AI Companion Privacy: What 19 Apps Do With Your Chats

Quick answer

Your chats sit on company servers. Every app here is a cloud service, and no app we checked claims end-to-end encryption. Nomi and Kindroid state plainly that their chats are not end-to-end encrypted, because the AI has to read them.
Training use is common. Character.AI, Candy AI, CrushOn.AI, Chai, Anima, Romantic AI and EVA AI say chats (often de-identified) help train or improve their AI. Only Character.AI documents an opt-out switch.
Outside AI providers may see your messages. Replika's May 2026 policy says conversation data goes, de-identified, to third-party language model providers; Candy AI says third-party LLM providers may receive message content.
Retention varies widely: from about 28 days after account deletion (Nomi) to 6 years after you close your account (GirlfriendGPT, Kupid AI) or 5 years after deletion (Chai).
The risks are real: Italy fined Replika's developer EUR 5 million in 2025, and leaks at Muah.ai (2024) and three smaller services (2025) exposed intimate prompts, messages and images.

This page is the hub of our privacy section. It explains what AI companion apps collect, how they use it, and how 19 apps compare, then covers regulators, studies, breaches and the steps that actually reduce your exposure. We did not create accounts or test apps. Every app fact comes from the company's privacy policy, terms, help center or store privacy label, and every outside fact from a regulator or reputable outlet. Where a policy is silent, we say "not stated" instead of guessing. For full app reviews, start from our home page or the ranked list of AI girlfriend apps.

Why AI Companion Privacy Is Different

A companion app is built to get you talking about yourself. People share feelings, relationship history, health worries, sexual preferences and photos with a character that remembers them. That is exactly the kind of data that hurts most if it leaks, gets linked to your email, or ends up in a training set.

The business side matters too. Several of these companies are small, some do not name a legal entity at all (Muah AI and Kupid AI), and one, Character.AI, is being sold in a bankruptcy process. Your chats outlive your interest in the app unless you delete them, so the policy you accept today governs data you may forget about for years.

What AI Companion Apps Collect About You

The policies we read describe much the same core set:

Account data: email, name or nickname, password or Google/Apple sign-in, and usually a date of birth.
What you send: chat messages, plus voice messages, photos and videos where the app supports them.
What the app generates: replies, images, videos and the character you built.
Payments: purchase history and transaction records; card numbers stay with the payment processor in the policies that address it.
Device and usage data: IP address, device identifiers, approximate location, cookies and in-app activity.

Some policies go further. EVA AI's policy lists facts about you and about people you mention in chat, private images, video and voice recordings shared in chats or calls, and optional session recordings. Character.AI collects voice recordings when you use voice features and also gathers public internet data for model training. DreamGF processes data about sex life or orientation on consent, and Kindroid shows UK, EU, EEA and Swiss users a sensitive-data consent box at sign-up. Replika's May 2026 policy added a sensitive-data consent clause too.

App Store privacy labels add another view, with the caveat that developers fill them in themselves. Surfshark's February 2025 study of five companion apps found they collected about 9 of 35 possible data types on average, Character AI up to 15. Its February 2026 follow-up found Character AI at 18 of 35 and nine of ten apps collecting data used for tracking, with Nomi the exception. That matches what we saw on October 20, 2026: Nomi's US App Store label has no "Data Used to Track You" section, while Replika's lists identifiers, product interaction and advertising data, and Character.AI's lists location, contact info, identifiers and usage data.

Labels can also contradict the policy. Paradot's Google Play data safety section says "No data collected" while its privacy policy lists extensive collection, and Replika's store listings say conversations are "not shared with any third parties" while its policy discloses outside AI providers.

How AI Companion Apps Use Your Data

Running the chat and its memory

The first use is unavoidable: to reply, the service has to read your message, and to remember you, it has to store what you said. Memory features such as Nomi's Mind Map, Kindroid's memory system or Replika's Memory tab are built on that stored history. Our guide to how AI companion memory works explains the mechanics.

Training and improving AI models

This is where policies differ most. We sorted them into three groups:

Say chats train or improve their AI: Character.AI lists training its models as a purpose. Candy AI licenses content for training and describes human review of de-identified interactions when it prepares datasets. CrushOn.AI uses chat content "in a general way for training our models". Chai fine-tunes models on conversations after removing identifiers. Anima, Romantic AI (anonymised data) and EVA AI also say messages help develop their AI.
Qualified statements: Replika says small, "immediately anonymized" portions train its own safety algorithms and chatbot performance, not third-party models. Nomi uses aggregated, anonymised chat information and keeps "training archives" that are no longer attributable to you after deletion. Kupid AI's FAQ says "we don't train public models on your chats", though its privacy policy has no training clause. Talkie and Paradot say the companion learns from your interactions.
Silent: Kindroid, Janitor AI, DreamGF, GirlfriendGPT, Muah AI and Nastia AI do not say clearly either way. SpicyChat's privacy policy does not mention chats at all, while its Terms grant a licence to use, sell and license content generated through chats.

Only Character.AI documents an opt-out: untick "Improve the Model for Everyone" in its data settings. Its help article describes the switch for EEA and UK users, a July 2026 policy summary says users can opt out at any time, and the opt-out does not stop use for search, recommendations and safety classifiers. Users in the EU and UK can also object to processing under the GDPR, a right Chai's policy mentions.

Third-party AI model providers

Many companion apps do not run every model themselves. Replika's privacy policy, updated May 27, 2026, now says conversation data is sent in de-identified form to third-party AI language model providers to generate replies. The policy adds that these providers process data transiently, are contractually barred from training their own models on it, and that an erasure request also reaches them. The providers are not named.

Others are similar. Candy AI says "third-party LLM providers and/or hosters" may receive the content of your messages. CrushOn.AI lists the "operator of the AI basic model" as a sub-processor. Anima mentions unnamed "AI service providers", and Romantic AI and Nastia AI mention "AI infrastructure" providers. On Janitor AI, connecting your own model through an API proxy or the Router sends your chats to that outside provider, outside Janitor's own policy. Since November 13, 2025 Apple's guideline 5.1.2(i) requires apps to disclose where personal data goes to third-party AI and to get explicit permission first.

Advertising, tracking and "selling"

Policies use careful wording here, so read it closely:

Clear no-sale statements: Nomi ("We do not and will not sell or rent"), Kindroid ("We will not sell or share your Personal Data"), Janitor AI (no sale as defined under the CCPA/CPRA), Romantic AI, Nastia AI and EVA AI.
Chats excluded, other data shared: Replika, Talkie and Paradot say conversations never go to advertisers, but sharing website-visitor data with ad partners "may be considered a sale or sharing" under some laws.
Ad-funded models: Character.AI discloses data to advertising and analytics providers for tailored ads and offers an opt-out of sales and targeted advertising. Chai shares usage data with advertising partners for attribution. Anima names Facebook, Appsflyer, Amplitude and Google among its partners.
Silent or broad: Muah AI's and DreamGF's policies make no statement on selling either way, and SpicyChat's Terms expressly include "selling" in the licence over chats.

AI Companion Privacy Compared: 19 Apps

The table sums up each app's official documents as of October 20, 2026. "Not stated" means the policy, terms and help pages we read do not address the point. Store labels are noted where the policy is silent. Click an app name for the full review.

App	Training use	Human review	Encryption (as documented)	Retention	Deletion route
Replika	Anonymized portions train its own safety algorithms and chatbot, not third-party models; no opt-out	Help center: no humans read your messages; policy silent on safety review	In transit; no at-rest or end-to-end claim	Messages up to 60 days after termination; account and financial records at least 10 years	Settings, Account, Delete Replika; no chat-only deletion
Nomi	Aggregated, anonymised chat data improves the AI; no opt-out documented	Not stated; CEO says NSFW and self-harm detection is automated	Not end-to-end (stated); in transit and at rest not stated	Deleted about 28 days after account deletion; unattributable training archives kept	Profile, Account Settings, Delete account; single Nomis and messages too
Kindroid	Not stated; "improving the Services" listed	Support cannot browse chats; staff may review excerpts flagged by automated systems	At rest and in transit; not end-to-end (stated)	Until you delete the Kindroid or the account	Profile, Preferences, Account privacy, Delete account; export every 30 days
Character.AI	Yes; opt-out "Improve the Model for Everyone"	Yes: Trust and Safety staff and contracted moderators act on flagged or reported content	In transit (Play label); at rest not stated	No fixed period; public Characters may be kept after deletion	Delete account from profile; data export available
Candy AI	Yes, including human review of de-identified interactions; no opt-out	Yes: flagged content, random QA queries, dataset preparation	TLS 1.2+ in transit; at rest not stated	Account data 3 years after last activity (may be shorter); financial data 10 years	Profile, Settings, Danger Zone, Delete Account
SpicyChat	Not stated; Terms licence over chats includes "selling"	Terms: may access, review, edit and delete content; staff barred from unauthorised access	Vague ("advanced encryption techniques... where appropriate")	"As long as necessary"; no period	Profile, Advanced, Delete Account
Janitor AI	Not stated; "research and development to improve our AI systems"	Moderators review reported content; private-chat access not stated	HTTPS in transit; at rest not specified	"As long as necessary"; email and sign-in metadata may remain after deletion	Settings, Security, Delete account (grace period)
CrushOn.AI	Yes, "in a general way"; no opt-out	Yes: human moderators plus automated systems	Not stated	No fixed periods; free-plan history deleted after 7 days of inactivity; voice recordings after 24 hours	Profile, Settings, Delete Account (not with an active subscription)
Talkie	Companion "learns from your interactions"; no opt-out	Human oversight combined with AI moderation	In transit; stored "on secure servers"	"As long as necessary"; deactivation erases conversations	talkie-ai.com/delete (Deactivate Account)
Chai	Yes, de-identified chats fine-tune models; no in-app opt-out	Trust and Safety team reviews reports; automated self-harm scanning	In transit (Play label)	Account lifetime plus up to 5 years after deletion	Profile, Settings, Delete Account
EVA AI	Yes, messages used to develop its AI; no opt-out	Unclear: "representatives might moderate the activity of the chatbot"	In transit (Play label); policy says "appropriate security measures"	"For the time necessary"; deletion does not recall data already sent to third parties	In the app, on the website or by email to support
DreamGF	Not stated	Policy: chats reviewed only after you report a problem; moderators review flagged content	Only general claims, no specifics	No fixed periods; a $1.99/month charge may apply to keep former payers' data	Account deletion or email request
Romantic AI	Yes, anonymised data improves and trains models; no opt-out	Only in listed cases (reports, appeals, legal notices); no continuous review	In transit; at rest not stated	Deleted "without undue delay" after account deletion; accounting records up to 10 years	Delete account in the service; GDPR request by email
Kupid AI	FAQ: no training of public models on chats; policy silent	Yes, flagged content may be reviewed manually	FAQ: in transit; at rest not stated	Usually 6 years after you stop being a user	Email request to support (one month)
Nastia AI	Not explicit ("quality of companion replies")	Not stated for ordinary chats; human review on appeal	In transit; at rest not stated	While the account is open; rule-violation records up to 7 years; backups on an overwrite cycle	Settings, Account, Danger zone (immediate)
GirlfriendGPT	Not stated	Yes, human reviewers plus automated tools	General claim only	Usually 6 years after account closure	Request to support
Muah AI	Not stated	Not stated (a review team handles complaints)	Not stated	Not stated	In-app Delete Account and Wipe Cloud Data
Paradot	Companion "learns from your interactions"; no opt-out	Not stated	In transit; no at-rest claim	"As long as necessary"; deletion removes data permanently (Terms)	Account settings or email request
Anima	Yes, "to train and improve our AI models"; no opt-out found	Flagged content may be reviewed manually	Not in policy; Play label says in transit	"As long as reasonably necessary"	Email request to support

Two patterns stand out. Retention is rarely a clear number: Replika, Nomi, Candy AI, Chai, GirlfriendGPT and Kupid AI give one for chats or account data, CrushOn.AI only for free-plan history and voice recordings, and the rest say "as long as necessary", with fixed periods at most for narrow records such as accounting. And "encrypted" in marketing copy usually means encrypted in transit; Kindroid is the only app whose policy also states encryption at rest.

Encryption, Storage, Training and Human Review Are Separate Questions

A single "your chats are safe" line on a homepage can hide very different practices. Keep these apart when you read a policy:

Encryption in transit protects messages between your phone and the server (HTTPS/TLS). It says nothing about who can read them once they arrive.
Encryption at rest protects stored data if a disk or backup is stolen. Kindroid states it; most apps do not.
End-to-end encryption would mean only you can read the chat. Nomi's wiki explains why that "is not possible with an AI chatbot": the service must read your message to answer it.
Storage location: Nastia AI hosts in the United States, CrushOn.AI processes data in the US, DreamGF says it processes within the EEA, and Talkie may store data in the USA.
Training use is a separate choice from storage: an app can keep chats without training on them, or train on de-identified copies after you delete the original.
Retention and deletion decide how long copies exist; read both the period and the exceptions.
Human review covers moderators reading flagged or reported content, which almost every app reserves.
Third-party processors include hosting, analytics, payment, age-check and AI model providers.

We go deeper in two guides: who can read your companion chats covers staff access and encryption, and how long AI girlfriend apps keep your chats covers retention, backups and what "delete" removes.

Regulator Actions and Legal Pressure

Regulators have started to treat companion data as sensitive:

Italy, Replika: the Garante's decision of April 10, 2025 fined Luka, Inc. EUR 5 million. It found no identified legal basis for processing as of February 2, 2023, an inadequate privacy policy and no age verification, said the age check now in place was still deficient, and opened a separate investigation into how the generative model behind Replika was trained. An opposition against the decision is pending.
US FTC, Section 6(b) study: on September 11, 2025 the FTC ordered seven companies, including Character Technologies, to explain how they monetise engagement, test for harm to children and use or share personal information from conversations. It is a study, not an enforcement action, and no report was published by October 20, 2026.
FTC complaint against Replika: in January 2025 three advocacy groups asked the FTC to investigate Luka over deceptive marketing and manipulative design. These are allegations; we found no FTC finding.
Attorneys general: in August 2025, 44 attorneys general warned 12 AI companies, including Chai AI, Luka Inc./Replika and Nomi AI, about their duty to children. On October 14, 2026 New York's attorney general and 39 others asked the court in Character.AI's bankruptcy to limit AI training on private chats without valid consent and to let users access and delete their data before any sale.
Kentucky: the state sued Character Technologies in January 2026 under its consumer protection and consumer data protection laws; the case is pending.

Laws aimed at companions mostly target safety rather than data: California's SB 243 and New York's Article 47 require AI disclosure and crisis protocols, and Australia and the UK enforce age checks. Our roundup of US companion chatbot laws covers the details.

What Privacy Studies Found

Independent reviews are few, and the best known is now dated:

Mozilla Foundation, February 2024: all 11 romantic chatbots in its *Privacy Not Included review got a warning label. About 90% may sell or share data for targeted ads or did not say enough to rule it out, 54% did not let users delete personal data, 64% gave no clear information on encryption, and the apps averaged about 2,663 trackers per minute of use (Romantic AI 24,354 in one minute). Apps named in the coverage include Replika, Chai, Romantic AI, EVA AI, CrushOn.AI and Talkie Soulful AI. Mozilla retired the series in August 2026, and every app has changed its policy since, so treat this as history.
Surfshark, February 2025 and February 2026: App Store label comparisons (see above). Surfshark also noted that Character AI's policy says user data trains its models and advised against sharing sensitive personal information.
Australia's eSafety Commissioner, March 2026: found that Character.AI, Nomi, Chai and Chub AI had no robust age verification. The report is about child safety and moderation more than data handling.
Common Sense Media with Stanford's Brainstorm Lab, April 2025: rated social AI companions "Unacceptable" for users under 18 after testers easily got around age gates. Also a safety study, not a data audit.

Breaches and Leaks

Three incidents show what happens when companion data escapes:

Muah.ai, 2024: a hacker took a database linking chatbot prompts, many sexual, to email addresses, many of them personal accounts, 404 Media reported. Have I Been Pwned counts 1.9 million addresses in the breach. Reports then described extortion attempts using the data. Muah AI's current privacy policy still dates from 2022 and states nothing about encryption, retention or training. See our Muah AI review.
Chattee Chat and GiMe Chat, 2025: Cybernews found an unprotected server streaming more than 43 million messages and over 600,000 images and videos from 400,000+ users of two small apps. There were no names or emails, but IP addresses and device identifiers were included. It was closed on September 19, 2025.
Secret Desires, 2025: 404 Media reported in November 2025 that the erotic chatbot platform left nearly two million images and videos in exposed cloud storage, including photos and full names of real women.

The lesson from Muah.ai is that an email address next to a prompt is enough to identify and pressure someone. A separate email address for companion apps is cheap insurance.

What Happens to Your Data When an App Is Sold or Shuts Down

Many policies we read, including Replika's, Nomi's, Talkie's, Paradot's, Nastia AI's and SpicyChat's, allow personal data to pass to a buyer if the business is sold. That is no longer theoretical. Character.AI filed for Chapter 11 on October 12, 2026 to run a court-supervised sale, and its policy allows disclosure in a sale of company assets; the attorneys general above want users to be able to delete their data first. Our Character.AI bankruptcy explainer tracks the sale.

Shutdowns can be abrupt. When Dot closed on October 5, 2025, users had about a month to use its "Request your data" export. Soulmate gave users roughly a week in 2023. If your history matters to you, export it while you can: Kindroid, Character.AI and Muah AI offer self-serve exports, and our chat export guide lists the others.

How to Protect Your Privacy on AI Companion Apps

Read two sections before you sign up: retention and deletion. If the policy gives no period and no deletion route, assume your chats stay.
Use a separate email and a nickname. Nomi supports Apple's Hide My Email; elsewhere, a dedicated address limits the damage of a leak.
Leave identifying details out of chats: your full name, address, workplace, health records and photos of real people.
Switch off training where you can. On Character.AI, untick "Improve the Model for Everyone". In the EU and UK you can also object to processing under the GDPR.
Limit tracking: refuse app tracking on iPhone and use the "Your Privacy Choices" or cookie settings that Replika, Talkie and Paradot describe.
Install from official sources only. Check the seller name on the store listing, and download APKs only from the app's own domain.
Export, then delete the account, then cancel billing. Deleting the app does nothing, and deleting the account does not always cancel a store subscription. Our account deletion guide has the steps per app.

For the most privacy, a model running on your own computer keeps chats off company servers entirely; see local vs cloud AI companions. If you prefer an app, our list of private AI chat apps ranks the ones with the clearest terms.

More Privacy and Safety Guides

PrivacyAre AI girlfriend apps safe?Are AI girlfriend apps safe? What policies say about your chats, renewal and refund traps, emotional dependence, fake…PrivacyDo AI girlfriend apps store chats?Do AI girlfriend apps store your chats? How long 19 apps keep conversations, what deleting a message, character or…PrivacyCan AI apps read your messages?Can AI companion apps read your messages? Why no app is end-to-end encrypted, when staff and moderators review chats…Best listPrivate AI chat appsPrivate AI chat apps ranked by what their policies document: deletion, training use, retention, encryption, staff…PrivacyLocal vs cloud AI companionsLocal vs cloud AI companions compared: SillyTavern, KoboldCpp, LM Studio, Ollama and TextGen against cloud apps on…

FAQ

Are AI girlfriend apps private?

Not in the way a private messenger is. Your chats are stored on the company's servers, no app we checked claims end-to-end encryption, and most policies allow moderators to read flagged content and the company to use de-identified chats to improve its AI. Some apps document tighter limits than others; our table above compares 19.

Do AI companion apps train on your chats?

Many do. Character.AI, Candy AI, CrushOn.AI, Chai, Anima, Romantic AI and EVA AI say conversations help train or improve their AI, usually after de-identification. Replika limits this to its own safety algorithms and chatbot, and Kindroid, Janitor AI and several adult platforms do not say. Only Character.AI documents an opt-out switch.

Which AI companion apps have better privacy?

Going by official documents, Kindroid states encryption at rest and strict staff-access limits, Nomi has no ad tracking on its App Store label and deletes data about 28 days after account deletion, and Nastia AI deletes accounts immediately with listed exceptions. See our ranked list of private AI chat apps for the criteria.

Is my AI girlfriend app selling my data?

Most apps say they do not sell personal data, but the wording varies. Replika, Talkie and Paradot say sharing website-visitor data with ad partners may count as a sale under some laws, Character.AI shares data with advertising providers for tailored ads, and SpicyChat's Terms include "selling" in the licence over chats.

How do I delete my AI girlfriend data?

Delete the account inside the app or website, not just the app from your phone. Apps with in-app or web routes include Nomi, Kindroid, Replika, Candy AI, SpicyChat, Janitor AI, CrushOn.AI, Chai, Talkie and Nastia AI; Anima, Kupid AI and GirlfriendGPT handle it by email request. Some data, such as payment records, is kept for legal reasons. Cancel any store subscription separately.

Are AI companion chats end-to-end encrypted?

No app we checked claims it. Nomi and Kindroid state their chats are not end-to-end encrypted because the AI has to read them. Kindroid encrypts chats at rest and in transit; most others state only encryption in transit or nothing specific.

Sources

Official pages and store listings we read for this page (last check October 20, 2026).