Let’s talk about the meeting that now shows up with a warning label on it.
Mid-November, Microsoft starts switching on deepfake detection inside Teams meetings. A certified outside provider scans the audio and video. Teams shows you the alert. That’s the shape of it.
And a detail almost nobody will quote. Microsoft changed the schedule on the twenty-second of October. The first notice said general availability from late November into mid-December. The revised one only promises Targeted Release, mid to late November, and tells everyone else: we will communicate via Message center when we’re ready.
So the thing launching this month is a preview for a slice of customers. Everyone else gets an email at some point in the future.
Now, the old scene.
Video calls became the one place in business where we still trusted our eyes. A face. A voice. A room full of people who all seemed to know each other. We built a whole layer of trust on that, and it worked for long enough that we stopped calling it a risk.
In 2024 a design and engineering firm lost about twenty-five million dollars this way. An employee in Hong Kong got pulled into a video conference. The people on the call looked like the people they claimed to be. The instruction was confidential, urgent, and came from a face he recognised. He moved the money.
Nobody in that meeting had done anything stupid. That’s the part worth sitting with.
So Microsoft bolting a scanner onto the meeting is, on paper, sensible. The network got defended. The login got defended. The meeting was the gap we filled with instinct, and instinct stopped being enough.
Here’s what people are going to miss.
First. A detector is a guess with a nicer interface. It will be right a lot of the time, and the case that ruins you is the one where it stays quiet. A clean badge does not mean a human is on the call. It means no signal fired. And nobody in this launch has published the number that would actually matter — how often it misses.
Read the fine print. The providers aren’t named. It isn’t clear whether you need to pay extra. Accuracy isn’t disclosed. And it isn’t disclosed how the provider handles the meeting data it’s looking at.
That last one is the one to sit with.
Because to check the audio and the video, the audio and the video have to go somewhere. To a third party. Which means the board call, the retrenchment conversation, the negotiation with the client you can’t afford to lose, all of it now passes through a company you didn’t choose, on terms you haven’t read. And the person on the other side of the call, the actual human being whose face is being analysed, almost certainly doesn’t know.
We are about to solve a privacy problem by shipping more private data to more vendors.
Second thing. Coverage. Deepfake detection lands on desktop, Mac and web. Not on the phone. There’s a separate impersonation warning feature that does reach Android and iOS in November, and that one checks whether a display name matches an account. It does not check whether the face is real.
Think about the attack. A finance director, in a car, on a phone, told to approve a payment by someone whose voice is perfect. That is where it happens. And that is exactly where there is no scanner.
Third. And this is the one that will still be true in five years.
Detection asks a question about appearance. Does this look generated? That question gets harder every quarter, because the generators get better and the detectors chase them, and the gap never closes, it just moves.
There’s another question. Did this person’s device sign this session? That one doesn’t decay. A record of who joined, from what device, with what proof, is a fact. A probability is a mood.
The industry keeps buying detection because detection ships this quarter. Provenance takes coordination, and coordination is boring, and boring doesn’t demo well.
So what do you actually do, as an operator, this month?
Set a callback rule and mean it. Any request to move money, share a credential, or change bank details gets confirmed on a number you already have. Not the number in the invite. That costs nothing and it works today.
Put the authority outside the meeting. Amount thresholds. Two people. A hard stop on any change of banking details, no exceptions, no urgency clause. The deepfake’s only job is to make you skip the boring step, so make the boring step impossible to skip.
Before you switch the scanner on, ask four questions. Which provider. What’s the miss rate on audio only. Where does the media live and for how long. What happens to the recording when the call ends. If nobody can answer, you’re not buying protection. You’re buying a feeling.
And start recording who was in the room, and how they proved it. Signed sessions. Verifiable credentials. A join record that travels with the meeting and can be checked later without calling anyone. That is the thing that will still hold when the detector is on version nine.
Last thing. Tell the other side. If a tool is scanning the call, say so at the top of the call. It’s their face too.
We are replacing the face with a score. And a score has never met anyone.
The machine can tell you that the person in the room might not be real. It cannot tell you who is.
So set the rule before the meeting starts, not during it. That part is still yours.