How to Spot a Deepfake: 12 Signs That Work in 2026

Most advice on how to spot a deepfake is stuck in 2019. "Watch for unnatural blinking" no longer works, because modern models blink fine. The signs below are the ones that still hold up in 2026, grouped by what you are actually looking at: video, image, voice, or a live call. I will also be honest about where your eyes fail and where you need a tool.

Most advice on how to spot a deepfake is stuck in 2019. "Watch for unnatural blinking" no longer works, because modern models blink fine. The signs below are the ones that still hold up in 2026, grouped by what you are actually looking at: video, image, voice, or a live call. I will also be honest about where your eyes fail and where you need a tool.

Quick answer: To spot a deepfake, check for mismatched lighting, lip sync drift, blurred face boundaries, odd skin texture, and unnatural voice cadence, then verify the source and run the file through a deepfake detector. No single sign is proof, so stack several checks before you trust or share anything.

If the technology itself is new to you, start with what is a deepfake and then come back for the checklist.

Why Spotting Deepfakes by Eye Keeps Getting Harder

Here is the uncomfortable truth that most "how to spot a deepfake" articles skip: on a good fake, people perform at roughly the level of a coin flip. A 2024 systematic review and meta-analysis in Computers in Human Behavior Reports, pooling 56 studies and 86,155 participants, found average human deepfake detection accuracy of about 55.5%, which is barely above chance (Diel et al., 2024).

That number matters because it sets the right expectation. The visual signs in this guide reliably catch low-end and rushed fakes, the kind that make up most of what circulates. They catch fewer of the top-tier fakes, the ones with manual cleanup and high-resolution source footage.

So treat your eyes as a fast first filter, not a verdict. When the stakes are real, money, reputation, a hiring decision, you stack the manual checks and then confirm with software that reads signals humans cannot see. The same study found that giving people detection strategies measurably improves their accuracy, which is exactly what the rest of this page is for.

How to Spot a Deepfake Video: 6 Visual Signs

Video is where most people first meet a deepfake, and it is where the most signs cluster. Work through these six in order. Each one includes a what-to-do step.

1. Face Boundaries and Hairline Blur

Face-swap deepfakes paste a generated face onto a real head, and the seam is the weak point. Look at the edge of the face, the hairline, the jaw, and the ears. You may see a faint blur, a slight color shift, or hair strands that dissolve into the skin instead of overlapping it.

What to do: Pause on a frame where the head turns. Boundary errors flare during motion, especially when hair crosses the face line.

2. Lip Sync Drift and Teeth Rendering

Lip-sync fakes drive a mouth from audio, and the mouth is the hardest region to fake. Watch whether the lips land exactly on the consonants. Then look at the teeth: deepfakes often render them as a flat white blur rather than individual teeth, and the inside of the mouth can look like a dark smear.

What to do: Mute the clip and watch the mouth alone. Without audio masking it, drift and teeth artifacts are easier to catch.

3. Lighting and Shadow Mismatches

Light is physics, and generators get it subtly wrong. Check that the light on the face comes from the same direction as the light in the room. Look for shadows under the nose and chin that do not match the scene, or a face that looks evenly lit while the background is dramatically shadowed.

What to do: Find the brightest light source in the frame, then check that every highlight and shadow on the face agrees with it.

4. Skin Texture That Is Too Smooth or Too Sharp

Generated skin often misses the middle ground of real skin. It can look waxy and poreless, like heavy retouching, or it can show oddly crisp, repeating texture that does not move naturally with expression. Foreheads and cheeks are the giveaway zones.

What to do: Compare the face texture to the hands and neck in the same clip. A mismatch between a flawless face and ordinary skin elsewhere is a red flag.

5. Glasses, Jewelry, and Reflection Errors

Reflective and fine-detail objects confuse generators. Glasses may show glare that does not move correctly with the head, or arms that warp where they meet the ear. Earrings can flicker, change shape, or vanish between frames. Reflections in the eyes may not match the room.

What to do: Track one small accessory across several seconds. Consistency over time is hard for fakes to maintain.

6. Head Position Extremes and Profile Glitches

Most face-swap models learn from front-facing footage, so they break down at steep angles. When the subject turns toward a full profile or tilts far back, watch for the face smearing, the swap slipping off the underlying head, or features briefly misaligning.

What to do: Look for moments of extreme head movement and study them frame by frame. If the clip never shows a profile, that absence can itself be a choice to hide the weakness.

What Are Some Signs of a Low-End Deepfake?

Cheap, fast fakes leave different fingerprints than polished ones, and they are far more common, so they are worth a section of their own. A low-end deepfake is made in minutes with default settings and no cleanup, and it tends to show several of these at once.

• Frame jitter and flicker. The face wobbles or shimmers slightly from frame to frame because the model is not stabilized across time.
• Static or limited expressions. The face holds an oddly fixed look while the head moves, or the same micro-expression loops.
• Hard edges and crops. A visible rectangle or halo around the face, or a blur ring where the fake region ends.
• Recycled or mismatched backgrounds. A clean, generic, or slightly warped background, sometimes lifted from the source clip.
• Watermark remnants. Faint logos or text from the generator or the stolen source footage, especially in corners.
• Audio that does not fit the room. Studio-clean voice over a casual setting, or no room echo at all.

High-end fakes erase most of these with manual work, which is why the absence of low-end tells does not prove a clip is real. It only means you are not looking at a lazy fake. That is the moment to escalate to source verification and a detector.

How to Spot a Deepfake Image

Still images cannot rely on motion tells, so the checks shift to logic and structure. AI-generated and manipulated images fail at the parts of a scene that require understanding, not just texture.

Hands, Text, and Background Logic

Hands remain a classic failure point: count the fingers, and look for bent or fused joints. Read any text in the image, signs, labels, name tags, because generators often produce garbled, dreamlike lettering. Then check background logic: doorways that lead nowhere, railings that merge, patterns that repeat unnaturally, and people in the crowd who melt together.

For a deeper image-specific walkthrough, see how to tell if an image is AI generated.

Reverse Image Search the Frame

Provenance often beats pixel-peeping. Run the image, or a screenshot of a video frame, through a reverse image search to find where it first appeared. If a "breaking news" photo has no earlier source, no other angles, and no credible outlet, treat it as suspect regardless of how clean it looks.

What to do: Search the image, then sort by date to find the earliest copy. A single orphaned image making a big claim is the pattern to distrust.

How to Spot a Deepfake Voice or Audio Clip

Audio is the easiest channel to fake convincingly and the hardest to judge by ear, because phone and app compression hides the artifacts that would give a voice away. A few seconds of someone's real voice is enough to clone it.

Cadence, Breathing, and Emotional Flatness

Cloned voices nail the timbre but stumble on delivery. Listen for an unnaturally even rhythm, words that all land with the same weight, and emotion that feels flat or slightly off for the situation. Natural speech includes breaths, small hesitations, and "um" sounds; many clones either omit these or place them oddly. Background noise that never changes can also signal synthesis.

What to do: Where you can, slow the audio down. Flat emotional range and missing breath sounds are easier to catch at reduced speed.

The Callback Test for Suspicious Phone Calls

For a live call, no listening test is reliable enough to bet money on, so the defense is procedural, not auditory. If a caller claims to be a relative in crisis or a boss demanding an urgent transfer, hang up and call the person back on a number you already know. Scammers exploit voices precisely because urgency makes you skip this step.

The U.S. Federal Trade Commission specifically recommends calling the person back on a known number and agreeing on a family code word in advance, a word only your real family knows, so you can verify a caller instantly (FTC Consumer Alert, 2023). For more on this attack, see voice cloning scams and our AI voice detector.

How to Spot a Deepfake on a Live Video Call

Live video impersonation is the newest and most alarming format, and in 2024 it cost the engineering firm Arup about $25.6 million when a finance employee joined a video call with what looked like the CFO and colleagues, every one of them a deepfake (CNN Business, 2024). Real-time fakes still strain under conditions they were not built for, so you can stress-test them.

• Ask them to turn fully sideways. Real-time face swaps degrade at a profile, where the model has little to work with.
• Have them wave a hand in front of their face. The swap can smear, flicker, or let the hand briefly pass behind the face.
• Change the lighting. Ask them to move toward a window or turn a lamp on and off. Synthetic faces relight unnaturally or lag.
• Watch for a tell mismatch. Audio that trails the lips, a face that floats slightly against the background, or blinking that feels mechanical.

The strongest defense is not visual at all: verify through a second channel. End the call and reach the person on a known number or in person before acting on any urgent or unusual request. For the corporate version of this attack, see deepfake video call scams.

When Your Eyes Are Not Enough: Run a Detector

Every sign above has the same limit: it catches what is visible, and the best fakes hide the visible cues. Detection software works on a different layer. Rather than squinting at pixels, it reads statistical fingerprints left by the generation process, frequency patterns and subtle inconsistencies across an image or clip that no human can perceive.

Our detector reports a clear, whole-file result: a verdict of Authentic, Likely Synthetic, or Inconclusive, paired with a TrustScore (0-100) so you can see how confident the model is. We detect output from today's leading generators across video, image, and audio, and we update our models as new generators appear. Independent of any single visual tell, it gives you a second opinion grounded in signals your eyes do not have.

The honest caveat: no detector is perfect, and a brand-new generation technique can briefly outrun any model, which is why source verification belongs alongside any software check. Used together, your eye, the source trail, and a detector form a stack that is far harder to fool than any one of them alone. For the full method, see our deepfake detection guide.

FAQ

What is a sign of a deepfake?

The single most reliable sign is a mismatch at the face boundary, blurring or color shift where the face meets the hairline, jaw, or ears, especially during head movement. But no one sign is proof. Stack several checks, then confirm with a detector before you trust or share.

Can you always spot a deepfake by eye?

No. On high-quality fakes, human accuracy sits near chance: a 2024 meta-analysis of 56 studies found average detection accuracy of about 55.5% (Diel et al., 2024). Visual signs catch low-end fakes well and top-tier fakes poorly, which is why tooling matters.

How can I check a deepfake for free?

You can run an image, video, or audio file through our detector free, up to 50 checks per month with no card required. Create a free account and upload the file for a verdict and TrustScore in under a minute.

Do deepfakes blink normally now?

Mostly, yes. Abnormal blinking was a real tell around 2018, but modern models fixed it, so "watch the blinking" is outdated advice. Rely instead on face boundaries, lip sync, lighting, and source verification.

How do I spot a deepfake voice on a phone call?

You usually cannot tell by ear, so do not try. Hang up and call the person back on a number you already know, and agree on a family code word in advance so you can verify any urgent caller instantly, as the FTC recommends.

Conclusion: Stack the Checks, Then Verify With a Tool

Knowing how to spot a deepfake in 2026 means dropping the stale tells and building a habit instead: scan the face boundaries, lighting, and mouth on video; check hands, text, and provenance on images; distrust urgent voices and call back on a known number; and stress-test live calls with a profile turn. Each sign is a filter, not a verdict, because on the best fakes your eyes are barely better than a coin flip.

So treat the checklist as step one and a detector as step two. Not sure about a file? Check it free in under a minute, no card required, and get a clear verdict and TrustScore before you trust or share.

Related reading