How ChatGPT works: what you type leaves letter by letter

Most people picture ChatGPT like an envelope: you write your message, you seal it by pressing Enter, and only then does it travel to the server. Reality is different. The conversation with the server starts as soon as you touch the keyboard — and even with text you may never send.

What actually happens when you type

Take a simple example. You start typing "The job of Sameh". Before you've finished the first word, the page has already opened a dialogue with the server.

1. The conversation is prepared on the first keystroke

As soon as the first letter appears, the page calls an endpoint that prepares a conversation server-side:

POST https://chatgpt.com/backend-api/f/conversation/prepare

{
    "action": "next",
    "parent_message_id": "client-created-root",
    "model": "gpt-5-5-thinking",
    "partial_query": {
        "author": { "role": "user" },
        "content": {
            "content_type": "text",
            "parts": ["T"]
        }
    },
    "client_contextual_info": { "app_name": "chatgpt.com" },
    "thinking_effort": "extended"
}

Notice the field partial_query → parts: it already contains "T", the first letter of your unfinished sentence. The server is told which model is selected, your timezone, and that a turn is about to begin — before you've written a single word.

2. Autocompletions are requested as you write

Right after, and then with every few keystrokes, the page asks the server to guess how you'll finish your sentence:

POST https://chatgpt.com/backend-api/conversation/experimental/generate_autocompletions

{
    "input_text": "The j",
    "num_completions": 3,
    "in_search_mode": false,
    "history_and_training_disabled": false,
    "is_follow_up_turn": false,
    "prefetch_only": false
}

Here, input_text contains "The j" — a live fragment of what you're writing. As the sentence grows ("The job", "The job of", "The job of Sameh"), this call repeats, each time carrying a longer slice of your text. That's what powers the grey "ghost" suggestions that complete your prompt.

The key detail: this fragment travels to the server whether or not you ever press Enter. Type a sentence, hesitate, erase it, rewrite it — the intermediate versions have already left your machine.

Why the site does this

To be fair, there are legitimate engineering reasons behind this design, and they are not inherently malicious:

• Inline autocomplete. The grey suggestions that finish your sentence need the server to see what you've typed so far.
• Latency reduction. Calling prepare early lets the backend warm up the model and the conversation context, so the answer feels faster once you do hit Enter.
• Smoother experience. Prefetching and buffering shave off the small delays that make a chat feel sluggish.

None of this is unusual for a modern web app — search engines have sent your keystrokes for live suggestions for years. What changes with a chatbot is the nature of what you type.

What it means for your privacy

A search box gets keywords. A chatbot gets your thinking out loud: a medical question with a real name, a draft resignation letter, a salary figure, a confession, a half-formed idea you decide not to send. With the behavior above, that material reaches the server even when you change your mind.

A few things worth knowing:

• Abandoned text counts too. Because fragments are sent as you type, a message you delete before sending has, in part, already been transmitted.
• The history_and_training_disabled flag. In the observed payload it is set to false. In plain terms: unless you've turned off the relevant settings, this data falls inside the scope your account allows for history and model improvement. The flag flips to true when those settings are off — but the network call itself still happens.
• You can't see it. None of this is visible in the interface. The only way to observe it is to open your browser's network inspector, as we did here.

So — do they build a model of you on their servers?

This is the honest part: from the outside, we cannot prove what OpenAI does with this stream once it arrives. Anyone claiming certainty is guessing. What we can say:

• It is technically possible. A continuous flow of partial queries, timing, hesitations and rewrites is rich behavioral data. Whether or not it is used to profile an individual, the raw material to do so passes through.
• It is governed by policy, not by your machine. What happens next depends on OpenAI's terms and your account settings — not on anything you control locally. Policies change; defaults change; what's true today may not be tomorrow.
• The only guarantee that holds is the one you enforce yourself. If sensitive text never leaves your computer in clear form, the question of what a server does with it becomes moot.

That last point is the whole idea behind Quieta.

How to protect yourself

You don't have to stop using ChatGPT. You have to stop sending it raw personal data.

Quieta masks personal information on your own machine, before it ever reaches the chat window — names, addresses, figures, identifiers are replaced by neutral placeholders. The original text never leaves your device; the AI works on the masked version; and when its answer comes back, Quieta puts the real details back in place for you, locally and reversibly.

The processing runs 100 % on your computer — there's no Quieta server in the loop. So even the keystroke-by-keystroke behavior described above only ever carries text that's already been stripped of what identifies you.

---

Transparency note — This page describes network traffic observable from a standard browser in June 2026, on chatgpt.com. Endpoints, payloads and field names are OpenAI's and can change without notice. We do not claim to know how the data is processed once it reaches the server — only what leaves your browser, and when. Spotted an inaccuracy? Tell us at hello@quieta.ai.