Deepfake

The word combines deep learning, the training of deep neural networks, with fake. It emerged in 2017 on an online forum where users shared videos with swapped faces. Today it covers videos, images and audio recordings in which people say or do things that never happened: a swapped face, altered lip movements, a cloned voice.

The EU AI Act defines the term in Article 3(60): AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful. The legal definition is broader than everyday usage. It requires no intent to deceive and covers places, objects and events as well as people. Text alone does not qualify.

For executives, the topic has two sides. Fraudsters use cloned voices and faces of managers to obtain payments or credentials. And companies that use realistic AI content showing real people or places in marketing or internal communication may be creating deepfakes themselves. Since August 2, 2026, they have had to disclose that in the EU.

How a deepfake is made.

It starts with material of the target: photos, videos, voice recordings. A neural network learns the face, expressions or voice from that material and transfers them to new content. Germany's Federal Office for Information Security (BSI) describes the basic principle: the networks learn to extract facial expressions and lighting from face images and generate matching new ones.

In a face swap, the model replaces a face frame by frame while the original head movements and expressions remain. In face reenactment, someone drives another person's expressions with their own face. Lip sync adjusts mouth movements to a new script. Image and video generators now create entire scenes from a text description.

Voices take especially little effort. Researchers are working on methods that clone a previously unheard voice from just a few seconds of speech; a study published in Scientific Reports in 2025 describes this as the aim of its field. Some methods work in real time, including in phone calls and video conferences.

Every public video, interview and podcast featuring executives is potential raw material.

How deepfakes hit companies.

A widespread pattern is fraud by a fake executive, often called CEO fraud. A voice on the phone or a face on a video call poses as senior management and demands an urgent, confidential transfer, usually to a new account. The technology replaces the fake email of the past with something that feels more real.

Authorities warn about this explicitly. In May 2025, the FBI reported a campaign in which, since April 2025, criminals had used AI-generated voice messages to impersonate senior US officials. AI-generated content has advanced to the point that it is often difficult to identify, the alert says.

Another target is the IT help desk that resets passwords after a phone call. Then there is reputational damage: a fake video of the CEO saying something offensive, an invented product defect, a fake press conference.

The damage rarely happens in the IT department. It happens where someone moves money, grants access or responds publicly on instruction. Protection therefore starts with those processes and the people who carry them out.

Why detection alone won't protect you.

Flickering edges, unnatural blinking, lips out of sync, wrong lighting: such artifacts are clues, not proof, and they become rarer with every generation of generators. On the phone, there is no image at all. Detection tools deliver probabilities. They are trained on known methods and lag behind new ones. The BSI notes that countermeasures cannot be applied in every situation and generally do not provide complete protection.

Provenance helps where it exists. Digital signatures based on the C2PA standard or invisible watermarks can show that content comes from a specific source or was generated by AI. Their absence proves nothing, because fraudsters don't add them.

The reliable check runs through a second channel. The FBI recommends verifying identities by calling back an independently verified number rather than the one in the message. Inside a company, that means payments to new bank details, password resets and unusual instructions go ahead only after a callback to stored contact details and approval by a second person. The rule holds even when the instruction seems to come from the top, and staff who approve payments practice it.

The disclosure duty under Article 50.

Article 50 of the EU AI Act has applied since August 2, 2026. Under paragraph 4, deployers of an AI system that generates or manipulates deepfakes must disclose that the content has been artificially generated or manipulated. Under Article 3(4), a deployer is anyone using an AI system under their own authority, such as a company that uses an image or video tool for its own content. Purely personal, non-professional use is exempt (Article 2(10)).

For evidently artistic, creative, satirical or fictional works, an appropriate notice that does not hamper the display or enjoyment of the work is enough. Unlike published AI-generated text, deepfakes get no exemption for human review. Under paragraph 5, the notice must be clear and distinguishable, appear at the latest at first exposure and meet accessibility requirements.

Separately, the providers of the tools must mark their outputs in a machine-readable way. For systems already on the market before August 2, 2026, that duty applies from December 2, 2026, under amending Regulation (EU) 2026/1744. Violations of Article 50 can draw fines under Article 99 of up to 15 million euros or 3 percent of worldwide annual turnover, whichever is higher; for small and medium-sized enterprises, the lower amount applies (as of November 2026).

Bans and criminal liability, as of November 2026.

Deepfakes are not banned in the EU as such. Amending Regulation (EU) 2026/1744, however, extends the prohibited practices in Article 5. From December 2, 2026, it is prohibited to provide or use AI systems to generate realistic intimate depictions of an identifiable person without explicit consent, or child sexual abuse material. For providers, the ban applies if the system is intended for this or foreseeably produces such content without adequate safeguards. For prohibited practices, Article 99 provides for fines of up to 35 million euros or 7 percent of worldwide annual turnover.

Germany has no specific criminal offense for deepfakes yet. Depending on the case, existing rules apply: the right to one's own image under Section 22 of the Art Copyright Act (KUG), which allows portraits to be distributed only with consent, general personality rights, data protection law and, for payment fraud, the fraud offense. On October 7, 2026, the federal cabinet approved a draft law against digital violence. It adds Section 184k of the Criminal Code on creating or distributing sexualized deepfakes without authorization and Section 201b on distributing reputation-damaging deepfakes. The draft is now before the Bundestag and Bundesrat.

In the US, the TAKE IT DOWN Act, signed in May 2025, criminalizes publishing non-consensual intimate depictions, explicitly including digital forgeries, and requires platforms to remove reported content within 48 hours.

Deepfakes in your own marketing.

The same technology has legitimate uses: synthetic narrators for training videos, translations in the voice of your own CEO, product videos without a shoot. It gets legally sensitive once real people, places or events are recreated realistically.

Real people need to give consent, ideally in writing, that covers purpose, duration, channels and withdrawal; that applies to employees and to your own leadership as well. The finished content is labeled under Article 50. A deepfake can exist without any person in it, for example a realistically generated photo of a real building or of an event that never happened. An invented AI character that resembles no real person, on the other hand, is not a deepfake under the regulation.

A short check before each publication works well in practice: Does the content show a real person, place or event? Would it pass as real? Is consent on file, and is the notice in place? Advertising with recreated celebrities who have not agreed will as a rule violate their personality rights.

Where the term stops working

A deepfake requires three things: creation or alteration with AI, a realistic impression and a link to real people, places, objects or events. Conventionally edited footage (shallowfakes), invented AI characters with no real counterpart and AI-generated text are not deepfakes. Synthetic media is the broader umbrella term for all AI-generated content; AI labeling covers the duties and methods for disclosing it.