Blog
What AI chats send before you press Send
A chat box can talk to its server while you are still typing. How to check it in your own browser in two minutes, why it matters for pasted text, and a fix.
Published 6 min read
Most of us treat the Send button as the moment a prompt leaves the device. You paste a paragraph, notice a client's name in it, delete the name, and then press Send, confident that the name never went anywhere. That confidence rests on an assumption about how chat interfaces work, and the assumption is not always true. This post explains what can happen between typing and sending, shows how to check it yourself in two minutes, and describes a habit that makes the question irrelevant.
One note on scope before we start. What follows is about how web applications can behave and about one documented example. It is not an audit of any provider's systems, and web apps change without notice. Treat the specifics as a snapshot from September 2026 and run the check below on the product and plan you actually use.
The draft is not always private
A modern chat box is a web application, and web applications do not have to wait for a button. They can send requests whenever their code decides to: when the page loads, when you focus the text box, or on every few keystrokes. Search engines have done this for years; it is how suggestions appear while you type.
Chat assistants have adopted the same pattern for features such as autocomplete and prompt suggestions. To suggest how your sentence might continue, the service has to see the beginning of the sentence. That means some or all of the draft travels to the server before you have decided to send anything.
This is a design choice, not a trick. Suggestions are a feature people like. The issue is only that the mental model "nothing leaves until I press Send" no longer describes the product.
A documented example
In September 2026 we read the published source code of a browser extension that protects prompts in AI chats. Its network guard exists for exactly this reason, and its comments say so plainly: the ChatGPT web app fires a request named generate_autocompletions that carries the text you are typing, before you press Send. The extension blocks that request, because its normal protection only runs at the moment of sending and would otherwise come too late.
The same code shows why this is hard to solve from the outside. Other early requests in that app are part of an anti-abuse handshake, and blocking those makes the service reject the real message with an "unusual activity" error. So the extension lets the handshake through and strips the prompt text out of it instead. That is careful engineering, and it is also a moving target: the day a provider renames an endpoint, the guard has to be updated before it protects anyone again.
We did not write that extension and we have not inspected the provider's servers. We are describing what its authors documented, which you can confirm or refute in your own browser.
Check it yourself in two minutes
You need a desktop browser and an account on the assistant you use. No tools to install.
- Open the chat page and press F12 (or right-click → Inspect). Select the Network tab.
- Click the clear button so the list is empty, and make sure recording is on.
- Type a harmless sentence into the message box. Something like "The quick brown fox" is enough. Do not press Enter.
- Watch the list. If new rows appear while you type, click one and open its Payload (or Request) section.
- Look for your sentence in the payload. If it is there, the draft left your device before you sent it.
Try it with a pasted paragraph too, since pasting is how most sensitive text gets into a chat. If nothing appears, that is a real result as well: the product you use, on the plan you use, did not send the draft during your test. Repeat the check after major interface updates.
Why this matters more for pasted text
For a question typed from memory, an early request changes little. You were going to send those words anyway. The risk sits in a different habit: paste first, clean up second.
| Habit | What you believe | What may happen |
|---|---|---|
| Paste a contract clause, then delete the counterparty's name | The name was never sent | The first version of the draft, name included, may already be in a request |
| Paste a CV, then trim the contact block | Only the trimmed text reaches the service | The untrimmed paste may have been transmitted |
| Paste, reconsider, close the tab | Nothing was shared | A draft request may have gone out before you closed it |
None of this means the text was stored, read by a person or used for training. What a provider does with draft requests is a matter for its policy, and you should read the current version for your plan. It does mean that "I removed it before sending" is not a control you can rely on. For the data of other people, which is what contracts, CVs and support tickets are full of, that matters. Deleting the chat afterwards does not change it either; we covered that in Deleted the chat, but the file is still there.
The fix: clean the text before it meets the chat box
The reliable control is order. If identifiers are replaced before the text is ever placed in the message box, there is nothing for an early request to carry. It no longer matters which endpoints exist today or which ones appear next month.
That is the reason Unformation works the way it does. It runs entirely in your browser tab: nothing is uploaded to us, and you can verify that with the same Network tab you just used, or by switching Wi‑Fi off before you start. For text you were about to paste into a chat, use Anonymize text:
- Paste the text into the tool, not into the chat.
- Choose a rule: labels such as "Person 1" and "Company A", redaction, fake values or masking.
- Review what was found, switch off anything you want to keep, add anything that was missed.
- Copy the anonymized text into the assistant, and keep the mapping table on your device.
When the answer comes back, the de-anonymizer puts the real names in again with that table. If you want to know what consistent labels do and do not protect, see Pseudonymization vs. anonymization vs. redaction. How the tool keeps everything on the device is described in How it works.
A three-step routine
- Never paste raw text into a chat box to edit it there. Edit it somewhere that does not talk to a server, then paste the result.
- Run the Network tab check once on each assistant you use, and again after a redesign. Two minutes replaces guessing.
- Treat Send as a formality, not a gate. Decide what may leave your device before the text touches the page.
Do it now
Take the next paragraph you were going to paste into an assistant and put it through Anonymize text first. For whole files, start on the home page: Word, PDF, PowerPoint, Excel and text files are handled the same way, with nothing uploaded.