I’ve been able to reliably reproduce the “ghost canvas overwrite/append” bug.
In my case, it happens when editing an open Canvas doc — even a simple word change — and then asking an unrelated question. Instead of answering in chat, ChatGPT overwrites or appends to the open document and is then in an “edit latched” mode.
Key points from my test:
-
It can be triggered by a normal Canvas edit (word change), though can also occur on a new canvas creation
-
Once triggered, all subsequent inputs, even unrelated ones like “LOL” or “End test,” went into the doc or can overwrite it entirely
-
Content could be overwritten completely or appended (start or end).
-
This “edit latch” mode would persist silently until I manually closed the canvas.
-
I’ve reproduced this in different browsers and workspaces.
Risk: Easy to lose work without noticing.
Mitigation idea: Add topic-shift detection, latch timeout, or lock toggle to prevent accidental edits.
If you’ve seen this happen, an easy way to reproduce is:
-
Open a Canvas doc with something (I used a Robert Frost poem)
-
In chat, tell the assistant to change one word.
-
Immediately ask something unrelated (“Tell me about calico cats”).
If it happens for you, please report it by writing to support@openai.com — the more confirmations, the faster this gets fixed!