Mia Gave a Wrong Answer
Why an answer from Mia in the Help panel or in a meeting can be wrong, and how to check it before you act on it.
Mia Gave a Wrong Answer
Mia is a language model, the same kind of technology as every other AI assistant. She writes the most likely answer from what she was given: the documentation, the screen you are on and, signed in, your meetings, chats and files. A likely answer is not always a true one. Treat it the way you would treat a quick answer from a colleague: usually right, worth checking before money, a deadline or a promise depends on it.
Why an answer can be wrong
| What you see | Why it happens |
|---|---|
| A confident answer that is simply not true | The model fills a gap with the most plausible text instead of saying it does not know |
| A date, a number or a name that is slightly off | Numbers and names are the easiest things for a model to reshape; a figure from a transcript may be rounded or mixed with a neighbouring one |
| An answer about a feature InterMIND does not have | The question described something familiar from another product and the model followed the description |
| An answer from the wrong meeting or channel | The panel answers about the screen that is open; the header names it. With nothing open it answers about your account as a whole |
| A fact from the web that is out of date | Web results are pages as they were when read; the address under the answer says where it came from |
How to check
- Read the mark under the answer. It names the meeting, channel or page the answer was drawn from. Open it: the transcript, the recap and the file are the source, the answer is a summary of them.
- Ask for the source. «Where does this come from?» makes Mia quote the line of the transcript or the passage of the document.
- Ask again, narrower. «What was agreed about the price on 12 October?» gives the model less room than «what did we agree?».
- Say what is wrong. A correction in the same conversation is used for the rest of it. A term that is translated wrongly is fixed for good in the glossary.
What is not a model error
- The panel answers from the documentation only. Either you are not signed in, or the organization has switched AI features off, or the daily budget for answers is spent; the panel says which.
- Mia does not answer aloud in a meeting. Aloud she answers licensed participants of paid plans; everyone can write to her in the Help panel. See Mia in meetings.
Nothing you write to Mia is used to train a model, and nothing is kept by the AI gateway after the answer is produced: see Security.