Deepfakes should not be used to manipulate, deceive or mislead people. That part is not controversial. But the more I look at the European Commission's draft guidance on deepfakes under the AI Act, the more I think the real issue is not whether we need transparency. We do. The real issue is whether the way we label AI content will actually create clarity — or whether it will create the next layer of confusion.
The reference point for this article is the European Commission's draft Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of the AI Act, published on 8 May 2026.
Deepfakes are no longer only about malicious deception
In public debate, the word "deepfake" still sounds like something clearly malicious: a fake politician, a fake CEO, a fake celebrity endorsement, a fraud attempt, a manipulation campaign. But the European Commission's draft Guidelines make the practical question much broader.
Article 3(60) of the AI Act defines deepfakes as 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. At first glance, that sounds close to what many people would expect. But the crucial point is this: intention is not decisive.
The guidance says the assessment does not depend on whether the deployer intended to deceive or mislead people. The question is whether the content may falsely appear authentic or truthful to the audience. That changes the conversation. Because the issue is no longer only: "Was this created to manipulate someone?" The issue becomes: "Could this appear real to the people who see it?"
If it looks real, it becomes a labeling question
For brands, agencies and communication teams, this is not a distant legal detail. It touches everyday production workflows: AI-generated campaign images, AI-edited video assets, synthetic people in advertising, AI-generated product environments, voice clones, AI-enhanced event visuals or social media assets that look realistic but were created or substantially changed with AI.
Marketing often works by creating believable worlds. We create scenes, people, places, product moments and atmospheres that are meant to feel real. AI makes this faster, cheaper and easier. The legal and ethical question is now catching up.
If AI was involved in generating or manipulating realistic content, teams will need to ask whether the audience could mistake it for something authentic. That is a very different workflow question from simply asking whether the asset looks good.
The benchmark is not the most AI-literate person in the room
The guidance makes clear that the assessment cannot be based only on highly informed or digitally sophisticated audiences. It has to consider the possible audience, including children, elderly persons and people with lower digital and AI literacy, because they may be more easily misled about the authenticity or truthfulness of content.
That matters. The benchmark is not the marketing team, the agency, the legal department, the creative director or the AI-literate person who instantly sees how the asset was produced. The benchmark may include people who have very little understanding of how easily realistic media can now be generated or changed.
For brands, this raises the bar. Communication is not judged only by production intent. It is judged by potential audience interpretation.
Only obvious fantasy stays clearly outside the deepfake logic
The examples in the draft Guidelines make the distinction more tangible. An AI-manipulated image of two professional footballers in front of a football stadium may be a deepfake. AI-generated audio using cloned podcast voices may be a deepfake. An AI-generated video of someone resembling a politician giving a speech may be a deepfake. An AI-generated depiction of a celebrity influencer in an advertising context may be a deepfake.
At the same time, certain obviously unrealistic examples may fall outside this logic: a sphinx flying over the Eiffel Tower, mice arguing in human language over cheese, or a radio broadcast where only technical audio parameters such as volume or noise reduction are adjusted.
But most professional AI visuals are not flying sphinxes. They are realistic people, realistic places, realistic product scenes, realistic voices and realistic situations. That is exactly why this will become relevant for marketing, brand and communication teams.
Minor touch-ups may be excluded — but the operational line is hard to draw
The guidance makes clear that minor technical manipulations may not turn content into a deepfake. Examples include background details, lighting adjustments, audio parameters, colour correction, noise reduction, accessibility improvements or file compression. That sounds reassuring. But where exactly does minor end?
A colour correction is probably minor. Noise reduction is probably minor. But what about AI-generated background extensions, AI-created product environments, AI-enhanced faces, synthetic models or realistic scenes that never existed?
The guidance points to context and the impact on people's perception of authenticity or truthfulness. That makes sense legally. Operationally, it is difficult. Companies will not assess one AI asset per quarter. They will assess hundreds or thousands of assets across social, paid media, product communication, employer branding, internal communication, websites, events and sales material. "Do we need to label this?" will become a workflow question. And workflow questions need clear operating principles.
The missing problem is a common language for AI labels
For me, the bigger issue is not the labeling obligation itself. It is the lack of a common label. The guidance says deployers should disclose that deepfake content has been artificially generated or manipulated. It also says disclosure methods should be understandable and perceivable by natural persons, for example through visible or audible labels, without requiring people to rely on a specific technical tool.
That direction is right. But it does not yet give the market one standard visual label. No shared icon, no common wording, no universal design system, no agreed hierarchy that tells people whether something was fully generated, partially edited, lightly enhanced, fictional, synthetic, satirical, promotional or realistic.
So everyone can basically get creative. And that is where transparency starts to become fragile.
When every AI label looks different, transparency becomes harder
Every platform, brand, agency, publisher and tool may create its own AI label. Some will say "AI-generated". Some will say "created with AI". Some will say "AI-assisted". Some will use a sparkle icon. Some will use a robot icon. Some will use a watermark. Some will put the disclosure in small print. Some will make it visible. Some will make it technically detectable but hardly meaningful for the average person.
The result may be technically compliant. But will it be understandable?
A label only creates trust when people understand what it tells them. Was the whole image generated? Was only the background changed? Was the person real? Was the voice cloned? Was the event fictional? Was the product actually photographed? Was the content edited for quality, or was reality materially changed? A generic AI label will not answer these questions. And if people do not understand the label, the label does not create transparency. It creates noise.
AI transparency needs meaning, not just disclosure
The AI Act is pushing the market toward more disclosure. That is the right direction. But disclosure alone is not the same as understanding.
A label will not stop a fake CEO video from spreading before anyone checks it. It will not prevent a synthetic political speech from circulating in the decisive hours of a news cycle. It will not make people more media-literate simply because a small icon appears somewhere on a visual.
Companies need to build internal judgment. What counts as minor editing? What counts as material AI manipulation? What needs visible disclosure? What can be handled through metadata or provenance? What should not be produced at all? These are not only legal questions. They are brand trust questions.
AI labeling is becoming a leadership and brand governance issue
The practical challenge for companies will not be writing one AI policy and moving on. The challenge will be embedding judgment into everyday communication workflows. Marketing teams will need clearer decision trees. Agencies will need clearer production documentation. Legal teams will need scalable review principles. Brand teams will need a point of view on what kind of AI-generated realism fits the brand. Leadership teams will need to decide where the company wants to be more transparent than the minimum requirement.
Because transparency is not just a legal checkbox. It is a trust signal. And trust signals only work when people understand them.
AI transparency will only create trust if it gives people relevant and understandable context. Not if it becomes another generic disclaimer nobody reads.