A message you already sent is already stored, and closing the tab does not reach it. An unsent draft was never sent anywhere, so it dies with the tab, a crash, or Shift and Escape exactly the same way. None of those four events is a request sent to ChatGPT or Claude.
A sent message and an unsent draft behave differently
The moment a message is submitted, it leaves the browser and is written to your account on the assistant's own server. From that point on, the tab is no longer the thing holding it. Closing the tab a second later, an hour later, or never, makes no difference to that message, because it was never sitting in the tab to begin with.
A draft is the opposite case. Text typed into the composer but never sent exists only in the browser's memory, the same way an unsaved line in a document editor exists only until you close it. Nothing about typing a message writes it anywhere durable. Submitting it is the action that does that, and until that action happens, the draft has no home outside the tab you are looking at. Close that tab and the draft goes with it, not because closing a tab is somehow destructive, but because there was never a copy of it anywhere else.
This is the distinction worth keeping straight, and it is a narrower, sharper question than the general local versus server split covered in browser history versus chat history. That article explains why the two systems are separate at all. This one is about what happens to a specific message depending on which side of sending it was on when the tab went away.
A crash or force quit is not a request sent to anyone
A browser crash, a forced quit, or the operating system killing the process outright all land in the same place as a deliberate close, and the reason is the same in every case: none of them is a request sent to ChatGPT or Claude. A crash is a local failure of the browser process. The account on the far end has no way of knowing it happened, because nothing about a crash reaches it.
That means the same split applies. Whatever had already been sent before the crash is already stored, unaffected by however ungracefully the tab went away. Whatever was still a draft at the moment of the crash is lost, for the same reason it would be lost by a calm, intentional close: it was never anywhere but the tab's own memory. A dramatic ending does not change which side of that line a given piece of text was on.
What Shift and Escape does, and does not, leave behind
Shift and Escape, which both hides the disguise and brings it back toggles the disguise, not the page underneath it. Pressing it changes what is drawn on screen and nothing else. It does not submit whatever is in the composer, it does not clear it, and it does not send any instruction to ChatGPT or Claude in either direction. A message you were mid typing when you pressed the shortcut is exactly where you left it once you toggle back, because the shortcut never touched the composer at all, only the layer drawn over it.
The shortcut is a display toggle, nothing sends or clears
Shift and Escape brings the real page back or hides it again. It behaves like a light switch for what is on screen, not like a save button or a clear button, and no draft or sent message is affected by using it in either direction.
What ends a session, what deletes it
| Action | Ends the on-screen session | Deletes a message already sent | Loses an unsent draft |
|---|---|---|---|
| Close the tab, deliberately | Yes | No | Yes |
| Browser crash or force quit | Yes | No | Yes |
| Toggle Docognito with Shift and Escape | No, only redraws it | No | No |
| Delete the chat from your history list | No, it may still be open | Yes | Not applicable |
The third row is the one worth noticing. Toggling the disguise does not even end the session the way the other rows do, since the real page and its composer are still there underneath, unsubmitted text included.
When a conversation is genuinely gone
Both ChatGPT and Claude offer a delete option inside their own history panels, usually one conversation at a time, and some accounts also expose a bulk option to clear the entire history at once. Either way, the action is triggered from inside the assistant's own interface, and it works by sending a removal request to the account, not by anything happening in the browser. That is the only route covered here that actually results in a sent message being gone rather than simply out of view, and it has nothing in common with any of the four rows in the table above.
A sent message and an unsent draft do not share a fate
Treat a sent message and an unsent draft as two different things with two different fates. A sent message is already safe on the account regardless of how the tab later ends, deliberately, by crash, or by toggling the disguise. An unsent draft is safe nowhere until it is sent, and any of those same four endings loses it the same way. If a conversation needs to be actually removed rather than merely out of view, that only happens inside ChatGPT's or Claude's own history panel.