Summary of the bug:
When using ChatGPT’s Canvas feature, any edit request made by the user—regardless of the specific Canvas intended—results in the model updating the most recently created Canvas instead. This leads to consistent, unintended overwriting of unrelated Canvases.
How to reproduce:
-
Create two Canvases: Canvas A and Canvas B.
-
Attempt to update Canvas A using a prompt or direct instruction.
-
ChatGPT updates Canvas B (the most recently created) instead of Canvas A.
-
No warning or feedback is given that the wrong Canvas is being edited.
Why this is critical:
-
If a Canvas has already been shared via URL and stored by the user, its content can be overwritten silently.
-
There is no way to confirm which Canvas will be edited based on user intent.
-
ChatGPT (or the custom assistant) does not detect or report this mismatch, making debugging extremely difficult for users.
Expected behavior:
The assistant should only edit the Canvas explicitly specified by the user, or at the very least, clearly confirm the Canvas ID before proceeding.
Suggested solutions:
-
Display the current active Canvas ID or title in the UI when editing.
-
Confirm with the user before overwriting a Canvas, especially when shared via link.
-
Prioritize “create new Canvas” behavior over implicit overwrites in cases of ambiguity.
This was discovered during an extended session analyzing collaborative workflows using Canvas, where we attempted to coordinate multiple Canvas documents. The issue significantly impacts confidence in using Canvas as a reliable long-term information structure.
Thanks in advance for looking into this — this issue has a significant impact on Canvas usability.