Where Do Your AI Companion Chats Go? Privacy Compared
Your AI companion chats go to the vendor's servers, stay there under a retention policy you have probably never read, and, as some policies state, may be used to improve the models. That's the plain answer. The interface feels like a diary with a heartbeat; legally and technically it's much closer to feedback flowing into a software service. This piece compares what five apps publish about training, deletion and retention, flags the red-line patterns, and gives you a five-minute method for auditing any companion policy yourself.
Intimacy is exactly why this matters. People tell companions things they tell no one else β that's the product working as intended. Which makes this the one software category where the gap between how private something feels and how private it is deserves a full article.
The journey of a message
- Your message leaves the device over an encrypted connection. Transport encryption is standard everywhere and tells you almost nothing.
- It lands on the vendor's servers, where the model must read it in plain form to reply. True end-to-end encryption is generally impossible for this product type β the question is never whether the company can access your chats, but who else can, and what the company does with them.
- It is retained. Retention windows range from clearly stated to entirely unspecified.
- It may be sampled for safety review, quality checks or β where policy permits β model training.
- Around the conversation sit analytics and, in free ad-supported apps, advertising SDKs that observe usage even if they never see message content.
- Backups and logs can outlive the conversation you deleted.
Memory features raise the stakes further: the better a companion remembers you, the more of you is retained by design. That trade-off is covered in our guide to how companion memory works, and the two articles are worth reading together.
What the five apps publish
Everything in this table comes from the vendor pages linked in each row, checked on 2026-11-12. Policies in this category change frequently (Character.AI's help center, for example, shows policy edits dated 9 to 12 November 2026), and stated commitments are not audited ones. "Not stated" means the sources we read do not say. Verify before you rely.
| App | Training on your chats (per policy) | Human review | Deletion path | Export | Notes |
|---|---|---|---|---|---|
| Nomi | Chat use for training not stated; its wiki says thumbs up or down feedback may be used to improve Nomi (wiki) | Not stated | Account Settings, with email verification; the wiki says full deletion removes all data (wiki) | Not stated | The privacy policy mentions training or communications archives retained after deletion |
| Kindroid | Not stated for chats; the policy says de-identified data is used "to analyze, build and improve the Services" (policy) | Not stated | Account and per-AI deletion; chats are kept until you delete the AI or the account | Not stated | Deleting an account does not cancel a paid service, per the same policy |
| Replika | Small anonymized portions of messages are used to train safety algorithms and enhance chatbot performance (policy) | Not stated | Account deletion in settings; no way to delete message history without deleting the account (help) | Not stated | Its policy describes sharing some personal data for targeted advertising, while a help-center article says it does not sell personal information; the two do not match |
| Character.AI | Retains data you share to help train and improve its models; opt-out toggle for EEA and UK users only (help) | Not stated | Permanent account deletion in profile settings (help) | Not stated | Policy pages are edited often; see the privacy policy |
| Talkie | Unknown | Unknown | Unknown | Unknown | Its site is a JavaScript page and we could not read a privacy policy |
Two fairness notes. First, this page compares what the policies say, not whether the vendors follow them, and we cannot verify practice. Second, publication is uneven: Talkie's privacy terms were not readable in the sources we checked, and Nomi and Kindroid do not say in the pages we read whether chats train models. A missing statement is not a promise either way, so the differences above are differences in what each vendor chose to publish.
Red flags in any companion privacy policy
- Silence on model training. A policy that never mentions training either way was written by someone who chose not to answer the question.
- Deletion instructions that only cover the app. Uninstalling deletes nothing; if account deletion isn't described, assume it's difficult.
- Advertising SDKs inside an intimacy product. Your message content may be safe while your usage patterns β when, how often, how long β feed ad systems.
- Sharing with unnamed partners and affiliates. Named recipients can be checked; unnamed ones cannot.
- No age-assurance process in an adult-capable app. Beyond the obvious problem, it signals weak governance overall.
- A free product with no visible business model. Servers and model inference are expensive; if you cannot see what pays for them, consider what might.
The five-minute policy check
- Open the privacy policy in a browser and search for train, improve and machine learning. You are looking for an explicit statement about whether conversation content trains models, and whether you can opt out.
- Search retain and delete. A good policy describes account deletion, what it covers, and a timeframe. Note whether chat deletion and memory deletion are treated separately.
- Search share, partners and advertising. Count how many recipients are named versus gestured at.
- Compare the app-store privacy label against the policy. Labels are self-reported, but a mismatch between label and policy is itself a finding.
- Find the privacy contact or data-request email. Rights created by GDPR and CCPA-style laws are often extended to all users in practice; if you can't locate the mailbox to exercise them, that's your answer about how seriously they're taken.
Five minutes, no legal training required, and it will place any companion app on the spectrum above.
Data hygiene for companion users
- Sign up with a pseudonym and a dedicated email address. The companion experience loses nothing.
- Share patterns, not identifiers. A companion needs to know you hate early meetings, not your employer's name; that you live near the coast, not your street.
- Treat photos and voice notes as the most identifying data you can send, because they are. Skip face photos and documents entirely.
- Remember that deletion means account deletion, requested through the proper path β not archiving a chat, and never just removing the app.
- Re-read the policy when something big changes: an acquisition, a new model, a feature overhaul. Policy updates ride along with product updates, and consent boxes are easiest to click through at exactly those moments.
Two scenarios your policy read should survive
The acquisition. In most privacy policies, user data is an asset that transfers in a merger, acquisition or bankruptcy β the clause is usually one bland sentence filed under business transfers. In a young, consolidating category, that sentence deserves more attention than it gets: privacy commitments made by a small, principled team are only as durable as that team's ownership. When an app you use changes hands, re-read the policy that same week, and treat any quiet policy update in the following months as a prompt to re-decide, not just to re-consent.
The account takeover. Your chats are also only as private as your login. Companion accounts are a soft target β rich personal content, rarely protected the way people protect a bank account. Use a unique password from a password manager, turn on two-factor authentication where it's offered β support varies across the category β and look for an active-sessions or sign-out-everywhere control. No vendor policy can protect a conversation from someone holding your credentials, and no amount of corporate good faith substitutes for basic account hygiene on your side.
The bottom line
As of 2026-11-12, no vendor in this comparison offers an audited privacy guarantee, and the published detail varies. Replika's policy describes limited use of anonymized message portions, Character.AI says it retains shared data to train its models with an opt-out only for EEA and UK users, Nomi and Kindroid do not say whether chats train models in the pages we read, and Talkie's terms were not readable. All of it can change, and your own hygiene is worth more than any vendor's promise.
Privacy rarely decides which companion app people choose, but it should be the tiebreaker. See our companion app roundup for the wider picture, and our pricing guide for why a subscription you pay for has a clearer business model than a free, ad-supported one: when you're the customer, you're less likely to be the product.
Policies in this category change often. One email a month, no noise: join the free monthly digest.
Frequently asked questions
Are AI companion chats private?
Less private than they feel. The sources checked describe chats being retained on vendor servers (Kindroid, for example, keeps chats until you delete the AI or the account), and some policies permit using chats to improve models. Treat the chat window as a service, not a diary.
Do companion apps train AI models on my conversations?
It varies. Character.AI says it retains data you share to help train its models, with an opt-out toggle for EEA and UK users only. Replika says small anonymized portions of messages are used for safety algorithms and chatbot performance. Nomi and Kindroid do not state in the pages we read whether chats train models. Stated policies are not audited ones, so verify the current text yourself.
Can I delete everything an AI companion knows about me?
Usually you can delete your account. Nomi says full account deletion permanently deletes all data, and Replika says there is no way to delete your whole message history without deleting your account. Uninstalling the app deletes nothing at all.
Is it safe to send photos or voice notes to a companion app?
They are governed by the same policies as text but are far more identifying. We recommend against sending face photos, documents or anything you would not upload to a stranger's server, whatever the app's stated posture.
Which AI companion app is best for privacy?
No vendor publishes an audited guarantee, so we do not rank them. Nomi and Kindroid do not state in the pages we read that chats train models, but they do not state that chats are never used that way either. Compare the policies in the table above and combine the result with your own data hygiene.