Do AI companion apps train on your chats?

Usually yes. Most AI companion apps train on your chats by default, using conversation logs to improve their models or passing them to the third-party model provider behind the app. A growing number now offer an opt-out, but it is almost never on by default and is rarely mentioned during signup.

The bigger problem is not that training happens. It is that you often cannot tell whether it is happening to you. When Mozilla reviewed a batch of relationship chatbots, the recurring finding was not aggressive disclosure of training practices but the absence of any clear disclosure at all.

Why do AI companion apps train on your chats in the first place?

AI companion apps train on your chats because conversation logs are the cheapest available source of the exact data they need. A general-purpose model knows how to write an email. It does not natively know how people talk at 1am when they are lonely, how they phrase an apology to a character, or which replies make someone keep typing. Real chats teach that, and nothing else does as well.

There is a second, blunter reason: retention. Logs reveal which responses end a session and which extend it. Training on that signal makes the product stickier, which is a business goal rather than a quality goal, and the two are easy to confuse from the outside.

Which parts of a conversation actually become training data?

The parts of a conversation that become training data are broader than most people picture. Typically it is not only your messages but the full exchange: your text, the character’s replies, edits and regenerations, ratings or thumbs you gave, and metadata like session length and time of day.

Two categories surprise people. First, content you deleted from your view may already sit in a training set, because deletion in the interface and deletion from a data pipeline are different operations. Second, anything you told the app to remember about you, the profile details that make a companion feel continuous, is structured personal data and is often retained separately from the chat log itself.

A 2025 incident cited in an August 2026 Congressional Research Service report makes the stakes concrete: more than 300,000 Grok conversations were exposed, some containing names, passwords, and other personal details users had typed into what felt like a private box.

How do you opt out of AI companion training?

Opting out of AI companion training is possible on most major apps, but the control is usually three or four screens deep. Look under Settings for a section named Data, Privacy, or Data Controls, and for a toggle worded something like “improve the model for everyone” or “help improve our AI.” Turning it off is the opt-out. There is often no confirmation and no visible change to the product.

Three practical notes. Opting out is usually forward-looking only, so it does not retroactively pull past chats out of a model already trained on them. Some apps tie the toggle to a paid tier, which is one of the less obvious things what a subscription actually buys you. And where a companion app is a wrapper around a third-party model, you may need to check both the app’s policy and the underlying provider’s. For comparison, OpenAI documents its own training toggle publicly, which is more transparency than most companion apps offer.

Can you tell whether a specific AI companion app trains on your chats?

Telling whether a specific AI companion app trains on your chats means reading the privacy policy for two phrases: “improve our services” and “train our models.” The first is elastic enough to cover training without saying so. The second is the honest version.

Assume the answer is yes when the policy is vague, because vagueness is not an accident in this category. Mozilla’s review of 11 relationship chatbots found most policies said remarkably little about how conversation content feeds their AI, and only one app confirmed it does not share or sell personal data. Ten of the eleven also failed Mozilla’s minimum security standards, which is a separate and equally relevant signal about how carefully your chats are being held.

Worth separating two ideas that often get merged: your chats being read by a person, and your chats being absorbed into a model. Companion conversations are generally private from other users but not from the company, and training is one specific thing the company may do with that access.

Does the law require AI companion apps to tell you?

The law is starting to require AI companion apps to disclose more, but coverage is patchy. Data-protection regimes like the GDPR and California’s privacy law give users rights to know what is collected and to request deletion, and those rights apply to chat logs. What they do not do is force a plain-language notice at the moment you start talking to a character.

Bills introduced in the 119th Congress would add companion-specific rules, including disclosure requirements and limits on processing user data for chatbot outputs. As of September 2026 those are proposals, not law, so the practical protection available to you today is still the settings screen rather than a statute.

What is still unknown about AI companion training?

What remains unknown about AI companion training is most of the detail that would let you make a fully informed choice. Companies rarely publish which conversations enter a training run, how long logs are retained before or after an opt-out, or whether human reviewers read samples. Independent audits of companion apps are scarce, so outside evidence comes mostly from privacy reviews and breach disclosures rather than from verified internal practice.

The direction of travel is also unsettled. Disclosure norms have improved since 2024, and several apps that had no training toggle now have one, but improvement is uneven and a policy can change with an update you never read. The reasonable posture is neither alarm nor trust: check the setting, assume anything you type could persist, and keep the genuinely sensitive material out of the box. At Vinfluencer we build AI conversational companion software and treat that last point as a design constraint rather than a disclaimer.

Frequently asked questions

Does deleting my account delete my chats from the training data?
Deleting an account usually removes chats from your view and from active storage, but a model already trained on that text cannot unlearn it. Deletion requests under GDPR or CCPA cover stored records, not weights inside a trained model. Delete early if you intend to delete at all.

Do free AI companion apps train on chats more than paid ones?
Free tiers are more likely to train on your chats, since data is the compensation when money is not. Some apps make the training opt-out a paid feature outright. Paying does not automatically stop training, though, so check the setting on any tier.

Can a human at the company read my companion conversations?
Sometimes yes. Many privacy policies permit human review of samples for safety, abuse detection, and quality checks. That access is usually limited and logged, but it exists on most platforms and is separate from automated training.

Is anything I type genuinely off limits?
Nothing is off limits by default, which is why the practical rule is to withhold passwords, financial details, medical record numbers, and identifying information about other people. A companion app does not need those to hold a conversation.