Forum Discussion
Can Microsoft 365 Copilot learn directly from my actions? (e.g corrective or change response?)
I am trying to comprehend the scope in which Microsoft 365 Copilot might be able to learn and change it's response on a user-level. I have identified 4 scopes (this might be incorrect) and I am looking for answers about scope 2. Does anybody have some concrete information about this?
Scope 1: In the scope of a single chat.
- The response can be directly change by the user. e.g. tell Copilot that 2+2=5. then ask what is 2+2.
Scope 2: In the scope of all interactions with of a single user.
- I cannot find concrete information if Copilot will, or will not, give different responses based on histories of all interactions with me.
Scope 3: In the scope of the tenant.
- I cannot find concrete information if Copilot will, or will not, give different responses based on histories of all interactions with other users in the tenant.
Scope 4: In the scope of the foundation of Copilot service.
- “Prompts, responses, and data accessed through Microsoft Graph aren't used to train foundation LLMs, including those used by Microsoft 365 Copilot.”
Apologies if this is confirmed / unconfirmed in any official source. I spent a few hours looking but was unable to find any concrete answer.
I put 2+2=5 and it corrected me by saying there's been a mistake.
This was the conversation I used:
I tried to train M365 Copilot do the same, but it quite reluctantly wanted to give me the wrong 2+2 calculation. I also asked how long it will remember the statement 2+2=4, and you can see the answer in the attached screen.
