Which AI companion apps train their models on your chats: 17 privacy policies read on October 29

Of the 17 companion apps with profiles on CrushRated, five say in their own policies that your chats help train or develop their AI, and none of the five states a switch to stop it: Candy.ai, Chai, CrushOn.AI, EVA AI and ourdream.ai. Two more, Character.AI and Grok, train by default and offer an opt-out; Character.AI’s written steps are for EEA and UK users. The other ten say something narrower or nothing clear about chats, and Anima is among them: its policy says it trains on “data” without naming chats.

Every policy below was read on October 29, 2026, with the version date the page showed, and read again on October 31. On the second read two pages carried newer dates, EVA AI (October 30) and Kindroid’s privacy policy (effective October 29); the wording quoted from both is unchanged, and the table shows the newer dates. Quotes are short and verbatim.

Where each policy stands

App	What the policy says about chats	Switch to opt out	Version read
Candy.ai	“train and develop our AI models and moderation technologies”; may include human review of de-identified interactions	None stated	Oct 19, 2026
Chai	“enhance and fine tune the artificial intelligence models” after personal identifiers are removed	None stated	Oct 12, 2026
CrushOn.AI	“We may use User Content to train AI models”	None stated	Mar 6, 2025
EVA AI	Text and voice messages used “to develop its AI”	None stated	Oct 30, 2026
ourdream.ai	Information used “to train our artificial intelligence/machine learning models”	None stated	Aug 26, 2026
Anima	“including to train and improve our AI models”, from “data”; chats are not named	None stated	Jan 9, 2025
Character.AI	“we retain certain data that you share with us to help train and improve our AI models”	Yes, “Improve the Model for Everyone”; steps written for EEA and UK users, Help Center says “at any time”	Jul 1, 2026
Grok	You choose whether your content trains the models, when logged in	Yes, in settings	Oct 27, 2026
Replika	“small portions of Messages and Content data” train its safety algorithms	None found	Oct 28, 2026
Nectar AI	Chats “are not used to train external AI models”; silent on its own models	Not applicable	FAQ undated
Kindroid	Never uses the word “train”; de-identified content usable “for any purpose”	None stated	Oct 29, 2026
Nomi	Mentions “training archives”; does not say chats train models	None stated	Apr 27, 2026
Janitor AI	“research and development to improve our AI systems”	None stated	Sep 17, 2026
Paradot	Your companion may “learn from your interactions”	None stated	Apr 21, 2025
Talkie	Near-identical sentence to Paradot’s	None stated	12/09/2025
PolyBuzz	Photos and precise location “to train our AI algorithm”; chat text not addressed	None stated	Oct 27, 2026
Secret Desires	Two documents that contradict each other	None stated	Oct 24, 2024 and Mar 9, 2026

Five apps that say yes and offer no switch

Candy.ai is the one that should worry you most in this group. Its notice says the training work includes “human review of de-identified and/or anonymized interactions with AI Companions”, so a person may read prepared chat datasets. Separately, it lists “third-party LLM providers and/or hosters” among its service providers and warns they “may receive the content of your messages exchanged with our chatbot”. If the company is sold, your data “may be among the assets transferred”, though the notice adds that it “will always remain subject to this Privacy Notice, as updated”. The notice names legitimate interest as the legal basis for training and lists a right to object, exercised by contacting the privacy team.

ourdream.ai and CrushOn.AI are blunt. ourdream’s policy puts training inside one clause about analyzing and improving the Services (“including to train our artificial intelligence/machine learning models”), so there is no separate section to read or object to. CrushOn’s page says chat content is used “in a general way for training our models” and is not linked “to specific users”. The sentence promises that chats are not tied to your account, and it confirms they are used. Its policy was last modified in March 2025, the oldest of the five.

Chai is the most explicit and the most usable of the five. The policy says conversation data goes into model improvement after personal identifiers are removed, and that all users, anywhere, can delete their account from the app (Profile, Settings, Delete Account). Retention is capped, and the cap is long: it keeps data no longer than necessary and, in any case, “not exceeding” the duration of the account and five years after deletion. EU and UK users get a stated right to object to processing that rests on legitimate interest, and Chai uses that basis for model improvement.

EVA AI is the only one of the five whose training wording names voice. Its table lists “Text and voice messages sent to a virtual friend” with the purpose “To improve the quality of text and voice messages of the virtual friend and to develop its AI.” If you talk to your companion by voice, that wording covers it.

Character.AI and Grok: opt-out, with a catch on each

Character.AI’s model-training page opens by saying it retains data “to help train and improve our AI models”. The opt-out steps on that page and in the Help Center FAQ are written for users in the EEA and the UK. A July 1, 2026 Help Center notice says you can opt out of AI training “at any time”, then sends you to that same FAQ. The pages do not say whether the setting appears for accounts elsewhere. The path on the app is Settings, Data & Privacy; on the web, Settings, Account, Manage Account & Data, Model Improvement.

Two limits apply after you opt out. The page says only “new content” stops being used, and the company “still may use your data” for “search, recommendations, and safety classifiers”. Chats already used are not addressed. With the company in Chapter 11 and a group of state attorneys general pressing for protections on chat histories, what happens to stored chats in a sale is an open court question; what is known so far is in our post on the filing.

Grok’s terms are the cleanest on the switch and the harshest on the exception. Logged in, “you can select whether or not you want us to use your User Content to improve our products and services and train our models.” Logged out, where the service allows it, “you grant us full rights to use any data you provide” for product development and model training. Private Chat conversations are deleted from its systems within 30 days, per the privacy policy.

Quiet policies: what the silence leaves open

Anima sits at the edge of this group. Its policy section 3.11, titled “To improve our Services, for example, to train the model”, says “We use data to enhance and optimize our Services, including to train and improve our AI models.” The policy never uses the words chat, message or conversation, so whether your chats count as “data” is not stated.

Kindroid is the one to watch among the quiet ones. Its Terms give it a license to “de-identify and/or aggregate your Private User Content” and “freely use such de-identified and/or aggregated content for any purpose, without acknowledgement or compensation to you”. The privacy policy adds that Kindroid is “not end-to-end encrypted” and that automated safety systems review chat content. The text never says chats train a model and never rules it out, and “for any purpose” has no stated limit.

Nomi’s policy never says whether chats train models, but it exempts “information that is in our training or communications archives” from account deletion, and says anything in those archives “would no longer be attributable to you”. Janitor AI lists “Conducting research and development to improve our AI systems” among its uses of data. Paradot and Talkie use a near-identical sentence about letting your companion “learn from your interactions to improve your conversations”, and both say conversation content is never used for marketing or advertising. That limit is about ads, not about training.

PolyBuzz names training only for photos and “precise geolocation”, and says it shares data with third parties “including the Chat-GPT platform”. It does not say what happens to chat text. Replika is a partial case: its policy says “small portions of Messages and Content data” are anonymized and used “to train our proprietary safety algorithms, enhance chatbot performance”, and that the data “is not used to train third-party large language models”. That is a narrower claim than the five above, and the policy was updated October 28.

Secret Desires is the clearest conflict of wording. The privacy policy (October 24, 2024) lists uses including “to train our artificial intelligence/machine learning models”. The Trust Hub page (March 9, 2026) says “Your conversations are never used to train AI models.” The pages read do not say which one governs. Nectar AI’s FAQ says chats “are not used to train external AI models”, which says nothing either way about Nectar’s own.

How we sorted them

CrushRated’s profiles store this as a field, “Used to train models”, with a source and a check date. As of November 2, the field reads yes for six apps, “yes unless you opt out” for two and unclear for nine; the sixth yes is Replika (small anonymized portions only), which the count above puts among the narrower ten. For the method behind the profiles, see the methodology page; for the rules that apply to these apps by country, see the laws page. Related coverage: four companion apps leaving the iPhone App Store and Kindroid ending its free Lite plan.