Are your chats training the model?

Privacy 2026-09-20 10 min

The three things to look for

1. Model improvement language

Look for phrases such as “improve our models”, “train machine-learning systems”, “enhance service quality” or “develop new features”. These often mean user interactions may inform model development in some form.

2. Human review and safety review

Some services allow human review for moderation, support or safety investigations. That is a separate issue from model training, but it matters just as much for privacy expectations.

3. Region-specific controls

A company may describe stronger opt-out or privacy rights for some regions than for others. That means a Malaysian user should not assume they receive the same control described for an EEA user unless the policy says so.

How the main apps differ

App	Broad privacy impression	Main caution
Replika	Better documented than many rivals	Strong privacy-enforcement history means the category risk is real
Character.AI	Explicit about model-improvement logic	Public controls may differ by region
SpicyChat AI	Simpler value proposition	Users should still assume cloud-service risk
CrushOn AI	Adult-first use case raises sensitivity	Corporate transparency is weak
PolyBuzz	Free and easy to enter	Premium and operator disclosure are not strong enough

The table is not a legal ranking. It is a reminder that “safe” and “private” should not be read from homepage tone alone.

The practical rule

The safest working assumption is this: if the policy leaves room for service improvement, moderation or training-related processing, do not type anything that would seriously matter if stored, reviewed or leaked.

This does not mean the app is malicious. It means AI companion chats are unusually intimate and therefore deserve a higher privacy standard from the user too.

Questions people ask

Are my chats private?

Private compared with a public social feed, yes. Private in the sense of a sealed diary, not necessarily.

Do humans read chats?

Sometimes policies allow human review for support, safety or moderation reasons.

Are chats used for training?

Often the answer is “possibly for improvement” rather than a clean no.

Which app should I trust most?

Trust should be based on clear policies, operator transparency and restrained data sharing habits, not on brand tone alone.