To detect a deepfake video, look for inconsistencies in facial movement, lip-sync, lighting, reflections, skin texture, background details, and frame-to-frame continuity. Then verify the source, search for earlier versions of the footage, check available provenance information, and use a deepfake detector when the authenticity of the video matters.

The most important rule is not to rely on one visual clue. A 2024 systematic review and meta-analysis of 56 deepfake studies covering 86,155 participants found overall human detection accuracy of 55.54%. For video specifically, pooled accuracy was 57.31%, and the confidence interval crossed chance.

Detection-improvement strategies such as feedback training and AI assistance raised pooled performance to 65.14%, suggesting that structured and tool-assisted verification can help, but still does not make human judgment infallible.

Automated detectors have limitations too. The Deepfake-Eval-2024 benchmark found that state-of-the-art open-source detector AUC dropped by 50% for video when models were evaluated on recent real-world deepfakes rather than older academic datasets.

That is why the strongest verification process combines visual inspection, audio-visual analysis, source checks, provenance, automated detection, and human review.

Key Takeaways

Check the face, lips, eyes, skin, hair, lighting, reflections, accessories, and background for inconsistencies.
Watch suspicious sections frame by frame because artifacts may appear only during movement.
Compare the audio with lip, jaw, facial, and body movement rather than evaluating the video silently.
Treat familiar signs such as abnormal blinking or distorted hands as clues, not proof because generators continue to improve.
Verify who originally uploaded the video and whether credible independent sources confirm the event.
Reverse-search distinctive frames to find older or original versions of the footage.
Check for Content Credentials, watermarks, or other provenance signals when available.
For consequential decisions, use a current deepfake detector and human review rather than depending on eyesight alone.

What Is a Deepfake Video?

A deepfake video is video that has been generated or manipulated using AI in a way that misrepresents a person, object, action, or event.

The UK Department for Science, Innovation and Technology describes deepfakes as a subset of synthetic media involving audio-visual content generated or manipulated with AI to misrepresent someone or something.

Manipulation can include:

Replacing one person's face with another
Changing facial expressions
Synchronizing new speech with existing footage
Generating a synthetic person
Altering a background or object
Manipulating only part of an authentic recording
Combining synthetic audio with real video
Generating an entire scene from a prompt

This means a deepfake is not necessarily a completely fabricated video. A clip might contain genuine footage with a synthetic voice. Another may preserve the speaker's body and environment while altering only the face and a third may be fully generated.

Detection is therefore less about finding one universal “AI mistake” and more about identifying inconsistencies between the different signals within the media.

How Good Are Humans at Detecting Deepfake Videos?

Current research offers an important warning against overconfidence.

The studies measure different populations, media, models, and experimental conditions, so their percentages should not be directly combined into one universal accuracy rate. They do, however, point toward the same practical conclusion: Neither “I can tell by looking” nor “the detector said fake” is a sufficient verification strategy by itself.

Visual warning signs are useful as an initial screening step. MIT Media Lab's Detect Fakes project recommends examining facial appearance, skin consistency, eyes, eyebrows, facial hair, blinking, and lip movement. Importantly, MIT explicitly notes that there is no single telltale sign that establishes a video is fake.

The FBI similarly advises looking for distorted features, unrealistic accessories, inaccurate shadows, unnatural movement, and audio or video inconsistencies when assessing suspicious media.

1. Look for Unnatural Lip-Sync

The mouth is one of the most useful places to inspect.

Watch whether:

Lip shapes correspond with the sounds being spoken
The mouth opens and closes at the right moments
Teeth remain consistent
The jaw moves naturally
Lip boundaries remain stable
Fast speech stays synchronized
Facial expressions match the words being spoken

A lip-synced deepfake may look convincing at normal speed but reveal inconsistencies when slowed down.

However, poor synchronization can also result from:

Dubbing
Video editing
Streaming latency
Conferencing software
Compression
Reposting

Lip-sync errors are therefore evidence to investigate, not proof of manipulation.

2. Examine the Face Boundary

Face-swapping systems need to blend an altered face with the original footage.

Pay attention to:

Jawlines
Cheeks
Ears
Hairlines
Forehead edges
Neck
Areas covered by hair
Eyeglass boundaries

Possible signs include:

Flickering
Blurry edges
Sudden changes in face shape
Skin tones that do not match the neck
Facial details disappearing during head turns
Hair merging unnaturally with the face

Transitions become especially informative when the subject turns their head or something passes in front of their face.

3. Check Skin Texture and Facial Detail

Synthetic facial skin may sometimes appear:

Too smooth
Waxy
Over-sharpened
Excessively uniform
Inconsistent between frames
Different in age from surrounding features

MIT specifically recommends comparing the cheeks and forehead with other facial characteristics and checking whether the apparent age of the skin is consistent with the person's hair and eyes.

Ask:

Do wrinkles remain in the same place?
Does skin texture change during motion?
Do pores or facial marks disappear?
Does the forehead match the cheeks and neck?

Modern generators have improved significantly, so smooth skin is not a dependable standalone test.

4. Watch the Eyes and Blinking

Early deepfake guidance often emphasized blinking because older systems sometimes reproduced it poorly.

Possible clues include:

Very little blinking
Excessive blinking
Unnatural gaze
Inconsistent eye shape
Strange reflections
Eyebrows that do not match expressions
Shadows around the eyes that conflict with lighting

But this tip requires an important 2026-era qualification. The U.S. Government Accountability Office specifically warns that advances in deepfake generation are expected to eliminate known artifacts such as abnormal blinking. So, normal blinking does not establish that a video is authentic. Use eye behavior as one clue among several.

5. Inspect Lighting, Shadows, and Reflections

AI-generated scenes must maintain the physics of light across objects and frames.

Look at:

Shadows around the nose
Highlights on the forehead
Reflections in the eyes
Eyeglass glare
Light on the face versus body
Reflections on nearby objects
Shadow direction
Changes in lighting as the subject moves

MIT recommends inspecting whether shadows around the eyes and eyebrows appear where expected because deepfakes can fail to reproduce the natural physics of a scene. The FBI also lists inaccurate shadows, lighting irregularities, and unrealistic accessories among possible signs of manipulated media. One strange reflection is not conclusive. Multiple inconsistent physical cues deserve more attention.

6. Look for Hair, Teeth, Hands, and Accessories That Change

Fine visual details can be difficult to preserve consistently over time.

Inspect:

Hair strands
Earrings
Necklaces
Eyeglasses
Teeth
Fingers
Tattoos
Moles
Facial hair
Clothing logos

A suspicious element may:

Change shape
Disappear
Merge with another object
Move independently
Switch sides
Become blurry
Look different after the subject turns

The FBI's current guidance specifically mentions distorted hands or feet, unrealistic facial features, irregular faces, glasses or jewelry, and unnatural movements as possible warning signs. Again, modern generators can reproduce these details correctly. Their absence does not establish authenticity.

7. Check Frame-to-Frame Consistency

A single screenshot may appear convincing while the full sequence contains inconsistencies.

Slow the video and inspect moments when:

The head turns
The subject speaks rapidly
Hands cross the face
Another person passes in front
Lighting changes
The camera moves
The person enters or exits frame

Look for:

Features changing between frames
Face flickering
Sudden texture changes
Objects appearing or disappearing
Identity features shifting
Impossible transitions
Background inconsistencies

Video detection systems also analyze temporal information because manipulation may become visible only when frames are considered together.

8. Compare the Audio With the Video

Do not evaluate a suspicious video only with the sound muted.

Listen and watch together.

Check whether:

Voice emotion matches the facial expression
Breathing corresponds with speech
Speech rhythm matches body movement
Pauses sound natural
Mouth movement follows spoken words
Room acoustics fit the visible location
Background noise remains continuous
Voice characteristics remain stable

This matters because a manipulated video may combine real footage with synthetic audio. A 2024 Nature Communications study involving 2,215 participants found that people performed better when more communication modalities were available. In one experiment, accuracy was 57% with transcripts, 64% with silent video, 81% with audio, and 86% with video plus audio. Those results came from a controlled political-speech experiment and should not be interpreted as an 86% universal human deepfake detection rate. The lesson here is to use all the information available rather than judging the face alone.

9. Examine the Background and Physical Environment

The face receives most of the attention in deepfake guides, making background inconsistencies easy to overlook.

Check:

Signs
Text
Doorways
Furniture
Screens
Reflections
Other people
Architecture
Vehicles
Logos
Repeating objects

Possible clues include:

Text changing between frames
Objects merging
People deforming in the background
Reflections showing inconsistent scenes
Impossible geometry
Objects appearing and disappearing
Background motion that does not match camera movement

Generative video models are improving at temporal consistency, so these remain warning signs rather than permanent weaknesses.

10. Ask Whether the Video Makes Sense in Context

Sometimes the strongest evidence has nothing to do with pixels.

Ask:

Who uploaded the video first?
Is the account authentic?
Does the person normally communicate through this channel?
Have reputable independent sources reported the event?
Does the statement contradict verified information?
Is the video asking you to act urgently?
Does it ask for money, credentials, or sensitive data?
Is someone discouraging you from verifying the request?

The FBI recommends independently verifying the identity of someone contacting you rather than trusting the apparent voice, image, number, or account. For a suspicious message, it recommends finding a known contact method independently and using that to confirm authenticity. This principle is especially important in financial fraud. A technically flawless video cannot defeat an independent verification process if the supposed executive is contacted through a separate, trusted corporate channel.

How to Detect a Deepfake During a Live Video Call

Live deepfakes create a different challenge because there may be no time for detailed forensic analysis.

During a suspicious call, watch for:

Facial features changing when the head rotates
Face boundaries breaking when hands cross the face
Delayed or unnatural lip movement
Inconsistent lighting
Voice-video timing differences
Sudden changes in facial texture
Unusual frame freezing
Audio that continues differently from facial motion

The FBI's 2025 guidance on fraudulent remote workers recommends, when virtual meetings are necessary, asking the person to keep the background visible and notes that asking them to wave a hand in front of their face may trigger failures in some AI-generated video systems. That is a useful challenge test, but it should not be treated as proof. More sophisticated systems may handle occlusion correctly.

For a consequential request, verification should move outside the video itself:

Call a known phone number
Require approval through an authenticated internal system
Confirm through a trusted device
Use a pre-agreed verification phrase
Require a second approver
Pause financial action until identity is verified

A Better 7-Step Process for Verifying a Suspected Deepfake

If a video affects money, identity, reputation, evidence, or public information, move beyond visible warning signs.

Step 1: Stop Before Sharing or Acting

Urgency is often part of the attack.

Pause if a video requests:

Money
Login credentials
Confidential information
Account changes
Investment
Emergency assistance
Immediate publication
Unusual authorization

Separately verify the request even if the person's appearance looks convincing.

Step 2: Find the Original Source

Determine where the video first appeared.

Check:

Original uploader
Upload date
Account history
Caption
Longer versions
Platform verification
Previous posts
Whether credible sources link to the same recording

Try to obtain the highest-quality version available. Heavy compression can remove useful forensic information or create artifacts that look suspicious.

Step 3: Inspect Multiple Frames

Do not diagnose a video from one screenshot.

Select frames where:

The head turns
The mouth is moving
A hand crosses the face
Lighting changes
Multiple people appear
Background details are visible

Compare them for continuity.

Step 4: Reverse-Search Key Frames

Capture distinctive frames and use an image-search service to locate previous appearances.

This may reveal that:

The footage is older than the claimed event
Genuine footage was reused with different audio
The original recording contains another statement
The clip came from a film, advertisement, or previous event
Fact-checkers have already investigated it

This step is important because not every misleading video is a deepfake. Authentic footage can be miscaptioned or taken out of context.

Step 5: Verify the Claim Independently

Search for confirmation outside the account sharing the video.

For public events, check:

Official statements
Full speeches
Established news organizations
Press offices
Other recordings
Independent fact-checkers

For private business interactions:

Contact the person using a known number
Use an authenticated internal system
Require dual approval
Verify through a separate trusted channel

The FBI explicitly recommends independently identifying contact information before verifying a suspicious sender.

Step 6: Check Provenance and Watermarks

Some media contains technical information about its source and modification history.

The C2PA specification provides a framework for cryptographically signed provenance information. A C2PA manifest can contain assertions covering creation, editing, capture details, software, and related information.

However, C2PA deliberately does not decide whether provenance information is “good” or “bad.” Its role is to determine whether signed assertions are associated with the underlying asset, properly formed, and free from tampering.

Therefore:

Missing Content Credentials do not mean a video is fake.
Valid Content Credentials do not prove every claim made about a video is true.
Provenance is strongest when used alongside source and content verification.

Step 7: Use Technical Deepfake Detection

When the consequences matter, analyze the media with an appropriate deepfake detection system.

Modern detectors may inspect:

Pixel-level artifacts
Spatial inconsistencies
Temporal patterns
Face-swap evidence
Lip-sync irregularities
Synthetic audio artifacts
Partial manipulations
Generator-specific statistical patterns

The result should be treated as additional evidence rather than unquestionable proof. The GAO warns that detector accuracy can decrease when lighting, facial expressions, video quality, audio quality, or generation methods differ from the data used to train the system.

Why Visual Deepfake Warning Signs Are Becoming Less Reliable

One of the biggest problems with old deepfake advice is that it treats current artifacts as permanent limitations of AI.

They are not.

Earlier guides emphasized:

Strange blinking
Blurry face boundaries
Incorrect teeth
Distorted hands
Smooth skin
Unnatural head motion

These clues can still appear, but generation models improve specifically around known weaknesses.

The GAO describes this as an evolving challenge: future generation techniques may eliminate recognizable hallmarks, while detection techniques themselves encourage attackers and developers to adapt. That creates an arms-race dynamic.

Real-World Deepfakes Are Harder Than Old Benchmarks Suggest

The 2024 Deepfake-Eval-2024 study offers one of the clearest examples.

Researchers assembled current in-the-wild media containing:

45 hours of video
56.5 hours of audio
1,975 images
Content from 88 websites
Media spanning 52 languages

When state-of-the-art open-source models were evaluated on this dataset, AUC decreased versus earlier academic benchmarks by:

50% for video
48% for audio
45% for images

Commercial systems and models fine-tuned on the new benchmark performed better than off-the-shelf open-source detectors but still did not reach the reported performance of specialist deepfake forensic analysts. This is why a detector's historical benchmark score should not automatically be interpreted as its real-world accuracy against new generators.

Humans and AI Detectors Make Different Mistakes

The answer is not simply to replace humans with machines.

A PNAS study involving 15,016 participants compared ordinary human judgments with a leading computer-vision deepfake detection model.

The researchers found that humans and the model achieved similar overall performance but made different types of mistakes. Participants who received model predictions performed better overall than either humans or the model alone, although incorrect model predictions sometimes caused people to change correct judgments into incorrect ones.

This supports a careful hybrid workflow:

Automated screening
Human contextual review
Source verification
Provenance checks
Independent verification when the signals disagree

Human-machine collaboration helps only when people understand that the model can also be wrong.

Visual Inspection vs. Technical Verification

The appropriate verification depth should match the risk. A funny social-media clip and a video authorizing a multimillion-dollar transaction should not receive the same level of scrutiny.

Detection and Authentication Are Different

Deepfake detection asks whether media contains signs associated with synthetic generation or manipulation.

Authentication takes a different approach: it tries to establish trustworthy information about the content from creation onward.

The NSA, FBI, and CISA recommend combining detection with technologies that help establish media provenance and authenticity, including real-time verification and authentication mechanisms.

The GAO similarly describes combined defenses as an opportunity for improving deepfake resilience.

A strong authenticity stack can therefore include:

Deepfake detection
Content provenance
Watermarking
Cryptographic signing
Trusted capture
Identity verification
Source verification
Human forensic review

How Organizations Should Detect Deepfake Videos

Enterprise detection needs a repeatable workflow rather than an ad hoc judgment.

1. Screen High-Risk Media

Run high-risk submitted or live media through automated detection.

Examples include:

Executive video calls
KYC submissions
Remote hiring
Claims evidence
Payment authorizations
Public-interest media
User-generated content flagged for impersonation

2. Preserve the Original

Where policy allows, keep the original uploaded media rather than relying only on screenshots or recompressed copies.

3. Record the Detection Context

Preserve:

Detection timestamp
Model version
Confidence score
Relevant frames
Detected modality
Decision threshold
Supporting forensic explanation

4. Verify Context Separately

Compare the media with:

Known identities
Trusted communications
Original publications
Prior recordings
Transaction history
Account behavior
Provenance evidence

5. Escalate Ambiguous Cases

High-impact or contradictory results may require specialist forensic review.

6. Require Secondary Authorization

A suspicious video should trigger a stronger control rather than simply a “fake” label.

Options include:

Out-of-band confirmation
Multi-person approval
Trusted-device authentication
Enhanced liveness verification
Manual identity review

The NSA's joint guidance with the FBI and CISA similarly recommends technological detection alongside response planning, protection of high-priority personnel and communications, information sharing, and employee training.

How Resemble AI Helps Detect Deepfake Videos

Visual inspection can expose obvious manipulation, but modern deepfakes increasingly require analysis below the level people can reliably perceive. Security teams may also need to understand why a video was flagged rather than receiving only a binary verdict.

Resemble Detect analyzes audio, images, and video through its multimodal DETECT-World model.

Multimodal detection: Analyze the video's visual stream alongside its audio.
Frame-level analysis: Evaluate temporal and visual signals across the media rather than relying only on one screenshot.
Face-swap and partial-edit detection: Screen for localized manipulation as well as fully synthetic media.
Live-call analysis: Extend detection into video-conferencing environments.
Flexible deployment: Use REST APIs, SDKs, on-premises, or air-gapped environments based on workflow requirements.

Resemble Intelligence adds the explainability layer. It returns a human-readable forensic breakdown identifying the artifacts that triggered a flag, the detected fraud type, and liveness information, along with a structured audit trail.

For live communications,Resemble Meetings currently supports Zoom, Microsoft Teams, Google Meet, and Webex. Its documentation describes real-time audio-video analysis and frame-by-frame forensic reporting for detections.

Why Provenance Matters More in 2026

Deepfake verification is gradually expanding from: “Can we detect the fake?” to: “Can we establish where trusted media came from?”

Detection is probabilistic. Provenance provides another kind of evidence. This has become especially relevant under the EU AI Act. The European Commission's final Article 50 transparency guidelines were published on July 20, 2026, and the relevant transparency obligations began applying on August 2, 2026.

For covered systems and circumstances, Article 50 includes obligations relating to machine-readable marking of AI-generated or manipulated outputs and disclosure of deepfake content.

This does not mean every video without an AI label is fake, and it does not make technical detection legally conclusive.

It does reinforce a broader trend toward combining:

Detection
Machine-readable marking
Watermarking
Provenance
Disclosure
Human-readable labeling

Common Deepfake Detection Mistakes

Mistake 1: “The Person Blinks Normally, So It Is Real”

Modern generation systems can reproduce natural blinking. The GAO explicitly warns that abnormal blinking may disappear as a useful artifact.

Mistake 2: “The Video Looks Weird, So It Is AI”

Real video can look strange because of:

Low bandwidth
Compression
Motion blur
Camera stabilization
Low light
Portrait effects
Filters
Frame interpolation

An unusual artifact warrants investigation, not an automatic verdict.

Mistake 3: Checking Only the Face

The manipulation could exist in:

Audio
Background
One video region
Individual frames
A short segment
Another participant

Evaluate the complete media.

Mistake 4: Ignoring the Source

Contextual verification can catch attacks that visual analysis misses. A convincing CEO video requesting a wire transfer can still be defeated by calling the CEO through a known internal number.