Skip to content Skip to footer

What is a deepfake? Here's how it works, and how to spot one

A deepfake is an AI-generated or manipulated video, photo, or audio clip in which a person’s face or voice is imitated. It is so convincing that it is almost impossible to tell it apart from the real thing with the naked eye.

What started as a crazy spaghetti-eating Will Smith has by now grown into an instrument for fraud. In January 2024, an employee of engineering firm Arup made fifteen transfers to five accounts, totaling about 25 million dollars, after a video meeting with his CFO and colleagues. Everyone on that call was fake. And on August 2, 2026, the legal playing field changes: that is when the transparency requirement for deepfakes under the AI ​​Act applies.

This article explains what a deepfake is, how the technology works, what it is used for, and why the human eye is losing this race.

In a nutshell

  • A deepfake is synthetic media: an AI-generated image, video, or audio of a person who never actually said or did that.
  • The term comes from deep learning plus fake. The technology is based on GANs and, more recently, diffusion models.
  • Less and less source material is needed. A few seconds of video or a handful of photos from a public profile are often enough.
  • Deepfakes are used for CEO fraud, job application fraud, and to circumvent facial recognition during identity checks.
  • The human eye is no longer a reliable detector. Recognition tips help, but technology is outpacing them.
  • As of August 2, 2026, Section 50 of the AI ​​Act requires that deepfakes be labeled as artificial. Fines can reach up to 15 million euros or 3% of global turnover.
  • Detection works in layers: liveness detection, AI analysis of the image, and verification of the chip in the identity document.

Table of Contents

What exactly is a deepfake?

A Deepfake is synthetic media: images, video, or audio generated or manipulated by artificial intelligence, in which a real person appears to say or do something that never actually happened.

The name is a contraction of deep learning, the AI technology behind it, and fake. The word first appeared in 2017 as a username on Reddit and has since become the umbrella term for anything realistically simulated using AI.

What’s important is what a deepfake is not. A cropped video, a slowed-down clip, or a misleading clip taken out of context is called a shallowfake. There's no AI involved—just an editing trick. The distinction matters, because a shallowfake can be debunked with the right context. A deepfake, on the other hand, requires technical proof.

In short: In the case of a shallowfake, the footage is real but has been edited or taken out of context. In the case of a deepfake, the footage itself is fake: that event never took place.

How is a deepfake created?

The classic technique is the Generative Adversarial Network (GAN): two AI models that work against each other. The generator creates fake images. The discriminator tries to distinguish the fake from the real. Over thousands of iterations, they correct each other until the generator creates images that the discriminator can no longer detect as fake.

Newer deepfakes also rely on diffusion models, the same family of techniques used in image generators. These produce sharper results with less training.

What has changed most of all is the amount of source material required. Whereas in the past you had to collect hundreds of photos from different angles, a few seconds of video is often enough now. The deepfakes of the Arup executives were created using publicly available material: recorded presentations, videos, and social media profiles—material that every executive at a large company has posted online.

From source material to real-time deepfake

Four steps, with less and less source material than a few years ago

01

Collecting source material

Recorded presentations, videos, public profiles. Often, just a few seconds of footage is enough.

02

Model Training

A GAN or diffusion model trains on the face and voice until the result is no longer noticeable.

03

Project a Face

The generated face is superimposed on the fraudster's own camera feed.

04

Live During the Conversation

The other person sees a colleague, a candidate, or a board member—not the person actually sitting there.

What types of deepfakes are there?

Face swap. Person A's face is superimposed onto Person B's body in an existing video. The classic version.

Real-time face swap. It's the same thing, but live during a video call. The scammer is sitting in front of his own camera; what the other person sees is a different face. This is the technique behind the Arup scam and behind job interview scams during video calls.

Voice cloning. A voice is imitated using a recording that lasts just a few seconds. This technique is used in CEO fraud and in scams targeting private individuals.

Fully synthetic individuals. Not a replica of an actual person, but a face that has never existed. Combined with real or stolen personal data, this results in a synthetic identity: a person who appears legitimate on paper but does not exist. That is a form of fraud in itself, and it deserves its own story.

What are deepfakes used for?

Against private individuals

By far the largest volume of deepfakes is pornographic and non-consensual. Other examples include scams involving cloned voices of family members and damage to one’s reputation. The damage is often permanent, even after the image has been debunked.

Against companies

This is where the most severe financial damage occurs, and the Arup incident is a textbook example.

In January 2024, a finance employee at the Hong Kong office of the British engineering firm Arup received an email from the CFO in the United Kingdom regarding a confidential transaction. He didn’t trust it. He did exactly what every security training course recommends: he asked for a video call to confirm.

The CFO was on that call. So were a few colleagues he recognized. They were all AI-generated, created from publicly available image and audio material.

He made fifteen transfers to five Hong Kong accounts, totaling HK$200 million, or approximately $25.6 million. It wasn’t until he later called headquarters that the fraud came to light. Arup confirmed afterward that no internal systems had been hacked. There was no malware, no exploit. Just a video call and a credible story.

That is exactly what makes this threat so difficult to deal with. The security measures in which organizations invest the most—multi-factor authentication, endpoint detection, and email filters—have little effect on this type of attack.

Opposed to recruitment and verification processes

And then there's the category we're in DataChecker deal with on a daily basis.

As soon as identity verification is done remotely—which has now become the norm during onboarding—the face on the screen becomes proof of identity. It is precisely that face that can now be generated.

Job Application Fraud. A candidate goes through a fully digital onboarding process: video intake, ID upload, selfie check. The face on camera is not the candidate’s. The ID document has been stolen or forged. The result could be an employee who is not authorized to work, with all the associated liability under the Chain Liability Act.

Bypassing facial recognition. This is not just a theory. During the same period as the Arup case, the Hong Kong police reported that fraudsters had used eight stolen ID cards to submit 90 loan applications and open 54 bank accounts. In at least twenty of those attempts, an AI-generated deepfake was used to deceive a facial recognition system.

Multiple uses of a single identity. Criminal networks link stolen identities to a generated face and use the same legal identity simultaneously with different staffing agencies.

In short: A video call was never intended to be a means of identity verification. Yet it was used that way, because for decades, a face on the screen was considered sufficient proof. That is no longer the case.

How can you spot a deepfake?

There are signs. They're just getting weaker every year.

Eyes and blinking. Early deepfakes barely blinked at all. Modern models are much better at this, but unnatural eye contact or a fixed gaze remains a telltale sign.

The edges of the face. Around the hair, ears, and jawline, you sometimes see faint transitions, color differences, or pixels that don't quite line up.

Lip-sync. Teeth, tongue, and lip movements do not always correspond to the sound. Consonants such as "f" and "v" are technically difficult.

Light and shadow. The face is lit differently than the surroundings, or the shadows are falling in the wrong direction.

The voice. Cloned voices often sound just a little too smooth: not enough breath, not enough variation in pitch.

Skin. Suspiciously smooth. No pores, no imperfections.

And here’s the honest conclusion: this is no longer a reliable defense. A well-made deepfake, viewed by a recruiter who conducts thirty interviews a week, will slip through the cracks. The Arup employee was observant enough to be suspicious of the email. He was still taken in, however, because he saw and heard his CFO.

The Eye Versus Technology

Two Ways to Look at the Same Picture

What the Human Eye Sees

  • Does that person blink normally?
  • Do the lips match the sound?
  • Does the shadow on the face look right?
  • Does the skin look real?

It's becoming less reliable every year. A well-made deepfake can slip through here.

What Detection Technology Measures

  • Is there a living person in front of the camera?
  • Does the image show artifacts caused by AI generation?
  • Is the face geometry consistent from frame to frame?
  • Is the signature on the document chip valid?

Measures properties that the eye does not see.

Tip: When dealing with unusual requests, never trust the face on the screen, no matter how convincing it may seem. Always verify the information through a second, independent channel: call back a number that you look up yourself, not one that was mentioned during the call.

Deepfake, spoofing, or synthetic identity?

These terms are often used interchangeably. The difference determines which defense works.

What It IsWhat it attacks
DeepfakeAI-generated image or audio of a personYour trust in what you see and hear
SpoofingImpersonating someone else using a fake contact method, such as a phone number or email addressYour trust in the sender
Synthetic IdentityA person who does not exist, made up of real and fictional detailsYour confidence in the case file

In practice, they go hand in hand. The Arup fraud began with a spoofed email and escalated into a deepfake video call.

Is creating a deepfake a criminal offense?

There is no specific law in the Netherlands called “deepfake.” What does exist are existing criminal provisions that are applied to such cases.

For sexual deepfakes, Article 139h of the Dutch Criminal Code serves as the legal basis; since 2020, this article has made the production and publication of visual material of a sexual nature without consent a criminal offense. In 2023, the Amsterdam District Court ruled that a deepfake porn video falls under this provision, provided that the footage looks so real that a reasonable person could mistake it for authentic. In addition, fraud, defamation, coercion, and identity theft may come into play, depending on the case.

The WODC had previously concluded that the problem lies not so much in the legislation as in enforcement: proving that something is a deepfake requires technical expertise that is still scarce.

What Will Change on August 2, 2026

As of 2 August 2026, Article 50 of the AI ​​Act (Regulation (EU) 2024/1689) applies. That article introduces an explicit transparency obligation for deepfakes: anyone who publishes AI-generated or manipulated images, audio, or video that is realistic enough to be mistaken for real must make it clear that it is artificial.

Two things are important in this regard. First, this obligation has not been postponed. The high-risk obligations under the AI ​​Act have been carried forward via the Digital Omnibus package, but the transparency obligation for deepfakes remains set for August 2. Second, malicious intent is not required: a well-intentioned or commercial deepfake also falls under it. A lighter form of the obligation applies to clearly artistic, satirical, or fictional work. Fines run up to 15 million euros or 3% of global annual turnover.

That is, of course, a transparency obligation for those who create deepfakes, not protection against those who use them maliciously. Fraudsters will not label their deepfakes. For organizations that verify identities, therefore, little changes legally on this point; the defense remains technical in nature.

For more information on how the AI Act affects biometric applications, see our article on The AI Act and Biometric Identification.

What should you do if you're the victim of a deepfake?

  1. Document the evidence. Take screenshots or download the content before it disappears. Make a note of where and when you found it.
  2. Report it to the platform. Under the Digital Services Act, social media platforms are required to remove unlawful content and have reporting procedures in place for that purpose.
  3. File a report. In cases of sexual or identity-related deepfakes, it is possible and advisable to file a police report, even if gathering evidence is difficult.
  4. Ask for help. Victim Support Netherlands provides support to victims of online abuse.
  5. Consider filing a complaint with the Dutch Data Protection Authority. Personal data is used for a deepfake of an existing person, usually without a legal basis. This can give rise to a complaint with the Dutch Data Protection Authority (AP).

How Technology Detects Deepfakes

If the human eye can no longer detect it, the verification process must take over. Effective detection works in layers, because each layer on its own can be circumvented.

For example, DataChecker set it up.

Four-Layer Detection

Each layer can be bypassed individually. Together, they make it much more difficult.

Layer 1 · Liveness Detection

Active and passive at the same time. Is there a living person in front of the camera, or is it a photo, a recording, or a computer-generated face?

Layer 2 · AI Detection of Synthetic Media

Analysis of texture patterns, facial geometry, and generation artifacts. Provides a risk score, not a simple yes or no.

Layer 3 · The chip in the document

The chip data is digitally signed by the issuing authority. This establishes that the document is authentic, independent of what happens in front of the camera.

Layer 4 · Where the Data Goes

ISO/IEC 27001:2022 certified by BSI. All verification data is processed within the EU.

Layer 1: Liveness Detection

FaceVerify It does not just compare a selfie with the photo on the ID document. It checks whether there is a living person in front of the camera, and not a photo, a screen recording, or an AI-generated face.

This happens in two ways simultaneously. Active liveness requires the user to perform an action. Passive liveness analyzes the image in the background, without the user noticing. This combination is intentional: an attack that bypasses one layer usually stumbles over the other.

Layer 2: AI-based detection of synthetic media

On top of the liveness layer, it analyzes DataChecker the video frames for signs of AI generation: unusual texture patterns, inconsistencies in facial geometry, and artifacts typical of generated images.

The outcome is not a simple yes or no, but a real-time risk score. This is a deliberate choice. A hard yes or no forces the system to always choose one side in cases of doubt, and therefore sometimes be wrong. A score allows the assessor to make the assessment themselves, taking into account the context of the application.

Layer 3: the chip in the document

This is a strong layer, and the reason why document verification is fundamentally different from facial verification.

A face can be generated. The data on the chip of an ICAO-compliant passport is digitally signed by the authority that issued the document. DataChecker reads the original document data from the NFC chip and verifies whether the signature is correct.

This makes chip verification one of the strongest checks on the authenticity of an identity document. Facial comparison and liveness detection determine whether the person in front of the camera is indeed the holder of that document. It is precisely this combination that makes it difficult for a deepfake.

Layer 4: Where the Data Goes

DataChecker is ISO/IEC 27001:2022 certified by BSI (ISC 365 certificate) for the development and implementation of the verification of personal data. All verification data is processed within the EU.

Schedule a demo and see how deepfake detection works in a real onboarding process.

Last updated: November 2026. The entry into force of Article 50 of the AI ​​Act is scheduled for 2 August 2026; the European Commission is still finalizing the guidelines and the code of conduct.

Frequently Asked Questions

Frequently Asked Questions About Deepfakes

What does "deepfake" mean?

"Deepfake" is a portmanteau of deep learning and fake. It refers to an image, video, or audio clip generated or manipulated by AI in which a person appears to say or do something that never actually happened.

In the case of a shallowfake, the footage is real but has been manipulated through editing or misleading context. In the case of a deepfake, the footage itself is generated by AI.

Yes. During digital onboarding and video calls, a fraudster can use a real-time face swap. The Hong Kong police found that AI deepfakes were able to fool facial recognition systems in at least twenty cases.

There is no specific legal provision addressing deepfakes. However, existing provisions are applied: in 2023, the Amsterdam District Court ruled that a deepfake porn video falls under Article 139h of the Dutch Criminal Code. Depending on the case, charges such as fraud, defamation, or identity theft may also apply.

In that case, Section 50 of the AI ​​Act applies, with a transparency obligation for deepfakes. Anyone publishing realistic AI imagery must disclose that it is artificial, even without malicious intent. Fines run up to 15 million euros or 3% of global turnover.

They're becoming less and less reliable. Signs such as unnatural blinking, blurry facial edges, and off-sync lip movements help, but well-made deepfakes still slip through. Technical detection is necessary.

Liveness detection alone is not enough to defend against the most sophisticated attacks. That is why it combines DataChecker Active and passive liveness detection, including AI-based detection of synthetic media and reading of the NFC chip in the identity document. That chip data is digitally signed by the issuing authority and is independent of how convincing the face appears on camera.

Go to Top