How to Detect a Deepfake: 8 Visual Tells That Reveal AI Face Swaps

The seam between a swapped face and the original head is the hardest area for face-swap algorithms to get right. Look for a semi-transparent, slightly blurry halo around the perimeter of the face — particularly along the forehead, jaw, and where the face meets the ears. In real photographs, there is no visible boundary around a face. In a deepfake, the compositing step that blends the face into the scene typically leaves a 2–5 pixel zone of softness or colour inconsistency. This is most visible at high magnification, or when the subject turns their head and the boundary angle changes.

When a face is swapped, the source face comes from a different video or image with different lighting, camera settings, and colour grading. Even after colour correction, there is often a subtle but visible difference between the skin tone of the swapped face and the original person's visible neck, ears, or hands. The face may appear slightly warmer, cooler, more or less saturated, or have a different level of contrast than the surrounding skin. This mismatch is particularly visible in the transition zone between the jaw and the neck — look for a colour discontinuity where the face ends.

Every real photograph has a consistent light source — you can trace where the light is coming from by looking at the direction of shadows on the face, the position of catchlights in the eyes, and the highlights on the nose and cheekbones. In a deepfake, the source face was captured in different lighting conditions. Even after correction, the face often has highlights and shadows that don't match the rest of the scene. The give-away: the light on the face appears to come from a slightly different direction than the light on the shoulders, clothing, or background. Check which side of the nose has the highlight, and compare it to the shadow direction on the background.

This tell applies primarily to video deepfakes, but it's one of the most reliable. Human beings blink 15–20 times per minute in natural conditions, with each blink taking about 150–400 milliseconds. Early deepfake models (2018–2020) famously struggled to reproduce blinking naturally — many produced faces that almost never blinked. Modern models handle basic blinking better, but look for: one eye blinking more than the other (asymmetric blink); blinks that happen mid-sentence at inappropriate moments; or blink timing that feels robotic or slightly off-rhythm. The eye area is also where boundary artifacts are most visible when the face is in motion.

In deepfake video, the face is generated frame-by-frame and composited back onto the original footage. Because each frame is processed independently (in most face-swap pipelines), there can be slight variations in face position, scale, or colour between adjacent frames. The result is a subtle flickering or shimmer on the face that is absent on the body and background. This is most visible when the video is paused and scrubbed frame-by-frame, or when viewed at half speed. Look for: the face appearing to subtly shift position by 1–2 pixels between frames; hair at the face edges flickering; or a slight texture or colour variation in the face region that doesn't appear elsewhere.

Ears are frequently the least-well-reconstructed area in face swaps. The source face typically doesn't include the full ear, and the deepfake algorithm must either use the original person's ear or blend the two. Signs: an ear that looks different from the rest of the face in skin texture or colour; an ear that has a visible boundary on the inner edge where the swapped face ends; an ear that shows signs of being digitally reconstructed (too smooth, lacking the usual cartilage texture detail); or earrings that appear on the face side but not the ear side, or vice versa. If the subject is wearing earrings, compare both ears carefully.

In video deepfakes that also alter what someone appears to be saying (sometimes called "audio-driven deepfakes"), lip sync is a common failure point. The generated mouth movements may not perfectly match the phonemes of the audio — a bilabial consonant (p, b, m) requires the lips to touch, but the deepfake mouth may not fully close. The transition between phonemes may be slightly too fast or too slow. The lip movement may also not match the expected animation for the visible teeth and tongue. This is hard to detect casually, but if something about a video speech feels "off" — particularly around the mouth — lip sync is worth examining closely.

The absence of deepfake artifacts doesn't confirm a video is real, but some elements are very hard to fake. Highly specific and consistent anatomy: visible pores or skin texture detail that is consistent at scale across the face and neck. Natural micro-expressions — real faces have subtle muscle movements between expressions that deepfake models flatten out. Consistent spectacle reflections: if the subject wears glasses, the reflection in the lenses should correspond to the background scene, and should shift naturally as the head moves. Highly detailed backgrounds with readable text that remains consistent between frames. And body language that is anatomically natural — deepfakes sometimes struggle with the connection between the swapped face and the body's head rotation.