Does my AI content need a label?

How the Guidelines decide whether AI content is a deep fake

Article 50(4) of the AI Act says deployers of an AI system that generates or manipulates image, audio or video content constituting a deep fake shall disclose that the content has been artificially generated or manipulated. Article 3(60) defines a deep fake 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. The Commission's Guidelines on Article 50 (C(2026) 5054 final, published July 20, 2026) work through that definition in paragraphs 113 to 116 and give lists of examples. We read those paragraphs and lists in the Guidelines' own text on October 22, 2026.

• It has to look like something real. Persons, objects, places, entities and events must be realistic, and it is enough that the content resembles something that exists, can plausibly exist or could have plausibly existed (paragraph 113). Content that defies the laws of nature or physics, such as humans flying without mechanical aids, dragons or elephants driving cars, and that has no potential to mislead is unrealistic and outside the obligation.
• It has to falsely appear authentic or truthful. The Guidelines assess this as a whole: the level of resemblance, the message, the places where the content is deployed, the environment it is shown in, and the foreseeable audience and what it expects. The test is objective, so no intention to deceive is needed. A high degree of photorealism makes a deep fake more likely but does not decide it alone (paragraph 114).
• The audience is the foreseeable one, not an average person. Where children, the elderly or others with lower digital and AI literacy may see the content, it can be enough that it falsely appears authentic to them (paragraph 115).
• Minor edits are not deep fakes. AI-supported changes to insignificant aspects of existing content, such as lighting, color correction, noise reduction, file compression or removing a passerby, may be of minor relevance, and for product advertisements and packaging the Guidelines name AI color correction, background extensions, aesthetic background replacement, arranging existing products and rescaling as likely to have only a minor impact. Whether a given edit is minor depends on the context and its effect in the specific case (paragraph 116).
• The examples. Two lists follow paragraph 116: examples of deep fakes, and examples that are not. Two more follow paragraph 124: deep fakes that are creative, satirical or fictional works, which get a lighter duty, and deep fakes that are not. The table below shows the ones this tool uses.

What the four results mean

• A listed example of a deep fake. The Guidelines list this kind of content as a deep fake.
• A listed example of something that is not a deep fake. The Guidelines list this kind of content as not a deep fake, on conditions the result shows. A few edits are named in the text of paragraph 116 rather than in a list, and the result says so.
• The closest listed example. No listed example is the same. The result shows the nearest ones, each with the direction the Guidelines give it. Which is nearest is our reading, not the Commission's.
• Not settled by the Guidelines. The Guidelines do not answer it, or a listed example turns on a fact you marked as unsure. The question then falls to the definition in Article 3(60) and the factors in paragraph 114.

None of the four is a finding about your business. Each is a reading of published examples, shown with the words it rests on.

The examples this tool works from

The table lists the examples the tool draws on, in the Guidelines' own words, with the list each one sits in. It is not every example in the Guidelines. The last row comes from the Commission's page on the EU icons, which was last updated on October 13, 2026 and is not part of the Guidelines; the page says the use of its icons "does not establish legal compliance by itself".

What the Guidelines do not list

The lists have no testimonial or review video, no AI model in a still product photo, no virtual staging of a listing photo, no synthetic voice that imitates nobody, and no AI-generated image of a place. For those the tool shows the closest listed examples, or says the question is not settled. For virtual staging, the Commission's icons page lists an AI-furnished empty apartment as a use case for its Partially AI-Modified icon; the Guidelines themselves do not settle it.

Rules that are not about deep fakes

Some US rules turn on other tests. They may apply to the same content, and the rules finder lists them with their sources. We verified these on October 21, 2026.

• New York General Business Law 396-b has applied since June 9, 2026. It asks a business that creates an advertisement, with actual knowledge that a synthetic performer is in it, to conspicuously disclose that. The test is whether the person is synthetic, not whether the ad would falsely appear authentic. Audio-only ads are exempt.
• California SB 1050 takes effect on January 1, 2027. It asks for a clear and conspicuous disclosure in an audio, video or audiovisual ad that prominently includes a synthetic performer, in wording substantially similar to "this performance features a synthetic performer" or "no human performer is depicted".
• California AB 723 has applied since January 1, 2026. For real estate brokers and salespeople advertising property for sale, a digitally altered image needs a statement that it was altered and a link, URL or QR code to the original. Photo editing software counts, not only AI.
• The FTC rule on consumer reviews and testimonials (16 CFR Part 465) bans writing, creating or selling a review or testimonial that materially misrepresents that the reviewer exists, that they used the product, or what their experience was, and the FTC says that reaches reviews from someone who does not exist, such as AI-generated fake reviews. No AI label fixes a fake review.

Whether each rule reaches you depends on where your audience is and what your business does. The rules finder asks, and our guides on synthetic performer ads and virtual staging disclosure go deeper.

Why a record, and what it is

Paragraph 148 of the Guidelines says providers and deployers that have not signed a code of practice the Commission found adequate are expected to demonstrate how they have complied through other adequate means, and gives a gap analysis against such a code as an example. The record this tool prints is smaller than that. It notes, with a date, what the content is, what AI did to it, which listed example it matched and the paragraphs behind the reading, and it leaves lines for the decision, where the label appears and who reviewed it. Keep it with the content's files. It is not proof that a rule was met, and it is not legal advice.

If you decide to label

The Guidelines say a label should be understandable and perceivable by people, for example visible or audible, without special tools, and that a business cannot rely on the machine-readable marking the provider embeds (paragraph 117). Our accessible AI label generator drafts label wording, and the guide to accessible AI labels covers what a label can do for screen reader users. The AI image metadata checker reads the markers a file declares. None of them decides whether an image is AI-generated, and a missing marker proves nothing.

What this tool does not do

It does not look at your content, detect AI, or tell you whether you need a label. It covers deep fakes under Article 50(4), not the chatbot duty in Article 50(1), the machine-readable marking duty on providers in Article 50(2), or the duty for AI-written public-interest text. It does not cover consent, personality rights, data protection or copyright, which the Guidelines say a label does not displace (paragraphs 127 to 129). The Guidelines are non-binding, and only the Court of Justice of the EU gives an authoritative interpretation of the Act (paragraph 5). Our page on EU AI Act Article 50 and the AI disclosure laws overview cover the rest. No tool makes content meet a law: what we check lists what our scans cover.