Do your conversations train the model?

Three questions that get treated as one

Does the text go into training? A yes or no about one specific use. It has nothing to do with whether a human ever reads the text, and nothing to do with how long it is stored.

How long is it kept? A retention question. A service can train on nothing at all and still hold your messages for years, and a service can train on everything and delete the raw logs in a month.

What disappears when you delete the account? The one people actually care about, and the one least often answered. Deleting an account and deleting the data behind it are two different operations, and only some providers say whether the second follows from the first.

Keeping them apart is what makes the five comparable. Merged together, every answer sounds the same.

What the five say, in their own words

Each row is taken from the document named in the first column, read on 20 September 2026.

The pattern in that table is not a privacy ranking. It is a documentation ranking: some of these companies wrote a policy about their product, and some installed one.

The one that says it plainly

Candy AI's notice is the most explicit document in the set, and reading it is the fastest way to understand what training actually means in practice.

It lists training as a purpose in its own right rather than folding it into service improvement. It says the datasets prepared for that training may include human review. And it names the legal basis: a legitimate interest in providing the best Service possible and improving our Services.

That last detail matters more than it looks. Consent and legitimate interest are not interchangeable. A purpose built on consent stops when you withdraw consent. A purpose built on legitimate interest keeps running until someone objects to it and the objection is upheld, which is a different process with a different outcome.

The same notice is the only one of the five that publishes a number for logs: 30 days, for files containing users' interactions with the platform, user-IDs, and timestamps. That is a short window by the standards of this category, and it is verifiable in a way that as long as necessary never is.

The ones that say nothing

Silence in a privacy policy is not the same as a denial, and on SpicyChat the two documents point in different directions.

The privacy policy, still dated 4 May 2023, never mentions AI training. The terms of service, updated 23 July 2026, grant the company permission over your chatbot and the chats from it that is perpetual and irrevocable. Neither document says that permission is used for training. Neither says it is not. A permission that broad does not require a separate training clause to cover training.

The same policy contains a detail that says something about how carefully it was assembled: the company is defined as SpicyChat.AI, Corporate Address, with the template placeholder left in, and the country of operation is given as Canada with no entity named.

Infatuated AI is the other silence, and it starts earlier. There is no privacy policy at the address where one normally lives; the document sits at a different path entirely, which is why our review of Infatuated AI previously recorded that none was published. It exists, it is dated 13 September 2024, and it says nothing about training.

It also still describes the company as certified under the EU-US Privacy Shield framework. The Court of Justice of the European Union invalidated that framework on 16 July 2020 in case C-311/18. The policy hedges the point in the next paragraph, which means the contradiction was known and left in.

How long the text lives

Four of the five answer this with a version of the same sentence: we keep it as long as we need it. That is legally standard and practically useless, because there is no way to check it from outside and no date on which anything is guaranteed to be gone.

Candy AI is the exception, with four published periods. Whether three years of account retention is reasonable is a separate argument; at least it is an argument you can have with a number in front of you.

OurDream publishes one retention rule the others do not have, and it is specific enough to plan around. If a public character you created becomes popular enough, the company reserves the right to preserve that Character's characteristics and to keep that Character active on the Services, even if you otherwise delete your data and your account. Its own advice is to keep identifying details out of the character creation form for that reason. Nothing in the clause touches your chats, but it is the one case in the whole set where something you made outlives the account you made it on.

What deleting the account actually deletes

Two of the five let you do it yourself from settings. One requires an email. All five reserve the right to keep some records regardless, which is normal and mostly about tax and legal obligations.

The operational details are where they diverge, and they are the kind you only discover at the moment you want out:

• CrushOn AI will not delete an account that has an active subscription. Cancel first, then delete, in that order. There is also a three-day window in which logging back in undoes the deletion, which is a safety net if you change your mind and a trap if you were counting on it being final.
• Infatuated AI takes up to 90 days and only by email. A different section of the same policy promises a response within 30 days, so the two numbers describe different things and the document does not say which applies to what.
• None of the five says what happens to the chat history specifically. Deleting the account is described as deleting account data. Whether the text of your conversations is inside that phrase is not stated by any of them.

That last point is the honest headline of this section. The question people actually want answered is the one nobody in this category answers in writing.

What they know after the first conversation

OurDream publishes the only plain-language summary of this in the set, and it is short enough to be useful: account information including email and age confirmation, payment details, the characters, chats, images and videos you create, usage telemetry, and device information including browser, operating system and approximate region.

The same page answers a question the others leave open. On who can see your private chats, it lists third-party AI providers with a yes: the message you send is processed by the model provider to generate the reply, governed by their data-processing agreement. Your conversation reaches a second company by design, not by accident, and that is true of every product in this category that does not run its own models.

Candy AI's notice adds the piece most people would not think to look for. Alongside your content it records metadata regarding your Content — time and date sent, and originating IP address. The conversation is one record; when and from where you had it is another.

SpicyChat's collection list is the shortest, and it has a wrinkle: the four social logins it supports are Google, Discord, Twitter and LinkedIn. Signing in through any of them hands over whatever name and address are attached to that account, which is the opposite of what most people choose a social login for.

A free account and one conversation is enough to create all of this. There is no threshold you cross later.

What none of them publish

Worth stating plainly, because an absence is a finding:

• No opt-out from training. Not on any of the three that say they train. The only opt-outs in any of these documents concern marketing email, cookies and analytics.
• No retention period for chat text. Candy AI's 30 days covers log files, not conversations.
• No statement that deleting an account removes your messages from data already used for training. Once text is in a training set, no policy here describes a route out of it.

We do not know what happens inside these companies, and neither does anyone outside them. What is checkable is what they committed to in writing, which is why this page quotes the documents rather than grading the companies. The same rule governs the rest of the site and is written out on our methodology page: we use the apps ourselves, then check every claim against the provider's own pages, and say so when a claim cannot be checked at all.

What to do with this

Nothing here requires technical skill. It requires doing it before rather than after.

1. Use an address that is not your name. The single most effective step, and the only one that still works after a breach. What leaked in the Muah AI incident was email addresses attached to prompts, which is only identifying if the address identifies you.
2. Skip the social login. SpicyChat accepts four of them; each one hands over a real name you did not have to give.
3. Check the date on the policy before you trust it. A document last touched in 2023 or 2024 predates most of what these products now do.
4. Cancel before you delete. On CrushOn AI the deletion button does not work while a subscription is running.
5. Ask in writing, and keep the reply. Every one of these companies names an address for data requests. A dated email is the only evidence you will have that you asked.
6. Assume the text is permanent. Not because these companies are dishonest, but because none of them has written down a commitment that it is not.

We hold your data to the same standard we are applying to theirs, and our own privacy page says exactly what this site collects and how to switch it off. It is a much shorter document, for the straightforward reason that we do not run a model.

The reviews carry the rest: what Candy AI and OurDream charge, what their free tiers actually allow, and what their terms say about the money rather than the data. If you want the shortlist first, the ranking is on the front page.