Skip to content
Security guide · 2026

Deepfake Detection: How to Spot AI-Generated Video, Voice & Call Fraud

Deepfakes are AI-generated video and audio that impersonate a real person — and they now run in real time, during a live call. That breaks an assumption most organizations never wrote down: that seeing a face or hearing a voice proves who you are talking to. This guide explains what the technology can and cannot fake, how the fraud actually plays out, and — honestly — why spotting artifacts matters far less than the process controls that verify identity before money or access moves.

Awareness demo: an urgent payment request on a video call. AI-generated; the people and organisations shown are fictional.
Awareness demo: a voice-clone call asking for a password reset. AI-generated; the people and organisations shown are fictional.

In one sentence: a deepfake is AI-made or AI-altered video, image or audio that makes a real person appear to say or do something they did not, and the best defense is to verify the request through a second, trusted channel. For the voice-call version of the same fraud, see the voice-clone fraud defense checklist; for payment-fraud response, the BEC wire-recall triage; to practise spotting a lure, play Pick the Phish; for the voice-call attack by name, see voice phishing (vishing) explained.

TL;DR

The short version

5 points
  • A deepfake is synthetic media — a face, a voice, or both — generated by AI to impersonate a real person. The techniques now run in real time, so an attacker can face-swap and voice-clone during a live video call, not just in a pre-recorded clip.
  • Seeing someone on a video call or hearing their voice is no longer proof of identity. The most damaging attacks are not viral fake videos; they are targeted finance fraud — a "CEO" or "CFO" on a call authorizing an urgent wire — and video-KYC bypass during account opening.
  • Trying to spot artifacts (odd blinking, blurry hairlines, lip-sync drift) catches older or rushed fakes, but these tells are disappearing fast and give false confidence. They are a weak last line, not a control.
  • What actually works is process, not eyesight: out-of-band callback verification on a number you already trust, a shared challenge question or code word, and dual authorization on every payment regardless of who appears to be asking.
  • Layer in liveness checks for identity verification, content provenance (C2PA) where your tools support it, and recurring training so staff expect the pressure tactics — urgency, secrecy, authority — that accompany almost every deepfake fraud attempt.

What US agencies say about deepfakes

In September 2023 the NSA, FBI and CISA published a joint information sheet, Contextualizing Deepfake Threats to Organizations. Its summary calls deepfakes a particularly concerning type of synthetic media that uses artificial intelligence and machine learning to create believable, highly realistic media. It names three of the most substantial threats to organizations: damage to brand, impersonation of leaders and financial officers, and fraudulent communications used to gain access to networks, communications and sensitive information.

It found limited signs of significant use by malicious state-sponsored actors, but said less capable malicious actors are increasingly able to use these techniques, which will likely increase their frequency and sophistication.

Source: NSA/FBI/CISA Cybersecurity Information Sheet, executive summary (copy hosted by the AHA), accessed 3 November 2026. Summary only; we did not verify the sheet's full recommendations.

What a deepfake actually is

"Deepfake" covers a family of generative-AI techniques that synthesize or alter media to make a person appear to say or do something they did not. In practice there are three capabilities worth separating:

  • Face synthesis & face-swap. A model generates a target person's face, or maps it onto a performer's head, frame by frame. Pre-recorded versions have existed for years; the newer development is that it runs live, swapping a face onto a webcam feed during a video call with low enough latency to hold a conversation.
  • Voice cloning. From a short sample of someone's speech — often scraped from a webinar, earnings call, or voicemail greeting — a model reproduces their voice saying arbitrary new text, including over a phone line where audio quality already hides imperfections.
  • Lip-sync and reenactment. An existing video of a person is driven by new audio so their mouth and expressions match words they never spoke — useful to an attacker who has footage but wants a new "message" from a known figure.

Why "I saw them on the call" is no longer proof

For decades, a live video or a familiar voice was treated as strong evidence of identity — good enough to authorize a payment or waive a check. Real-time face-swap and voice cloning quietly removed that guarantee. The face on the screen and the voice on the line are now things an attacker can produce, which means they are claims, not proof. The defensive mindset shift is to stop treating presence on a call as authentication and start treating every sensitive request as something that must be independently verified — the same way you would not accept an email "from the CEO" as sufficient authorization just because the display name matched.

We covered this shift in depth in Real-Time Deepfakes Are Here: Video-KYC and the CEO on the Call.

The attack pattern

The headline-grabbing deepfakes are fake celebrity or political clips, but the ones that cost organizations money are quieter and more targeted:

  • Executive-impersonation (CEO/CFO) fraud. A finance or operations employee gets a call, voicemail, or video meeting from someone who looks and sounds like a senior leader, pressing for an urgent, confidential wire transfer or change of banking details. It is business email compromise with a face and a voice attached, which makes the victim far more likely to comply.
  • Video-KYC bypass. During remote identity verification for account opening — banks, crypto exchanges, telecoms — an attacker presents a synthetic face to pass the "show us your live video" step under a stolen or fabricated identity, enabling fraud and money laundering downstream.
  • Help-desk & recovery social engineering. A cloned voice calls IT or a help desk to reset credentials, approve an MFA re-enrollment, or unlock an account, using the impersonated person's authority to skip the normal identity checks.

Almost every one of these rides on the same three levers: urgency ("this has to happen now"), authority (a senior figure is asking), and secrecy("keep this between us"). Those pressure cues are a more reliable warning sign than any pixel-level artifact.

Detection and defense that actually helps

Order matters here: the strongest defenses are process controls that do not depend on anyone being able to see through the fake.

  • Out-of-band callback verification. For any sensitive request, hang up and call the person back on a number you already hold — from your directory, not from the call. This single habit defeats the overwhelming majority of impersonation fraud, deepfake or not.
  • Code words and challenge questions. A pre-agreed shared secret, not written down where it could leak, that a real colleague can answer and a stranger cannot. Cheap to set up for finance and executive teams, and very hard for an attacker to pass.
  • Dual authorization on payments and changes. Require a second, independent approver for wires, new payees, and banking-detail changes above a threshold — regardless of who appears to be requesting it. The control works even if the first person is completely fooled.
  • Liveness checks for identity verification. For KYC and high-assurance onboarding, use liveness detection (active prompts, depth/3D sensing, injection-attack resistance) rather than a plain "record a video" step, and keep the vendor's models current against new attack techniques.
  • Content provenance (C2PA). Where your tools support it, prefer and verify signed media so unsigned content earns extra scrutiny. It strengthens trust in published assets but does not verify a live call.
  • Training and realistic drills. Teach staff — especially finance, executive assistants, help desk and KYC reviewers — that a face and voice are not authentication, and rehearse the callback habit so it survives real-time pressure.

Visual and audio tells (use with caution)

These can flag a weaker or older fake, but they are a fading last resort, not a verification method: unnatural or absent blinking, lighting and shadows that do not match the room, a hairline or glasses edge that shimmers, warping around teeth and ears, hands or background distorting when they cross the face, and lip movements fractionally out of sync with the words. On audio: flat, even cadence, missing breaths, odd emphasis, or an inability to handle interruption and overlapping speech naturally. Asking a live caller to do something unscripted — turn fully to profile, wave a hand across their face, pick up and describe a nearby object — can still disrupt some real-time pipelines. Treat any of this as a reason to stop and verify out-of-band, never as a clean bill of health.

The one thing to take away: you will not reliably out-see a modern deepfake, and betting on your eyes is exactly the failure mode attackers exploit. Build the assumption that video and voice can be faked into your process — callback verification, code words, and dual authorization on anything that moves money or grants access — so the fraud fails even when the fake is perfect.

Frequently asked questions

Can you detect a deepfake just by looking at it?

Sometimes, but you should not rely on it. Older or hastily made fakes can show tells — unnatural blinking, mismatched lighting or shadows, a blurry or shifting hairline, teeth and ears that warp, lip movements slightly out of sync with the audio, or a voice with flat, even cadence and no breath sounds. On a live call, asking the person to turn their head fully sideways, pass a hand in front of their face, or stand up can still break some real-time face-swap pipelines. But these tells are vanishing as the models improve, and a confident "it looked fine to me" is exactly the false assurance attackers want. Treat visual inspection as a weak signal, never as verification.

How do I verify a suspicious video call or voice message?

Verify out-of-band: contact the person through a separate, previously trusted channel — a phone number from your own directory or address book, not one supplied in the call, email, or chat that triggered the request. Call them back and confirm the request directly. A pre-agreed challenge question or code word that is not written down anywhere an attacker could find it is a fast, strong check. If the request involves money, data, credentials, or access, it should pass your payment or change-control process no matter who appears to be asking. Urgency and secrecy ("do not tell anyone", "we must close this before end of day") are the tell to slow down, not to hurry.

What is a deepfake video-KYC or video-call scam, and who is targeted?

Two common patterns. In CEO/CFO fraud (a form of business email compromise), an attacker impersonates a senior executive on a call or voicemail and pressures a finance or operations employee into an urgent wire transfer, gift-card purchase, or data handover. In video-KYC bypass, an attacker uses a synthetic face during an identity-verification video check to open a bank, exchange, or telecom account under a stolen or fabricated identity. The targets are usually people who can move money or grant access: finance teams, executive assistants, help-desk and onboarding staff, and KYC/compliance reviewers.

Are automated deepfake detectors reliable?

They help, but they are not a guarantee and they are not a replacement for process. Detectors look for statistical artifacts of synthesis, and their effectiveness varies with the generator used, compression, resolution, and how recent the attack technique is — a tool trained on last year's fakes may miss this year's. They also degrade on short clips and over typical video-conferencing compression. Use them as one input alongside liveness checks and content provenance, and keep the human process controls (callback, dual authorization) as the decisions that actually release money or access.

What is content provenance (C2PA) and does it stop deepfakes?

C2PA (the Coalition for Content Provenance and Authenticity) is an open standard for attaching tamper-evident metadata to media — who or what created it, when, and what edits were applied — so a viewer or system can check whether a file carries a verifiable origin. It does not detect a fake directly; instead it helps you trust media that is signed and be more skeptical of media that is not. It is a useful part of a defense-in-depth posture, especially for published or archived content, but it does not verify a live, unsigned video call, so it never substitutes for out-of-band verification on a real-time request.