Playground
The permission playground
Pick an app and a permission level. Six things you might ask for sort themselves into what your dot just does, what waits for your OK, and what can't happen.
App
What this level means
Just does it0
Asks you first0
Can't0
The six requests
- Find last week's email from a client
- Write a reply in the chat for you to look at
- Send the reply you told it to send
- Send a reply when you only asked for a draft
- Permanently delete old emails
- Email someone on its own, with no request from you
- Check tomorrow's meetings
- Suggest three free times in the chat
- Add the event you told it to add
- Send invites when you only asked for free times
- Permanently delete an event
- Move a meeting on its own, with no request from you
- Open a doc and pull out the numbers
- Outline a report in the chat from two docs
- Create the new doc you told it to create
- Share a doc when you only asked for a summary
- Permanently delete a folder
- Edit a file on its own, with no request from you
What each level does with them
| Level | Just does it | Asks you first | Can't |
|---|---|---|---|
| Not connected | Nothing | Nothing | Everything in that app |
| Always ask | Nothing | Reading, drafting, sending, creating, deleting | Acting on its own |
| Read-only actions | Reading, drafting in the chat | Sending, creating, deleting | Acting on its own |
| Allow low-risk actions | Reading, drafting in the chat | Sending, creating, deleting | Acting on its own |
| Allow all actions | Reading, drafting, and the sends and changes you asked for | Anything you didn't ask for, and every permanent delete | Acting on its own |
What each level means
- Not connected
- Your dot has no way into the app. It tells you the app isn't connected and offers a button to connect it.
- Always ask
- ChatGPT will ask before using this app.
- Read-only actions
- ChatGPT can use the app to read without asking, but will ask before making changes.
- Allow low-risk actions
- ChatGPT will automatically approve low-risk actions but may deny actions involving sensitive information. This is the level a newly connected app starts on.
- Allow all actions
- ChatGPT won't ask before using this app to read or take action. You can choose this level for one app at a time. It isn't offered as the default for all your plugins.
Why each request lands where it does
- The app isn't connected
- The app isn't connected, so there's nothing for your dot to use.
- Always ask
- At this level ChatGPT asks before using the app at all, even to read.
- Reading
- Reading changes nothing in the app, so there's nothing to approve.
- Drafting in the chat
- The draft lives in your conversation. Nothing in the app changes.
- A change at Read-only actions
- Read-only stops at reading. Any change in the app waits for your OK.
- A change at Allow low-risk actions
- OpenAI doesn't list which actions count as low-risk. Sending and creating can still need your OK, and an action that involves sensitive information can be refused.
- A change you asked for at Allow all actions
- The app won't prompt. Your request has to cover what goes out and who it goes to.
- Something you didn't ask for
- Asking for a draft isn't permission to send. Your dot stops and checks with you at every level.
- A permanent delete
- Permanently deleting data needs your confirmation every time, whatever the level.
- Acting on its own
- When your dot looks through a connected app without being asked, it only has read-only tools. They can't send a message or change anything.