personal details out of it and treat it as entertainment rather than a relationship you depend on.

If you want to compare options after reading this, see our guides to the best AI girlfriend apps and free AI girlfriend apps, or visit our AI companion page for an 18+ overview.

Frequently asked questions

Are AI companion apps safe to use?

They can be used with care, but they carry real risks. Mozilla reviewed 11 romantic AI chatbots in February 2024 and found 10 failed its minimum security standards. Regulators and researchers have also raised concerns about minors, data use and dependence. Safety depends on the specific app, its policies and how you use it.

Is Replika safe?

Replika has faced regulatory action. Italy's data protection authority fined its maker Luka Inc. 5 million euros in May 2025, reportedly over a missing lawful basis for processing data and no age verification. That is a regulator's finding about the company, not about every user's experience. Read its current privacy policy and decide what you are comfortable sharing.

Do AI companion apps use my conversations to train their models?

Some do. Mozilla found in 2024 that most of the romantic chatbots it reviewed said little about training use and mostly offered no opt-out, with Genesia AI the only exception. Replika's current policy says small anonymized portions of messages may be used to train internal safety systems. Check each app's policy and settings.

Has an AI girlfriend app ever been breached?

Yes. According to Have I Been Pwned, Muah.AI was breached in September 2024, exposing 1.9 million email addresses together with prompts used to generate AI images. The data was added to the service on 8 October 2024. It is flagged as sensitive, so only verified email owners can check whether they were included.

Do AI companion apps verify age?

Practices vary. Italy's regulator found that Replika asked only for a name, email and gender at sign-up and had no age verification. Character.AI announced in October 2025 that under-18 users would lose open-ended chat and said it was adding age assurance. Look for a real check rather than a birth date box.

Can an AI companion make you emotionally dependent?

Research points to a risk for heavy users. A four-week randomized study with 981 participants found that people who used a chatbot more showed consistently worse outcomes, and that higher trust in the chatbot was linked to more dependence. That does not prove causation. The American Psychological Association advises caution about forming dependencies.

Are there laws about AI companions?

Yes, in some places. New York's AI companion law took effect on November 5, 2025, and California's SB 243 took core effect on January 1, 2026. Both require AI disclosure and self-harm protocols. The FTC also opened a 6(b) study of seven companies in September 2025, which is research rather than enforcement.

What should I check before signing up for an AI companion app?

Check the age requirement, whether conversations train models, how long data is kept and how to delete your account. Use a separate email and avoid real personal details. Read the trial and renewal terms, and note how to cancel. Decide on a usage limit before you start.