Does AI “Learn” From Your Conversations?
4 min read · updated August 3, 2026
“It remembered what I told it last week” and “my conversations are training the model” sound like the same observation. They are produced by different mechanisms, and only one of them is irreversible.
The weights do not change while you talk
Start with the mechanical fact, because it rules out the most common version of the worry. When you send a message, the model performs a forward pass: your tokens go in, a distribution comes out, a token is sampled, repeat. Nothing in that operation writes to the model’s parameters. Training is a separate process, run on separate infrastructure, on a schedule that has nothing to do with your session.
So the model cannot be learning from you during the conversation, in the sense of changing itself. If it appears to have adapted to you within a session, that is the context window doing its job — everything earlier in the conversation is being re-sent with every request, because the model has no state between calls.
Three mechanisms called learning
| Mechanism | Description |
|---|---|
| context | Text in the current request, including earlier turns. Disappears when the conversation ends or the window fills. Affects nobody else, ever. |
| memory features | The product extracts facts from your conversations, stores them in an ordinary database, and silently prepends them to future requests. Persistent, viewable and deletable — because it is text in a store, not a change to the model. |
| training | Your conversation is added to a dataset used in a future training or fine-tuning run. Affects the model everyone else uses. This is the mechanism people are actually worried about, and it is governed by policy rather than by architecture. |
The middle row is the one that produces the uncanny experience. A product that appears to remember your job, your writing style and your preferences across sessions is almost always doing retrieval and injection, not modification. That is testable from the outside: if the product exposes a memory list you can read and edit, then what “it remembers” is a row in a table, and deleting the row deletes the memory completely. A weight update would have no such affordance, because there is no way to remove one conversation from a trained model short of retraining. That asymmetry is also why erasure requests are hard to satisfy once data has reached the third row.
The question you actually meant
Whether your conversations enter a training corpus is not a fact about language models. It is a fact about one company’s policy for one product tier at one point in time, and it varies along at least four axes that people routinely conflate:
- Consumer product versus API. These frequently have different defaults, and the difference is usually the single largest factor.
- Free versus paid versus enterprise tier. Terms often differ by plan within the same product.
- Default versus setting. An opt-out that exists is not an opt-out that is on.
- Training versus retention versus abuse review. Three separate uses of your text with three separate policies. A provider that does not train on your data may still retain it for a period, and may still have staff review flagged content. See what providers actually do with request data and what zero data retention means.
Because all four move independently and change over time, any page that printed a specific answer for a specific provider would be wrong within months. This one deliberately does not. The structure of the question is stable; the answers are not.
Why the evidence people offer does not work
Two pieces of apparent evidence circulate, and neither one shows what it is taken to show.
“I told it something and later it knew it”
Within a session this is the context window. Across sessions it is a memory feature. Neither requires a weight update, and a weight update would in fact be a poor explanation: training runs are infrequent and expensive, and a single conversation among an enormous corpus would have almost no measurable effect on the resulting model’s behaviour.
“I asked whether it was trained on my data and it said yes”
A model has no privileged access to its own training pipeline. It has the text of its instructions and whatever generalisations about AI systems were in its training data. Asked a question about its own provenance, it produces a plausible answer, and plausible answers to questions about a system’s internals are exactly the case where fabrication is most likely, because the training data contains many confident statements about AI systems in general and none about this specific deployment. Self-report is not evidence here in either direction.
How to find the real answer for your account
The answer exists, it is written down, and it is not in the marketing page. Read the terms for the specific product and tier you use, looking for three separate things: whether inputs may be used to improve or train models, how long they are retained and by whom, and whether a setting exists that changes the first answer. If you are on an API rather than a consumer app, the same three questions have different answers and the enterprise agreement usually addresses them explicitly.
The practical rule that survives every policy change: text you cannot afford to have leave your control should not be sent to a service whose terms you have not read, and that rule was true of every SaaS product long before this one.