Home › News

How to spot deepfakes in video calls

26.09.2026

You are staring at a screen. The person on the other end is strikingly attractive, highly attentive, and seemingly eager to fulfil a specific fantasy. In the realm of virtual companionship and intimate online encounters, this scenario is common. It is also the perfect hunting ground for deepfake technology. When the goal is intimate virtual https://hotvirt.com/virt-sex/service/custom-photos interaction, the emotional and financial stakes are high, making the ability to distinguish a live human from an AI-generated puppet an essential skill rather than a luxury.

Why virtual intimacy invites deception

The intersection of desire, anonymity, and distance creates a lucrative opportunity for operators running deepfake schemes. Real-time face-swapping technology has evolved from obvious, glitchy novelties into sophisticated software overlays. An operator can now route a generated face through a virtual camera, broadcasting it seamlessly to almost any video platform. The trade-off is clear: the barrier to entry for deploying this software is dropping rapidly, while the potential payout for convincing someone they are interacting with a genuine partner remains high. Understanding this motivation is the first step in evaluating the authenticity of a call.

The mechanics of a real-time deepfake

To spot the deception, you need to understand how the software functions. A typical setup involves an operator sitting at a computer. They use a source video—either a pre-recorded loop of a model or a live feed from a third-party camera—and apply a deepfake model trained on images of the desired "girl." The software maps the source's facial expressions onto the target face, then outputs this composite to a virtual camera driver. The video call application then reads this virtual camera as if it were a physical webcam. Because the processing happens in real-time, the computational load forces the software to make compromises in rendering quality.

Essential visual checks

Before demanding complex proofs, observe the baseline presentation. Current real-time deepfake models, while improving, still struggle with micro-expressions, consistent rendering, and physics.

• Edge bleeding: Watch the boundary where the face meets the hair, neck, or background. You will often see a subtle blur, a shimmering line, or hair that moves as a solid block rather than individual strands. The algorithm struggles to define a clean edge in motion.
• Blink rate and eye movement: Humans blink roughly 15 to 20 times per minute, and their eyes dart naturally. Deepfakes often blink at unnatural intervals—either too rapidly or with long, unsettling stares. The pupil dilation also rarely reacts to changes in screen brightness or ambient light.
• Lighting mismatches: Compare the lighting on the face to the environment. If the room has a warm overhead light but the face has a cool, flat studio illumination with no corresponding shadows on the neck, the face is likely an overlay sourced from different conditions.
• Jewellery and accessories: Earrings, necklaces, and glasses frames often deform, merge with the skin, or change shape when the head turns. Algorithms lack the spatial understanding to render these rigid objects consistently during movement.

Audio-visual synchronisation

A common flaw in live deepfakes is latency. The operator must speak or type, the AI must generate the corresponding facial movement, and the mapping software must project it onto the video feed. This pipeline introduces delays.

• Lip sync: Watch the mouth carefully. Are the phonemes matching the sounds? Early models produce a generic mouth flap; newer ones are better but still lag during rapid speech or emotional outbursts.
• Voice generation: Many operators use real-time voice changers. Listen for a metallic resonance, a lack of natural breath sounds, or a voice that stays at a perfectly consistent volume and pitch regardless of the emotional context.
• Processing delay: If there is a noticeable pause before the person reacts to a complex question or a sudden event, the operator might be typing a prompt for an AI language model to generate the response, which is then fed to the voice and face software.

Interactive challenges: forcing the system to break

Passive observation is useful, but active testing is definitive. Real-time rendering software has limited computational overhead. If you force the avatar outside its trained parameters, it will deform. This is where you distinguish a genuine connection from a puppet.

1. The object test: Ask the person to hold up a specific number of fingers, or an everyday object like a pen or a piece of paper with a word written on it. Deepfakes struggle immensely to render hands interacting with the face or holding objects in real-time. If they refuse, claim their camera cannot focus that close, or if the hand turns into a mangled claw, you have a strong indicator of a fake.
2. The occlusion test: Ask them to cover part of their face with their hand. When a hand crosses the generated face, the software often fails to understand the depth hierarchy. The face will bleed onto the hand, the hand will become transparent, or the tracking will glitch entirely.
3. Rapid motion: Ask them to turn their head quickly, look up at the ceiling, or stand up and spin. Fast or extreme movement stresses the tracking algorithms, often resulting in the face sliding off the head, lagging behind the body's rotation, or melting temporarily before re-locking.

Evaluating the platform and software trade-offs

Where the call happens matters, though not always in the ways you might expect. End-to-end encrypted platforms like Signal or standard conferencing tools like Zoom do not filter deepfakes, but they also do not typically apply beauty filters that might mask deepfake artefacts. Camming sites or dedicated virtual dating platforms sometimes employ detection algorithms, though these are routinely bypassed by virtual cameras.

The critical factor is the software routing the camera. If the other person's video feed lists a virtual camera (such as "OBS Virtual Camera") rather than a hardware device (such as "FaceTime HD Camera" or "Logitech Webcam"), it is a significant red flag. While OBS is used legitimately by streamers, in a private, one-on-one intimate call, its use strongly suggests an intermediary software layer where a deepfake can be inserted.

A pragmatic approach to verification

Not every hesitation is proof of a deepfake, and not every perfect image is AI. A poor internet connection can cause audio lag and visual artefacts that mimic deepfake glitches. Someone might refuse to show their face because of privacy concerns, not because they are a bot. The goal is to accumulate evidence rather than jumping to conclusions based on a single artefact.

Run your interactive challenges casually. Frame a request for them to hold up an object or turn around as part of the flirtatious dynamic rather than an interrogation. If the person consistently refuses all interactive challenges, exhibits multiple visual artefacts, and the interaction feels overly scripted or delayed, the probability of a deepfake is high. In the pursuit of virtual connection, protecting your wallet and your trust requires treating extraordinary encounters with ordinary scrutiny.