Your Interview Call Is Getting Deepfake Detection

Both major interview platforms are adding synthetic-media detection to live meetings this quarter. Microsoft's message center item MC1509787, published 15 October 2026 and updated 22 October 2026, says Teams will support certified third-party detection of synthetic audio and video in meetings, rolling out to Targeted Release from mid-November 2026. Zoom announced "Deepfake Risk Detection" at Zoomtopia 2026, listing it as expected December 2026 and describing it as flagging possible synthetic media in "hiring, CX, and meetings" — hiring named first (Zoom product news).

If you interview over video, this is a new signal attached to your camera and microphone feed, scored while you talk, and shown to someone on the employer's side. It is not an identity check and not a proctoring tool. It is a probability estimate about whether your stream looks generated. This article covers what each platform has actually committed to, what the signal can and cannot establish, and the ordinary webcam conditions that make a genuine candidate look synthetic.

What the two platforms have committed to

The announcements are narrower than the headlines suggest, and they work differently.

Zoom describes Deepfake Risk Detection as real-time analysis of video and audio signals, looking for participant media that may have been "altered outside of Zoom," and flagging suspected deepfakes with a confidence indicator (Zoom newsroom, Zoomtopia 2026). Zoom has not published who sees the indicator, which plans include it, whether an administrator must switch it on, or any accuracy figures. Expected availability is December 2026.

Teams is not building a detector. Per that message center item, a certified third-party provider analyses the meeting media and sends detection signals to Teams; Teams supplies the integration and the in-meeting surfacing. An administrator has to choose a provider, deploy its app through the Teams admin center, grant permissions, and configure meeting policies. Microsoft's own compliance note tells admins to review the provider's terms and data handling, because the provider may analyse meeting audio and video.

So on Teams, the question "what is scoring me" has a different answer at every employer, and the answer is whichever vendor their IT team licensed.

This is prediction, not matching

The distinction matters for what a flag can prove. A verification check compares you against something you enrolled earlier — Zoom's World ID integration, for example, only shows a verified-human badge when an enrollment image, a liveness selfie, and a live video frame all match, as covered in our walkthrough of what Zoom's verified-human badge requires. That check can be wrong, but it is anchored to a prior reference.

Deepfake risk detection has no reference. It inspects the stream for artefacts that generated media tends to leave and returns a likelihood. A high score is a statement about pixel and waveform statistics, not a finding that you are someone else. Nothing in either announcement claims otherwise, and neither company has promised that a flagged participant is removed automatically.

It reads the stream, not your computer

These features sit on the meeting media path. They do not enumerate your running processes, your windows, or your monitors, which is the territory of assessment lockdown software and proctoring extensions. They are also distinct from the vendors that scan stored recordings after the interview is over — a separate and growing practice we examined in how interview recordings get scanned months later. A live detector sees what your camera and microphone send; a post-hoc scanner sees the file that was kept.

Why a real candidate can trip it

This is the part worth preparing for, because the failure mode is asymmetric: a false positive lands on you during a hiring decision, and neither platform has documented an appeal path.

Detector accuracy drops sharply outside the lab. In Deepfake-Eval-2024, a benchmark of deepfakes actually circulating in 2024, open-source state-of-the-art detectors lost roughly half their discriminative power: the authors report AUC "decreasing by 50% for video, 48% for audio, and 45% for image models compared to previous benchmarks," and found that commercial and fine-tuned models, while better than off-the-shelf open-source ones, "do not yet reach the accuracy of deepfake forensic analysts" (arXiv:2503.02857). That is measured on detecting fakes. The complement — how often authentic video gets called synthetic — is not something either platform has published for these features at all.

Several ordinary interview conditions push a genuine stream toward the artefacts these models key on:

• Aggressive compression. A weak uplink makes the platform drop bitrate and resolution. Compression removes the fine-grained texture a detector uses as evidence of authenticity, which is the mechanism behind the long-standing finding that detectors degrade on user-generated video relative to their own test sets.
• Beauty filters, smoothing, and background replacement. These are generative edits to your face and surroundings. A virtual background in particular puts a synthesised boundary around your head for the whole call; we covered the adjacent question of whether virtual backgrounds get candidates flagged.
• Virtual camera and audio software. OBS, camera-routing utilities, noise suppressors, and voice processors modify media before the platform receives it. Zoom's stated scope — media "altered outside of Zoom" — describes this category precisely.
• Low or mixed lighting. Backlighting and a single warm lamp produce the blotchy, low-detail regions that detectors frequently misread.
• External webcams and phone-as-webcam apps. Each added encode and re-encode step is another layer of processing between your face and the detector.

There is an irony in Zoom's own roadmap here. "Translate Me," expected October 2026, renders a speaker's words in a synthesised version of their own voice, and "Touch up my Voice," expected December 2026, applies real-time accent smoothing. Both generate audio — but inside Zoom, which is the opposite of the "outside of Zoom" scope the detector is described as targeting. Whether a detector trained on synthetic-speech artefacts cleanly separates in-platform from out-of-platform generation is not something Zoom has addressed publicly. If you use accent smoothing in an interview, you are relying on an untested boundary. Voice-side identity checks have a related but separate logic, which we broke down in how voice interviews verify the speaker.

What to do before an interview call

None of this requires a different setup from a good interview setup. It requires removing processing.

1. Use the physical camera and microphone directly. Close virtual camera software, routing utilities, and voice processors for the call. If you use a phone as a webcam, test whether the laptop camera is adequate instead.
2. Turn off filters, touch-up, and virtual backgrounds. A plain wall or a real room is a lower-risk background than a synthesised one. If your room is genuinely not presentable, say so in the call rather than generating a replacement for it.
3. Prioritise a stable uplink. Wired or close to the access point, with other uploads paused. Stable bitrate matters more than peak resolution.
4. Light your face from the front. One diffuse source in front of you beats a bright window behind you, for both human and machine readers.
5. Ask what a flag triggers, if the employer mentions detection. Reasonable questions: who receives the signal, whether it is logged against your application, and what the process is if it fires. Microsoft requires the administrator to pick the provider, so the employer — not Zoom or Microsoft — is the party that can answer.

What remains unknown

As of 26 October 2026, both features are announced and neither is generally available. That means the following are genuinely open, and anyone telling you otherwise is guessing:

• No published accuracy, false-positive, or false-negative rates for either implementation.
• No documented statement of who sees a flag — organiser, all participants, or administrators only.
• No documented default state or licensing requirement for Zoom's feature.
• No stated candidate-facing process for contesting a flag.
• For Teams, no public list of which third-party providers are certified, so the detector behind any given interview is unknown to you.

Microsoft's roadmap language is explicit that detection signals are probabilistic. "Certified" is a statement about integration, not about accuracy.

Where this leaves a candidate

Treat platform deepfake detection as one more reason to put an unprocessed, well-lit, stably-encoded version of yourself on the call, and to keep generative processing out of the camera and microphone path. That is the whole of the practical response. The detector is looking at your media stream; it is not an audit of your machine, your preparation, or the tools you used to prepare.

Control is a desktop AI interview assistant for Windows and macOS that transcribes the interviewer in real time and answers through an overlay driven by global hotkeys. Per Control's own product documentation, it is a native desktop overlay rather than a browser extension, and it does not inject anything into your camera or microphone feed — a different surface from the meeting media path these detectors inspect, though that is Control's own description and not an independent test of any detector. If you want to see how the workflow behaves before an interview, the free allowance covers 5 messages and 2 minutes of voice transcription with no card required.