A Label Can Tell You Where Content Came From—Not Whether It Is True
This series examines when AI-generated or AI-manipulated content must be labeled under the EU AI Act, how provenance systems such as C2PA can make origin and editing history more visible, and what those signals still cannot establish. The practical question is not simply whether a file has an “AI” badge. It is what the badge says, who is responsible for saying it, how the statement can be checked, and whether it changes what people believe or do.
That distinction matters because the same piece of content can carry several different kinds of information. A conversational system may need to tell a person that they are interacting with AI. A provider may need to mark synthetic text, audio, images, or video in a way software can detect. A deployer may need to disclose a deepfake or certain public-interest text to human readers. A provenance record may add a signed account of how an asset was created or changed. None of these statements is automatically a judgment that the content is accurate, complete, harmless, or unpersuasive.
The word “label” hides several different claims
Imagine an image that was generated from a prompt, edited by a person, uploaded to a platform, and used to support a claim about a real event. “AI-generated” could be a useful disclosure, but it leaves several questions unanswered.
First, what is being disclosed: that a model generated the pixels, that a model assisted with an edit, or that the image passed through a tool without substantially changing the input? Those are different descriptions of origin and transformation. Second, who has the duty to disclose? The answer may depend on whether an organization supplied the AI system, used it under its authority, or simply encountered the resulting file. Third, where should the information appear? A machine-readable mark can help software detect an output, while a clear human-facing notice serves a person encountering it. Fourth, what evidence supports the description? A provenance record may connect a file to a signed declaration, but that is not the same as independently checking the claim depicted by the file. Finally, what effect is the disclosure supposed to have? It may support transparency without making an audience reject the content.
These questions are easy to collapse because everyday language treats “labeling” as one action. In practice, they belong to different layers. Legal disclosure assigns responsibility and gives people relevant context. Machine-readable marking makes that context detectable at scale. Provenance records preserve a declared history. Fact-checking examines whether the content’s substantive claims are true. Audience research asks whether any of these signals changes judgment or behavior.
The layers can reinforce one another, but they do not substitute for one another. A visible notice may be understandable but easy for a platform to lose when a file is copied. A machine-readable mark may survive in a technical field without being presented clearly to a person. A signed history may be useful evidence about the file while saying nothing about whether a quoted statistic is correct. A factually accurate message may remain persuasive after its AI origin is disclosed.
Three sources, three kinds of authority
The EU AI Act is the normative anchor for this subject. Article 50 separates several transparency situations rather than creating one universal rule for every file touched by AI. It addresses direct interaction with an AI system, provider obligations for outputs that generate synthetic audio, images, video, or text, and deployer obligations for deepfakes and certain public-interest text. The same article also contains exceptions and conditions involving standard editing, substantial alteration, creative or satirical work, human review, editorial responsibility, and accessibility. The legal answer therefore depends on the actor, the content, and the context.
C2PA approaches a different problem. Its Technical Specification 2.2 describes a way to store and access cryptographically verifiable information about an asset’s provenance. Assertions can describe actions, software, ingredients, or bindings to content. Claims can be signed, and manifests can travel with an asset or be stored elsewhere. The specification describes this as an opt-in provenance architecture and says that its purpose is to let consumers assess trust signals—not to make value judgments about whether the content itself is good or bad. A validator can check whether a signed record is well formed, whether content bindings match, and whether a signing credential is trusted. The consumer still has to decide what the record means and whether the underlying content deserves belief.
The experimental source adds the audience layer. In a survey experiment with 1,601 U.S. participants, researchers showed people the same AI-generated persuasive message about one of four relatively non-polarized policy proposals. Participants were randomly told that the message came from an expert AI model, a human policy expert, or no labeled author. The messages moved policy support by 9.74 percentage points on average. Most participants in the labeled conditions believed the authorship information: 94.6% believed the AI label and 89.3% believed the human label. Yet the authors found no statistically significant difference in support between the AI-label and human-label conditions, or between the AI-label and no-label conditions. They also found no significant differences in reported accuracy, confidence, or sharing intention.
That result does not show that labels are useless. It shows something more precise: in this design, a label could communicate its intended origin without neutralizing the message’s persuasive effect. The study tested text messages, a particular population, current model capabilities as of 2024, and a particular wording of authorship disclosure. It did not test every warning design, deepfake disclosure, provenance system, platform interface, or long-term behavior. But it makes the central tension concrete. Transparency and influence are related questions, not the same outcome.
This sequence is not a claim that every workflow moves in a neat line. Legal duties and technical records can be created at different points, and a person may encounter content before any provenance is inspected. The diagram is a safeguard against category errors: it keeps a duty, a data format, a validation result, a factual judgment, and a behavioral effect in view without making any one of them stand in for the others.
A better question than “Is there a label?”
A more useful analysis begins with five questions.
What happened to the content? Was a person interacting with an AI system, was new content generated, or was existing material merely assisted or reformatted? The difference matters because the EU AI Act’s duties are tied to particular kinds of systems and transformations, not to an abstract suspicion that AI was somewhere in the workflow.
Who is responsible in this context? The Act distinguishes the provider of a system from the deployer using it under their authority, and it can reach some third-country activity when the output is used in the Union. Personal non-professional use is treated differently. Responsibility cannot be inferred from the file alone.
What signal is required? The relevant signal may be a notice at first interaction or exposure, a machine-readable mark on synthetic output, or a disclosure to people who encounter a deepfake or certain public-interest text. A short visible badge and a machine-detectable record solve different communication problems.
What does the provenance support? If a C2PA manifest is present, the reader or validator may be able to inspect declared actions, software agents, ingredients, cryptographic bindings, signatures, and trust status. That can make the history more legible. It does not guarantee that every action was recorded, that every ingredient is public, or that the asserted account is true in the broader sense.
What still needs independent judgment? The content may make claims about people, events, statistics, or public policy. Those claims require evidence appropriate to their subject. The presence of a label does not prove deception, and its absence does not prove human authorship. A valid provenance record does not prove that a depicted event occurred. A trusted signer does not automatically become an expert witness for every sentence in the asset. And a disclosure that audiences believe may still leave the content influential.
What this distinction makes possible
Keeping the layers separate does not weaken the case for labeling. It improves the case by assigning each tool a defensible job. Legal disclosure can give people context they would otherwise lack. Machine-readable marking can help platforms and other software identify synthetic or manipulated output. Provenance can preserve a declared chain of creation and modification across tools. Independent verification can test the substance. Audience research can reveal whether the combination is changing the decisions that matter.
The EU AI Act supplies the first question: under these facts, who must disclose what, and when? C2PA supplies one possible way to make parts of a creation history inspectable. The experiment supplies a warning against assuming that visibility automatically produces skepticism. Together, they support a restrained conclusion: an AI-content label is valuable as a disclosure signal, but its meaning ends where provenance, factual verification, and audience judgment begin.
The next step is to read the legal rule with that discipline. Instead of asking whether “AI content” must always be labeled, we can ask which actor is involved, what kind of interaction or output occurred, which transparency duty applies, and what the rule leaves for technical standards and context-sensitive judgment.