Rules
What a deepfake is, which rules apply now, which prohibition is coming and how organisations tackle fraud using a fabricated voice or video.
A deepfake is image, audio or video content created or manipulated with AI that falsely appears authentic. Creating one is not automatically prohibited. It depends on the content, who makes it and what happens to it. A sexual deepfake without consent is a criminal offence. Other deepfakes are subject to rules on transparency, personal data and portraits.
That is the position on 13 October 2026. Four kinds of rules overlap: the EU AI Act, GDPR, portrait rights and criminal law. This article is an explanation, not legal advice.
The AI Act provides a definition in Article 3. A deepfake is AI-generated or manipulated image, audio or video content resembling existing persons, objects, places, organisations or events that would falsely appear authentic or truthful to someone. Three points stand out.
On its deepfakes page, the Autoriteit Persoonsgegevens (AP), the Dutch data protection authority, describes a deepfake as a video or audio fragment manipulated by an algorithm, making it appear someone said or did something that never happened. Improved lighting or a clearly drawn character is excluded: the boundary is whether someone takes it for real.
The AI Act does not prohibit deepfakes generally. It opts for transparency, with one prohibition added in December 2026.
Article 50 has applied since 2 August 2026. Paragraph 4 says deployers of an AI system generating or manipulating a deepfake must disclose that the material was artificially generated or manipulated. The information must be clear and given no later than first exposure to the material.
On 10 June 2026, the European Commission published a code of practice on marking and labelling AI-generated content. Its second part covers labelling deepfakes, with icons usable as labels. Participation is voluntary; the Article 50 duty is not.
Paragraph 2 targets developers of generative AI systems. They mark output in a machine-readable form as artificially generated or manipulated insofar as technically feasible. Systems already on the market before 2 August 2026 have until 2 December 2026. What such marking does and does not solve is in our article about the AI detector.
The digital omnibus on AI extended Article 5's list of prohibited practices. The adopted text contains two new points.
The text limits the prohibition. Offering a system is prohibited only if such creation is its intended purpose or a reasonably foreseeable outcome against which the system lacks adequate safeguards. Use is prohibited only if the user employs it for that purpose. Under amended Article 113, these points apply from 2 December 2026. See the AI Act in autumn 2026.
Article 99 provides fines up to €35 million or 7% of worldwide annual turnover for breaching Article 5 prohibitions, and up to €15 million or 3% for breaching Article 50 transparency duties. The implementing legislation determines who imposes fines in the Netherlands. As far as we could establish, this had not yet become law on 13 October 2026.
The GDPR defines personal data as all information about an identified or identifiable person, including identification through physical identity characteristics. This is in Article 4 GDPR. The AP says that if your face or body is recognisable in a deepfake, its creator processes your personal data. For a voice identifiable as one person's, the same follows from the definition.
A photograph or recording is not thereby special category data. Recital 51 GDPR says photographs count as biometric data only if processed with particular technical means to identify someone uniquely. When that is the case is explained in our article on special category data.
The AP gives four steps. Ask the creator to delete the deepfake; the AP says they are usually obliged to because a deepfake can be regarded as inaccurate personal data. If you do not know the creator, ask the platform. If refused, you may complain to the AP. If you suffer damage, you can go to court. The same rules apply to well-known people, but a court balances their interests against the creator's, including satire.
Portrait rights are in Articles 19 to 21 of the Dutch Copyright Act. For a portrait not commissioned by its subject, Article 21 says publication is not permitted insofar as that person's reasonable interest opposes it. The AP notes portrait rights can play a part in deepfakes. The statutory text refers to a portrait. We could not establish from the law whether this also covers a fabricated voice; ask a lawyer.
A private member's bill giving everyone their own right over deepfakes of their voice and appearance, the Act on neighbouring rights in deepfakes of persons, was consulted on from 30 October to 31 December 2025. On 13 October 2026, we found no source showing it had been submitted to the Dutch House of Representatives. It is not law in force.
The Dutch Criminal Code has no article using the word deepfake. Existing articles may apply.
The public prosecution service and courts assess whether a specific case is criminal. Victims should report it to police.
In CEO fraud, a finance employee receives an instruction apparently from the director: transfer a large amount quickly. The Fraudehelpdesk lists recurring features: emphasis on authority, confidentiality and time pressure. It says smaller businesses, foundations and associations are affected too.
The deepfake makes the instruction more credible. In 2019, the Fraudehelpdesk saw the first reports using an imitated director's voice in a spoken message. This appears in its overview of fraud over the past decade.
In a November 2023 interview at the National Cyber Security Centre, a police specialist describes a failed attempt at a Dutch bank: an employee was asked in a video call by the supposed chief executive to transfer money and called the real chief executive back. She says detection tools lag behind developments and behaviour solves this: calling back or asking a verification question must be normal.
We did not find figures on the scale of deepfake fraud in the Netherlands in a primary source.
An organisation can also be the creator: an instructional video with a digital colleague or a voiceover using a cloned voice. You then need a lawful basis to use face and voice, document purpose and duration, and disclose that the material is artificial. Consent is a difficult lawful basis in employment because employees do not always feel free to say no. Include synthetic voice and image use in your AI policy.
| Situation | Which rule | What you do |
|---|---|---|
| Your organisation publishes a video with a fabricated person or voice | AI Act, Article 50(4); GDPR | Clearly disclose artificial material, arrange a lawful basis and document agreements with the depicted person |
| A satirical or artistic production | AI Act, Article 50(4), limited duty | Disclose the presence of artificial material; have the assessment checked if real people feature |
| An employee or client appears in a deepfake without consent | GDPR; portrait rights | Help request deletion from creator and platform, point out the AP complaint option |
| A sexual deepfake of a recognisable person | Dutch Criminal Code, Article 254ba; GDPR; from 2 December 2026 also Article 5 AI Act | Report to police, request deletion, retain evidence without further sharing |
| A payment instruction with a manager's voice or face | Dutch Criminal Code, Article 326 | Do not pay, call back on a known number, report to bank and police |
| A fabricated recording damaging someone's reputation | Dutch Criminal Code, Articles 261 and 262; GDPR; portrait rights | Record where it is, request deletion, seek legal advice |
The table is an aid. The statutory text and circumstances determine what applies.
PrivaSearch does not detect deepfakes or prevent fraud. It works with documents, not images or audio you want to authenticate.
It does address the underlying point: data about people should not enter an AI service unchecked. PrivaSearch sits between your documents and the AI model. Personal data is detected and replaced with pseudonyms, and a colleague approves the result before a document becomes searchable. The steps are on How PrivaSearch works. What happens when such documents go straight into a chat window is on the page about AI chatbots and the GDPR.
Employees also complete short AI literacy training, repeated periodically. Without valid training, the workspace does not open. It does not replace fraud training for your finance department.
Two limits apply. Detection is never complete, so human review is mandatory. And pseudonymised data remains personal data for your organisation. How the measures relate to the rules is on the page about GDPR and the AI Act. To see how it works, book a demo.
Not always. The Dutch Criminal Code does not mention the word deepfake. Creating a sexual image of someone without consent is an offence, and using a deepfake to obtain someone's money may be fraud. Outside your personal circle, the GDPR and portrait rights also apply. The public prosecution service and courts assess whether a specific case is criminal.
The AI Act does not generally prohibit deepfakes. Since 2 August 2026, anyone professionally creating or manipulating a deepfake must disclose that the material is artificial. Generative AI providers must mark output in machine-readable form. On 2 December 2026, an additional prohibition takes effect on AI systems creating sexual imagery of recognisable people without consent and systems creating child sexual abuse material.
Yes, if your organisation or a professional creates and distributes it. This follows from Article 50 AI Act. Disclosure must be clear and given no later than first exposure to the material. For clearly artistic, satirical or fictional work, a limited disclosure suffices.
Ask the creator to delete it. According to the Autoriteit Persoonsgegevens, the creator is usually obliged to do so. If you do not know who created it, ask the platform hosting it. If refused, you may complain to the Autoriteit Persoonsgegevens. A sexual deepfake may be reported to police.
Do not rely on your eyes and ears or a detection tool. Focus on the request: urgency, secrecy, an unknown account and a route differing from normal procedure. Call back on a number you already had and have a second person approve payment.
A voice identifiable as one person's is information about an identifiable person and therefore falls within the GDPR's personal data definition. An organisation recording, cloning or using someone's voice needs a lawful basis and must inform them. A recording is not automatically special category data. It becomes that only through technical processing to identify someone uniquely.
The AI Act timeline after the digital omnibus: what applies already, what has moved to 2027 and 2028, and what AI chatbot users need to arrange now.
What the Dutch Algorithm Register and IAMA are, what is mandatory and what is not, and what a municipality arranges when staff use AI chatbots.
When you must conduct a DPIA for an AI chatbot or AI search feature, what it needs to contain and how to approach the assessment in seven steps.
We show it with fictitious client files from your sector: upload, detection, review, search and the connection with your AI assistant. Afterwards you'll know exactly what your team can use safely.
Book a demo →