Deepfakes: How to Spot Them

• A. A type of encryption used to protect video files
• B. Synthetic media in which a person's face, voice, or body is convincingly replaced or altered using artificial intelligence ✓
• C. A search engine algorithm that ranks images by quality
• D. A hardware component used to speed up video rendering

Why: The term 'deepfake' refers to synthetic media created with AI, typically deep learning, that realistically swaps or manipulates a person's face, voice, or body so they appear to say or do things they never did.

• A. Decision trees
• B. Linear regression
• C. Generative Adversarial Networks (GANs) ✓
• D. K-means clustering

Why: Generative Adversarial Networks, introduced in 2014, pit two neural networks against each other: a generator that creates fake images and a discriminator that tries to detect them. This competition produces increasingly realistic outputs, which is why GANs became central to early deepfake tools.

• A. It compresses the video file to reduce its size
• B. It adds background music to the generated video
• C. It tries to distinguish real images from fake ones, pushing the generator to improve ✓
• D. It stores the training data in a database

Why: The discriminator acts as a critic: it evaluates whether an image is real or generated. Its feedback helps the generator learn to produce outputs that are harder to tell apart from authentic footage.

• A. The creators intentionally add these flaws as a watermark
• B. Training data may lack varied examples of blinking or face angles, and blending the synthetic face onto the original frame can leave visible artifacts ✓
• C. Deepfakes are always recorded in low resolution by law
• D. Video cameras automatically blur faces for privacy

Why: Deepfake models learn from limited training images, which can underrepresent behaviors like blinking, and the process of blending a generated face onto the target footage often leaves edge blurring or color mismatches. These imperfections are common clues, though they are shrinking as the technology improves.

• A. By recording a person speaking and sending the audio through a noise filter only
• B. By training a model on samples of a target's voice so it can generate new speech in that voice saying arbitrary content ✓
• C. By lowering the pitch of an existing recording
• D. By transcribing speech into text and reading it aloud with a standard system voice

Why: Voice cloning models learn the acoustic characteristics of a specific speaker from sample audio, then synthesize new speech that sounds like that person saying words they never spoke. Simple pitch shifting or filtering does not create convincing new content.

• A. Photogrammetry
• B. Double exposure and photo retouching ✓
• C. Holography
• D. Stop-motion animation

Why: Photographers and filmmakers used double exposure, darkroom retouching, and painting over negatives to alter images well before computers. These analog methods show that manipulating visual evidence is far older than deepfakes.

• A. Forrest Gump ✓
• B. Jurassic Park
• C. The Matrix
• D. Titanic

Why: Forrest Gump (1994) digitally placed Tom Hanks alongside real figures such as John F. Kennedy and Elvis Presley. It was a landmark demonstration of convincing digital human insertion in footage.

• A. A 1980s academic paper on neural networks
• B. A 2017 Reddit username that posted face-swapped videos ✓
• C. A 2001 Hollywood visual effects manual
• D. A 2010 government report on cybersecurity

Why: The term emerged in late 2017 when a Reddit user posting face-swapped adult videos used the name 'deepfakes'. The word combines 'deep learning' with 'fake' and quickly became a general label for AI-generated synthetic media.

• A. Convolutional Neural Networks only
• B. Generative Adversarial Networks (GANs) ✓
• C. Decision Trees
• D. Support Vector Machines

Why: Ian Goodfellow and colleagues introduced GANs in 2014, pitting a generator against a discriminator. This adversarial training produced increasingly realistic synthetic faces and images, becoming the foundation for many early deepfake tools.

• A. Audiences in the 1990s could not perceive realistic faces
• B. Computing power, training data, and face-modeling algorithms were far more limited ✓
• C. Governments banned realistic digital humans until 2015
• D. Film cameras could not capture enough detail for compositing

Why: Early digital humans were limited by slow processors, small datasets, and crude models of skin, lighting, and movement. Advances in GPUs, large image datasets, and deep learning later made realistic synthetic faces far more achievable.

• A. The image has a high resolution
• B. Hair edges look blurry, smeared, or unnaturally merged with the background ✓
• C. The image was taken on a smartphone
• D. The subject is wearing bright colors

Why: AI-generated or face-swapped images often struggle with fine details like individual hair strands, which can appear blurred, smeared, or blend unnaturally into the background. Resolution, camera type, and clothing colors do not indicate manipulation.

• A. Shadows on a face fall in a different direction than shadows on the objects or background ✓
• B. The image is in black and white
• C. The photo has a visible watermark
• D. The subject is smiling

Why: When elements from different sources are combined, their lighting often does not match. If shadows on a person point one way while shadows elsewhere point another, the image may be composited or altered.

• A. The eyes and the reflections within them ✓
• B. The number of freckles
• C. The length of the subject's eyebrows
• D. The shape of the ears only

Why: Deepfakes often produce eyes with mismatched or missing catchlights (reflections), irregular pupil shapes, or inconsistent reflections between the two eyes, since these details are hard for generation models to replicate accurately.

• A. Text is always distorted in digital photos
• B. AI image generators often struggle to render coherent letters, signs, and logos accurately ✓
• C. Text is added only in printed photos
• D. Distortion happens only when the image is compressed for social media

Why: Many image generation models produce garbled, misspelled, or nonsensical text on signs, clothing, and packaging because they learn visual patterns rather than the rules of written language. Such errors are a useful warning sign.

• A. Immediately share it to see whether others agree
• B. Assume it is real because it looks professional
• C. Use a reverse image search and check whether the image appears in trusted sources with its original context ✓
• D. Delete it so it cannot spread

Why: Reverse image search can reveal the earliest appearance of an image and whether it was taken out of context or altered. Checking reliable sources and original context is a practical verification step, while sharing or assuming authenticity can spread misinformation.

• A. The speaker uses a regional accent
• B. Unnatural pauses or breathing patterns that don't match the speech rhythm ✓
• C. The recording includes background music
• D. The speaker mentions a specific date

Why: Voice clones often struggle to reproduce natural breathing, pacing, and pauses. Speech that sounds rhythmically off, or breaths that are missing or placed in odd spots, can signal synthetic audio. Accents, background music, and dates are not reliable indicators on their own.

• A. A recording of a person reading a script word for word
• B. A synthetic voice model trained to mimic a specific person's voice ✓
• C. A voice assistant like those built into phones
• D. An audio file that has been compressed to a smaller size

Why: A voice clone is an AI-generated model trained on samples of a real person's voice so it can produce new speech that sounds like them. Voice assistants use generic voices, and compression changes file size, not identity.

• A. The volume level of the recording
• B. Emotional tone and inflection across different sentences ✓
• C. The language spoken
• D. The file format of the audio

Why: Cloned voices often sound flat, or show emotional shifts that don't fit the words being said, such as cheerfulness on serious statements. Volume, language, and file format are easy to change and are not specific markers of cloning.

• A. Phone calls always have lower audio quality
• B. Scammers can now use short audio samples from social media to clone a voice ✓
• C. Phone companies are required to record all calls
• D. Familiar voices are always fake in urgent situations

Why: Modern tools can create a convincing clone from only a few seconds of publicly available audio, such as a video post. Urgency combined with a request for money is a classic social engineering pattern, so familiarity alone should not be trusted.

• A. Ask the caller to speak louder
• B. Hang up and call back using a number you already have on record, or agree on a private code word in advance ✓
• C. Ask the caller to repeat the sentence they just said
• D. Listen for whether the caller sounds tired

Why: Calling back on a trusted, independently obtained number breaks the attacker's control of the channel. A pre-agreed code word gives you a verification method an impersonator cannot know. Asking for repetition or judging tiredness can be easily faked by a clone.

• A. A person blinking at a natural rate
• B. Inconsistent or unnatural blurring along the edges of the face, especially around the hairline and jaw ✓
• C. Clear, well-lit footage with a steady camera
• D. Audio that is perfectly synchronized with lip movements

Why: Face-swapping algorithms often struggle to blend the synthetic face with the original head, hair, and background, producing blurring, shimmering, or mismatched edges around the hairline and jaw. Clean, well-lit footage and good lip sync are not reliable signs of authenticity by themselves, since they can be produced intentionally.

• A. Deepfake software deliberately blinks more to appear human
• B. Training datasets often contained few images of people with closed eyes, so the generated faces blinked rarely or unnaturally ✓
• C. Blinking is impossible to simulate digitally
• D. Cameras automatically remove blinks from recordings

Why: Many face-generation models were trained on photos where subjects' eyes were open, since still photographs rarely show people mid-blink. As a result, early deepfakes often showed too little blinking or blinking that looked mechanical. Newer models have improved, so blink analysis is a clue rather than a guarantee.

• A. Lighting that is always perfectly even across the whole scene
• B. Shadows and highlights on the face that do not match the direction and color of light in the surrounding environment ✓
• C. Lighting that changes every time the camera moves
• D. Any bright highlights, which always indicate manipulation

Why: A common inconsistency in manipulated video is that the lighting on the synthetic face does not match the scene, such as highlights coming from a different direction or a skin tone that looks slightly off compared with the neck and ears. Checking whether shadows agree with the light sources in the environment is a useful inspection step.

• A. Increasing the playback speed to at least four times normal
• B. Pausing and zooming in on the mouth, teeth, and reflections in the eyes to look for distortions or inconsistencies ✓
• C. Converting the video to a lower resolution to hide artifacts
• D. Only listening to the audio track without watching the video

Why: Slowing down and zooming in lets you inspect fine details that are easy to miss in real time, such as teeth that blur or merge, unusual reflections in the eyes, and mouth shapes that do not quite match the sounds. Frame-by-frame review is a standard part of video verification.

• A. Every real video contains manipulation artifacts from compression
• B. Compression, low resolution, poor lighting, and camera motion can produce similar artifacts in authentic footage ✓
• C. Deepfake detection is only possible with audio, never with video
• D. Artifacts only appear in videos recorded after the year 2020

Why: Video compression, low bitrate uploads, motion blur, and poor cameras can all create blurring, blockiness, and odd edges in genuine recordings. Good verification combines several cues, checks the source and context, and looks for corroborating evidence rather than relying on one visual clue.

• A. Judge it by how realistic the lighting and skin textures look
• B. Find the original source and the earliest known version of the image
• C. Ask whether the image makes you feel angry or emotional
• D. Check whether the image has a watermark in the corner

• A. To automatically delete edited copies of a photo from the internet
• B. To find other places a photo appears online, including earlier or original versions
• C. To measure the exact pixel resolution of an image
• D. To confirm the identity of every person shown in the picture

• A. The photo is in color
• B. The photo has no clear source, and its metadata or earlier appearances point to a different date or event
• C. The photo was posted by an account with a profile picture
• D. The photo is a close-up rather than a wide shot

• A. Real photos always contain hidden manipulation markers
• B. Photos can be taken out of their original context, such as a real image from a past event presented as recent
• C. Cameras automatically add false details to every image
• D. Genuine photos cannot be shared without being altered

• A. The first repost you see with a dramatic headline
• B. An anonymous comment thread discussing the image
• C. The original publisher or a verified eyewitness, cross-checked with reputable news or fact-checking outlets
• D. A post from an account that recently gained many followers

• A. Share it quickly so friends can give their opinions
• B. Check whether reputable news outlets or fact-checkers have already reported on it
• C. Assume it is real because it looks very convincing
• D. Delete it so it cannot spread further

• A. Improving the video resolution
• B. Finding earlier appearances of an image or video frame to check whether it was taken out of context or altered
• C. Removing the audio track from a clip
• D. Confirming the exact time a video was uploaded by its creator

• A. Original sources always have higher resolution files
• B. The earliest uploaded version is often closer to the original recording and can reveal edits, missing context, or a misleading caption
• C. Only official government channels can host authentic video
• D. Source checks are only needed for text articles, not video

• A. Trusting content more when it confirms what you already believe
• B. Pausing to consider your emotional reaction before sharing, since content designed to provoke strong feelings is a common manipulation tactic
• C. Sharing only content that is longer than five minutes
• D. Relying on a single account that you follow regularly

• A. Asking the person who posted it whether it is real
• B. Finding independent sources, witnesses, or records that support the same claim before accepting it as true
• C. Checking how many likes or shares the post has received
• D. Confirming that the video has subtitles in your language