How LLMs actually handle your data

Most privacy concerns about AI come from a misunderstanding of how language models work at the inference level. Understanding the mechanics behind your interaction with LLM services will make the privacy boundaries more clear and making the choice to put your data first easy.

A (massive) game of hot or cold

Every time you send a message to an LLM, the model performs a forward pass, a single mathematical computation through billions of weighted parameters. Your input goes in, the LLM counts the probability of the most likely answer, and the response is returned.

That's the entirety of what the LLM model does. Nothing is written anywhere or stored inside the model as the models themselves are inert and does not change.

your message → [ model weights ] → response

Stateless forward pass

Each forward pass is independent. Stateless means the model has no memory between requests. When your next message is sent, the model has no inherent recollection of the previous one.

The "context" or "conversation history" you experience is an illusion created by the application layer, just like any other application. A chat interface that assembles your prior messages into a prompt and sends the entire thing with each new request. The model sees it all fresh every time.

Request 1:

[ "Hello" ] → model → response

Request 2:

[ "Hello", response, "Follow up" ] → model → response

the model doesn't remember request 1 — it's re-read as part of request 2's prompt

Where data actually lives

If the model is stateless, where does the privacy risk come from? It's in how the service providers treat and trade on your data, your inputs, and interactions with the model.

Provider logging

The API provider may log your requests on a server. This is where your data is stored and then sifted through by other application systems.

Training pipelines

Some providers use logged conversations as training data for future model versions. This is opt-out, not opt-in, and often buried in terms of service.

Advertising use

With most free products, your data is used or sold to third party platforms as a revenue source.

How Chili is different

Chili puts your data first and our revenue model does not rely on trading or selling your data.

—Session turns are stored in your account, scoped to your user ID. You can delete them.
—No training pipeline. Chili does not log conversations for model training. Each request is a fresh inference call with no downstream retention.
—We rely on direct pricing as revenue source.Not selling your data to other platforms.

The short version

LLMs don't remember anything. Each inference is a one-way computation as a forward pass. Your message goes in, a response comes out, and nothing changes inside the model. Privacy risks are entirely in how the surrounding application handles your data, not in the model itself. Chili is built to make that handling explicit, minimal, and yours to control.