Deepfakes are AI-generated media that appear to be real but are in fact fake.
Deepfakes can be hard to spot, especially if they’re edited on top of being AI-generated. But if you examine them closely and carefully, you can recognize them.
In this blog, we will take a look at how to recognize deepfakes and tell them apart from real images. The signs listed in this guide will help you do it.
What are Deepfakes?
Deepfakes can be defined as AI-generated visuals that are, in fact, fake but look too real and believable to be considered fake.
The visual formats could include images, videos, and audio, and by extension, anything of visual form, like GIFs, and even live cam recordings.
Deepfakes are often made to represent real-life people, like celebrities or politicians, and are potent tools for stirring up discussions on social media and the internet.
For example, a deepfake might appear to be a “recording” of a celebrity making critical remarks about another celebrity, while the event in fact never occurred.
Some deepfakes, on the other hand, may feature ordinary people and be used for scams and deceptions.
For example, deepfakes of people from a made-up company video-calling a potential client and securing a monetary contract, while in reality, those people don’t actually exist and are AI deepfakes.
Common deepfake formats are images, videos, and audio. A deepfake image could be a selfie of an individual who never took it.
A video deepfake could be a clip of an individual doing something that they never did. Whereas an audio deepfake might simply be a scam call.
Why Are Deepfakes a Concern?
Deepfakes are a concern because they show real-life people, places, or events that never occurred.
These can be incredibly convincing, and an average person is likely to find them believable and take them for granted.
Deepfakes open up a door that was previously hard to open. These have made propaganda, scams, and blackmail way easier to do.
Previously, creating a fake image, video, or audio needed hours of complex editing and a high level of skill, but now, faking someone’s actions and words only takes an AI tool, some model training, and a bit of editing to smooth out the flaws.
Although the deepfake technique itself isn’t evil, it’s the intent of the person using it.
Someone might create a deepfake to just fool a friend or colleague, but the broader use of the technology is for deception, because that’s what deepfakes boil down to: Fake visuals that appear real. (The colleague gets fooled = gets deceived.)
This means we can expect cyber crimes like impersonation and scams to increase and also become harder to control. This is why deepfakes are a concern.
Additionally, the use of deepfakes is also considered illegal under some laws against cybercrimes in many countries, and can get you into legal risk if you’re caught.
How to Spot a Deepfake?
You can spot a deepfake by examining the content and spotting unusual flaws, like glitches or limb distortions in the people depicted.
Spotting Deepfake Videos
Here are some signs to look for in deepfake videos:
Unnatural facial movements: One of the flaws in deepfakes is unnatural facial movements, like unnecessary or involuntary tensing of lips and strange eye blinking.
Lip-sync issues: If a deepfake video has audio, there’s a chance the audio might not be synchronized well with the person’s lisping and mouth movements. If the video is convincing but the lip-syncing is off, it might be a deepfake you’re looking at.
Unusual eye movements: Eyes and their movements in particular can point to AI’s involvement.
Lighting and background mismatches: Deepfake content tends to have imperfect lighting, such as the foreground lighting not matching the background lighting.
Flickering or glitches: AI-generated visuals tend to have unrealistic flickering and glitching of pixels.
Strange yet smooth transitions or movements: Some deepfake visuals might have peculiar yet overly smooth transitions of scenes or movements. These might be subtle, but can be noticeable.
Spotting Deepfake Images
Look for the following signs in deepfake images:
Smooth skin texture: Skin textures in deepfake images often appear too smooth and perfect.
Distortions in facial features: Faces tend to look realistic in deepfake images, but, at closer inspection, might reveal strange features, like asymmetrical features and strange shadows.
Irregularities with eyes and glasses: Eyes tend to be a major giveaway in deepfakes. Look for pupils that are misaligned or are in different sizes. They might also have an unusual shape. Plus, light reflections in the eyes might not make sense. Generating flawless glasses, especially with the challenge of light reflections, can also be a challenge for AI. The glasses’ frames might be uneven, crooked, unequal, or have parts that disappear or blend into the face.
Unusual teeth: Teeth are also difficult for AI to render like eyes. If you see an image in which the teeth of the person are visible, look to see if they appear distorted or misaligned.
Unusual mouths: Pay extra attention to the mouth and lips. The mouth area can also look kind of smudged.
Distorted hands and limbs: Hands, in particular, are challenging for AI models to draw flawlessly and remain a common flaw in deepfake images, especially the fingers. Fingers might be the wrong number (six fingers) or distorted and strangely lengthy or short.
Inconsistent lighting and shadows: The lighting on the subject's face might not match the lighting in the background. Shadows may fall in illogical directions, as if there are multiple light sources where there shouldn't be.
Unsetting background: An AI model might put most of its focus on drawing the subject and leave out the background as if it’s rushed through. Deepfake background might appear sort of low-quality and can be a good indicator of the image’s origin. Look for weird, distorted shapes, repeating patterns, or examine if it feels like "patchworked." Objects and people in the background may also appear revised, strange, or unfinished.
Blurriness or lack of detail in specific areas: While the face might be a high-resolution deepfake, other parts of the image, like the hair, clothing, or background, can be blurry or smooth and lack the detail or real-life feeling of real images.
Spotting Deepfake Audios
Look for the following signs in deepfake audios:
Robotic wording: The speech of a person might have robotic wording.
Unnatural rhythm of speech: AI-generated speech doesn’t have the same natural flow and rhythm as a human voice. Its pacing sounds steady and robotic. The speaker tends to speak at an overly even tempo, which is why sentences might end without the natural rise or fall in pitch, making it feel flat.
Missing pauses and breaths: If you’ve ever video-called someone, you’d know how real people talk. They take breaths and sometimes pause to think. Sometimes their speech is punctuated with filler words, like "um" and "uh." AI-generated voices and speech don’t have these nuances. The speaker might not take a pause or breathe while speaking. They might also not use any filler words, although this can be changed by including filler words in the script. But the overall speech sounds too smooth to be real.
Pronunciation errors in some other languages: Due to training limitations, AI tools aren’t good at speaking in some languages other than English.
Perfect pronunciation: Humans tend to mispronounce words as they speak in an everyday manner. For example, native English speakers won’t necessarily pronounce every word perfectly. In fact, most people don’t force perfect pronunciation and might even skip some of the syllables. But AI-generated speeches tend to pronounce every word with near-perfect accuracy.
Faked emotional depth and lack of nuance: AI voices lack the understated depth and authenticity of a human's emotional expression. AI-generated speech tends to sound a little stilted, as if the speaker is Strong-willed on pronouncing every word with perfection, which sounds stiff and emotionless. Their emotion might sound forced or overstated and not match the context of the words being spoken. For example, a voice might sound emotionless while delivering a happy message, unlike a human’s voice that would sound as if the speaker is happy.
If you’re examining deepfake content, take into account how many signs or flaws stack up. If a piece of content has too many signs, chances are it is a deepfake.
Conclusion
“Deepfake” means AI-generated content that looks real but is actually fake, showing people, places, or events. Deepfakes can be hard to spot because they are made to look real. Sometimes, deepfakes are edited so that they don’t have errors and look convincing.
However, you can inspect deepfakes closely and try to spot some of the signs and common flaws. For deepfake videos, look for strange facial movements, lip sync issues with the audio, unusual eye movements, lighting and background mismatches, subtle glitches or flickering, and smooth transitions.
Upgrade Now
