A speech is linguistically polished. A political decision is summarized for Instagram. The accompanying image is also generated by AI, because the image database, once again, only shows people staring intently at a tablet.
AI in political communication has long since become more than just an experiment. It supports political parties, government agencies, associations, and communications agencies. This raises a practical question: When should the public be told that artificial intelligence was involved?
The answer depends on what kind of content is being created, how it is published, and who actually reviews it. A blanket statement like “created with AI” is just as inadequate for determining this as a blanket statement like “someone reviewed it.”
In simple terms, the rule for operators of AI systems is this: Deepfakes must generally be disclosed as artificially generated or manipulated. There is also a disclosure requirement for published AI-generated texts on matters of public interest. However, this requirement may be waived if human review or editorial oversight has taken place and a natural or legal person bears editorial responsibility. This exception applies to texts, not to the labeling of deepfakes. The legal basis is Article 50(4) of the AI Act.
The relevant transparency requirements have been in effect since August 2, 2026. On July 20, 2026, the European Commission published guidelines on the practical interpretation of these requirements. The guidelines explain the regulation but do not replace the text of the law.
A statement on energy policy, an explanation of a local government resolution, or an article about a legislative process can inform the public about matters of public interest.
This does not apply exclusively to political parties. Associations, companies engaged in public affairs, and contracted agencies should also review their disclosure processes accordingly.
However, not every text published by a political organization is automatically subject to the same requirement. An internal scheduling discussion is different from a publicly distributed article on pension policy. What matters are the content, purpose, and context of publication—not just the logo above the text.
The exception for reviewed texts is important in practice. It allows for responsible use of AI without requiring every publication to include an AI disclosure.
However, this requires two conditions to be met: actual human review or editorial oversight, and clearly assigned editorial responsibility.
According to the European Commission’s guidelines, superficial, purely formal reviews are not sufficient. The reviewer must be able to evaluate the content from a technical perspective. Editorial review includes, in particular, the ability to modify or reject statements based on their content.
For a communications team, this means, for example: verifying figures, tracing sources, checking quotes, and ensuring that statements align with the actual position taken.
A correctly placed comma is a welcome sight. However, it does not yet confirm the pension calculation.
Even a second AI system does not replace the need for human review. Responsible publication does not result from two language models finding each other convincing.
For images, audio recordings, and videos, a separate check must be conducted to determine whether they are deepfakes. This refers to AI-generated or manipulated content that resembles real people, objects, places, institutions, or events and falsely appears to be authentic.
This does not necessarily require the impersonation of a well-known person. According to the authorities’ explanations, depictions of plausibly existing persons or situations may also be covered. The decisive factor remains the false impression of authenticity in the respective context. The Federal Network Agency explains this distinction based on the Commission’s guidelines.
A photorealistic image of what purports to be a town hall meeting may therefore be evaluated differently than an obviously abstract illustration.
Important: In the case of a deepfake, human approval does not eliminate the disclosure requirement. For works that are clearly artistic, satirical, or fictional, the regulation provides for an adapted form of disclosure, not a blanket exemption.
So the conclusion that “the image is artistically successful” does not yet answer the question “What should we say about it?”
The following examples are fictional.
The Reviewed Feature Article: An association commissions an article on a draft bill. A knowledgeable staff member verifies sources, corrects statements, and approves the final version. Editorial responsibility has been assigned. The exception for reviewed texts may apply.
The automated news channel: An organization publishes AI-generated summaries of political decisions without verifying their content. If the criteria for texts of public interest are met, the use of artificial intelligence to generate or edit the content must be disclosed.
The Fabricated Event Photo: An agency creates a photorealistic image of a political event that did not actually take place. The image appears to be a documentary photograph. In such cases, the possibility that it is a deepfake must be examined. Approval by the client does not replace the requirement to label the image as such.
Not every legally permissible publication is automatically a wise choice from a communication standpoint.
Political actors, in particular, can therefore openly state how they use AI—for example, for research support, translations, or language editing. Such a statement should describe how the technology actually works.
One possible optional wording is:
"We use AI to assist with research and text editing. Our team reviews the content of published statements and takes responsibility for them."
This is not a universal template and does not guarantee legal compliance. Anyone who publishes content automatically should not claim that every statement is reviewed in advance.
Furthermore, a general website notice does not replace the required labeling of specific content. Legally required information must be clearly and distinctly visible no later than the first time the user encounters the relevant content. A label that is machine-readable only is not sufficient for this purpose. The details are explained in the Commission’s Transparency FAQ.
Organizations should not have to reinvent the wheel when it comes to AI transparency for every single post. A straightforward process helps:
This document is a practical governance recommendation, not a blanket legal requirement to log every single post.
K11 helps organizations integrate such processes with data protection, communication, and internal responsibilities. AI training can help embed the necessary distinctions within the team. An AI Officer can coordinate the implementation. However, this does not imply a general legal obligation to appoint such a role.
Labeling answers the question of whether the content is synthetic. It does not automatically clarify whether the content may be published at all.
Data protection, privacy rights, and copyright must be considered separately. In the case of political advertising, the requirements of the Regulation on Transparency and Targeting of Political Advertising may also apply. These include, in particular, information regarding the advertising nature of the content and the sponsor. A notice stating that AI was used does not replace this information.
Our article “EU AI Act: When AI Must Be Visible” explains the general differences between visible disclosure and technical markings .
AI in political communication requires neither a blanket ban nor a free pass.
The key factors are proper classification, a genuine review of content, and clear accountability. Where labeling is required, it must be clearly visible and understandable. Where it is not required, it still makes sense to make a conscious decision to provide additional transparency.
AI can help with drafting. It should remain up to the organization to decide what position it takes publicly.
Effective Date: November 2, 2026.