How to spot a deepfake: the key signs of AI-generated photos and videos

What is a deepfake and how does it differ from a regular fake?

Deepfakes are content (photos, videos, or audio) created or altered using neural networks to appear real: one person's face is «transferred» to another's body, a voice is reproduced as if the person actually spoke, and AI image generation creates a situation that never existed. The term comes from a combination of «deep learning» and «fake».

The difference from a classic fake is fundamental. Good old Photoshop is a manual process, and the traces of such manipulation are sooner or later noticeable: inconsistent shadows, sharp edges, odd proportions. A deepfake, on the other hand, is created by a neural network trained on thousands of real images and reproduces facial expressions, lighting, and camera angles so naturally that the line between reality and fiction blurs much faster than people can get used to.

Why do people create deepfakes?

Don't immediately assume collusion — most deepfakes are used harmlessly: as entertaining filters, dubbing an actor's voice in a different language, or restoring old photos. But the same technology is actively used for fraud. In one high-profile case, a Hong Kong company employee transferred approximately £25 million to attackers after a video call in which a deepfake impersonated his boss personally ordering the transfer. This is deepfake phishing — when a fake face or voice is used not for entertainment, but to trick the victim into believing something that didn't happen and making a decision that harms them.

Therefore, the question, «Is it possible to verify a video's veracity?» is no longer just an abstract curiosity, but a practical cybersecurity skill that everyone should have.

Physical errors are what neural networks produce most often so far

Physical errors in AI are the most reliable guide for the untrained eye, because this is where generative models still stumble most often. What to look for:

• Hands and fingers. Classic signs of AI-generated photos include an extra or missing phalanx, fused fingers, or an oddly curved hand. Neural networks are significantly worse at «understanding» hand anatomy than faces.
• Teeth and eyes. Teeth that are too even, as if painted on; reflections in the eyes that don't correspond to the light source in the frame; asymmetrical or «frozen» pupils.
• Ears and jewelry. Earrings of different shapes on the left and right ears, chains that end in nothing — details that humans typically reproduce symmetrically, but neural networks not always.
• Hair and object boundaries. Strands of hair that seem to «dissolve» into the background, or a sharp, unnatural outline around the head, are another common AI anomaly in images.
• Shadows and lighting. A shadow falls in the opposite direction than it should, based on the logical direction of the light source in the frame, or is completely absent where it should be.
• Background. Often, it's the background that reveals the most: blurred or «melting» textures, lines that don't converge, or sign writing that devolves into a meaningless jumble of symbols.

In addition to obvious signs, there are more subtle, psychological criteria that can be trained to notice even when the image appears flawless:

• Symmetry. A natural face typically has slight asymmetry — a slightly drooping eyelid, a smile that leans more to one side. AI tends to «smooth out» these minor imperfections, so an overly symmetrical, perfectly balanced face is a reason to take a closer look.
• Proportionality. You've probably noticed that real faces are rarely perfectly proportioned: they can have a large nose, protruding ears, or other features that deviate from the canons of beauty. Generated faces almost never have such features.
• Attractiveness AI faces, on average, look more attractive than real ones — the model strives for an aesthetically pleasing result, rather than the random, sometimes «uncomfortable» truth of a real face.
• Uniqueness. A real face has something that makes it stand out from the crowd. Generated faces often gravitate toward average parameters and therefore appear somewhat typical, lacking a distinct personality.
• Emotional expressiveness. Real faces convey emotions naturally and unevenly. AI faces often appear emotionally muted, as if «frozen» in a neutral expression.
• Memorability. AI-generated faces are, on average, less memorable — precisely because they lack those unique, “imperfect” details that make a real face recognizable.

There's another pattern worth being aware of: AI is noticeably worse at reproducing the faces of older people, children, and people of color, primarily because the datasets used to train these models are historically skewed toward young, white faces. The likelihood of visual defects in such images is statistically higher, and this should also be taken into account during verification.

Added to this list are signs of a deepfake in video: unnatural blinking (too infrequent or absent), frozen facial expressions, lips out of sync with the sound, or facial sharpness that is noticeably sharper than the rest of the frame — as if the head were «glued» onto the body. These are the inconsistencies in deepfakes that are the hardest to disguise, even for the newest models.

Technical verification methods

Besides a careful examination, there are also instrumental methods for recognizing the results of AI creativity:

1. Reverse image search (for example, via Google Images or TinEye) helps determine whether a photo is actually an old shot taken from a different context and passed off as something new — this is a different type of forgery, but this method often works as a first check for originality.
2. Checking the file's metadata (creation date, camera model, editing program) is another effective method, although metadata can be easily removed or replaced, so it is more of an additional, rather than conclusive, proof.
3. Specialized deepfake detection tools are also emerging — services that analyze an image or video using a neural network trained specifically to recognize synthetic content and provide a percentage probability of forgery. None of these tools provide a 100% guarantee, but they are useful as an additional filter, especially when visually distinguishing a fake is already difficult.




4. It's worth knowing specifically about the Content Authenticity Initiative, founded by Adobe and The New York Times, and the related open standard C2PA, which is supported by Google, Microsoft, Meta, OpenAI, and other major players. The idea is simple: cryptographically signed data about a file's origin — what device or program it was created on, what edits were applied — is attached to it from the moment it's created. If a platform supports this standard, users can verify the file's «history» rather than just guessing at it. Social networks like Instagram, Facebook, TikTok, and YouTube are also gradually implementing their own labels for AI-generated content, although these systems aren't always effective.

Why you shouldn't rely solely on your eyesight

AI-driven fake news is a separate risk category that overlaps with all of the above. A fake photo or video embedded in a plausible text context evades even the most attentive people precisely because we tend to trust our own eyes more than text. Research shows an interesting pattern: without training, people recognize deepfakes slightly better than chance, but after a short training session with examples, recognition accuracy increases significantly. Simply put, this is a skill that can and should be developed, not an innate talent that you either have or don't.

The practical approach that works best is to not rely on a single feature, but to check holistically — physical details in the image, context (where and by whom it was first published), reverse image search, and, if possible, technical verification tools. If something in a photo or video evokes even the slightest feeling of «something's not right», that's a reason to take a closer look, not to ignore it.

A stable Internet connection as an additional verification factor

Cybersecurity and deepfakes aren't just about recognizing fakes on screen, but also about the security of the connection delivering the content. Image verification using reverse image search services, metadata analysis, or downloading videos for frame-by-frame inspection — all of this only works quickly and smoothly on a stable connection. Maxnet's fiber-optic Internet provides precisely this kind of speed and connection stability — useful not only for streaming or video calls, but also when you need to quickly verify suspicious content before believing it, let alone acting on it.

Fraud using AI-powered content is unlikely to disappear anytime soon — in fact, technology continues to improve. But the habit of verifying rather than automatically trusting, along with a basic knowledge of where to look, already significantly reduces the chances of becoming the next victim of a convincing fake.