As of August 2, 2026, significant transparency obligations of the European AI Act apply. For Austrian companies this affects not only developers of large AI systems. Teams that deploy chatbots, publish synthetic voices, use deepfakes or distribute AI-generated texts on matters of public interest also need to examine their specific use case.
The practical challenge is:Label AI contentwithout making every document unintelligible with a blanket notice. Decisive are the role, content, purpose and form of publication. This guide classifies the most important cases for communications, marketing, customer service and internal editorial teams. It is a practical aid and does not replace legal advice on individual cases.
Why 'created with AI' alone is not yet a concept
A notice can be too late, hidden or unintelligible. Conversely, not every spelling correction is AI content that requires labeling. Companies therefore need a process that clarifies before publication:
- Is the company acting as a provider or operator of an AI system?
- Do people interact directly with the system?
- Was text, an image, audio, or video generated or substantially manipulated?
- Does the content resemble real people, places, objects, or events and could it be mistakenly believed to be real?
- Does a text inform the public about a matter of public interest?
- Was there actual expert and editorial oversight?
The AI Service Center of the RTR erklärt that Article 50 contains several separate transparency obligations. They can, depending on the system and output, also be relevant at the same time. Information must be clear, unambiguous, accessible, and provided no later than at the first interaction or perception.
Four cases that teams should review separately
1. Direct interaction with an AI system
For a chatbot, voicebot, or other interactive system, people should recognize that they are communicating with AI, unless this is already obvious from the context. A customer service bot therefore needs a clear notice at the outset. Labels such as "digital assistant" can be too vague if they suggest a human contact person.
The notice should be placed near the interaction, not exclusively in terms and conditions or the privacy policy. For a voice system, a spoken notice is appropriate. For longer or sensitive conversations, a reminder may be necessary. In addition, an easy way to reach a human should be offered if the bot cannot reliably resolve a question.
2. Technical marking of synthetic content
Providers of systems that generate synthetic or manipulated text, audio, image, or video content must, as a rule, label outputs in a machine-readable way and enable detection. Examples include metadata, watermarks, cryptographic provenance proofs, or other technical methods. This provider obligation is to be distinguished from a visible disclosure by the publishing operator.
A company that purchases a third-party tool should ask during procurement which marking is embedded, how it can be verified, and whether it is preserved during export, cropping, or compression. A visible label does not automatically replace the provider's technical provenance information.
3. Deepfakes in image, audio and video
According to the AI Act, a deepfake is an AI-generated or manipulated image, audio, or video content that resembles real people, objects, places, institutions, or events and can falsely appear genuine. Anyone who, as an operator, publishes such a medium must disclose the artificial generation or manipulation.
The labeling should be perceptible at first contact: for an image, directly on the image or in its immediate vicinity; for audio, audible at the beginning; for video, visible or audible. There are relaxations for clearly artistic, satirical, or fictional works, but transparency does not disappear entirely there either. The specific context determines.
4. AI texts on matters of public interest
Special rules apply to AI-generated or manipulated texts that are published to inform the public about matters of public interest. The RTR explains an important exception: disclosure can be omitted if a genuine substantive human review or editorial control has taken place and a natural or legal person assumes editorial responsibility.
A mere spell check is not sufficient. Facts, sources, context and possible impacts must be professionally reviewed. Companies should document this review and make the organization editorially responsible easy to find.
A labeling matrix for day-to-day editorial work
A short matrix prevents teams from only discussing immediately before publication. It can contain the following columns:
- Content type: Text, image, audio, video or interaction,
- Generation method: fully generated, substantially manipulated or only technically assisted,
- Realistic reference: yes or no,
- Purpose and audience,
- human editorial oversight,
- technical labeling present and verified,
- visible or audible disclosure,
- responsible approval and publication date.
Example: A fully synthetic product photo without a real person is categorized differently than an artificially generated video message that appears deceptively realistic of the CEO. An internally corrected paragraph should be treated differently than an automatically generated news item without editorial review.
Separate supportive editing from substantial manipulation
Not every AI feature changes the meaning of content. Noise reduction, file compression, color correction, or a pure accessibility improvement may be assessed differently under the guidelines than a manipulation that alters authenticity or message. The team should not decide based on the product name, but should document the difference between the source material and the final version.
Three questions can help:
- Was a new statement, person, action, or situation added?
- Could the audience therefore judge the content as something other than real or true?
- Was only the technical quality improved, without substantially changing semantics and context?
In cases of uncertainty, disclose cautiously and examine the case with qualified scrutiny. Voluntary transparency can also be useful beyond a mandatory obligation, as long as the notice is not phrased in a misleading way.
This is how a good notice should be designed
A transparency notice only fulfills its purpose if people actually notice and understand it. Good notices are:
- early: on first interaction or perception,
- near: directly with the content instead of on a distant subpage,
- unambiguous: clear words like "AI-generated" or "AI-manipulated",
- appropriate to the modality: visible for images, audible for audio-only content,
- accessible: also accessible for assistive technologies and different forms of perception,
- consistent: used in the same way across comparable channels.
In its guideline on transparency and labeling, the WKO recommends that companies define how AI-generated content should be labeled. A central template prevents creative but unclear variations such as "digitally optimized" or "smart content".
Plan for accessibility from the start
A small icon alone is not accessible to everyone. An image reference needs a textual equivalent that can be captured by screen readers. Contrast, size, and placement must also work on mobile devices. For audio, the information should be spoken, and for video, it should also be visually available. Subtitles must not obscure the notice.
Test the initial perception: Does the notice appear before playback? Is it still recognizable when displayed at a reduced size? Is it described meaningfully in the alt text? Is a chatbot label reachable by keyboard and screen reader? Article 50 paragraph 5 explicitly links clear information with applicable accessibility requirements.
Make editorial oversight verifiable
Anyone who relies on human editing for AI-generated texts should store more than one name in the workflow. The review log can be brief:
- primary sources used and retrieval date,
- verified key statements and figures,
- significant changes compared to the AI draft,
- person or organization professionally responsible,
- time of approval and published version.
Full prompts or unnecessary personal data do not need to be permanently stored for this purpose. The goal is traceability of content responsibility. The process can be combined with the existing AI checklist for managers: expert review, right to stop, and escalation are determined before publication.
Don't forget procurement and export
Labeling can be lost during export. Social media platforms alter images, editorial systems remove metadata, and videos are re-encoded. Therefore, test the entire path from the generator to the public display. Check the published file, not just the original in the working folder.
When procuring, companies should at least ask:
- Which content does the system technically mark?
- Which detection tool is available?
- Does the labeling remain after normal edits?
- Which responsibilities does the provider have and which does the operator have?
- How are changes to the labeling solution communicated?
For automated publication by AI agents in the officeThe transparency review should be its own approval gate before every external Write. An agent must not skip a missing disclosure for the sake of speed.
An implementation plan over ten working days
Day 1-2: inventory channels
List chatbots, image generators, video and audio tools, text assistants, and automated publications. Capture purpose, audience, operator role, and responsible team.
Days 3 to 4: Classify cases
Classify direct interaction, synthetic content, possible deepfakes, and texts of public interest separately. Flag unclear cases for expert review.
Days 5 to 6: Design notices
Develop clear text, image, audio, and chatbot notices. Check accessibility, mobile presentation, and when users will first encounter them.
Days 7 to 8: Test technology
Check machine-readable markings, export, compression, and publishing. Document which platforms receive or remove metadata.
Days 9 to 10: Embed accountability
Train editors, marketing, customer service, and approvers. Publish the matrix, define an escalation path, and set a date for the first sample.
Use spot checks to verify whether the rule is reaching the public
A process description does not prove that notices remain visible across all channels. Each month, draw a small sample from the website, newsletter, social media, audio, video and chatbot interactions. Check the version actually published for the notice, timing, accessibility and technical tagging. For responsive pages, include both desktop and mobile views.
Document the URL or case ID, content type, outcome and correction. If a channel repeatedly makes the same mistake, don’t just fix the single post. The team investigates the template, export or publishing process. A missing image notice can, for example, occur because a platform cuts off the caption; lost metadata can be due to automatic compression.
Complaints and follow-up questions are also valuable signals. If people misunderstand a notice or continue to mistake a bot for a person, the wording must be improved. Labelling is successful when it supports actual perception — not when merely a field in the CMS has been filled in.
Practical example: voicebot in customer service
An Austrian energy service provider uses a voicebot for simple questions about opening hours and meter reading reports. At the start of the call, a short announcement clearly informs callers about the AI interaction. The bot always offers the option to switch to a staff member.
The announcement is tested not only in the documentation but in every actual call. The team also checks whether it needs to be repeated after transfers and whether users understand it well. Synthetic voice, call recording and data protection are handled as separate topics. A visible transparency step therefore does not replace other obligations.
Frequently asked questions about AI labeling
Does every text edited with AI need to be labeled?
No, not automatically. What matters are the purpose, the extent of the change, the publication and the conditions of Article 50. For texts of public interest it is particularly relevant whether genuine editorial control was exercised and responsibility was assumed.
Is a notice in the imprint sufficient?
Usually not for a specific interaction or output that requires labeling. The information must be provided in a timely manner and be visible close to the content. A general AI statement in the imprint can supplement but cannot replace every specific notice.
Is every AI image a deepfake?
No. A deepfake presupposes a realistic resemblance that could be mistakenly taken for genuine. Providers obligations to technically mark synthetic content may nevertheless apply. Operators must separately assess their specific publication case.
What applies to older content?
The RTR explains differentiated transitional questions. Content published before the cut-off date generally does not have to be labeled retroactively; if an older AI-generated item is only published afterwards, the transparency obligation may apply. Therefore the creation and publication times should be documented.
Conclusion: Transparency must be made visible in the workflow
Good labeling is not created at the last minute. It starts with inventory and clarification of roles, distinguishes content types, and is preserved through export and publication. Human editorial oversight is not a label but verifiable substantive control.
Take a real channel — for example a chatbot, an image campaign, or editorial text — and assess it with the labeling matrix. If notice, technical marking, accessibility and responsibility are verifiable in the public version, the process is ready for the next sample.
