IT teams that want Microsoft Teams to catch a cloned voice or a face-swapped video in a meeting will need an outside vendor to do the catching. Microsoft plans to start rolling out support for certified third-party deepfake detectors in mid-November, and its admin notice spells out who does the work: “Synthetic media detection is performed by the third-party provider.”
The notice, MC1509787, was revised on Oct. 22. Targeted Release customers should see the feature between mid-November and late November on Teams for desktop, Mac and web. General availability no longer has a date; Microsoft says it will post again in the Message center “when we are ready to proceed.” The first version had aimed for general availability from late November to mid-December, Windows Latest reported.
How a detection bot flags someone
The mechanics are already in Microsoft’s Graph beta documentation. A provider’s bot has to be admitted to the call, and the tenant administrator has to grant it the Calls.ReportSyntheticMedia.All permission. The bot then reports a participant with a confidence score from 0.0 to 1.0, an optional low, medium or high severity, and a yes-or-no judgment on malicious intent. Microsoft’s sample requests use model tasks named voiceClone, faceSwap and lipSync. The result reaches every participant in the meeting through roster updates.
Admins will pick a provider, deploy its app from the Teams admin center and set meeting policies. The two policy settings, SyntheticMediaDetection and SyntheticMediaDetectionAppId, already sit in the Teams PowerShell reference with a note that they have no effect until release. The default app ID, Microsoft writes there, resolves to no provider “when 1P is not available, or Microsoft Detection when 1P is available.”
What admins still don’t know
Microsoft hasn’t named any certified providers or said what certification involves, though Sensity, for one, already sells a deepfake detection app for Teams. The notice asks admins to “review provider terms, data handling, and compliance documentation before enabling,” since an outside service will be looking at meeting media. That differs from Microsoft’s own AI processing of meeting audio and video, where it says a meeting’s audio and video data “never leave Teams.”
The FBI said in a July public service announcement that scammers generate videos for real-time video chats with “alleged company executives, law enforcement, or other authority figures,” one tactic among the AI impersonation attacks hitting U.S. companies. Zoom is taking the in-house route, with its own deepfake risk detection expected in December.