How to Spot a Deepfake: Video, Audio and AI Images

How to spot a deepfake comes down to two things: (1) checking for specific “tells” in video, audio, or images, and (2) verifying the file or source using basic deepfake detection tools. If you’re thinking “is this a deepfake?” you’re not alone—research shows people correctly identify high-quality video deepfakes only 24.5% of the time, so using a repeatable checklist matters.

• What is a deepfake?
• How does a deepfake work?
• Why does deepfake detection matter now?
• How to spot a deepfake video (visual checklist)
• How to spot a deepfake voice or audio (sound checklist)
• AI generated image detection: how to tell if an image is fake
• Deepfake detection tools: what works in 2025–2026 (and limits)
• A 10-minute “Is this a deepfake?” verification routine
• Real-world examples: deepfakes that fooled people
• Is it legal? Deepfake laws and platform rules
• What you can do today (reporting, documentation, safety)
• Conclusion: how to spot a deepfake without panic

What is a deepfake?

A deepfake is synthetic media—video, audio, or images—made or altered with AI to convincingly show someone doing or saying something they didn’t do or say. Some deepfakes are obvious. Others are close enough to reality that a quick scroll can fool almost anyone.

That’s why people often search variations like how to spot deepfake, how to tell if a video is fake, or ai generated image detection: the problem isn’t that deepfakes exist, it’s that they spread fast and can be hard to verify.

How does a deepfake work?

Deepfakes generally work by generating or altering pixels (for faces and bodies) or generating/altering sound waves (for voices) so they match a target person’s appearance or voice. The practical result is a piece of media that looks “recorded,” even when it’s partly or fully synthesized.

Two details matter for everyday detection:

• Modern deepfakes (2024–2026) are better at “details” than early ones. For example, abnormal blinking was an early giveaway, but by 2025 leading models have largely fixed it. You can’t rely on one old trick.
• Deepfakes still struggle with consistency. Many tells show up when a head turns quickly, when lighting should match across a scene, or when hands appear.

Why does deepfake detection matter now?

Deepfakes aren’t just internet pranks. They’re being used for fraud, harassment, and manipulation—and documented cases show real financial and civic harm.

• Fraud: In February 2024, a worker at Arup in Hong Kong was tricked in a deepfake video call and made 15 wire transfers totaling $25 million, after seeing what looked like the CFO and colleagues.
• Election interference: In January 2024, an AI-cloned voice robocall impersonating President Biden reached 5,000–25,000 voters in New Hampshire, telling them not to vote in the primary.
• Mass harassment: Explicit deepfakes of Taylor Swift in January 2024 racked up 47 million+ views before removal.

And there’s a hard reality check: on high-quality deepfake video, humans are right only 24.5% of the time. In the DeepFake-Eval-2024 benchmark (published on arXiv in 2025), researchers also found a big lab-to-real-world performance drop of ~45–50% for video, audio, and image detection models. So, it’s normal to feel uncertain—use process, not gut instinct.

How to spot a deepfake video (visual checklist)

If you’re asking how to tell if a video is fake, start with the parts AI still trips over: lighting, motion consistency, mouth/teeth, and hands.

1) Eyes and gaze (blinking isn’t enough anymore)

Early deepfakes often blinked oddly, but by 2025 that’s “largely fixed” in leading models. What still shows up is a mismatch between eye movement and the emotional context—like eyes that look fixed or “staring,” or gaze shifts that don’t line up with what a person is reacting to.

Also check reflections: mismatched reflections in the eyes can signal compositing problems, especially if the reflections don’t match each other or the visible light sources in the room.

2) Lighting and shadows (do they agree with the room?)

A common tell is lighting mismatch between the face and the background. Another is shadows that behave oddly as the head moves. These issues can be subtle in high-quality fakes, but they’re still worth checking, especially around the nose, under the chin, and near the ears.

3) Hairline, jaw, and edges during motion

Look for blurring, flickering, or a “halo” around hairlines, ears, or jawlines—especially during head movement. Fast Company reported that by 2025 some classic edge artifacts like blurred ears and glitchy shadows had “all but disappeared” in high-quality deepfakes, so treat this as a sometimes clue, not a guarantee.

4) Mouth, lip sync, and teeth (quick, high-value checks)

Audio-video mismatch is still one of the most practical deepfake tells:

• Lip sync: Watch for lips that don’t fully close on “p,” “b,” and “m” sounds, or mouth shapes that don’t match the audio.
• Teeth: Teeth sometimes render as a uniform white mass instead of distinct teeth. It’s an easy checkpoint—pause and zoom.

5) Hands and fingers (still a reliable tell in 2025–2026)

Hands remain one of the most reliable visual tells. AI often produces malformed hands: extra fingers, merged digits, or impossible joints. If the video includes gesturing, slow it down and inspect the hands frame-by-frame.

6) Temporal consistency (flicker and “texture pop”)

Deepfakes can look great in a still frame but fall apart across time. Watch for:

• Outlines that flicker when the person turns quickly
• Sudden changes in skin texture during motion (a “pop” from smooth to textured)

Note: older guidance often says “look for overly smooth skin.” That can still happen (an airbrushed look), but modern models (2024–2026) synthesize skin texture more convincingly, so you need multiple checks.

7) Text inside the video frame

If there’s a sign, label, badge, or on-screen text, zoom in. AI-generated text is frequently jumbled, nonsensical, or misspelled. This is especially useful for “street interview” style clips with storefronts, name tags, or posters in the background.

How to spot a deepfake voice or audio (sound checklist)

Deepfake audio can be incredibly persuasive—so rely on patterns. If you’re wondering is this a deepfake based on a call, voicemail, or voice note, listen for cadence, background noise, and emotion.

1) Cadence and rhythm that feels “clipped”

Voice clones often have an unnatural cadence: odd emphasis, pauses in the wrong places, or a “read aloud” rhythm. NBC News noted the Biden New Hampshire robocall had a “clipped cadence, particularly toward the end, that seemed unnatural and robotic.” That kind of pacing issue is a practical red flag.

2) Too-clean audio (missing normal messiness)

Be cautious when audio is unnaturally clean—no room tone, no background noise artifacts, no mic handling sound—especially if it supposedly came from a phone call or a live environment. The absence of expected noise can be a tell.

3) Emotional flatness

Voice clones may fail to modulate emotion naturally. The words might sound “right,” but the emotion doesn’t rise and fall the way a real person’s would in the situation.

4) Don’t assume you can spot it by ear

Pindrop reported that only 1 in 4 people can identify a high-quality deepfake voice. Treat your own confidence as unreliable, and verify using a second channel (text confirmation, call-back number you already have, or an in-person check).

AI generated image detection: how to tell if an image is fake

AI generated image detection is often easier than video because you can slow down and inspect details. But image fakes can still be convincing—especially on small screens—so use a simple two-part approach: visual inspection + file/provenance checks.

1) Hands, again

The same rule applies: inspect fingers, joints, and how hands interact with objects. As of 2025–2026, malformed hands are still a notoriously common failure mode.

2) Text in the image

AI-generated text on signs, labels, shirts, and documents is often misspelled, garbled, or inconsistent. If the image is trying to “prove” something via a screenshot or a sign, this check is essential.

3) Check EXIF metadata (camera info)

If you have the actual file (not just a screenshot), look at its EXIF metadata. Real photos often include camera make/model, lens, ISO, and shutter speed. AI images frequently lack authentic camera data. Missing EXIF isn’t proof by itself (platforms strip metadata), but it’s a useful signal.

4) Use error-level analysis for manipulation boundaries

Error Level Analysis (ELA) can reveal compression inconsistencies where parts of an image were manipulated. One free tool that supports this is FotoForensics (error-level analysis and metadata tools).

Deepfake detection tools: what works in 2025–2026 (and limits)

There are real deepfake detection tools, and some are helpful—but none are magic. In the DeepFake-Eval-2024 benchmark (arXiv 2025), detection systems saw a ~45–50% performance drop when moving from lab conditions to real-world samples. The same benchmark reports leading AI detection models at about ~84% on real-world samples, with ~5–7% false positives on human content.

So treat tools as “decision support,” not a final verdict—especially for high-stakes situations.

Comparison: deepfake detector options you can actually use

• Deepware Scanner (deepware.ai): Free, user-friendly consumer tool for scanning media.
• AI or Not (aiornot.com): Free browser-based ai generated image detection.
• FotoForensics: Free ELA + metadata examination (best for images).
• Microsoft Video Authenticator: Frame-by-frame analysis producing confidence scores per frame.
• Reality Defender: Enterprise tool with a small free tier (50 scans/month).
• Hive Moderation: Notable industry tool; the U.S. Department of Defense invested $2.4M.
• Sensity AI: Claims 95–98% accuracy; offers cloud/on-premise; ISO 27001; used by law enforcement and media.
• Intel FakeCatcher: Uses blood flow signals (PPG) across 32 facial locations; claimed 96% real-time accuracy (as reported by VentureBeat).

Provenance and watermarking: the tools that can make verification easier

Detection is one path. Provenance is another: it asks, “Where did this come from, and can we trace edits?”

• C2PA (Content Credentials): A provenance standard that embeds information about creator/date/tool into files. It’s adopted by Meta, YouTube, X, and TikTok. Version 2.2 was published in May 2025.
• Google DeepMind SynthID: Watermarks AI-generated content and reportedly survives compression and cropping, with 98% detection rate after those transformations.

If you can find C2PA credentials or a watermark signal, that’s often stronger evidence than “it looks weird.”

A 10-minute “Is this a deepfake?” verification routine

If a clip is going viral or pushing you to act fast, slow down. Here’s a repeatable routine that doesn’t require tech expertise.

1. Save evidence first: Take screenshots, copy URLs, note usernames and timestamps. (Platforms can remove posts quickly.)
2. Do a 60-second visual/audio scan: Check lip sync on “p/b/m,” teeth detail, reflections in eyes, and hands.
3. Look for “too clean” audio: Ask whether the background sound matches what’s being shown.
4. Run one consumer tool: Use Deepware Scanner for video or AI or Not for images.
5. If it’s an image file, check metadata/ELA: Use FotoForensics to review EXIF and ELA patterns.
6. Check for provenance: Look for Content Credentials (C2PA) indicators where available, or any labeling that the platform provides.
7. Verify via a second channel: If it’s about a real person you can contact, confirm through an independent route (a known phone number, a separate email thread, or in person).
8. Decide what to do next: Don’t share if you can’t verify. If it targets someone, report it and preserve your documentation.

If you’re building a workplace or school policy for synthetic media, Ban the Bots has templates you can adapt: /no-ai-policy-template/ and /human-made-policy-template/.

Real-world examples: deepfakes that fooled people

It helps to remember that smart, careful people get tricked—because deepfakes exploit normal trust and time pressure.

• Arup deepfake fraud (Hong Kong, Feb 2024): A finance worker joined a video call that appeared to include Arup’s CFO and colleagues. The worker made 15 wire transfers totaling $25M. This is widely cited as the largest confirmed deepfake fraud case.
• Biden robocall (New Hampshire, Jan 2024): 5,000–25,000 voters received an AI-cloned voice message telling them not to vote. NBC News pointed to a “clipped cadence” that sounded robotic—an example of an audio tell you can listen for.
• Taylor Swift explicit deepfakes (Jan 2024): The content reportedly reached 47M+ views before removal, showing how fast synthetic media can spread on major platforms.
• Marco Rubio voice clone (April 2025): A voice clone was used to contact foreign officials, illustrating that deepfake audio isn’t just a celebrity problem—it’s a governance and security issue.
• Israeli Defense Minister deepfake broadcast (April 2025): A deepfake was inadvertently aired on Channel 14, showing that even professional media environments can get fooled.

Looking ahead, University at Buffalo and Fast Company have pointed to a 2026 outlook where real-time interactive deepfakes (able to respond in video calls) may emerge. That makes “call-back on a known number” and other second-channel verification habits even more important.

Is it legal? Deepfake laws and platform rules

Legality depends on what the deepfake does (fraud, defamation, harassment, election interference) and where you live. But regulation is moving toward mandatory disclosure in at least some contexts.

EU: disclosure requirement is coming

Under the EU AI Act, Article 50 requires disclosure labeling on all deepfakes. The requirement is set to be effective in August 2026. (If you want the plain-English breakdown, see /explainers/eu-ai-act.)

Platforms: labeling and enforcement are uneven

• YouTube (2026): Requires clear labeling of synthetic media depicting real people and uses active detection algorithms.
• Meta: In 2025, Meta’s Oversight Board said the company’s deepfake detection was “fundamentally broken” and relied almost entirely on self-disclosure.
• X: Blocked searches for Taylor Swift briefly in January 2024, but follow-through on deepfake removal has been mixed.

If you need a deeper legal overview (including what can be reported and what remedies might exist), see /explainers/deepfake-laws and our running list of related disputes at /ai-lawsuits/.

What you can do today (reporting, documentation, safety)

You don’t need to become a forensic analyst to reduce harm. You need a few habits and a plan.

Before you share: do the two-step check

• Step 1: Visual/audio tells. Use the checklists above (hands, teeth, lip sync, cadence, too-clean audio).
• Step 2: Tools + provenance. Run one scan (Deepware / AI or Not / FotoForensics) and look for labeling or Content Credentials (C2PA).

If the deepfake targets you or someone you know

• Document first: screenshots, URLs, and timestamps.
• Report to the platform: include your documentation and note that it’s synthetic or impersonation content.
• Tell people likely to be tricked: especially workplaces and family group chats where scams spread fast.

If it looks like fraud (especially a call asking for money)

Use the “known-channel” rule: hang up and call back using a phone number you already have saved or that you obtained independently. The Arup case shows why: deepfakes can be good enough on a video call to push someone into urgent wire transfers.

Get involved beyond your feed

If you’re feeling overwhelmed, it can help to connect deepfakes to the bigger picture: how AI systems are being deployed and governed (or not). Ban the Bots tracks public pushback and accountability efforts at /ai-backlash/ and practical steps you can take at /fighting-back/.

And because AI has real-world infrastructure and costs, our /data-center-map/ and background explainers like /explainers/data-center-impact can help you understand the “behind the scenes” side of the AI boom.

Conclusion: how to spot a deepfake without panic

How to spot a deepfake isn’t about finding one perfect giveaway—it’s about stacking signals: visual tells (hands, teeth, lip sync, lighting), audio tells (cadence, emotion, too-clean sound), and verification using deepfake detection tools plus provenance signals like C2PA and watermarking.

If you want to go deeper, start here: track accountability efforts at /ai-backlash/, learn how people are responding at /fighting-back/, and follow real disputes at /ai-lawsuits/. And if you’re feeling AI’s impact at work, we’re also tracking it at /ai-layoffs/.

When in doubt, don’t share—verify.