AI Tools

Deepfakes: How They're Made, How to Spot Them, Why They're Risky

A video where a famous politician says something they never said. A call "from your boss" in a voice indistinguishable from the real one. These are deepfakes, and by 2026 they are within reach of literally anyone. Let's cut through the panic and the myths: how the technology works, the signs that give a fake away, and how not to become a victim.

A white humanoid robot in a dynamic pose on a blue-and-pink background — an image of artificial intelligence
Behind deepfakes is the same AI that paints and animates faces — the only question is the intent. Photo: Pexels

What a deepfake is in plain English

A deepfake is a photo, video or audio clip in which artificial intelligence has swapped a person's face, expression or voice so that they appear to say and do things that never happened. The word itself is a blend of deep learning and fake.

It is important to grasp: a deepfake is not a crude cut-and-paste montage. The neural network re-synthesises the face or voice, fitting it to the expression, lighting and articulation of the original. That is exactly why a good deepfake is so convincing — it has none of the seams that gave away old Photoshop jobs.

In short

A deepfake is "synthetic plausibility." The AI does not splice ready-made pieces; it generates a new face or voice to fit someone else's expression and speech. Hence both the power of the technology and its danger: telling a fake by eye is getting harder and harder.

How deepfakes are made: GANs, diffusion, autoencoders

Under the hood, deepfakes run on three families of generative models. You do not need to master them, but it helps to know: each has its own "signature" and its own artifacts.

  • Autoencoders. The classic face-swap. The model compresses a face into a compact "code," then rebuilds it — but in the likeness of another person, preserving the source's expression.
  • GANs (generative adversarial networks). Two rivals go to work: a generator creates a face and a discriminator tries to tell fake from real. The generator improves until the discriminator can no longer catch it. GANs give a sharper picture but are finicky to train.
  • Diffusion models. The same family that paints AI images. They deliver the highest quality and the best identity preservation in a swap, but demand more compute.

By type of manipulation, deepfakes fall into face swap, facial reenactment (someone else's face mimics your movements) and lip-sync (the lips are fitted to a new audio track). A separate and especially dangerous branch is voice cloning, which we covered in detail in our piece on AI voiceover and voice cloning.

A robot arm hands a white mug to a bearded man on a grey background — human-AI interaction
AI mimics the human ever more convincingly — from a gesture to a face and a voice. Photo: Pexels

How to spot a deepfake: 6 signs

The good news: even high-quality deepfakes often give themselves away. Here is what to look at first — in 2026 these artifacts remain the most reliable clues.

1

The face edge jitters

Along the face contour, the hairline and the neck you can see shimmer, a "floating" edge or slight jitter between frames.

2

Unnatural blinking

Too rare, too frequent or "mechanical" blinking is a classic sign of synthetic footage.

3

The light does not match

Lighting on the face does not fit the scene: shadows fall the wrong way, highlights on the skin look "borrowed."

4

Lips and sound drift

Articulation lags or leads the audio slightly — most visible on plosive consonants.

5

Color shifts at the neck

A sharp change in skin tone at the boundary of the "pasted" face and the real neck or hair.

6

Small details warp

Teeth, ears, earrings, glasses and reflections in the eyes often come out blurry or asymmetric.

Beware overconfidence

The better the models get, the more these artifacts shrink. In experiments, people identified a synthetic voice only about 60% of the time — barely better than a coin flip. The takeaway: do not rely on eye and ear alone. Suspicious content must be checked through its source and provenance, not by "gut feeling."

Provenance and C2PA: a digital seal of authenticity

Since the artifacts keep getting subtler, the industry is betting on provenance — a traceable origin for the file. The key standard here is C2PA: cryptographically signed metadata that records what created or edited the content, and when. It is like a tamper seal on the file: you cannot forge it without breaking the signature.

The second layer of defense is invisible watermarks (such as Google's SynthID): a hidden mark inside the image, video or audio itself that survives compression, cropping and format changes. Under the new European code of practice, robust marking combines both approaches — signed C2PA metadata and an imperceptible watermark at the same time.

The same provenance principle underpins detecting AI images — we broke down C2PA, SynthID and metadata checks in detail in our guide on how to detect an AI image. For deepfake video and audio, the exact same mechanisms apply.

~60%human accuracy by ear
C2PAprovenance standard
Aug 2, 2026EU labelling rules begin

Risks: fraud and disinformation

Deepfakes are not just entertainment clips. Their dark side already costs people real money and reputation.

  • Financial fraud. A cloned voice of a "boss" asks for an urgent transfer; a "relative" calls with a fake emergency. According to the FBI, AI-enabled scams cost victims hundreds of millions of dollars in 2025 alone, and older people are hit especially hard.
  • Disinformation. Fake videos of politicians and "breaking news" are pushed into social feeds to sway opinion or wreck a reputation.
  • Smearing and harassment. Face-swapping into pornographic or defamatory content without consent is one of the most destructive uses.
  • Bypassing biometrics. Real-time deepfakes try to fool video-identity systems during remote verification.

Law and labelling: what changes in 2026

Regulators are catching up with the technology. The big event of 2026 is the transparency requirements of the EU AI Act taking effect. From August 2, 2026, deepfakes must be clearly labelled as generated or manipulated content — even without intent to deceive. If the content looks or sounds like a real person, it must be marked.

On top of that, the European Commission published a Code of Practice on marking AI content, and the marks must be machine-readable — so platforms can automatically detect what was generated. A single labelling icon is being introduced for simplicity. In parallel, many countries enforce laws against fraud, defamation and non-consensual deepfake pornography, and in the US more than a dozen states have already passed voice-protection laws.

What this means for you

Gradually, "honest" AI content will more often arrive with a label, and its absence on suspicious material is itself a reason to be wary. But you cannot rely on labelling alone: bad actors do not add it, so healthy skepticism and checking the source remain your main defense.

How to protect yourself

There is no hundred-percent armor, but simple habits noticeably cut the risk of falling for a deepfake.

Code word
Agree on a password with family to verify calls
Distrust urgency
Any urgent money request is a stop signal
Fewer raw data
Do not post long clean recordings of your voice and face
Verify the channel
Call the person back yourself on a known number
Provenance
Check for a C2PA mark / watermark on key content
Reverse search
Run the frame through reverse image search

Working with images and frames?

To run a frame through reverse image search or prep it for analysis, the file often needs to be re-encoded to PNG or JPG. The free FormatZ converters do that in a couple of seconds — right in the browser, no install.

Open all converters

Deepfakes are the flip side of the same technologies that power AI video generators and voice cloning. Understanding how they work makes it easier to both create honestly and defend yourself. For a full map of generative tools, see our guide to the best AI tools of 2026.

A deepfake is dangerous not because it is perfect, but because we stop doubting. Healthy skepticism is the best detector.
A deepfake is a photo, video or audio clip where AI has swapped a person's face, expression or voice so they appear to say and do things that never happened. The name comes from deep learning and fake. It is built on generative models: autoencoders, GANs and diffusion networks.
Look for tell-tale artifacts: jitter along the edge of the face, unnatural blinking, lighting on the face that does not match the scene, lips out of sync with the audio, and sharp color shifts at the neck and hairline. Provenance checks help too: honestly made content often carries a C2PA mark or a hidden watermark.
The main risks are fraud and disinformation. A cloned voice of a boss or relative is used to extract money, and fake videos discredit people and spread false news. According to the FBI, AI-enabled scams caused victims hundreds of millions of dollars in losses in 2025 alone.
The technology itself is not banned, but its use is increasingly regulated. Under the EU AI Act, from August 2026 deepfakes must be clearly labelled as generated or manipulated content — even without intent to deceive. Fraud, defamation and non-consensual pornographic deepfakes are punishable in many countries.
There is no perfect defense, but you can cut the risk: limit public access to large amounts of your photos and video, avoid posting long clean recordings of your voice, agree on a code word with family to verify suspicious calls, and treat urgent money requests critically even if the voice sounds familiar.