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.
+1 also having the same issue. its frustrating. i have also asked ChatGPT for each of my canvases ID number and during my commands i will write the name of the canvas and the canvas ID number but it still it has a tendancy to overwrite the same, most recently created, canvas.
Chatgpt also noted:
What happened was a misstep in how the updates were applied. When you asked me to sync specific canvases using individual canvas IDs, I should have respected those IDs independently for each update. Instead, after the first sync action, I mistakenly applied subsequent updates to the last-used canvas, not the intended separate canvases, even though the correct IDs were listed in your request.
Why it happened:
The tool tracks the last active canvas unless I explicitly reselect a new one each time.
I failed to reassert the intended canvas IDs for each separate update block during rapid-fire syncing.
I have seen this behavior as well. ChatGPT will get “edit latched” onto one canvas and not let go. The only way out of this is to manually close that canvas.
I’ve also been able to find an easy way to force the bug to occur (for bug-report purposes):
Open a canvas with some text
From chat, ask the assistant to change one word of the text
The canvas will now be “edit latched”. Like if I ask, “Tell me about calico cats”, it will replace the entire contents of the canvas with info about calico cats. If I then say, “lol”, then the entire canvas will be replaced with “lol”, etc.
If this happens to you, please report it to support@openai.com. The more reports, the faster it will get fixed!
+1
It is bafling that this basic Core Feature is not solved till now? I guess team is busy working on the teased ‘romantic/spicy’ mode than actually fixing a basic core feature!