Faceswaps are outpacing human detection. Manual review can’t keep up.

We asked people to spot deepfakes. They knew fakes were in the set, and they were looking for them. Most still struggled, and feeling sure of an answer didn’t make them any more likely to be right.

That result, from our Deepfakes Quiz, matters well beyond a quiz. Across more than 8,000 quiz sessions, players got just 63% of visuals right on average, not far above the 50% you’d get by guessing. Three in 10 did no better than a coin flip, and only 1 in 40 got all 12 right. Confidence didn’t help either: more than 1 in 4 people who said they were “completely certain” scored no better than chance.

Faceswaps are the sharpest edge of this problem. The following explores why faceswaps are becoming more effective, why manual review struggles to keep up, and how a layered, real-time defence can help detect and prevent them.

Three things changed at once

In the last year, three shifts landed together. Any one of them would strain a verification flow. Together, they break the assumptions most flows were built on.

1. The tools got better and cheaper. A year ago, a convincing faceswap took real skill. Now consumer apps do it in real time on an ordinary phone.

2. The attack moved inside the camera. Holding up a photo or wearing a mask is a presentation attack, and liveness checks were designed for it. Injection attacks are different. A virtual camera or emulator feeds synthetic video straight into the camera pipeline, so the liveness model never sees a physical artefact. It sees a clean digital feed.

3. Fraud became an operation. Fraud rings now run sessions like a production line, using messaging-app bots to coach real people through verification step by step. The person is live, the face is real, the device is real. Liveness passes. Often, so does manual review.

The volume makes this an operational problem, not just a technical one. An estimated 8 million deepfakes were shared online in 2025, up from 500,000 in 2023, according to UK government figures. In contact centres alone, deepfake fraud attempts rose more than 1,300% in 2024, according to Pindrop’s analysis of 1.2 billion calls. At that scale, you can’t review your way out.

Why faceswaps are so hard to catch

Faceswaps hide best because almost everything in them is real. Fully synthetic faces and voice clones still leave tells: hair boundaries, edge consistency, teeth. A faceswap starts from a real video of a real person and replaces only the face.

The background is real. The lighting is real. The body language is real. The only synthetic part is the face, and that’s exactly where generation models have improved most.

Automated systems face the same squeeze. A year ago, compression artefacts and blending boundaries gave most faceswaps away. Today the best tools produce output that sits within the noise of a low-quality phone camera. Any single detection signal is now being gamed, so detection has to draw on several at once.

Detect deepfakes. Protect against coercion.

Discover how Veriff Presence Verification helps detect deepfakes and identify signs of coercion through advanced, real-time session analysis.

Explore Presence Verification

Confidence isn’t accuracy: the limit of manual review

The quiz result matches what we see in live verification flows. Confidence and accuracy have come apart. An experienced reviewer builds strong pattern recognition, but it was trained on last year’s fakes, and the tools move faster than intuition can follow.

Even trained specialists don’t always agree. When experienced fraud and quality teams review the same sessions, they sometimes reach different conclusions on whether a face is synthetic. If specialists disagree, asking a bank’s operations team to spot faceswaps at scale is unrealistic.

Volume and speed make it harder still:

• Capacity: attack volume grows faster than any review team can.
• False confidence: reviewers miss fakes even when they believe they’re catching them.
• Latency: by the time a human looks, the payout, transfer or account creation has often already happened.

Where manual review still earns its place

People bring context and judgement that models don’t. When an automated decision is inconclusive, a reviewer can weigh the document, the session and the account history together. Review is also worth keeping for edge cases where a false decline is costly, such as locking a genuine customer out of their own account during recovery.

What review shouldn’t be is the main line of defence in high-volume, real-time flows. There, it should handle the exceptions that automation flags, not carry the load.

Closing the gap with layers

No single check can stop faceswaps. That’s why effective detection relies on multiple signals working together in real time. Presence Verification brings these checks together, including injection attack detection, to identify and prevent faceswap attempts.

Liveness answers one question: is there a live human face in front of the camera right now? Presence Verification answers a bigger one: is this a real person, acting on their own, right now? It looks at the whole session, including the device, the camera feed, the video, the audio and patterns across sessions.

Faceswap detection sits inside Presence Verification, as part of injection attack detection. It aims to catch synthetic media before it reaches the liveness model.

Three tiers of signals

1. The gate: device integrity and injection detection. These checks run before the selfie is captured. Is this a real device or an emulator? Is the camera feed genuine, or is software injecting synthetic video? If they fail, the session is declined straight away, with no image to process.

2. Capture: passive liveness and deepfake detection. We analyse 100 frames of passive video, with no instructions to the user, in under two seconds. This catches presentation attacks and synthetic media that got past the gate.

3. Session: intent, coercion and cross-session patterns. Audio and video cues show whether someone is acting freely or being coached. Cross-session signals, such as velocity, multi-device activity and account-farm patterns, show when one operation sits behind many accounts.

Why the gate matters

Stopping emulators and virtual cameras before the selfie does two things. It’s cheaper, because there’s no image to process or store. More importantly, it breaks the economics of bot farms. If every attempt needs a real device instead of one of hundreds of emulator instances, the cost of each fraudulent account goes up sharply.

The takeaway

The threat model has changed, and most defences haven’t caught up. Faceswaps and coached sessions get past liveness. Manual review can’t scale against AI-generated attacks, and a reviewer’s confidence is a poor guide to accuracy.

The answer is layered: device integrity, injection detection, passive liveness, session-level signals and cross-session intelligence, all running in real time. Before the payout, before the transfer, before the account is compromised.

That’s Presence Verification. Not a better liveness check, but a different kind of defence.

See how Presence Verification fits your flow. Book a demo.