New transparency obligations for businesses under the EU’s AI rules came into effect in August. In certain cases, content generated or modified using AI must now be labelled accordingly. What does this mean in practice for day-to-day corporate communications and marketing?
The transparency obligations under Article 50 of the EU AI Act became applicable on 2 August 2026. Their purpose is to help people recognise when they are interacting with AI or encountering AI-generated content, thereby helping to prevent fraud and misinformation.
It is worth noting from the outset that most ordinary marketing and communications content does not need to be labelled simply because AI has been used to draft or edit it.
This is because an AI-generated or AI-modified image, video or audio recording is not automatically considered a deepfake under the AI Act. A deepfake is AI-generated or manipulated content that resembles existing persons, objects, places, entities or events and could falsely appear to a person to be authentic or truthful. If a company publishes this kind of content, it must be disclosed as having been artificially generated or manipulated.
For AI-generated text, the labelling requirement depends on whether the text is published for the purpose of informing the public about matters of public interest. However, such text does not need to be labelled if it has undergone human review or editorial control and a natural or legal person holds editorial responsibility for its publication. If these conditions are not met, the text must be disclosed as AI-generated or manipulated.
These obligations have only just begun to apply, so practices are still evolving. Below, we take a closer look at the requirements and use practical examples to consider when marketing and communications content may need an AI label.
When should an AI-generated image, video or audio recording be labelled?
In the definition of a deepfake, authenticity essentially refers to whether the content could appear to be a genuine photograph, video or audio recording. For example, a speech generated using a cloned version of a company executive’s voice may contain entirely accurate information but still falsely appear to be a genuine recording. In this case, an AI label is required.
AI service providers and companies publishing content have different obligations
Providers of generative AI systems must ensure that synthetic content generated by their systems can be detected in a machine-readable format as artificially generated or manipulated. In practice, this can be achieved through metadata, watermarks or other technical markings embedded in the file.
The deployer of an AI system – for example, a company publishing content – is responsible for providing a disclosure that is visible to people when publishing a deepfake or text that falls within the transparency requirements.
Machine-readable and human-visible disclosures are therefore two different things. The fact that an image or video file already contains a technical marker does not remove the publisher’s obligation to provide a visible disclosure where one is required.
Truthfulness, meanwhile, concerns whether the content could create a false impression of what actually happened, who was present or, for example, which product was used.
When assessing whether content could falsely appear authentic or truthful, four aspects can be considered in line with the European Commission’s guidance:
- How closely does the content resemble the person, object or event depicted?
- What does the content state or imply about reality?
- In what context is the content used, or can it reasonably be expected to be used?
- Who will see the content, and what can that audience reasonably expect in the relevant context?
For example, cinema audiences expect to encounter artificially created images and sounds. By contrast, images accompanying a company case study could reasonably be expected to depict real people, products and situations.
An AI label would therefore be required, for example, when:
- a realistic image of a fictional person is presented as an employee of the company
- AI is used to place a product in a realistic-looking environment that is presented as an actual use case
- an expert’s speech is translated and the video is manipulated to make it appear as though the person is actually speaking another language.
The most difficult cases are likely to involve minor modifications to genuine photographs or videos. Small technical adjustments, such as correcting exposure, cropping an image or removing minor visual imperfections, will not generally turn content into a deepfake if they do not substantially alter its content or meaning.
But what if a bystander is removed from a genuine photograph to protect their privacy? In this case, it is worth considering whether removing the person changes the factual message of the image or the viewer’s understanding of what actually happened. In borderline cases such as this, transparency may be best served by adding a short disclosure, such as: “This image has been edited to protect individuals’ privacy.”
It is also worth remembering that adding an AI label does not give companies permission to use AI to create misleading marketing – for example, by giving a false impression of a product’s features.
GEO is part of the changing landscape of online visibility
AI does not eliminate the need for high-quality web content or search engine optimisation. Instead, it expands the ways in which online visibility is created. For businesses, the key is to ensure that their expertise is clearly presented online and supported by a strong technical foundation.
Tulus is your expert partner for developing both the content and technical aspects of your online visibility as you expand from SEO into GEO.
When should AI-generated text be labelled?
AI-generated or AI-modified text may be subject to disclosure requirements if it concerns a matter of public interest. According to the European Commission, such topics include public safety, environmental protection, consumer safety and public health. However, an AI label is not required if the text has undergone human review or editorial control and editorial responsibility is taken for its publication.
For example, a sustainability article drafted using AI would not need to be labelled if the company’s sustainability expert checks its facts and claims and the communications manager approves it for publication. By contrast, an incident notification that is automatically published without human review would need to be labelled as AI-generated.
Human review must cover the substance of the text, including figures, sources and claims. The reviewer must also have relevant expertise and be capable of exercising professional judgement. Proofreading alone is therefore not sufficient. In the case of editorial control, the responsible party must have the authority to approve, amend or reject the content.
The AI Act does not prescribe a specific method for proving that a published text has genuinely undergone human review. In practice, however, it is advisable to retain a record of the review and approval process. Evidence could include comments and edits made to the text, version histories, and emails or Teams messages in which the relevant expert and communications manager confirm that they have reviewed and approved the content.
How should AI-generated content be labelled?
The AI disclosure must be clear and distinguishable at the latest when the audience first sees or hears the published content. A machine-readable marker alone, or a disclosure hidden behind a link, is not sufficient.
The EU provides three icons in PNG and SVG formats: a general AI content icon, an icon for content generated entirely by AI, and an icon for content partially modified using AI.
The icon can be accompanied by text explaining, for example, that an image has been partially modified using AI or that a video contains AI-generated images and audio. Use of these icons is voluntary, and the disclosure requirement can also be fulfilled in other ways.
Companies that breach the AI Act’s transparency obligations may face fines of up to €15 million or 3% of their total worldwide annual turnover, whichever is higher. For SMEs, the lower of these maximum amounts applies. In Finland, compliance with the AI Act is supervised by several authorities, with Traficom, in Finland, acting as the national single point of contact.
Checklist for labelling AI-generated content
Images, video and audio
- Has the image, video or audio been generated or modified using AI?
- Does it closely resemble a real or plausibly real person, place, object, entity or event?
- Could the audience mistake the content for something authentic or truthful?
If you answered yes to all three, the content will probably need to be labelled.
Text
- Does the AI-generated text concern a matter of public interest?
- Has the text been reviewed by a human, and does the company take editorial responsibility for its publication?
If you answered yes to question 1 and no to question 2, the text must be labelled.
Still unsure whether your AI-generated content needs to be labelled? Get in touch with our expert and let’s look together at what the new requirements mean for your communications and marketing.
Lähteet
- EU AI act
- AI Act Service Desk: Article 3 – Definitions
- AI Act Service Desk: Article 50 – Transparency obligations for providers and deployers of certain AI systems
- AI Act Service Desk: Article 99 – Penalties
- European Comission: Transparency obligations under Article 50 of the AI Act – Questions and Answers
- European Commission: EU pictograms for labelling AI-generated content
- Finnish Competition and Consumer Authority: Misleading practices in marketing and customer relationships
- Traficom: New rules strengthen trust in artificial intelligence in Europe (in Finnish)
