AI companion platforms send your message to a service, where a language model predicts a likely reply; Google for Developers’ introduction to large language models explains that the model estimates likely pieces of text within a longer passage. A service may then store the chat: Aiamis’s privacy page says its messages and replies are kept in a database and deleted within 24 hours of the conversation starting. Some personalization settings can do more than keep the conversation record; Microsoft’s Copilot privacy FAQ says they can remember details such as your name, interests and goals.
What happens after you send a message?
At Aiamis, the service sends your current message and earlier messages from the same conversation to the company running the AI model. That company processes the text on its own servers under its own terms, as explained in the Aiamis privacy page.
The conversation so far is also kept in your own browser tab until you close the tab or press “New chat.” The stored service record includes a random conversation ID, the time, the page address, the language and a two-letter country code, according to the same Aiamis privacy page.
Conversations are not linked to an identity. Maintainers have technical access for no longer than the deletion window, although the page says conversations are not forwarded to staff or reviewed as they happen.
In plain terms, the model makes the reply, the service keeps a temporary record, and your browser may hold the visible chat. Those are separate storage points to consider when you care about privacy.
What can a service collect?
One company’s published policy gives you a concrete example. Replika’s privacy policy lists “messages and content” among the data it collects. That includes messages you send and receive, facts you provide about your life, and any photos, videos, and voice or text messages you provide.
Sensitive information included in a message is processed as part of the conversation data, including by outside companies that run language models. The stated purpose is to generate conversational responses and provide the app’s core features, according to the same Replika privacy policy.
That leaves you with two useful questions. What does a service collect, and where does that information go? A friendly tone does not answer either question. Check the policy and settings for the service you use before deciding what to share.
Why can a chat feel personal?
The eSafety Commissioner says companion chatbots adapt to what users type and learn to respond in ways that feel personal and realistic. That can make the exchange seem tailored to your mood and conversational style.
Its article also warns that companion chatbots are often designed to encourage ongoing interaction. This can feel addictive and may lead to overuse or dependence. The article was written about children and young people.
For you, the practical check is simple: notice whether the conversation is crowding out people or plans you had chosen. If it is, pause the chat and make contact with someone you know.
Personal-feeling replies and remembered details are also separate features. One concerns how a chatbot responds to your wording. The other concerns whether a service saves information that it can use later.
Why can a reply be wrong?
The US National Institute of Standards and Technology says plausible but inaccurate statements can arise naturally from this design because a language model predicts the next word or phrase.
Prediction can produce smooth wording that still contains an error. The sentence may sound certain because it fits the pattern, rather than because every part has been checked. You can treat the reply as conversation, but check important factual details against a reliable source.
What should you check before sharing?
Look for clear answers to a few practical questions:
- What does the privacy policy say it collects? Messages may contain more than you expect, including personal facts, images, video or voice content.
- Does the service send the current message and earlier chat text to a company running the language model?
- Is personalization turned on? If it is, what details does the service remember?
- How long are conversations kept, and can you delete them?
- Are there usage limits or costs? Check the current terms rather than assuming the service is free or unlimited.
Do not share anything you would be uncomfortable having another service process. The Aiamis privacy page gives that same practical warning, and its advice is useful beyond any one platform.
Keep the basic picture simple
The language model supplies likely text, while the surrounding service manages data, memory and interaction design. That is why a reply can feel warm without telling you much about what will be stored or remembered later. Go at your pace, share only what feels comfortable, and check important details elsewhere.
A chat like this is company in words, and that is all it is. If you would rather talk to a person right now, the Aiamis crisis help page lists services run by trained people you can reach.