Do AI Companions Train on Your Messages? Four Meanings
Does AI train on chats? Separate foundation-model training, provider improvement, companion evaluation, and runtime memory before reading any privacy promise.
When someone asks, “Does AI train on chats?” a yes-or-no answer is usually incomplete. Training can refer to building a general foundation model, improving a model provider’s systems, evaluating one companion product, or simply using stored memory to personalize your next reply. Those activities are not interchangeable.
The only reliable answer for a specific service comes from its current privacy policy, settings, and provider contracts. Here is how to read them.
Meaning 1: training a general foundation model
A foundation model learns broad statistical patterns from a large training dataset. If a company uses consumer conversations for this purpose, material from many chats may influence later versions of a model used across products and customers.
This is what many people mean by “my chats train the AI.” But policies vary by service and account type. OpenAI’s current data-use policy, for example, says content from individual services may be used to train models unless the user opts out, while business products and the API are excluded by default unless an organization opts in. That describes OpenAI’s products, not every app built with an AI model.
If a companion company uses an external model through a business API, the model provider’s consumer-chat defaults may not apply. Read the companion’s policy and the provider terms governing that business relationship.
Meaning 2: a model provider improving its own systems
A companion may send messages to a third-party model provider to generate replies. A separate question is whether that provider may retain or use the inputs and outputs to improve its models.
Look for language such as:
- “used to train our models”;
- “used to improve our services”;
- “not used for model training by default”;
- “may be retained for abuse monitoring”;
- “customer may opt in to data sharing”;
- “processor acting only on our instructions.”
The word provider matters. “We do not train models” from the companion company does not automatically answer what its vendors may do. Conversely, using a model provider does not automatically mean that provider trains on the content. Contract and product tier determine the answer.
Meaning 3: evaluating and improving the companion product
A company can promise not to train a publicly available foundation model and still use some conversations to improve its own companion.
Product improvement may include:
- measuring whether responses stay in character;
- evaluating memory accuracy;
- finding safety failures;
- reviewing a support complaint;
- testing a revised system instruction;
- labeling examples of good or poor replies;
- calculating aggregate quality metrics.
Some of these activities involve human review. Others can be automated. Some create training examples for a narrow classifier or product component without retraining the underlying foundation model.
This use may be reasonable, optional, necessary for safety, or too broad for your comfort. The point is to identify it accurately. “No foundation-model training” does not mean “conversation content is never used for quality work.”
Meaning 4: runtime memory and personalization
When a companion remembers that your sister is visiting, it may retrieve a stored memory and place it into the context for the current response. That is inference-time personalization, not necessarily model training.
The model’s general parameters do not have to change. The product is supplying information about you at runtime.
This distinction does not make memory non-sensitive. A stored inference can still be personal data, can be wrong, and needs correction, deletion, access, and retention controls. Our guide to AI memory versus a context window explains the architecture.
A fifth phrase to watch: “improve our services”
This broad phrase can cover multiple activities. Do not guess. Look for definitions, examples, settings, retention periods, and service-specific terms.
Google’s Gemini Apps Privacy Hub, for example, describes how settings affect activity, model improvement, human review, and retention. It also distinguishes temporary chats and safety processing. The details are specific to Gemini and can change, which is exactly why summaries should link to the live source.
A useful policy answers:
- Which product or model is being improved?
- Is the use automatic, opt-in, or opt-out?
- Does it apply to all accounts or only consumer accounts?
- Are conversations sampled, minimized, or de-identified first?
- Can a person review them?
- How long are source and derived records kept?
- What happens after you change the setting or delete data?
“We do not sell your data” is a different promise
Sale, advertising, model training, product improvement, service delivery, safety review, and legal disclosure are separate purposes.
A company can avoid selling personal information while still processing it to provide and improve the service. Another company could avoid model training but use data for targeted advertising. Read each purpose rather than treating one reassuring sentence as a complete privacy policy.
The same applies to de-identification. Removing direct identifiers can reduce risk, but conversation details can still be distinctive. Ask what “de-identified” means, whether re-identification is prohibited, and whether the original copy remains.
How Warmth answers the training question
Warmth’s Privacy Policy says:
- Warmth does not sell conversations;
- Warmth does not use conversation content to train publicly available foundation models;
- its model providers are contractually prohibited from using the content to train their own models;
- Warmth may use conversations to operate, evaluate, and improve Mia, including reviewing a small number for safety and quality;
- Warmth minimizes and de-identifies data where possible.
The policy also describes specific human-review triggers and the other provider categories involved. That full text controls over this summary.
Mia’s memory is separate again: Warmth extracts details and inferences so future conversations can refer back to them. Users can request correction or deletion. Mia is AI, available only to adults, and is companionship and entertainment—not therapy, emergency response, or professional care.
A policy-reading template
Search the policy for these terms:
train, improve, model, provider, processor, review, memory, inference, retain, delete, de-identify, and opt out.
Then write four separate answers:
| Question | Answer to find |
|---|---|
| Does the companion company train a general model on chats? | Yes, no, or conditional |
| Can its model provider train or improve models on chats? | Contract and account tier |
| Can chats be used to evaluate or improve this product? | Purpose, scope, people, controls |
| Does the product store memories for future replies? | What, how long, correction, deletion |
If the policy collapses all four into “we value your privacy,” it has not answered the question. If it distinguishes them, you can decide which processing you accept instead of relying on a slogan.