When your company turns on PII detection, chat messages are checked for personal data before they reach the model. This page covers the check you can run yourself while you write, and what happens when a message is held back because of what it contains.
On desktop, the check is the PII chip in the options row under the message box (open the row with + if it is collapsed). Its tooltip reads "Check PII - highlights personal data in your draft". On a phone, the same control sits next to the send button.
email · 92%. On a touch screen these tooltips do not appear.The colours follow your edits. They shift with your changes, and a coloured part whose own text you change loses its colour, because it no longer matches what was found. Run the check again for a fresh reading.
The control is unavailable while the draft is empty, while a check is already running, or while the message box itself is disabled. While global PII detection is off, the control is not shown at all: no checks run anywhere then.
Your message is checked before it reaches the model. While the check runs, the chat shows Detecting PII... in place of its usual working phrases; see Chat states and recovery.
If the message is clean, it is sent as usual. If personal data is found, the outcome depends on the policy an administrator set in PII detection:
The flagged text comes back to the message box with the parts still coloured, so you can edit or mask them and send again. Clicking the notice opens a dialog, Personal information detected, that shows the message with the flagged parts marked and offers Edit message and Close. A blocked message is final: it can be sent only after the personal data is removed or masked.
Where the policy only warns, you can continue past the check. This works only with the Bypass PII Detection permission and only for the exact text that was flagged:
Every bypass is recorded in the PII log; see AI, socket and PII logs.
A check can also open the full result in a PII scan dialog, as happens when you check a training card on a Mind's Training Hub. The scan starts as soon as the dialog opens, so there is no separate scan button inside, and Scanning… shows while it runs. See Train a Mind with training cards for the card check.
email - 2.. instead of a percentage.It is hidden while global PII detection is off, because no checks run anywhere then. An administrator can turn it on in PII detection.
The check needs something to read. It is unavailable while the draft is empty or holds only spaces, while a check is already running, or while the message box itself is disabled.
No. The check reads the draft and colours what it finds. Nothing is posted to the chat, and no answer starts until you send the message yourself.
The saved PII detection policy for chats, the same rules used when a message is sent. Changes an administrator has made but not saved yet do not apply. See PII detection.
The colours follow your text. A coloured part whose own text you change loses its colour, because it no longer matches what was found. The parts you did not touch keep theirs and shift with your edit. Run the check again for a fresh reading.
No. The check reads the message text. Excel and CSV files go through their own review, described in Attach files and Excel data to a chat.
Only where the policy is Warn User and only with the Bypass PII Detection permission: choose Send it anyway on the notice or in the dialog. It applies to the exact flagged text. Where the policy blocks the message, it can be sent only after the personal data is removed or masked.
Three reasons: the policy is Block User Traffic, so the blocked message cannot be sent as it is; you do not hold the Bypass PII Detection permission; or you edited the message after it was flagged. Editing cancels the bypass for that text, and the edited message is checked again when you send it. See Who can do what with Minds.
The send button carries a small triangle while a warned message is armed to skip the check: it means the next send will bypass the check for the exact flagged text. Its tooltip reads Send it anyway with the explanation "This message contains flagged personal data. Sending now skips the check - or edit the highlighted parts first." It appears only for people with the bypass permission, and any edit clears it.
The tooltips that name a label and its confidence score are for pointer devices; on a touch screen they do not appear. The colours, the result notice and the dialog still show.