Crash TestLver
Safety · 9 min read

What AI girlfriend apps do with your chats

Five services say your chats can train their AI, two say they don't, one sends mixed signals, and most don't say either way. Here is the full list, plus who can read your messages and how to delete them.

Updated

Of the 20 adult AI chat services I track, five say in their privacy policies that your chats can be used to train or improve their AI: Candy.ai, CrushOn.AI, DarLink AI, OurDream AI and Secrets AI. Only two say they don’t: Chub AI and, since a policy update on Nov 14, 2026, Promptchan. Dream Companion sends mixed signals: its homepage says it does not currently train on your chats, but its privacy policy updated Nov 2, 2026 allows anonymised data to be used for improving its AI models. The rest don’t say clearly either way, which in practice means you should assume your messages are stored and may be used to “improve the service”.

Everything here comes from official privacy policies, terms and help pages, checked Oct 25, 2026; Dream Companion, Promptchan and Swipey AI were re-checked Nov 15, 2026 after policy updates. This is a test drive: I read the policies, I did not sign up and inspect the data each service holds, and policies can change after the date I checked them.

Why this matters more for adult chat

A normal chatbot conversation is mostly recipes and code. An adult companion chat can hold your fantasies, your relationship history, your name, and whatever you typed at 2 a.m. That text sits on someone’s server. Three questions decide how exposed you are:

  1. Training. Can the company feed your chats into its AI models?
  2. Human access. Can staff, moderators or contractors read them?
  3. Deletion. Can you delete them yourself, and what stays behind?

There is also a fourth question nobody likes to ask: what happens if the company is breached? Muah AI is the cautionary example. According to Have I Been Pwned, its September 2024 breach exposed about 1.9 million email addresses together with users’ AI prompts. I don’t recommend Muah AI for that reason, and its privacy policy, effective 2022, does not specifically address chat content or AI training.

The full table

“Says yes” means the policy names training or developing AI models as a use of your data. “Says no” means it explicitly rules it out. “Not stated” means the policy neither confirms nor denies it. All checked Oct 25, 2026.

ServiceChats train AI?Staff can read chats?Delete yourself?
Candy.aiSays yesHuman review of de-identified chats possibleNo, by email
Chub AISays noNot without your written permissionNo process published
CrushOn.AISays yesNot statedNo, by email
DarLink AISays yes (anonymised)Not statedYes, in settings
Dream CompanionMixed: homepage says no, Nov 2026 policy allows anonymised data for model improvementOnly limited moderation or safety casesYes, Settings > Account
DreamGFNot statedOnly if you reported a problemNo, by email
GirlfriendGPTNot statedNot statedNo, on request
JanitorAINot verifiedNot verifiedYes, Settings > Security
KindroidNot statedFlagged excerpts, on request, or by lawYes, immediate for messages
Kupid AINot statedModerators check complianceNo, by email
LovescapeNot statedHelp centre says chats are visible only to youUnclear
Nectar AINot statedNot statedNo, via Discord or email
NomiNot stated (“training archives” mentioned)Not statedYes, Account Settings
OurDream AISays yesNot statedYes, Profile > danger zone
PromptchanSays no (policy of Nov 14, 2026)Not statedYes, in Settings
Secrets AISays yesNot statedYes, profile settings
SoulGenNot statedNot statedYes, Delete Account
SpicyChatNot stated (broad licence)Operators may read messagesYes, account settings
Swipey AINot statedFlagged chats may get manual reviewYes, Settings > Profile > Delete Profile

Muah AI is left out of the table on purpose. See above.

Kindroid screenshot
Kindroid · official site, blurred by us where needed

Services that say chats train their AI

These five are at least upfront about it.

Candy.ai lists “train and develop our AI models and moderation technologies” as a purpose in its privacy notice, last updated Oct 19, 2026. It also says this may include human review of de-identified or anonymised interactions, and that third-party language model providers or hosts “may receive the content of your messages”. So your chats leave EverAI’s own servers to be processed.

CrushOn.AI says chat-generated content is used “in a general way for training our models” without linking it to specific users. Its policy also says user content may be used for business purposes.

DarLink AI says chats and feedback may be analysed to improve its models, and that “we might train our AI on anonymized conversation data”. Its refund policy adds something unusual: after a chargeback, account records including private messages may be disclosed to the payment processor or bank. Dispute a charge and your chats could become evidence.

OurDream AI lists improving the services “including to train our artificial intelligence/machine learning models” in a privacy policy last modified Aug 26, 2026. The user content it collects includes chat communications.

Secrets AI lists “train and improve our AI models” under service improvement. It says text you submit in chats may be stored, and that voice calls are transcribed with the audio deleted immediately.

Being honest about training is better than silence. It still means your messages are raw material. In my paper check, a clear statement either way counts as “chat data explained”. It doesn’t count against a service that says yes. It counts against a service that won’t say.

Services that say they don’t

Chub AI has the strongest wording I found. Its privacy page says private characters and chats are accessed only programmatically to provide the service, may not be viewed by staff without the user’s written permission, and will not be used to train AI models. Its docs say the training data for its own Mars model does not include public or private Chub chats. For API use outside the Chub interface, it says inputs and outputs are not logged, only metadata such as length, model and timing.

The catch: Chub names no legal entity, and I found no published deletion process. A strong promise from a company you can’t identify is worth less than it sounds.

Dream Companion says on its homepage: “We do not currently use your chats to train our own AI models.” Note “currently” and “our own”. It also says staff don’t review private conversations except in limited moderation or safety cases, and it offers a delete button in settings. But its privacy policy, modified Nov 2, 2026, says anonymised or aggregated data may be used for “improving our AI models”. The homepage promise covers your chats as written; the policy leaves room for anonymised data. Checked Nov 15, 2026.

JanitorAI screenshot
JanitorAI · official site, blurred by us where needed

Services that don’t say

The majority. Here is what their policies say instead, which is sometimes more telling than the silence.

  • SpicyChat: the terms grant SpicyChat a perpetual, irrevocable, sublicensable licence to use your chatbots, chats and generations. They also say operators may read messages sent to the service, with employee access restricted by internal policy. Training isn’t named, but that licence would cover a lot.
  • Kindroid: chats are encrypted at rest and in transit, but not end-to-end. Aggregated, de-identified data may be used to build and improve the services. Support can’t browse chats.
  • Nomi: chat content is used to provide and improve the service. Training isn’t stated outright, but the policy refers to “training archives” whose data is no longer attributable to you after deletion. Read that as: something may stay in the archives.
  • GirlfriendGPT and Kupid AI: their policies say personal data may be used to adapt the services and “develop new tools”. No mention of training. Both keep data for up to six years after you stop being a user.
  • Swipey AI: no mention of training. Its privacy policy (dated 08/10/2026, checked Nov 15, 2026) says chat content is deleted or de-identified within 30 days of account closure; billing records are kept 6 years.
  • DreamGF: submitted content may be used to personalise and improve services and for aggregated analysis. Training and selling are not mentioned.
  • Lovescape: the privacy policy, last modified June 4, 2024, does not mention training. The terms grant a licence to use your content “to provide and improve our Services”. The help centre says chats are private and visible only to you.
  • Nectar AI and SoulGen: no mention of chats or training at all.
  • Promptchan: its privacy policy effective Nov 14, 2026 says it does not use chats or uploads to train AI models, “ours or anyone else’s”, and that it does not sell data. You can delete your account or download your data in Settings.
  • JanitorAI: its terms and privacy pages returned errors when we tried to read them, so I can’t tell you what they say.

If a policy doesn’t say, I assume the broadest reading. “Improve the services” can include model training, and nothing in these policies rules it out.

Who can read your chats

Training is a machine reading your messages. Staff review is a person doing it. The policies split three ways.

Limited or no human access, stated: Chub AI (only with written permission), Dream Companion (limited moderation and safety cases), DreamGF (only if you reported a problem), Kindroid (excerpts flagged by automated safety systems, on your request, or when legally required). Lovescape’s help centre says chats are visible only to you.

Human access allowed, stated: Candy.ai (human review of de-identified interactions), SpicyChat (operators may read messages), Kupid AI (moderators check compliance), Swipey AI (an automated filter may flag chats for manual review).

Not stated: everyone else.

On adult services, some moderation is unavoidable and probably good. Secrets AI, for example, replied to a Trustpilot reviewer that it had removed characters designed to appear underage, and that kind of enforcement needs someone to look. The question is whether the policy tells you when it happens. Kindroid’s version, flagged excerpts only, is the clearest.

Deleting your chats

This is where policies get practical. A delete button you can press tonight is very different from an email you send and hope gets answered.

Self-serve deletion

ServiceWhereWhat happens
Secrets AIProfile settings > Delete AccountPersonal info and chat history deleted immediately; transactions kept 7 years, security logs up to 12 months
KindroidIn appMessage deletion immediate; account or Kindroid deletion 24 hours after request; limited records kept to prevent ban evasion
OurDream AIProfile > danger zoneProcessed in 24–48 hours; deletes generations, unpublished characters, chat history and unused coins
Dream CompanionSettings > Account > Delete AccountRetention “as long as necessary”
JanitorAISettings > Security > Delete accountConfirmed by email code; posts, characters, personas and profile removed; chats not explicitly mentioned
NomiAccount SettingsPersonal info removed in about 28 days, except training archives and legal holds
SpicyChatAccount settings, or supportDeleted accounts cannot be restored
DarLink AISettings or supportBackups and required records may remain; deleting forfeits refunds
SoulGenDelete AccountRemoves all app data including purchase records; unposted creations auto-deleted after 7 days

Deletion by request

Candy.ai, CrushOn.AI, DreamGF, GirlfriendGPT, Kupid AI and Nectar AI describe no self-serve deletion in their policies. You email or message support and wait. Some of the retention periods are long:

  • Candy.ai: account data kept up to 3 years after last activity, transaction data 10 years. EverAI aims to respond to requests within one month.
  • GirlfriendGPT: data usually kept 6 years after the account is closed.
  • Kupid AI: data may be kept 6 years after you stop being a user.
  • DreamGF: erasure by email for EEA, Swiss and UK residents, with a 30-day response target.
  • Nectar AI: deletion via Discord or support email.

CrushOn.AI has one accidental privacy feature: free-tier chat history is deleted automatically after 7 days of inactivity. Chub AI publishes no deletion process at all. Lovescape says you may delete your account at any time but doesn’t say how, beyond an erasure address.

What I’d do before typing anything personal

  • Read the training line. If it says yes, decide whether you’re fine with that. If it says nothing, assume yes.
  • Use a separate email. A breach exposes your email next to your prompts, as Muah AI showed.
  • Don’t type real names, places or photos of anyone. Not yours, not an ex’s. That is also the rule on most of these services; see AI chat apps that ban deepfakes.
  • Prefer a delete button. Secrets AI, Kindroid, OurDream AI and Dream Companion describe the most complete self-serve deletion.
  • Don’t count on deletion being total. Almost every policy keeps transaction and legal records, and some keep “archives”.

Privacy is one of the ten checks in my paper check. For the shortlist that passes the content rules, see best AI girlfriend apps and best NSFW AI chat. If price is your next question, start with what AI girlfriend apps really cost.

FAQ

Do AI girlfriend apps use your chats for training?

Some say so openly. As of Oct 25, 2026, Candy.ai, CrushOn.AI, DarLink AI, OurDream AI and Secrets AI list AI training or model improvement as a use of your data. Chub AI and Promptchan (since Nov 14, 2026) say they don't. Dream Companion's homepage says it doesn't, but its Nov 2026 policy allows anonymised data for model improvement. Most others don't say either way.

Can staff read my AI girlfriend chats?

On some services, yes. SpicyChat's terms say operators may read messages, Candy.ai mentions human review of de-identified interactions, and Swipey AI says flagged chats may get manual review. Chub AI says staff can't view private chats without your written permission. Kindroid limits review to flagged excerpts, user requests and legal requirements.

How do I delete my chats from Candy.ai?

Candy.ai's privacy notice describes deletion by email to its support or data protection addresses, or by post to EverAI in Malta. No self-serve deletion is described. It aims to answer within one month. See the Candy.ai test drive for details.

Which AI girlfriend app is most private?

On paper, Promptchan now gives the clearest answers: no training on chats or uploads, no data sales, and deletion and data download in settings (policy of Nov 14, 2026). Dream Companion promises no training on chats "currently" and has a delete button, though its policy allows anonymised data for model improvement. Chub AI has a strong no-training promise but names no company and publishes no deletion process. Neither claim has been tested by me.

Are AI girlfriend chats end-to-end encrypted?

None of the policies in my research files claim end-to-end encryption. Kindroid says its chats are encrypted at rest and in transit and specifically says they are not end-to-end encrypted.

Crash Test LoverThe pen name of our tester. Runs every billing and cancellation check on this site.

Keep reading