The Cases That Change the Answer
The EU AI Act’s transparency duties become difficult at the edges: a tool may correct a sentence, substantially rewrite it, generate a new image, or produce a video that looks like a real event. This chapter focuses on the facts that move a workflow between those categories—what changed, what the audience is likely to believe, who controls publication, and whether the work is personal, editorial, artistic, or public-facing. The goal is not to turn every borderline case into a verdict, but to show why “AI was involved” is too crude a test.
The useful comparison is not between products marketed as assistants and products marketed as generators. It is between the function the system performed and the meaning or context of the resulting content.
Assistance, alteration, and generation are not the same event
Article 50(2) describes an exception to the provider-side output-marking obligation where an AI system performs an assistive function for standard editing or does not substantially alter the input data provided by the deployer or the semantics of that input. The wording directs attention to effect. A system’s label, interface, or commercial name does not by itself determine whether the exception applies.
Consider a document editor that corrects spelling, adjusts punctuation, or suggests a less awkward sentence while leaving the author’s proposition intact. That may look different from a system that replaces a paragraph with a new argument, invents supporting details, or changes the apparent meaning of a statement. The regulation does not offer a simple percentage threshold in the selected text. It asks readers of the rule to consider the extent of alteration and the semantics of the input.
The same distinction applies to images. Resizing, format conversion, or routine correction is not conceptually identical to inserting a person who was never present, changing an expression, or creating a scene from a prompt. A platform may use the same editing application for all three jobs, but the legal analysis cannot be inherited from the application’s brand. It follows the operation and its effect.
That does not mean an assistive workflow has no transparency concerns. It means the specific provider-side exception in Article 50(2) is tied to the extent of the system’s role. Other duties, contractual commitments, platform rules, or national requirements may still matter. A careful explanation should say that the Article 50(2) marking obligation may not apply to the stated extent, not that “human assistance makes the content safe” or that every other rule disappears.
The comparison can be made with four questions:
- What was the starting material? A new blank canvas, a person’s photograph, a draft article, or a recorded event creates different provenance questions.
- What did the system change? It may have corrected form, changed semantics, added content, or generated a new asset.
- What does the final work invite the audience to believe? A private design mock-up and a public-looking record of a real event do not create the same risk of mistaken authenticity.
- Who is responsible for the use? The provider’s technical obligation and the deployer’s publication duty are not interchangeable.
These questions make the rule more precise without pretending that every case can be resolved from a headline description.
A deepfake is not simply any synthetic image
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. Two elements are important: resemblance and the likely false appearance of authenticity or truth.
A completely imaginary illustration may be synthetic without fitting that definition. A generated video that imitates a real person delivering words they never spoke raises a different question. So does an altered recording that appears to show a real place or event. The definition is not limited to famous people or elections; its terms include persons, objects, places, entities, and events. At the same time, it does not say that every generated image is a deepfake, or that a disclosure proves the event did not happen. It defines a category relevant to a transparency duty.
The audience’s likely interpretation is therefore part of the problem. A clearly labeled fictional animation and a realistic clip presented as documentary material may use similar generative tools but communicate different claims about reality. The content’s surrounding presentation can affect whether it would falsely appear authentic or truthful. That is why a file-level detector alone cannot settle the entire legal analysis.
For deployers, Article 50(4) requires disclosure when an AI system generates or manipulates image, audio, or video content constituting a deepfake. The paragraph also says that where the content forms part of an evidently artistic, creative, satirical, fictional, or analogous work or program, the obligation is limited to disclosure of the existence of generated or manipulated content in an appropriate manner that does not hamper display or enjoyment.
This is a limitation on form and context, not a free pass. A satirical program may not need a warning that dominates the screen or destroys the work’s presentation, but the existence of generated or manipulated content still matters. The point is to preserve the audience’s ability to understand what they are seeing without treating every creative use as deceptive or every disclosure as a visual obstruction.
Public-interest text needs more than a human touch
The text branch of Article 50(4) has a different structure. It covers AI-generated or manipulated text published with the purpose of informing the public on matters of public interest. The deployer must disclose the artificial generation or manipulation unless the use is authorised for criminal-offence work or the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for publication.
The combination matters. “A human clicked publish” is not the same as the conditions stated in the provision. The text points to a process of review or editorial control and to a person or legal entity holding responsibility for the publication. That is more demanding than inserting a superficial human step after an automated output. It also does not mean that human review guarantees the text is correct. It describes a condition in the transparency rule, not a factual certification.
A hypothetical newsroom illustrates the distinction. Suppose an AI system proposes a draft about a public-health policy. If editors substantively review the draft, check its claims, exercise editorial control, and a responsible person or organization stands behind publication, the stated exception may be relevant. If a system generates a finished article that is automatically posted to inform the public, the absence of that review and responsibility structure changes the analysis. Whether a particular workflow meets the legal conditions requires facts about what review occurred and who held responsibility; the label “human edited” is not enough.
The same principle applies outside journalism. An organization publishing a generated explainer about a public controversy should not assume that adding a human byline resolves the question. A byline can identify editorial responsibility, but it does not by itself prove that meaningful review or control occurred. Conversely, an AI-assisted draft that was substantially reworked by an accountable editor is not identical to an unreviewed automated publication merely because both began with a model output.
Personal use and cross-border use pull in different directions
Context can change the actor analysis before the content analysis begins. The Act excludes obligations of deployers who are natural persons using AI systems in purely personal, non-professional activity. A person experimenting with a private image for personal use is not automatically in the same position as an organization publishing a realistic image to inform or influence the public.
But the personal-use boundary should not be stretched into a universal consumer exemption. It concerns deployer obligations in the stated context. It does not erase the provider’s role, other operators’ responsibilities, or other applicable laws. Nor does it answer whether a use is professional merely because no money changed hands. The facts surrounding authority, purpose, publication, and organization matter.
Cross-border use creates the opposite warning. A system provider can be outside the Union while the output is used in the Union. The location of the model therefore cannot be used as a shortcut for dismissing the regulation. An organization should first identify where the provider and deployer are situated and where the output is used, then examine the applicable content category and exception.
These two cases show why the same technical output can sit inside different legal analyses. A person using a tool privately, a company deploying it in a customer service channel, and a publisher distributing a realistic public-interest video may all use the same model. Their responsibilities do not follow from the model alone.
The diagram uses “may apply” deliberately. It represents the legal conditions that need examination, not an automated classifier. “Substantial alteration” is a factual judgment about the relationship between input and output. “Falsely appears authentic or truthful” is a condition in the deepfake definition, not a property that can be read from a file extension. “Human review” and “editorial responsibility” must be understood together in the public-interest-text branch.
Creative context changes the form, not the need for clarity
The artistic and satirical limitation is easiest to misunderstand when disclosure is treated as punishment. The provision does not say that a fictional work becomes unlawful because it uses generative tools. It says that when the work is evidently artistic, creative, satirical, fictional, or analogous, the transparency obligation is limited to an appropriate disclosure of the existence of generated or manipulated content that does not hamper the display or enjoyment of the work.
That creates a design trade-off. A large warning placed over every frame may make a work inaccessible or destroy its intended presentation. No disclosure may leave a realistic audience member unable to tell whether a scene is documentary. The legal language points toward a disclosure that is noticeable enough to convey the relevant fact while remaining proportionate to the work’s context. It does not supply a universal font size, screen position, or interface pattern in the material examined here.
A satirical video can therefore be both obviously stylized to some viewers and still benefit from a clear indication of artificial manipulation. An evidently fictional film may present the disclosure in program information or credits rather than as an intrusive overlay, depending on how people encounter it. The classification depends on the work and the mode of exposure, not merely on the creator’s private intention.
Where a general explanation must stop
The statutory text supplies categories and conditions, but it does not resolve every factual dispute. It does not tell a general reader exactly how much rewriting is “substantial,” when an audience would regard a work as evidently fictional, or what level of editorial review is sufficient in every public-interest workflow. It does not name one technical standard as the universal implementation of machine-readable marking.
Those gaps are not evidence that the rules are meaningless. They are the ordinary boundary between a general explanation and a case-specific legal assessment. Current official guidance, codes of practice, enforcement decisions, national law, and the facts of a workflow can change the analysis. A responsible organization should preserve records of what the system did, what humans reviewed, who held publication responsibility, and how the disclosure reached the audience. A responsible reader should resist both extremes: “AI was involved, so disclosure is always required” and “a human was involved, so disclosure is never required.”
The most defensible conclusion is conditional. Assistive editing, substantial alteration, generative creation, deepfake-like resemblance, creative context, editorial responsibility, personal use, and cross-border use are not interchangeable facts. They are the facts that determine which part of Article 50 is relevant and whether an exception or adapted form of disclosure deserves consideration. The next synthesis will put those categories together and show why even a legally appropriate disclosure leaves separate questions about machine detectability, provenance, and factual verification.