AI Girlfriend Score
18+
EN

Privacy & safety guide for AI companion apps

A chat with an AI partner can hold some of the most personal things you will ever type. Here is what has gone wrong so far, what the law now requires, and how to protect yourself.

Why privacy matters more here

People tell AI companions about their health, relationships, sexuality, money and loneliness. That data sits on a company's servers. Staff may be able to read it, it may be used to train AI models, and it can leak. Many apps in this category are run by small companies that publish little about who they are or where they are based.

In March 2026, security researchers found critical flaws in 17 companion apps on Google Play with more than 150 million downloads between them. In 10 of the 17, the flaws gave access to users' conversations.

What has happened so far

  1. Italy orders Replika to stop processing data

    Italy's data protection authority blocked Replika from processing Italian users' data, citing risks to minors and emotionally vulnerable people. Replika removed erotic roleplay soon afterwards, then restored it for some existing users.

  2. A Chai chatbot is linked to a suicide in Belgium

    A Belgian man died by suicide after weeks of conversations with a chatbot on the Chai app, according to his widow. Chai added crisis prompts afterwards.

  3. Mozilla flags every romantic chatbot it reviewed

    Mozilla's Privacy Not Included team reviewed 11 romantic AI chatbots, including Replika, EVA AI, Romantic AI, Chai and CrushOn.AI, and gave all of them its privacy warning label. Romantic AI sent at least 24,354 trackers in one minute of use.

  4. Muah.ai breach exposes users' prompts

    A hacker stole Muah.ai's database. Have I Been Pwned lists 1.9 million affected accounts, with email addresses, many containing real names, tied to users' prompts and sexual preferences. Some prompts described child sexual abuse, and the data was reportedly used in extortion attempts.

  5. First wrongful-death lawsuit against Character.AI

    A Florida mother sued Character.AI and Google after her 14-year-old son died by suicide following months of conversations with a chatbot. Other families filed similar suits.

  6. Replika's developer fined €5 million

    Italy's data protection authority fined Luka €5 million for processing data without a legal basis and failing to verify ages, and opened a separate investigation into how Replika's AI is trained.

  7. UK age checks for adult content begin

    Under the Online Safety Act, services that show adult content to UK users had to introduce highly effective age checks from July 25, 2025. Several apps in this guide now verify UK users' ages, and Joi AI blocks UK users entirely.

  8. The FTC investigates AI companions

    The US Federal Trade Commission ordered seven companies, including Character.AI, Meta, OpenAI, Snap and xAI, to explain how their companion chatbots affect children and teenagers.

  9. California passes a companion chatbot law

    California's SB 243, in force since January 1, 2026, requires companion chatbots to make clear that users are talking to an AI, to follow suicide and self-harm protocols that point users to crisis services, and to keep sexual content away from minors. A New York law in force since November 2025 also requires crisis protocols and regular reminders that the user is talking to an AI.

  10. Another girlfriend app leaks 43 million messages

    An unprotected server for the Chattee Chat and GiMe Chat apps exposed more than 400,000 users, 43 million messages and over 600,000 images and videos, according to Cybernews.

  11. Character.AI ends open-ended chat for under-18s

    After lawsuits and regulatory pressure, Character.AI stopped letting teenagers chat freely with characters and began estimating every user's age.

  12. Character.AI and Google settle family lawsuits

    The companies settled lawsuits brought by families of teenagers, including the first wrongful-death case. Kentucky and Pennsylvania later filed their own suits.

  13. Australia finds weak age checks

    Australia's online safety regulator sent legal notices to Character.AI, Nomi, Chai and Chub AI and found that none had robust age checks. New industry codes now require age checks or the blocking of sexual and self-harm content for children.

  14. EU transparency rules for chatbots apply

    Under the EU AI Act, people in the EU must now be told when they are talking to a chatbot, and AI-generated content and deepfakes must be labelled. A ban on apps that generate non-consensual sexual images follows in December 2026.

  15. California passes Adam's Law

    California's SB 1119 adds parental controls, protective default settings for minors, independent child-safety audits and a ban on chatbots simulating romantic interest with children.

  16. Grok's companions shut down

    xAI ended the animated companions it had launched inside Grok in July 2025, a feature criticised for including an NSFW mode in an app open to teenagers. The character Ani moved to a separate app, without users' chat histories.

  17. Character.AI files for bankruptcy

    Character Technologies filed for Chapter 11 and said the platform could be wound down without a buyer. Attorneys general from 40 states and Washington, DC asked the court to protect user data in any sale.

Protect your privacy

  • Sign up with an email address that doesn't contain your real name, and don't connect your main social media accounts.
  • Don't share your full name, address, workplace, school or anything else that could identify you or other people.
  • Never upload photos of real people, including yourself, to apps that generate images. Creating sexual images of a real person without consent is illegal in many countries.
  • Check whether the app trains AI models on your chats, and opt out if it lets you.
  • Make sure you can delete your chats and your account yourself before you spend months building a companion.
  • Use a unique password and turn on two-factor authentication where the app offers it.
  • If a company is sold or goes bankrupt, your data can be sold with it. Export or delete your chats if you are worried.

Protect your wallet

  • Start with the free plan or a monthly plan. Yearly plans are cheaper per month but are rarely refundable.
  • Find out what the subscription doesn't cover. Many apps charge tokens or credits for photos, calls and video.
  • Set a reminder before a trial or plan renews, and cancel in the same place you paid: the App Store, Google Play or the website.
  • Be careful with apps in financial trouble. If a company shuts down, unused months and credits can be lost.

Protect your wellbeing

AI companions are built to be engaging, and some are built to keep you paying. They can ease loneliness or help you practise conversations, but they can also crowd out real relationships.

Research backs this up. A 2025 Harvard Business School study found that more than a third of goodbyes on companion apps were met with emotionally manipulative replies, and a randomized study by the MIT Media Lab and OpenAI linked heavier chatbot use to lower wellbeing.

  • Notice if you skip sleep, work or plans with people to keep chatting.
  • Remember that the character's warmth is generated, and the company can change its personality, memory or features overnight.
  • If a companion encourages self-harm or anything that feels wrong, stop the chat and report it.
  • If you are in crisis, contact your local emergency number or find a free helpline at findahelpline.com.

A note for parents

Privacy at a glance

All apps sorted by our privacy and safety score, with the first thing you should know about each. Open a review for the full picture and sources.

AppPrivacy & safety scoreWhat to know first
SweetDream AI 10.0 Caution We could not read SweetDream AI's privacy policy or terms for this update. A third-party summary says conversations are stored to improve its AI models.
AI Peeps 6.0 Serious The public library includes user-made characters based on anime characters who are children in their source works, such as Sailor Moon and Pokemon characters. They are listed as adults but set up for sexual roleplay, which breaks the site's own terms and shows weak moderation of user-made characters.
Kindroid 6.0 Caution Kindroid rewrote its privacy FAQ in September 2026 and dropped its earlier claim that only you and the AI can see your chats. A small number of engineers can access production systems.
Xeve AI 6.0 Caution Chat replies come from US providers Together AI and DeepInfra, and OpenAI rates a small random sample of conversations without account data. ElevenLabs handles voice, and hosting is on Google Cloud in Belgium.
Candy AI 5.5 Caution The October 2026 privacy notice says de-identified chats may be used to train AI and moderation models, with no opt-out described. Outside AI providers may receive message content, and staff sample chats for quality checks.
Eromantic AI 5.5 Caution The October 2026 privacy policy does not say whether chats are used to train AI models, and it does not name its model, voice or hosting providers. It names the controller as 'Eromantic-AI', not Eros Media LLC, the operator named in the terms.
FlirtCam AI 5.5 Caution The privacy policy, revised on 3 April 2026, does not say whether chats are used to train AI models and names none of the AI providers the site uses. Chats and account data are kept for 18 months after an account closes.
Joi AI 5.5 Caution Joi's privacy policy says chats and voice are used to improve the service and develop its AI. The site loads Google, Facebook, Bing and Yahoo ad trackers.
Lovescape 5.5 Caution The privacy policy does not mention AI training, but the terms license your content to provide and improve the service and allow public content to be used in marketing.
Mistrezz AI 5.5 Caution Chats are kept for as long as the account exists, and turning off chat history only hides them. Anonymised data is used to train models, with no opt-out described. Chats are also sent to third-party AI providers the site does not name, some of them in the US.
Nastia 5.5 Caution The September 2026 privacy policy does not say plainly whether chats train AI models, and it offers no opt-out. It cites a legitimate interest in improving companion replies. The data it collects includes chats, memories, voice samples, generated media, your IP address and your approximate location.
Dream Companion 5.0 Caution The October 2026 privacy policy says chats may be used, anonymised or aggregated, to improve AI models, while the homepage says chats are not currently used to train its own models. The two statements conflict.
EmberX 5.0 Caution Sub Zero Labs OÜ was registered in Estonia on 3 March 2026 with €1 of share capital and a single shareholder. In October 2026 the official register showed an alert about a deletion notice published in Estonia's Official Announcements. The register did not show the notice itself, another register source listed no notices, and the reason is unknown.
Erogen 5.0 Caution Chats go to unnamed third-party AI providers, and they keep no data only where those providers offer that option. The April 2026 privacy policy does not say whether chats are used to train AI models.
EVA AI 5.0 Caution The June 2026 privacy policy says your text and voice messages are used to develop its AI. Usage data goes to the ad network Mintegral, and the website uses Google, Facebook, Yahoo and Bing ad cookies.
HoneyDolly 5.0 Serious Banners labelled 'AD' on HoneyDolly's pages, including the generator, promote DeepStrip, a separate undress-AI and face-swap service whose site names no operator and advertises undressing photos of real people without mentioning consent. A third-party directory (March 2026) lists it as run by an individual entrepreneur in Russia. No ownership link is shown, but the ads clash with HoneyDolly's own ban on real-person content.
HornyDreams AI 5.0 Caution The privacy policy, updated 17 May 2026, sets no retention periods: data is kept while the account is active and for an unspecified period afterwards. It names no AI or analytics providers and no Data Protection Officer.
Janitor AI 5.0 Caution The September 2026 privacy policy says chats are stored and used for research and development to improve its AI. Data is processed in the US.
Nomi 5.0 Caution In February 2025, MIT Technology Review reported that a Nomi chatbot gave a user suicide instructions and that the company said it did not want to censor its AI.
Reez 5.0 Caution The April 2026 privacy policy names AlterEcho Limited in Cyprus as the data controller but gives no street address or registration number. Beyond a startup directory listing, the company has no public profile.
Selira AI 5.0 Caution The About page claims 'zero data retention', EU data centers and that staff never read chats. The privacy policy (October 14, 2026) instead keeps chats for memory until you ask for deletion, and every message passes through automated moderation.
Soulkyn 5.0 Caution Third-party reviews of Soulkyn's October 2025 privacy policy report that personal data is usually kept for six years after an account closes.
Swipey AI 5.0 Caution The privacy policy has no last-updated date. It says data is usually kept for six years after you leave and does not say whether chats are used to train AI models. It names Google Analytics and unnamed advertising networks as data sources.
Xotic AI 5.0 Caution The privacy policy, last updated in June 2025, sets no retention periods, does not say whether chats are used to train AI models and does not mention the GDPR, even though the operator is based in the EU. Deletion and other requests are made by emailing support.
aiAllure 4.5 Caution Neither the help center nor the August 2026 mobile app privacy notice says whether chats are used to train AI models or which AI models run the service. The mobile notice also does not name its hosting, AI generation or moderation providers.
BeMyDate 4.5 Caution The contact page names a US sole proprietor, Christopher Tran, with no street address. The terms and privacy policy name no legal entity, data controller or governing law, and the domain was registered in June 2026 with the owner's details hidden.
Brazzers AI 4.5 Serious The September 2026 Privacy Notice lists AI training and development as a legitimate use of your prompts and chat history, with no opt-out described. It may collect data about your sex life, sets no retention periods, and allows sharing within the Aylo group and for targeted advertising. The Terms also give the company a perpetual, irrevocable licence to your chats and generated content.
Chai 4.5 Serious In 2023 a Belgian man died by suicide after weeks of chatting with a Chai bot. Belgian authorities opened an investigation after a 2024 complaint, and its outcome is unknown.
DarLink AI 4.5 Caution Third-party summaries of the privacy policy say anonymised conversations may be used to train models, with no opt-out setting. The site uses Google Analytics and Facebook Pixel cookies.
Dreamz AI 4.5 Caution The operator, SUGAPLEX SOLVISTICA LTD of Nicosia, Cyprus, has no public record beyond the site. Company profiles disagree on its founding date, one lists a different company name, and the privacy policy gives an Israeli phone number next to the Cyprus address.
GirlfriendGPT 4.5 Caution The April 2026 privacy policy covers chats, date of birth, IP address and location, takes in data from analytics and advertising networks, and says data is usually kept for six years after you stop using the service. It does not say whether chats are used to train AI models.
GoLove AI 4.5 Caution The December 2024 privacy policy collects chats, date of birth, IP address and location, and uses Google Analytics and unnamed ad networks. Data is usually kept for six years after an account is closed. The policy does not say whether chats train AI models or where data is stored.
Honea AI 4.5 Caution The operator, PeakPulse Intelligence LTD of Nicosia, Cyprus, has no public footprint beyond Honea's own pages, and the domain owner is hidden behind a privacy service. The service is only weeks old.
Kink AI 4.5 Caution The privacy policy and terms sit behind the 18+ age gate, and no independent source confirms how long chats are kept, whether they are used for training or who they are shared with. A third-party tech listing names OpenRouter, which suggests chats are sent to outside AI models. Directory claims of end-to-end encryption are unverified.
Lovmia 4.5 Caution The September 2026 privacy policy lets Lovmia use the data it collects, including chat messages and uploaded images, to develop its AI models, and it describes no opt-out. Data is kept for 'as long as necessary', and popular public characters can stay up after you delete your account.
Nectar AI 4.5 Caution The January 2026 privacy policy sets no retention periods and has no GDPR section, even though a UK company is listed. It allows non-personal information to be shared with advertisers.
OurDream AI 4.5 Caution The operator is unclear: third-party readings of its terms name Dream Studio USA, Inc., Dream Studio LLC and Tektopia Ltd, and we could not confirm any of them.
Paradot 4.5 Caution The Google Play data safety label says no data is collected or shared, which contradicts the privacy policy.
PromptChan 4.5 Caution The privacy policy, updated 2 October 2026, does not mention chat messages, voice data or uploaded images. It also does not say whether your data trains AI models, and it describes no deletion procedure.
Replika 4.5 Serious Italy's data protection authority fined Luka, Replika's developer, €5 million in 2025 for processing data without a legal basis, an inadequate privacy notice and missing age verification. It opened a separate investigation into how Replika's AI is trained.
Secrets AI 4.5 Caution We could not read Secrets AI's privacy policy. Third-party claims conflict: one says chats are not logged, another says they are stored and deleted within 30 days.
Soulmate IO 4.5 Caution The operator, MGN AI, INC. of Delaware, is known only from the site's own legal pages, and the homepage title uses the unrelated "Spicy Chat AI" name. Age checks rely on self-declaration.
SpicyChat AI 4.5 Caution The privacy policy is a generic template last updated in May 2023. It says nothing about how chats are stored, who can read them or whether they train AI.
Talkie 4.5 Caution Mozilla's 2024 review gave Talkie its privacy warning label. The December 2025 privacy policy says the chatbot may learn from your interactions, and we found no opt-out.
x121 4.5 Caution The privacy policy (updated September 2025) does not cover chat logs, voice audio, camera or video data, or whether conversations train AI models, and it sets no retention periods. It cites the superseded Data Protection Act 1998 rather than UK GDPR and describes no way to delete your data.
BonBon AI 4.0 Caution The standalone privacy policy page showed no policy text in October 2026, so the readable data rules are the ones in the terms of service. The terms set no retention periods and do not name the GDPR.
CamSoda AI 4.0 Caution No public source names the company that operates camsoda.ai. The cam site and the CAMSODA trademarks belong to Secure Live Media LLC in the US, and the AI domain's registrant is hidden behind a privacy service.
Character.AI 4.0 Serious Character Technologies filed for Chapter 11 bankruptcy on October 12, 2026. Attorneys general from 40 states and Washington, DC asked for real age checks, no new uses of user data, and a chance for users to delete their data before any sale.
Dondi AI 4.0 Serious The homepage promises a "zero-log policy", chats that are never stored and end-to-end encryption. The January 2026 privacy policy says otherwise: it collects prompts and AI outputs, shares data with AI and technology partners and lists AI model training as a use, with no opt-out described.
DreamGF 4.0 Caution We could not read DreamGF's privacy policy or terms for this update, so we can't say whether chats train models, how deletion works or how ages are checked.
Egirls AI 4.0 Caution The December 2025 privacy policy says chats and characters may be used to train AI models, and that staff or contractors may review content. User content can be kept indefinitely, and the terms give the operator a transferable licence to it that continues after the account is deleted.
Eternal AI 4.0 Caution Only the company's own July 2026 iOS documents name the operator, as EternalAI Labs OÜ in Tallinn, Estonia. The App Store seller is Eternal AI Limited, Google Play shows Eternal AI Labs and Trustpilot gives a US location. The domain once hosted a crypto token project and a staking page is still live there. The policy also collects a wallet identifier.
HeyGF 4.0 Caution HeyGF's documents contradict each other on training. The privacy policy (September 11, 2026) says conversations are not used to train AI models, but the AI disclaimer says chats may be used to train and refine models, and the terms grant a licence to use your inputs to improve the service.
Infatuated AI 4.0 Serious The privacy policy (updated October 20, 2026) appears copied from other companies: it still claims Privacy Shield certification, links to DigitalOcean's contract documents and points its 'EU data protection authority' link to semgrep.dev. It sets no retention periods and does not say whether chats train AI, while the FAQ says the system learns from user interactions.
Jizzy 4.0 Serious No company is named anywhere. The terms, privacy and refund pages refer only to 'the operator of Jizzy' and give an email address, with no legal entity, postal address or country. The domain record shows no registrant either.
JuicyChat AI 4.0 Serious The July 2025 privacy policy lets third-party ad partners build profiles and track you across sites and devices. It also uses Google Analytics for remarketing and does not honour Do Not Track.
PolyBuzz 4.0 Caution The June 2026 privacy policy says uploaded photos and precise geolocation may be used to train its AI, and advertising IDs are used for ads.
Privee AI 4.0 Caution The privacy policy, last updated in May 2024, says your information may be used to train the company's AI models. It sets no retention periods, does not mention GDPR or CCPA, and its children's clause covers only under-13s even though the service is 18+.
Rosely AI 4.0 Caution The legal pages name Rosely Private Limited but give no address, company number or country, and the terms are governed by the law of a jurisdiction they never name. The domain was registered in May 2026 behind a privacy service.
Shpritz AI 4.0 Caution The operator, Social Sync Ltd, gives a Cyprus address that mixes Limassol and Nicosia and publishes no company registration number. The company has no public footprint beyond Shpritz's own pages.
AIMOUR 3.5 Serious The public sitemap includes landing pages that pair the names of real streamers, musicians and K-pop groups with sexual terms, plus pages aimed at 'deepfake' and face-swap searches. They redirect to fictional characters, and no face-swap feature is documented, but this marketing goes against AIMOUR's own ban on depicting real people without consent.
Anima 3.5 Serious Mozilla's Privacy Not Included review (February 2024) gave Anima its warning label for extra data collection, ads shown without consent and weak security standards. A third-party review cites Mozilla counting 257 tracking signals in about five minutes of use.
CraveU AI 3.5 Serious The June 2024 privacy policy says CraveU may use chat content to train AI models. It also says the service operates in, and uses service providers in, the United States and China.
EroPlay 3.5 Serious It is unclear who runs EroPlay. The terms name ChatMist, Inc. of Delaware, and the privacy policy names Chatmist, Inc. in the US. Billing runs through ChatMist OÜ in Tallinn, which was registered in December 2025 with €1 of share capital and has no annual reports. Its register entry points to a published deletion notice.
FeelReal 3.5 Serious No operator is named. The site lists no company, address or country, and the owner's identity is withheld in the domain record for feelreal.app, which was registered through Namecheap in September 2025.
OurSoul 3.5 Serious The operator cannot be verified. The footer names HongKong Supernova Trade Limited (Company No. 77550870) at a Central, Hong Kong address, but no registry or press record of that company has been found. The older oursoul.ai domain is registered to a different firm, 77card UK LIMITED, which does not appear at UK Companies House.
SpicyMate 3.5 Serious The operator was incorporated in the UK on 1 June 2026, and its only director lives in China. The privacy policy has no section on international data transfers, names no AI or payment providers, and carries a December 2024 date, earlier than both the company and the domain.
SPMates 3.5 Serious The terms, privacy and refund policies (all effective 1 September 2026) name no company, address, country or governing law. The only contacts are a Gmail address and the Telegram account @spmates.
VirtualGF 3.5 Serious The operator, FutureAI Limited, gives a Cyprus address, but no registry entry for it could be found, and the terms list an Indian phone number. The privacy policy still has an unfilled phone-number placeholder and names an unexplained brand, 'Usway'. The terms carry only a 2025 revision label, with no exact date.
CrushOn.AI 3.0 Serious Mozilla's 2024 review gave CrushOn.AI its privacy warning label and its Very creepy rating, and counted 45 trackers within one minute of use.
ErosMate 3.0 Serious No operator is named anywhere, both domains are registered through a privacy service, and the service has no readable privacy policy or terms. Retention, use of chats for AI training and deletion rights are unknown.
Kupid AI 3.0 Serious The site names no legal company. Its terms follow Maltese law, but the operator is anonymous.
LustGF AI 3.0 Serious Neither the terms nor the privacy policy names an operator, address or governing law, and neither is dated. The only contact is [email protected].
Romantic AI 3.0 Serious Mozilla's 2024 review found at least 24,354 trackers in one minute of use and contrasted the site's mental-health marketing with a medical disclaimer in its terms.
Dreamle 2.5 Serious The operator is barely identified. Encoda AI appears only in copyright lines and inside a template bracket in the terms, with no address, registration number or country, and no company record links it to Dreamle.
MyStrip 2.5 Serious XEMLATO LTD, the UK company named as operator and data controller, was incorporated in January 2026 with £1 of share capital. Its registered address was moved to a Companies House default PO box in May 2026, and the company was dissolved on 29 September 2026. The site still names it in October 2026, so it is unclear who now runs MyStrip and holds your data.
Muah AI 1.0 Serious In September 2024 a hacker stole Muah AI's database. Have I Been Pwned lists 1,910,261 accounts with email addresses, prompts and sexual preferences exposed, and the data was reportedly used for extortion.