# How to Spot a Deepfake Video: What Still Gives It Away

> Check the source before the face: find the first upload and see whether trusted outlets carry the same clip. Then watch the face edges, teeth, eyes and profile turns at reduced speed, and listen for flat or delayed audio. In a live call, ask the person to turn sideways or call them back on a known number.

- URL: https://fake-ai.com/images-and-video/how-to-spot-deepfake-video/
- Language: en
- Updated: 2026-10-19
- Publisher: Fake AI (OnlineBuilders LLC, https://onlinebuilders.biz)

Germany's federal security agency BSI uses a blunt number to explain why software won't save you. Meta ran a Deepfake Detection Challenge in 2020, and the winning model reached 65.18 percent on a hidden set of videos it had never seen before. Guessing would have scored 50 ([BSI](https://www.bsi.bund.de/EN/Themen/Unternehmen-und-Organisationen/Informationen-und-Empfehlungen/Kuenstliche-Intelligenz/Deepfakes/deepfakes_node.html)). On the public data the same kind of models had looked far better, at 82.56 percent average precision.

That gap is the whole story of deepfake detection. Tools do well on the fakes they were trained on and badly on the next generation. Humans have the same problem. What still works is a mix of source checks, slow viewing and, in live calls, a few simple tests that current software handles poorly.

## Start with where the video came from

Most fake videos that cause real damage are spread by accounts, not discovered by forensic analysts. So the first questions are about distribution, not pixels.

Who posted it first? Search the platform for the earliest copy and look at that account: age, posting history, whether it has published original footage before. Then check whether a newsroom with its own reporters has the clip, or has written about it. Any video of a head of state saying something explosive that only exists on three anonymous accounts is a red flag regardless of how good it looks.

Pull a few keyframes and run them through a reverse image search. Some supposed deepfakes turn out to be real footage from another year or country, which is easier to prove than any AI manipulation. Our guide to [finding the original source of an image](https://fake-ai.com/images-and-video/find-original-image-source/) explains which search engines handle video stills best.

## The visual artefacts BSI lists

The BSI's deepfake page is one of the few official sources that lists concrete artefacts. For face swaps and reenactments it names:

- visible seams at the edge of the face, with shifts in skin colour or texture, sometimes doubled eyebrows
- washed-out detail in areas that should be sharp, especially teeth and eyes
- a limited range of facial expression and lighting that does not quite match the scene
- blurring when the head turns to a profile view

That last point matters most. Sharp turns to the side are where the mask tends to slip. Watch those moments at 0.25x speed. On a phone, screen-record the clip first so you can scrub frame by frame.

**Watching a suspect clip frame by frame**

- [ ] Play at **0.25x** and pause on every head turn
- [ ] Follow the jawline and hairline: does the edge flicker or shift colour?
- [ ] Compare teeth across frames: do they change shape or merge into one block?
- [ ] Look at glasses, earrings and hands passing in front of the face
- [ ] Check that the light on the face comes from the same side as the shadows in the room
- [ ] Watch lip shapes on hard consonants such as p, b and m
- [ ] Listen with headphones for a flat, metallic or slightly delayed voice

## Listen as carefully as you watch

Many fake videos pair a real face video with a cloned voice, or animate a real face to match new audio. The BSI lists audio tells too: a metallic sound, mispronounced words, monotone delivery, an accent or stress pattern that doesn't fit the speaker, and odd sounds at pauses. For real-time systems it adds a noticeable delay, because the original speech has to happen before it can be converted.

Lip sync is the place where face and voice have to agree. Mouth shapes for p, b and m require closed lips. If you hear a clear "b" while the lips stay apart, that is worth noting. One mismatch can be a dropped frame. Repeated mismatches across a sentence are more telling. There is more on the audio side in our guide on [how to detect an AI voice](https://fake-ai.com/voice-and-scams/detect-ai-voice/).

## Live video calls need a different approach

Pre-recorded deepfakes can be polished frame by frame. Live ones cannot, and that is your advantage. The FBI warned in December 2024 that criminals use generative AI for "real-time video chats" in which they pose as company executives or police officers ([FBI IC3](https://www.ic3.gov/PSA/2024/PSA241203)). The best-known case: in early 2024, a finance employee at the engineering firm Arup in Hong Kong joined a video call with what looked like the company's CFO and several colleagues and transferred about HK$200 million in 15 payments. Arup later confirmed that fake voices and images had been used. We break that case down in [deepfake CEO fraud](https://fake-ai.com/voice-and-scams/deepfake-ceo-fraud/).

**Tests in a live call and how a real-time fake tends to react**

| | Effort for you | Tends to expose a live face swap |
|---|---|---|
| Ask for a slow turn to full profile | low | yes |
| Ask them to pass a hand across the face | low | partly |
| Ask about a shared detail only the real person knows | low | yes |
| End the call and ring back on a number you already have | medium | yes |
| Judge by video quality or background | low | no |

The call-back is the strongest test because it does not depend on spotting anything. If the request involves money, passwords or confidential data, end the call politely and reach the person through a channel you set up yourself. Real colleagues accept that. Fraudsters push for urgency and secrecy.

## Cheapfakes are still the bigger problem

Not every manipulated video involves a neural network. Slowing a clip down to make someone sound drunk, cutting a sentence so it means the opposite, or relabelling footage from an old protest as yesterday's riot needs nothing more than a phone editor. These cheapfakes are quicker to make than deepfakes and they spread just as well.

They also need different checks. Artefact-hunting finds nothing, because every frame is genuine. What exposes them is the full-length original, the date of the first upload and the location. If a clip seems cut mid-sentence, look for the complete speech or broadcast. If it shows a crowd, check the weather, the signs and the season against the claimed date.

## Why detector apps disappoint

Online deepfake detectors give a percentage, often with a confident colour scheme. Treat that number as one opinion. Social media copies are compressed, cropped and re-encoded, which removes many of the subtle traces detectors look for. The BSI also notes that automated detectors are vulnerable to adversarial attacks, meaning a fake can be tuned to fool a specific tool.

There is a second trap. Any real video that scores "70 percent fake" can do as much damage as a fake that scores "clean". The liar's dividend is real: once people know deepfakes exist, anyone caught on camera can claim the footage was made with AI. Provenance is the better long-term answer. Cameras and apps that sign footage with [Content Credentials](https://fake-ai.com/images-and-video/content-credentials-c2pa/) let a publisher prove a clip is original, which is a stronger position than arguing about artefacts.

## What to do when a clip targets you or your company

Save evidence first: the URL, the uploading account, the date and a screen recording. Platforms remove content quickly once reported, and you lose the trail. Then report it under the platform's impersonation or manipulated media rules.

For companies the cheaper investment is upfront. Agree that payment and data requests over video always need a second channel, and publish official statements only through known channels so people can check. If you are building an internal tool to log and review suspicious media, a studio like [OnlineBuilders](https://onlinebuilders.biz/solutions/cyber-security) can help set it up.

Still images follow different rules, so if your clip came with a thumbnail or a screenshot, run it through our checklist on [how to tell if an image is AI generated](https://fake-ai.com/images-and-video/how-to-tell-if-image-is-ai/).

## FAQ

### Are there free deepfake detection apps that work?

Some give a useful second opinion, but none is reliable enough to settle a dispute. Detectors are trained on known fake-making methods and lose accuracy on new ones, on compressed social media copies and on short clips.

### Can a deepfake run live in a video call?

Yes. Face-swapping models can run close to real time, according to Germany's BSI, and the FBI has warned about criminals posing as executives in live video chats. Live fakes tend to break down on sharp head turns and when something passes in front of the face.

### What is the difference between a deepfake and a cheapfake?

A deepfake uses a neural network to swap or animate a face or voice. A cheapfake uses ordinary editing: slowing a clip down, cutting it out of context or relabelling an old video. Cheapfakes are more common and easier to debunk with a reverse search.

### Should I report a deepfake I find?

Report it to the platform using its manipulated media or impersonation option. If it targets you or someone you know, save the link, the account name and a screen recording before it disappears.

Photo: Loozrboy from Toronto, Canada, CC BY-SA 2.0 (https://commons.wikimedia.org/wiki/File:City_tv_control_room_Doors_Open_Toronto_2012.jpg)