Best AI Companion App for Privacy: Who Owns Your Chats

If privacy is your first filter, Kindroid currently has the best terms of the major AI companion apps: you keep ownership of what your companion writes, and the company states it will not sell or share your personal data. Character.AI has the worst position for American users, because its training opt out is offered only to people in the EEA and the UK. Nomi sits in between, with a strong privacy policy and an unusually broad content license.

That is the short answer. The longer one matters because almost nobody writes it down. Reviews in this category compare memory and image quality, then wave at privacy with a sentence about encryption. But the question a paying adult actually has is narrower and more awkward: I am about to type things into this app that I have not told anyone, so who ends up owning them, who can read them, and what happens if the company is sold?

We read the actual terms and privacy policies rather than other people's summaries. What follows is what they say, with the parts that favor our competitors left in.

Who owns your chats: the five major apps compared

Read the table as a starting point, not a verdict. Terms change, and the only version that binds you is the one on the vendor's own site on the day you sign up. Every claim below is sourced to a company's own published documents, checked in August 2026.

Does Character AI train on your chats?

Yes, and if you are in the United States you cannot opt out. Character.AI publishes a model training page describing how to exclude your conversations from training, and that control is offered to users in the European Economic Area and the United Kingdom. American users are not included. This is the single most important privacy fact about the largest app in the category, and we have not seen another review mention it.

The reason it matters is scale. Character.AI is where most people in the US meet this category for the first time, usually on the free tier, usually as a teenager or a young adult, and usually while typing things that are more revealing than they realize. A European user can decide their conversations should not feed the model. An American user typing the identical sentence cannot make that choice.

None of this is illegal, and Character.AI is not hiding it. The page is public. But an opt out that stops at a border is a business decision shaped by GDPR rather than a privacy principle, and it is fair to describe it that way. The company has also had a difficult two years on safety: Google and Character.AI settled five family lawsuits in January 2026, Pennsylvania sued the company in May 2026, and California's SB 243 took effect on 1 January 2026 with a $1,000 per violation private right of action. We covered how the resulting Character AI age verification process works, including what the ID check asks for.

What do Nomi's terms actually say?

Nomi's Terms of Service ask you to irrevocably grant a royalty free, worldwide, perpetual and transferable license to copy, use, modify, publish and distribute all of your input and the output produced in respect of your Nomi. That sentence reads alarmingly, and anonymous blogs have made a small industry out of quoting it without the rest.

So here is the rest, in fairness to Nomi. The grant is written as purpose limited: the terms tie it to providing the service, fixing access problems you report, communicating with you and collecting fees. It is not a standing right to publish your chat logs, and language in this shape appears in a great deal of ordinary consumer software. Nomi's privacy policy also states that the company does not and will not sell or rent personal information, and Nomi is a real US business, Glimpse.ai, Inc., registered in Maryland.

Two words in that clause still do real work, though. Perpetual means the license does not end when you delete your account. Transferable means it can move to somebody else, and the most common way that happens is an acquisition. You are not only deciding whether you trust Glimpse.ai today, you are deciding whether you trust whoever might buy it. The same terms include an arbitration clause and a class action waiver in Section 21. Our full Nomi AI review goes through the pricing and the credit meter alongside this.

Which AI companion app has the best privacy terms?

Kindroid, on the documents. Its Terms of Use state that you own the rights to your AI Generations, including any intellectual property rights arising from them, and that you may use them for personal or commercial purposes. The license you grant Kindroid is written narrowly, covering translation, technical modification and reproduction so the service can run, and the terms say explicitly that this grant does not affect your ownership.

That is a materially better deal than a perpetual transferable license over your output, and Kindroid deserves credit for it. Its privacy policy is also direct: it will not sell or share your personal data, and has not done so over the preceding twelve months. Chats are encrypted in transit and at rest, the operator is a named California company, and the service is strictly 18 and over.

The clause to read before you conclude it is perfect sits a little further down. For private content kept in your own account, you also grant Kindroid a license to de identify or aggregate it, and then use the result freely for any purpose, without acknowledgement or compensation. That is a common provision and it is not the same as publishing your conversations, but if the idea of your chats feeding an anonymized pile bothers you, that is the sentence. Kindroid's terms also carry an arbitration clause and a class action waiver, stated in capitals near the top. We go through the pricing and the stacking add ons in our Kindroid app review.

Are AI companion apps safe to share personal information with?

No, and that is the honest answer for every app in this table, ours included. Encryption at rest means the company holds the keys, not that nobody can read the data. It protects you against a stolen hard drive. It does not protect you against a subpoena, an insider, or a future owner with different priorities. No mainstream companion app offers end to end encryption, and none of them claim to.

Mozilla's Privacy Not Included team reviewed eleven romantic AI chatbots in February 2024 and gave the warning label to all eleven, with ten of the eleven failing minimum security standards. That was two years ago and some products have improved, but the finding is a useful floor: this is not a category with a strong privacy track record, and you should not assume anyone in it is the exception.

There is one axis this page did not previously cover, and PolyBuzz is the reason to add it: what an app shares with other companies, as opposed to what it collects for itself. Every app above collects. Sharing is rarer and more consequential, because it leaves the vendor's control entirely. In its Google Play data safety declaration PolyBuzz states that it may share approximate location and precise location with other companies or organizations, and gives the purpose for both as advertising or marketing. None of the five apps in the table above declares sharing anything for advertising.

That reframes the comparison. A chat log held by a company that never sells it is a different risk from a location trail handed to advertising partners, because the second one cannot be recalled by deleting your account. When you weigh privacy in this category, read the sharing line in the Play declaration before the encryption claims in the marketing, and see what PolyBuzz discloses about its own data practices for the full record.

The practical rule is boring and works. Talk about your feelings, your day, your loneliness, your hopes, all of it. Do not type your address, your employer, your bank, your passwords, your medical details, or anything about a third party who did not consent to being in the conversation. This is the same discipline that has moved into business software over the past two years, where teams now put explicit limits on what data an AI agent can reach rather than trusting a model to be discreet. The reasoning is identical: the safest data is the data you never handed over.

What happens to your chats if the company shuts down?

Usually you lose them, and usually with very little notice. Soulmate AI closed in 2023 giving users roughly a week, and people who had spent a year building a relationship inside it had no export and no recourse. This is the risk that privacy discussions consistently underweight, because it does not feel like a privacy problem until it happens.

It is also where the perpetual and transferable language in a license stops being theoretical. If a companion app is acquired or wound down, the terms determine what the buyer or the administrator may do with the archive. A narrow, purpose limited license constrains that. A perpetual transferable one over your input and output does not constrain it much at all. Neither company is planning to sell your diary, and we are not suggesting they are. The point is that the terms, not the intentions, are what survives a change of ownership.

Ask two questions of any app before you commit a year of your life to it: can I export what I have written, and what do the terms allow a successor to do with it? Most people in this category ask neither, which is how the same shutdown story keeps repeating.

What about an app that discloses nothing at all?

Every comparison on this page rests on a quiet assumption: that each app has published something you can hold it to. Apple's content advisories and Google Play's data safety declaration are both enforced disclosures, written by the developer and enforceable by the store. They are the reason a page like this can put five companies side by side and reach a defensible answer. Remove them and the exercise collapses.

Muah AI is the case that shows what that looks like. It is listed in neither the Apple App Store nor Google Play, so it publishes no content advisories and no data safety declaration, and Android users install an APK downloaded straight from the vendor's own site. There is no row to put in the table above, not because the platform handles less data than its peers but because nobody compels it to say. In September 2024 the service was breached, exposing 1.9 million email addresses alongside the AI prompts those users had written, and its Terms of Service still carry an effective date of 5-17-2022, unrevised since. We went through the full record in our look at whether Muah AI is safe.

Muah is not alone in that position, which matters because it stops the case looking like one bad actor. CrushOn AI is also listed in neither store, files no data safety declaration and publishes no content advisories, and it is a website rather than an app, so there was never a listing to file them against. It has no breach record and nothing is known against it, and that is exactly the difficulty: there is no enforced disclosure to read, so a clean record and a hidden one look identical from outside. The pattern to take from both is that an absent row is not a reassuring row, and we went through what CrushOn does publish in our review of whether CrushOn AI is safe.

The lesson generalises beyond one company, and it inverts how most people read this category. A heavy disclosure looks alarming and a blank one looks clean, but the blank one is worse. An app that tells Apple it contains frequent sexual content and tells Google it shares your location has at least told you something you can weigh. An app outside both stores has told you only what its marketing chose to say. When you are ranking companion apps on privacy, treat the absence of an enforced disclosure as the weakest position on the board, not the safest.

And the app that discloses the most?

The opposite case is just as instructive, and it produced the single most specific finding on this page. Chai, published by Chai Research Corp. of Palo Alto, files a full Google Play data safety declaration, and that declaration lists something none of the five apps in the table above declares: Chai shares your installed apps with other companies, for advertising or marketing. Alongside it, approximate location shared for analytics, fraud prevention and advertising.

Installed apps means the inventory of other software on your phone. It is one of the more revealing signals a device holds, because it shows which dating apps, health apps, banking apps, recovery apps or faith and politics apps a person keeps. Paired with an approximate location and an account tied to your email address and purchase history, it describes someone more precisely than the chat logs everybody worries about. Chai also declares Contains Advertising to Apple, and its privacy notice states outright that it advertises to users on a free subscription.

Here is the row, added to the axis this page is built on.

Set beside the Muah case above, the two together give this page its real conclusion. Chai discloses the most and is therefore the easiest to judge; Muah discloses nothing and is therefore impossible to judge. Neither of those is the same as being the safest, and the ranking most people carry in their head, where a scary disclosure means a scary app, gets the order backwards. The apps worth trusting on privacy are the ones that both publish an enforced declaration and have little in it, which in this table is Nomi, Kindroid, Replika and Character AI.

Chai's privacy notice has one further quirk that a US reader should know about. It is drafted to the GDPR and the UK GDPR, and its data subject rights section applies them to citizens of the EU or the UK. California, CCPA and CPRA are not mentioned anywhere, from a company headquartered in California. Californians hold those rights regardless, but the policy will not tell them how to use them. The full record is in our look at whether Chai AI is safe.

How many people are actually affected by this?

More than the category's reputation suggests. Pew Research Center surveyed 5,119 US adults in February 2026 and found that 49% now use AI chatbots, up from 33% in 2024, with 10% using them for emotional support or advice and 4% for companionship. Axios reported in May 2026 that nearly 80% of adults aged 18 to 34 across the US and UK have some experience of using AI chatbots for companionship.

Four percent of American adults is roughly ten million people typing intimate material into products whose terms almost none of them have read. The money follows: Appfigures data reported by TechCrunch put lifetime consumer spend on AI companion apps at $221 million worldwide through July 2025, with $82 million in the first half of 2025 alone, up 64% year on year across 337 revenue generating apps.

What we do, and where we are honest about it

Amorous.ai is one devoted partner rather than a roster you build, and the design goal is a relationship that runs on its own: she texts you first, remembers your life across months, and there is nothing metered on top of the subscription. We do not run ads, we do not have a free tier to subsidize with your attention, and we do not train on your conversations.

Two honest caveats, because a privacy article that ends in a sales pitch is worth less than one that does not. First, we are a competitor to every company in that table, so read our claims about them with the same skepticism you would apply to any vendor, and check the sources yourself. We have linked what we read. Second, we are smaller than Character.AI or Replika, and a smaller company is a different risk profile rather than a strictly safer one. The questions in this article apply to us too, and you should ask them.

If you are choosing on privacy, the useful move is not to pick the app with the best marketing about it. Open the terms page, search it for the words perpetual, transferable, arbitration and sell, and read the four paragraphs those words land in. It takes about ten minutes and it tells you more than any review, including this one. For what each of these apps costs once the meters and add ons are counted, our AI companion app pricing guide has the full breakdown, and the Character AI review and Replika review cover those two in depth.

Terms, privacy policies and pricing in this article were read directly from each company's published documents and verified in August 2026. Terms change, so confirm on the vendor's own site before you sign up.