Adult AI Companion Apps, Compared on What They Keep

OurDream's chat page says "Not even we can see it." OurDream's privacy policy says it collects your chat communications and may use them to train its models.

That's the pattern this page is about, so here's the direct answer for anyone choosing an adult AI companion on privacy. Of the five apps ranking for this search, three promise on their landing page that your conversations are private or encrypted, using nearly the same words. Two of those three have a privacy policy or terms that say something narrower and more specific, and in OurDream's case something close to the opposite. The third, Anima, has legal pages that return their own title and nothing else. The two that come out best are the two that don't make the promise. Kindroid says chats are encrypted at rest and in transit and then explains exactly when an automated scan reads them and when a human can. Nomi says end-to-end encryption is impossible for a chatbot that has to read your message to answer it, and describes what it does instead. If you're going to type things you'd never say aloud into one of these, the policy is the product, and the landing page is not.

Every quote below is from the app's own homepage, privacy policy, terms or help documentation, read on 15 September 2026. The age gate on each app, and what the content rules say once you're through it, is a separate question covered in the 18+ apps audit. This page is about the conversation after you've had it.

Five Promises, Five Policies

The results for this search are listicles ranked by impression, vendor pages, and one policy explainer aimed at parents. None of them quotes a privacy policy. Here's what the five apps say on the page you land on, and what they say in the document you agreed to.

Read from each vendor's own pages on 2026-09-15. "Not addressed" means the document was read and the topic is absent; "Unreadable" means the document would not load.

Two things stand out before the detail. Nobody in this set names the company whose model generates the replies, which is the single most useful privacy fact a companion app could publish. And the column that separates the five isn't the promise, it's whether the policy says anything at all about the thing you'd actually worry about, which is the text of the chat.

OurDream Says Not Even We Can See It

The chat page's FAQ is the strongest privacy claim in this article. "Your conversations are encrypted end to end. What happens in your chats stays in your chats. Not even we can see it."

The privacy policy, last modified 26 August 2026 and issued by Dream Studio USA, Inc., lists what the company collects, and the list includes "Your User Content (as defined in the Terms of Service), such as chat communications, posted images, and shared Characters". The uses it gives for that information include "Analyze, maintain, improve, modify, customize, and measure the Services, including to train our artificial intelligence/machine learning models". Its own description of the encryption is "encrypted both at rest and in transit using industry-standard AES encryption and TLS 1.3 protocols", which is a normal and good practice and is not end-to-end encryption, because the company holds the keys. A company that can train on your chats can read your chats. Both statements are on the same site, one written for the visitor and one for the regulator, and the second is the one that binds.

The rest of the policy is ordinary. Vendors for storage, payment and analytics. Google Analytics and ad networks named among the tools that collect device data automatically. A note that the company doesn't advertise "as of the date of this Policy" but may share data with advertising partners in future. Deletion through Settings, with the usual carve-out for legal compliance and security, and no numbered retention period that I could find. One clause I'd flag because it's rare in this category. Where the law requires it, age verification "is performed by a third-party age assurance provider and we receive the outcome of the check", which is the right way round for the user and is mentioned here only because it's a statement about what data the company holds.

Nastia's Notice Is Older Than Its Product

Nastia's homepage says "100% UNFILTERED CHAT", "Safe. 100% private conversations", and in its FAQ that it uses "industry-standard encryption to protect all conversations and personal data" and will "never share or sell your information to third parties".

The privacy notice was last updated on 28 January 2023. That's a document more than three and a half years old on the fetch date, and it reads that way. It never mentions conversations, chats or messages. The personal information it lists is names, email addresses, phone numbers, usernames, passwords, card numbers and billing addresses, which is the list a web shop would publish. It does, though, say the company processes "data about a person's sex life or sexual orientation" and "information revealing political opinions" and religious beliefs as sensitive categories, and that it collects "location data such as information about your device's location, which can be either precise or imprecise", including via GPS. It lists "personalized and relevant advertising content" among the purposes. And the only sharing it describes is a business transfer, which is where "never share" comes from.

So the notice is silent on the one thing the product is. The terms, updated 3 May 2024, are not silent, and this is the clause I'd want anyone considering Nastia to read. The definition of "Contributions" begins "The Services may invite you to chat" and includes text, audio, photographs and personal information you create through the service. Section 10 then grants the company "an unrestricted, unlimited, irrevocable, perpetual, non-exclusive, transferable, royalty-free, fully-paid, worldwide right, and license to: use, copy, reproduce, distribute, sell, resell, publish, broadcast" your Contributions, "including, without limitation, your image, name, and voice", "for any purpose, commercial, advertising, or otherwise".

I'll be fair about what that is. It's a template clause, the kind a terms generator produces for a site with forums and comment sections, and I don't think it was written with an intimate chat log in mind. But the definition includes chat, the licence includes sell, and the document is the one you agreed to. On a product whose homepage says "You can trust Nastia to keep your secrets safe", the paperwork says the company may publish them. The homepage's deletion instructions, Settings then Account then Delete Account, are the most concrete privacy control on the site, and the notice adds that some information may be kept afterwards to prevent fraud or assist investigations.

Anima's Policy Is Its Own Title

The Anima landing page ranking for this search says "All 18+ MyAnima AI chats are encrypted, private, and confidential" and, further down, that "all exchanges with MyAnima are fully encrypted and confidential".

Its privacy policy, at myanima.ai/legal/privacy, returned sixteen kilobytes of HTML to my request, of which the visible text is "Anima: AI Friend". The terms page is the same. A different privacy URL returned a 404. This is the result this site got when it read Anima's documents in July, and it hasn't changed. The pages presumably render for a browser running the site's JavaScript, and I'm not going to describe what I couldn't read. What I can say is that a privacy claim on a landing page, backed by a policy that a plain request can't retrieve, is a claim with nothing behind it that a reader can check. The app is on both stores, so its Data Safety panel and privacy label are the fallback, and those are the developer's own declarations.

Kindroid Encrypts, Then Says When It Reads

Kindroid's homepage, as served to my request, is a loading screen and a one-line description of the app, so there's no privacy promise to test, which after the three above is a relief. Its documents are the most precise in the set, and they took some getting. The help centre's public pages render the same loading screen without JavaScript, so I read the terms, privacy policy and moderation guidelines through the app's own API, the same source the page loads them from.

The privacy policy, effective 28 June 2024, says chats "and all other sensitive, application layer data will be encrypted in rest and transit, so we will not be able to view any of said data in our normal operation of the business". Chat content appears in the data table as "Identifying Information that you voluntarily choose to include in your chats", disclosed to service providers such as the cloud host and to parties you authorise. Communications you send the company are used for marketing "excluding the Chats". There is no training clause and no model provider anywhere in the terms or the policy. The one hedge is a sentence that appears in most US policies now, that depending on your state some disclosures "may constitute a 'sale'" as the law defines it. Retention is "for as long as necessary", without a number.

Then the moderation guidelines, updated 15 August 2026, say what "normal operation" excludes. Automated scans read "YOUR messages and inputs", plus backstory and memory fields, for three categories of real-world harm. "All scans are of current context, and historical chats/media are not scanned." "No human reads any content during the automated detection and enforcement processes." A human reads only if you appeal a lock, and appealing "gives explicit consent for a trained member of our Trust & Safety team to decrypt and review the specific content that led to the lock(s)", covering the previous two days. The design, and the trade it makes, is worked through in the sexting policy audit, and I won't repeat it. The point for this page is narrower. Kindroid is the only app here that tells you, in one document, that a machine reads the current chat, and in another, under what conditions a person can.

Nomi Says Encryption Can't Work and Explains What It Does

Nomi's homepage sells memory, not secrecy, and its wiki answers the encryption question in a way the other four landing pages should have to answer.

"End to end encryption is not possible with an AI chatbot as we would need to have the keys to decrypt the message, (so your Nomi can respond) which defeats the purpose of encryption." That's the whole problem with OurDream's sentence and Anima's, stated by a competitor. What Nomi describes instead is that the model has to read your message to reply, that chats are used for learning "in aggregate" with names removed so that a lesson reads "A Nomi said X to a human and that human didn't like it", and that the company tries to hold very little about you, "in most cases not even having your full name". The privacy policy, updated 27 April 2026, says the company doesn't sell or rent personal information and that deleting your account deletes your data within about 28 days, except for a training archive that survives in a form no longer attributable to you. The wiki adds that there are no plans for ads.

That's a training-on-chats admission, and I'd rather have it than OurDream's, because it's specific about the de-identification and it isn't sitting under a sentence that says the opposite. The rest of what Nomi publishes, including the memory limits that are the reason to use it at all, is in the Nomi review.

Who Reads It, and for How Long

The table the landing pages don't want you to build.

Compiled from each app's privacy policy, terms and help documentation, read 2026-09-15.

One app in five gives a numbered deletion window. One describes human review at all. Two publish a document that addresses the chat text as a category of data. And the one with the most confident landing page has the least behind it. The deletion mechanics for several of these, including which ones let you leave cleanly, are already in the free companion exit audit, so I've kept that column short.

SpicyChat, which the top-ranking listicle puts first, isn't in either table because its privacy policy wouldn't load. Both routes return an application shell with no text, and its documentation site has no legal section. I'd rather leave it out than guess, and the one thing its docs do say is that "accounts can not be restored once deleted".

Questions

Are adult AI companion chats private? On the landing pages, always. In the policies, it depends on the app, and on two of the five the policy says something the landing page doesn't.

Is anything here end-to-end encrypted? No, and Nomi's wiki explains why it can't be. Kindroid and OurDream both describe encryption at rest and in transit, which protects the data from outsiders and from casual internal access but not from the systems that generate the reply.

Do they train on my chats? OurDream says yes, in its policy. Nomi says yes, in aggregate, with names removed. Kindroid has no training clause. Nastia doesn't address it, though its terms licence would cover it. Anima can't be read.

Does a person read it? Kindroid is the only one that answers, and the answer is only if you appeal a lock and consent. Nomi's wiki says learning is anonymised before processing. The others don't say.

What happens when I delete? Nomi gives a number, about 28 days, with a de-identified training archive kept. Nastia and OurDream keep some data for fraud, legal and security reasons without a period. Kindroid keeps data "as long as necessary". SpicyChat says deleted accounts can't be restored.

Which one would I type into? Kindroid or Nomi, and for the same reason. They're the two that describe the mechanism instead of promising the outcome. What they say about age checks, which is a different document again, is in the 18+ apps audit.

What I'd Do Before Typing

Open the privacy policy and search it for the word "chat". If the word isn't there, as on Nastia, the document wasn't written for the product and the promise on the homepage has no paperwork behind it. If it is there, read the sentence that follows it, because on OurDream that sentence is a training clause.

Then find the encryption claim and ask what it protects against. "At rest and in transit" is a real protection against outsiders and a real limit on internal browsing. "End to end" on a chatbot is a sentence nobody can make true, and an app that makes it anyway has told you how carefully the rest was written.

Then decide what you'd accept. My own line is that I'd accept de-identified training if the company says so plainly, as Nomi does, and I'd accept automated scanning for a short list of real harms if the company says what it scans and when a person can look, as Kindroid does. What I wouldn't accept is a promise I can't check against a document, and three of the five apps ranking for this search ask you to.