You type something into a companion app that you would not say out loud, and a question forms behind it: who reads your AI girlfriend chats, apart from you and the character? The apps do answer this — in their privacy policies rather than on the pricing page, in paragraphs written for regulators. Here is what four of the apps on our ranking commit to, read from their own policies this month.
There are three readers, not one
Strip out the legal phrasing and the same three possibilities appear in every policy: the model itself, which may not belong to the app; a training process that turns your conversations into material for the next version; and a person — a moderator, under some condition the policy may or may not specify.
None of this is sinister on its own. Storage is not optional for an app sold on memory, which is why the apps built around memory keep the most. The question is not whether your conversations are stored, but how far they travel afterwards, and how much of that journey the policy will describe.
Your messages may not stay with the app
The clearest statement comes from Candy AI, whose policy says it may share data with third-party large language model providers and hosters at its discretion, and adds the sentence most policies leave out: those third parties may receive the content of your messages. Many companion apps are a character layer wrapped around a model somebody else runs, and when that is the architecture, your conversations leave the company you signed up with.
Compare the alternatives. Kupid AI and GoLove AI both describe sharing with service providers they work with to deliver the services, and never mention model providers. DreamGF lists infrastructure, IT, data storage, accounting and legal suppliers, and says they may process data only on its behalf and under its instructions. Nomi AI says it will not sell or rent your personal information to any third party, and notes that servers outside your country may be provided by third parties it does not name.
Notice what none of those is: a promise that your messages stay in-house. "Service providers" is broad enough to include a model vendor. An app that names the practice has told you something; a silent one has told you nothing.
Who reads your AI girlfriend chats to train the next model
Candy AI is again the most explicit. Its policy lists training and developing its AI models and moderation technologies among its purposes, says interactions with its characters may be aggregated, anonymised or de-identified for that work, and adds that preparing those datasets may include human review of the de-identified material. There is no training toggle. What it offers instead is the general right to object to processing based on legitimate interests.
Nomi AI never says it trains on chats, but one sentence gives it away. Describing deletion, its policy says information in its training archives would afterwards no longer be attributable to you. Training archives exist, then, and deletion changes what they are labelled rather than whether they are there. Kupid AI, GoLove AI and DreamGF do not mention training either; what they claim is the right to draw on your data to adapt the service, develop new tools, and run analysis and statistics.
The absence of the word "training" is not an undertaking not to train. "Improving the service" is broad enough to cover it, and it appears in all five.
The moderator, and the question of when
Human review is the part readers assume is happening and rarely check the terms for. The useful distinction is not whether a person can read your conversations, but what has to happen first.
DreamGF draws the narrowest line. Its policy says it has the right to review the chat correspondence you conduct with the AI only if you have reported a problem through its support centre: a specific trigger, and one you control. Candy AI describes human review of content and users flagged by its moderation controls or reported by you — two triggers, one of them automated and therefore not in your hands. Kupid AI and GoLove AI use the same sentence as each other — a moderation team checks you are complying with the law — with no trigger, frequency or scope.
A clause with a trigger in it is a limit. A clause without one is a permission, and reads the same on the page.
Deleting the account does not empty the archive
This is where the policies diverge most. Deletion is presented as one switch and is closer to several clocks running at different speeds.
Candy AI sets account data at three years after your last activity and log files at thirty days, with no separate period for conversation content. Kupid AI and GoLove AI both say they usually hold your personal information for six years after you stop being a user — a long time, but a legible number that typically reflects commercial record-keeping rather than chat. Nomi AI's deletion removes your personal information, with the training and communications archives carved out of it. DreamGF commits only to keeping data as long as necessary, reviewed regularly.
The lever here is a written request rather than a button. Where an app relies on legitimate interests — the basis Candy AI points to for training — UK and EU readers can use the right to object: the ICO's guidance explains that you can ask an organisation to stop using your personal information, naming the use and the reason, with a month for a reply. Our guide to what cancelling does and does not do covers the billing side, a separate request again.
What this changes about which app to pick
Three of these are worth singling out, for different reasons, and none is the right answer for everyone.
- Candy AI has the most expansive data use of the four and describes it the most plainly, which is useful: you can tell what you are agreeing to. It allows adult conversations, generates images and video, and at $3.99 a month billed annually as of October 2026 it is the cheapest here. There is no free plan, only a trial. Its Visit has the current tiers and its card the rest.
- DreamGF has the tightest human-review clause read here, tied to a report you make yourself. The trade-off is in the product, not the terms: image generation is limited, there is no video, and it runs on a trial rather than a free plan, from $5.99 a month in the first year as of October 2026. Its Visit shows what follows that year.
- Nomi AI is the one that admits an archive survives deletion — a mark against it on the face of it, and better understood as a disclosure most competitors omit. It is the dearest of the three, from $8.33 a month billed annually as of October 2026, its free tier is narrow and it produces no video. Its Visit has the plans.
If the data question is deciding this, read the policy before the pricing page, retention and sharing first. Our privacy checklist is the ten-minute version to run beforehand, and the ranking shows what each app costs on its card. The free plans list narrows it to the ones you can try without handing over a card.
Frequently asked questions
Do AI girlfriend apps use my chats to train their AI?
Some say so directly: Candy AI lists training its models among its purposes. Most say nothing about training and claim a broad right to improve the service instead, which is wide enough to cover it. Treat silence as unresolved rather than as a no.
Who reads your AI girlfriend chats at the company itself?
A moderator can, at every app checked here. What varies is the trigger: DreamGF ties review to a problem you report, Candy AI adds content its automated moderation flags, and Kupid AI and GoLove AI describe a moderation team with no stated trigger.
Does deleting my account remove my chats from the training data?
Usually not. Nomi AI's policy says deletion makes archived information no longer attributable to you, which is de-identification rather than removal. If you want more, object in writing and keep the reply.
Disclosure. NectarHush may earn a commission if you sign up through a Visit link on this page, at no cost to you. It does not change what we write. How we earn.