How to Spot AI Deepfakes: Voice Clones, Fake Videos, and AI Scams in 2026

TL;DR: Deepfakes stopped being a weird internet novelty a while back. Now they show up in scam calls, fake investment ads, political disinformation, even romance cons. This guide covers how to spot fake videos (eyes, ears, lighting are your best clues), cloned voices (flat emotion and weird breathing), and AI-generated images (hands and text are still giveaways). Plus a list of free tools and a step-by-step plan for what to do when you actually encounter one.

What are deepfakes and why they matter in 2026

A deepfake is any video, audio, or image that AI has manipulated or completely fabricated to look or sound like a real person. The term popped up around 2017 on Reddit. Back then the output was glitchy. Easy to catch. Not anymore.

Fast forward to 2026. A teenager with a decent laptop can make a convincing face-swap video in under ten minutes. Voice cloning tools need just 30 seconds of someone's audio to produce a near-perfect copy. The barrier to entry? Gone.

Why does this matter? Deepfakes are now the fastest-growing category of fraud. The FBI reported losses from deepfake scams in the US hit $3.6 billion in 2025. Globally the number is way higher. Political deepfakes have shown up in elections across Europe, Asia, and South America. And if you have any social media presence at all, your face and voice are probably already sitting in a training dataset somewhere.

Here's the good news though: most deepfakes still have tells. Flaws you can catch if you know where to look. New to how AI generates content in the first place? My introduction to AI for beginners covers the basics.

How to spot AI-generated videos

Video deepfakes have gotten way better. But they still mess up certain physical details. Here's what I look at, roughly in order of how fast I can check:

Eyes and blinking. Early deepfakes barely blinked. Modern ones do blink, but the rate feels off. Real people blink 15-20 times a minute. Count the blinks in a 30-second clip. Too few? Too regular? Suspicious. Also check eye reflections. In real life, light reflects in both eyes at the same angle. Deepfakes sometimes place the reflection at slightly different spots in each eye. Small detail. Dead giveaway.

Ears and hairline. Generative models still struggle with ears. Look for asymmetry, weird shapes, or ears that change size between frames. Hairline is another weak spot. Real hair has flyaways and imperfections. Deepfake hair tends to look too smooth, or it blends into the forehead in a way that just feels off.

Lighting and shadows. Does the shadow on the face match the light sources in the room? AI often slaps a face onto a body without matching the background lighting. If the left side of the face is lit but the ambient light comes from the right, something is wrong.

Background jitter. Watch the area right around the person's face and neck. Deepfake algorithms sometimes warp or shimmer the background near the edges of the face mask. You'll notice it most around the jawline and chin.

Mouth and teeth. Lip-sync errors are common. Mouth movements might lag behind the audio. Teeth might look blurred, too uniform, or just be a single white blob instead of actual individual teeth.

Hands near the face. When someone's hand passes in front of their face, deepfakes often glitch. The hand distorts, merges with the face, or flickers. Real life does not do that.

One quick tip: pause the video at random moments. Deepfakes tend to look worst in frozen frames, even when the motion seems smooth.

How to detect AI voice clones

Honestly, audio deepfakes scare me more than video ones. Most scam calls happen over the phone. You can't see the person. A 2025 McAfee survey found that 1 in 4 people globally had already received a call using a cloned voice. That number has probably grown since.

Here's what to listen for.

Flat emotional tone. AI voice clones get the words and cadence right, but they miss the subtle emotional shifts in real speech. Your "family member" is calling about an emergency but sounds oddly calm? Monotone? That's a red flag.

Breathing patterns. Real people breathe at natural pauses. Cloned voices sometimes skip breaths entirely or put them in weird spots. Listen for long stretches of speech with no breathing at all. Humans don't do that.

Background noise. In a real phone call, background noise is consistent. Natural. AI-generated calls sometimes have zero background noise (unnervingly clean) or a looping, repetitive pattern that sounds fake.

Unusual phrasing. The voice might sound right, but the script might not. AI clones trained on limited data may use phrases the real person would never say. If your boss suddenly calls with overly formal language they'd normally avoid, pause. Verify.

The callback trick. Simplest defense there is. Get a suspicious call from a family member or colleague? Hang up. Call them back on the number you already have saved. A cloned voice can't pick up when you call the real person's phone.

Want to learn more about how AI can produce convincing but false stuff? See my guide on how to spot AI hallucinations.

AI text and image deepfakes

Deepfakes aren't just video and audio. AI-generated text and images are part of the same problem.

AI-generated text. Articles, emails, social media posts written by AI tend to share certain patterns:

Overly balanced structure. Every paragraph is roughly the same length. Real human writing is messier than that.
Generic filler phrases. "It is important to note that..." and "In today's fast-paced world..." Classic AI tells.
No specific personal details. AI text speaks in generalities. A human writer will include personal anecdotes, real dates, named sources.
Odd word choices. AI sometimes picks a synonym that's technically correct but sounds wrong in context. "Utilize" where anyone would say "use," for instance.

AI detection tools aren't perfect though. Better to cross-check claims with trusted sources than to rely on a detector to flag things for you.

AI-generated images. Despite real progress, AI images still have recurring problems:

Hands. Extra fingers, merged fingers, fingers bending at impossible angles. This has gotten better since 2024, but it's still the single most common flaw.
Text in the image. AI models generate letter-like shapes but often produce gibberish. Check signs, labels, book covers. If the text doesn't make sense, the image is probably AI-generated.
Symmetry of accessories. Earrings that don't match. Glasses with different arm thickness. Buttons slightly off-center.
Background logic. A staircase that leads to nowhere. A window not aligned with the wall. AI generates details that look fine up close but break down at the architectural level.

Want to dig into the AI terminology behind these tools? My AI terms explained glossary covers that.

Real-world deepfake scams to watch out for

These aren't hypothetical. They're happening right now.

Investment fraud. Fake celebrity endorsement videos are a huge problem. In 2025, a deepfake video of a well-known financial influencer promoting a crypto platform scammed people out of an estimated $14 million before anyone took it down. These videos run as social media ads. The celebrity appears to endorse a specific platform, pushing viewers to invest right now. If you see a celebrity promoting an investment in a video ad, verify it through their official channels first. Always.

Kidnapping and emergency scams. This one is genuinely terrifying. A scammer clones your child's voice, calls you, claims they've been kidnapped. The voice sounds real. The "child" is crying. The scammer wants money immediately. Variations include fake car accidents, arrests, medical emergencies. The common thread: urgency plus a demand for money right now. Always hang up. Call the person directly.

Political manipulation. Deepfake videos of politicians saying inflammatory things have shown up in at least 14 countries during recent elections. In some cases, the deepfakes went live within 24 hours of voting. No time for fact-checkers to respond. Stay skeptical of any politically charged video in the final days before an election, especially if it's shared without a link to a real news source.

Romance and social engineering. AI-generated profile photos are now standard in romance scams. Video calls can be faked in real-time with face-swap tools. Someone you've only met online refuses to meet in person? Asks for money? Seems too good to be true? Be careful.

Tools that help you detect deepfakes

You don't have to rely on your eyes and ears alone. These tools help.

Microsoft Video Authenticator. Analyzes a video and gives you a confidence score on whether it's been tampered with. It looks for blending boundaries and pixel-level artifacts that the human eye misses. Currently available for media organizations and public figures, with a consumer version coming.

Hive Moderation. Originally built for content platforms. Now offers a free browser-based tool where you upload an image or video to check for AI generation. One of the most accurate detectors out there, with reported accuracy above 95% for images.

Illuminarty. Free tool focused on AI-generated images. Upload a photo and it highlights areas most likely to be AI-generated, plus an overall probability score. Works especially well for Stable Diffusion and Midjourney outputs.

AI Voice Detector. Several browser-based tools analyze audio clips for voice synthesis. Resemble AI's Detect tool and Pindrop's audio analysis platform both have free tiers. Upload a short clip, get a probability assessment.

Reverse image search. Sometimes the simplest tool is the best one. Suspicious image looks like a news photo? Drag it into Google Images or TinEye. If it only shows up on random websites with no credible news coverage, treat it as unverified.

One caveat. No detection tool is 100% reliable. They produce false positives and false negatives. Use them as one input in your judgment, not the final word.

What to do if you encounter a deepfake

Finding a deepfake is unsettling. Especially when it involves you or someone you know. Here's a concrete action plan.

1. Don't share it. Every share makes it worse. Even sharing with a "this is fake" caption can spread it further.

2. Save the evidence. Screenshot or screen-record the deepfake. Include the URL, the account that posted it, the date. If it's a video, record it before it gets deleted. You'll need this documentation for any report you file.

3. Report it to the platform. Every major social media platform now has a specific reporting path for AI-manipulated media. Use it. Platforms respond faster to deepfake reports than to general content complaints.

4. Report it to authorities. In the US, file a report with the FBI's Internet Crime Complaint Center (IC3). In the UK, use Action Fraud. In the EU, contact your national cybercrime unit. If the deepfake is being used for fraud, local law enforcement should be involved too.

5. Contact a lawyer if you're the subject. If a deepfake uses your likeness without consent, you might have legal options. Several US states and EU countries have passed specific deepfake laws. The legal situation changes fast, so talk to someone who specializes in digital media law.

6. Warn people you know. If a deepfake of you is circulating, tell your contacts before they see it. A quick message saying "a fake video of me is going around, it's not real" can prevent a lot of confusion and damage.

FAQ

Can deepfakes be detected by the human eye? Often, yes. If you know what to look for. The tips in this article (eye reflections, ear shapes, lighting mismatches, background jitter) catch most consumer-grade deepfakes. Professional-grade deepfakes made by skilled editors are harder to catch and may need detection tools.

Is it illegal to create a deepfake? Depends on where you live and what you do with it. In many US states, creating deepfake pornography without consent is illegal. Several countries have laws against deepfakes used for fraud or election interference. Making a deepfake for personal entertainment in a private context might not be illegal in many places, but laws are changing fast in this area.

How long does it take to make a deepfake? A basic face-swap video takes 10-15 minutes with free tools on a standard computer. A high-quality voice clone needs about 30 seconds of clear audio. More sophisticated deepfakes that can survive scrutiny take longer and need more skill. But the floor is shockingly low.

Can AI deepfake detectors be wrong? Yes. All detection tools produce both false positives (flagging real media as fake) and false negatives (missing actual deepfakes). The best tools hit around 90-95% accuracy for images, lower for video, even lower for audio. Helpful, but not infallible.

What's the difference between a deepfake and regular photo editing? Regular photo editing (adjusting brightness, removing a background, retouching skin) has been around for decades. Deepfakes specifically use AI to replace or generate faces, voices, or entire bodies with the intent to deceive. The key difference is intent plus the use of generative AI models.

Should I be worried about deepfakes targeting me personally? If you have any public online presence (social media photos, YouTube videos, podcast appearances), your face and voice data are already out there. Public figures face higher risk, but ordinary people are increasingly targeted in financial scams and harassment. Knowing the detection tips in this guide goes a long way.