AIHow AI Works

Deepfake

AI-synthesized video, image or audio that makes a real person appear to say or do something they never did — how deepfakes are made, and how they are caught.

Definition

A deepfake is a video, image or audio recording that has been synthesized or altered by AI so
that a real, identifiable person appears to say or do something they never actually said or did.

The word is a blend of “deep learning” and “fake,” and the distinction it draws matters: the harm
is not that the media is synthetic, but that it impersonates someone real. An AI-generated
landscape is synthetic media. An AI-generated video of a sitting head of state announcing a
military strike is a deepfake. The EU AI Act encodes exactly this line, defining a deep fake as
AI-generated or manipulated content that resembles existing persons, objects, places, entities or
events and “would falsely appear to a person to be authentic or truthful.”

Note what that definition turns on: appearance, not intent. Content that would pass as
authentic is a deepfake in the legal sense even if the creator never meant to deceive anyone.

How It Works

The techniques differ by medium, but they share a shape: learn a compressed representation of a
person’s appearance or voice, then reconstruct it under someone else’s control.

Autoencoder face swapping is the classic method and still the most common for video. Two
autoencoders are trained with a shared encoder but separate decoders —
one per identity. The shared encoder learns a general-purpose compression of “a human face at some
angle with some expression”; each decoder learns to rebuild one specific person from that code. To
swap, you run person A’s face through the encoder and rebuild it with person B’s decoder. The
result carries A’s pose, expression and lighting, wearing B’s identity. This is why deepfakes have
historically needed a lot of footage of the target: the decoder has to learn that one face
thoroughly.

Generative adversarial networks (GANs) train a generator against a discriminator whose only job
is to spot fakes. The generator improves until the discriminator cannot tell the difference — an
objective that, read carefully, is literally the objective of defeating a detector. This is the
root of why deepfake detection is structurally an arms race rather than a solvable problem.

Diffusion models now drive the highest-quality video generation,
starting from noise and iteratively denoising toward an image conditioned on a text prompt or a
reference face. They removed the old requirement for hours of target footage: a handful of
reference images can be enough.

Voice cloning encodes a speaker’s timbre into an embedding vector from a short sample, then
conditions a speech synthesis model on it. A few seconds of audio — a voicemail greeting, a
conference talk, an Instagram story — is now sufficient.

Lip-sync models take a different route entirely: they leave the original video intact and
re-animate only the mouth region, driven by a new audio track. Because so little of the frame
changes, these are among the hardest to catch by eye.

Types

  • Face swap: one person’s face is replaced with another’s, keeping the original body, pose and
    scene.
  • Face reenactment (puppeteering): the target’s own face is kept, but their expressions and
    head movements are driven by a performer — the target appears to speak words chosen by someone
    else.
  • Lip sync: only the mouth is re-animated to match new audio, leaving the rest of the video
    untouched.
  • Voice cloning: audio-only impersonation of a specific person’s voice, the form now most used
    in live fraud.
  • Full synthesis: an entirely fabricated scene or person. When the person does not exist, this
    is generative media rather than a deepfake in the strict sense — but it is used for fake
    profiles and fabricated “witnesses” at scale.

Real-World Applications

Deepfake technology is dual-use, and the legitimate side is not marginal.

Sanctioned uses:

  • Film and localization: de-aging actors, replacing stunt performers’ faces, and dubbing films
    into other languages with the actor’s own voice and matching lip movements.
  • Accessibility: reconstructing a synthetic voice for people who have lost speech to conditions
    such as ALS, using recordings made before the loss.
  • Corporate media: producing training and marketing video in dozens of languages from a single
    recorded presenter, with their consent.
  • Satire and art, which the EU AI Act explicitly accommodates: where a work is evidently
    artistic, creative, satirical or fictional, disclosure must be made in a way that does not hamper
    enjoyment of the work.

Malicious uses:

  • Non-consensual intimate imagery, which remains the most widespread documented harm to
    individuals and overwhelmingly targets women.
  • Financial fraud: a cloned executive voice on a live call authorizing an urgent wire transfer,
    or a synthetic “family member” manufacturing an emergency.
  • Identity verification bypass: injecting synthetic video into a KYC or liveness check to open
    accounts under a stolen identity.
  • Political disinformation, particularly fabricated audio, which is cheap to produce, easy to
    distribute and lacks the visual tells people are learning to look for.

Challenges

  • Detection is a race against the generator, and the generator sets the pace. A detector is a
    classifier trained on the models that existed when it was built. Every new generator is, by
    construction, out of distribution for it — and nothing in the detector’s confidence score tells
    you that it has left the domain it was trained on.
  • Compression destroys the evidence. Forensic detection leans on low-level signals — sensor
    noise, frequency-domain upsampling artifacts, frame-to-frame inconsistency. Uploading to a social
    platform recompresses the video and attenuates precisely those signals. The forensics work best
    on the pristine file that a real investigator almost never has.
  • Physiological signals are elegant but fragile. Remote photoplethysmography recovers a pulse
    waveform from the faint colour shifts real skin undergoes with each heartbeat — a synthesized
    face has no heartbeat. But low light, heavy compression and makeup degrade the signal in genuine
    video too, which produces false alarms on real people.
  • The liar’s dividend. The mere existence of convincing deepfakes gives anyone caught on
    camera a ready defence: it was faked. The damage is not only that fakes are believed, but that
    authentic evidence becomes deniable — and this harm grows even if every deepfake were detected
    perfectly.
  • Real time changes the threat model. A voice clone on a live call cannot be sent away for
    forensic analysis. Detection has to run inside the call, in milliseconds, or it does not run.
  • A score is not evidence. A detector that outputs “87% likely fake” with no explanation cannot
    be presented in a newsroom or a courtroom. In July 2026 the Fraunhofer IOSB institute published
    RealOrRender, an explainable classifier reporting 85–91% accuracy that also states why it
    reached its verdict — an acknowledgment that the field’s usable output was never the number.
  • Provenance is displacing detection. Rather than interrogate the pixels after the fact, the
    industry is moving to sign media at the point of creation — see
    Content Provenance (C2PA). Verification becomes a cryptographic
    check instead of a statistical guess, and it does not degrade every time a better generator
    ships.
  • Disclosure becomes law. China has required both visible and embedded labels on AI-generated
    content since 1 September 2025. From 2 August 2026, Article 50 of the EU AI Act requires deployers
    to disclose deepfake image, audio and video content as artificially generated or manipulated, and
    requires generative systems to mark their output in a machine-readable format. Disclosure
    obligations now turn on how the content appears, not on whether anyone meant to deceive.
  • Detection moves into the pipeline. Because voice fraud happens live, detection is being built
    into the call and streaming stack rather than offered as an after-the-fact upload tool.
  • Explainability becomes the requirement. Verdicts that cannot be justified are not actionable,
    and the tooling is shifting accordingly.

For how these detection methods actually perform in practice, and what the shift to signed
provenance means, see
How AI-Generated Video, Photo and Text Are Actually Detected.

Common questions

What exactly makes something a deepfake?

A deepfake is synthetic or manipulated media depicting a real, identifiable person doing or saying something they did not do. Fully invented people and abstract AI art are generated content, but they are not deepfakes — the defining feature is the resemblance to someone real.

Can you spot a deepfake by looking at it?

Increasingly, no. Visual tells such as malformed hands, odd blinking and mismatched shadows are being engineered out with each model generation, and relying on them is a strategy with an expiry date. Forensic tools and provenance signatures are more durable than the naked eye.

Are deepfakes illegal?

It depends on use and jurisdiction. The technology itself is not banned, but specific uses are criminalized: the US TAKE IT DOWN Act, signed on 19 May 2025, makes non-consensual intimate deepfakes a federal crime. From 19 May 2026 its notice-and-removal duty also requires covered platforms to take down reported non-consensual intimate images — AI-generated deepfakes included — within 48 hours of a valid request. Separately, disclosure is becoming mandatory — China has required labeling since September 2025, and the EU AI Act requires it from 2 August 2026 regardless of whether there was intent to deceive.

How are deepfake voices used in fraud?

A few seconds of a person's recorded speech is enough for modern voice-cloning models to reproduce their voice. Attackers use this on live calls — impersonating an executive to authorize a payment, or a family member to manufacture an emergency — which leaves the target no time for offline forensic analysis.

Is there a reliable deepfake detector?

No detector is reliable enough to be treated as proof. Detectors are classifiers trained on the generators that existed when they were built, and their accuracy degrades against newer models and against ordinary compression and re-encoding. Treat a detector score as one signal, not a verdict.