Transparency and labelling of AI-generated communication
This page provides guidance for staff working in communications as to when AI-generated or AI-manipulated content needs to be labelled under the EU Artificial Intelligence Act. It also explains how and where Lund University recommends placing labels when the University is the sender, in both internal and external channels.
As a public authority, Lund University bears editorial responsibility for all content published in the University’s name. When the University is listed as the sender, the recipient should feel confident that the content is accurate, reviewed and clear. That is why communications professionals must be transparent about when generative AI has been used if its use could affect how the content is perceived, for example, if the content could be mistaken for real content or content created entirely by humans.
Shortcuts to page content
Basic principles and responsibility
- Basic principles for labelling AI-generated content
- What do ‘human review’ and ‘editorial responsibility’ mean?
Types of content and labelling
Basic principles for labelling AI-generated communications
The recommendations are based on the following principles:
- A balanced approach to labelling is necessary. AI-generated content is to be labelled where required by the applicable legislation. However, excessive labelling should be avoided, as it may undermine confidence in the University.
- The recipient must never be at risk of being misled as to the authenticity of the content. Content that could be mistaken for authentic material (known as deepfakes) or for material that has been created entirely by humans must always be clearly labelled.
- AI-based tools should be regarded as a support for communication work and are not to replace human responsibility, judgement and quality assurance. Labelling is never to be used to disclaim this responsibility.
- The labelling is to be concise and factual. Follow the recommended wording for each type of content as far as possible.
- Labelling must be placed immediately next to the content. It must be clearly distinguishable and easy for the recipient to recognise immediately in the relevant context (by sight or sound), without having to click anywhere. This also applies to AI-generated content in printed material, and the example placements given under each content type should be adapted and applied to printed material as well.
These guiding principles have been drawn up on the basis of the regulatory document “Policy on Principles for the Use of Generative AI at Lund University” and are in line with the EU Artificial Intelligence Act (in particular Article 50). The transparency requirements set out in the Act are most stringent where content may be perceived as authentic, documentary or human-generated, whilst meaningful human review and clear editorial responsibility may, in certain cases, mean that labelling is not required.
Responsibility for the content published lies at various levels.
The person producing and publishing the content is responsible for ensuring that human review of the content is done correctly. Human review means that a natural person reviews and edits the content linguistically, and that a natural person with relevant expertise approves the content before it is published. What these processes entail varies depending on the different types of content, as explained on this page.
Editorial responsibility means that a legal entity – or, in the case of Lund University, a public authority – bears ultimate and overall responsibility for ensuring that the information contained in any content is accurate.
There are instances where content is published without having undergone a full human review, or where such a review is entirely absent. Although it is always advisable to review content, there may be exceptions, for example, in the case of automated summaries, etc. However, there are situations in which human review is always required.
Human review must always be carried out before the publication of content that may affect a person’s rights, obligations or expectations in relation to the University. This applies in particular to information where errors may cause the recipient to act on incorrect grounds. If the content sets out terms and conditions, provides information or could be interpreted as a promise made by the University, it must be reviewed and approved by a person with the relevant expertise before it is published, for example when information is provided regarding:
- students’ rights and obligations
- entry requirements
- application processes
- credit transfers
- support measures.
Types of content and labelling
Text
When generative AI is used to produce drafts, refine wording, summarise content or improve the language of texts, these do not generally need to be labelled. This applies provided that the texts are reviewed, edited and approved by one or more natural persons with relevant expertise.
Recommendations and legal requirements
Avoid publishing texts produced with the aid of generative AI where human review is entirely or partially absent. If such a text is nevertheless published, for example in the form of an automatically generated summary of an information page, it is treated as AI-generated text and must be labelled as required by law.
This text was created using AI and has not undergone full human review/Parts of this text were created using AI and have not undergone full human review.
- Check factual details, names, titles, dates, figures, references, decisions and contact details against reliable sources.
- Ensure that the AI tools have not added, omitted or drawn their own conclusions.
- Adapt the content, tone and terminology of the text to suit the target audience, channel and context, in accordance with Lund University’s language policy.
- Review the language in terms of plain language and accessibility, for example headings, link text, abbreviations and consistent terminology.
- Check English names, titles and terms against the LU glossary.
- Make sure that the text does not contain any stereotypes, exclusionary language, unauthorised personal data or other sensitive information.
- Have someone with relevant expertise read and approve the content and verify the reliability of the sources before publication.
Labelling must be placed immediately next to the content. The requirement is that the recipient must be able to view the labelling when directly viewing the content, without having to click anywhere. Use italics to distinguish the labelling from the rest of the text.
Recommended label placement
- On web pages or in articles where large parts of the text are AI-generated, the label should be placed directly below the introduction.
- On web pages or in articles where only a limited part (a paragraph, a fact box or similar) of the text is AI-generated, the label should be placed immediately before the text, for example under the heading for the paragraph in question or just before the paragraph if there is no heading.
- In PowerPoint presentations where all or a large proportion of the text is AI-generated, this information should be included as a visible footnote on every slide.
- In PowerPoint presentations where only certain slides or images are AI-generated, only those slides or images need to be labelled.
These examples can also be applied in other contexts and formats where AI-generated texts are published with Lund University as the sender. Use common sense and adapt the placement of the label according to the context.
Translation
When generative AI or AI-based translation tools are used to translate texts, this does not usually need to be indicated. This applies provided that the translated texts are reviewed, edited and approved by one or more natural persons with relevant expertise.
Recommendations and legal requirements
AI tools designed specifically for translation should be used. Avoid publishing translated texts where human review is entirely or partially absent. When such a text is nevertheless published – for example, in the case of an automatically generated translation – the use of AI must be clearly labelled, as required by law.
Parts of this text have been translated using an AI-based translation tool and have not undergone full human review/This text has been translated using an AI-based translation tool and has not undergone full human review.
- Compare with the original text and ensure that the meaning, nuances and purpose have been preserved.
- Check that nothing has been added, omitted, distorted or rendered misleading in the translation.
- Review factual information, names, titles, organisation names, dates, figures and contact details.
- Check that technical terms, LU terminology and English terms comply with the LU glossary and the recommended LU style guide.
- Adapt the language, tone and idioms so that the text reads naturally in the target language and for the target audience.
- Have the translation approved by someone with relevant expertise and language skills before publication.
When labelling AI-translated texts, you can follow the same principle and use the same examples as for other texts. See ‘Where to place text labels’ under the section ‘Text’.
Visual content
When generative AI is used to create or edit, either wholly or in part, visual content such as images, videos, animations, illustrations or other graphic material, this may affect how the recipient perceives the authenticity of the content. Transparency is therefore particularly important in visual contexts.
Deepfakes and legal requirements
Under the EU Artificial Intelligence Act, an image or video that resembles real people, objects, places or events and which could be mistakenly perceived as authentic or truthful is known as a ‘deepfake’. By law, these must always be labelled, regardless of whether a human review has been carried out.
Recommendations
Where AI-generated visual content is communicated with Lund University as the sender, it should, as a rule, always be labelled. This is because it can be difficult for an individual employee to predict whether the recipient will perceive the content as genuine (a deepfake) or not.
Do not use deepfakes for any meaningful purpose, such as when communicating about real places, people or events. However, AI may be used to add decorative elements or to generate genre-based images or video clips, as well as to streamline standard post-production work – provided that the content is not distorted or risks being interpreted as a representation of reality.
Examples: If a filmed segment deals specifically with the Skrylle Nature Reserve, the images and video clips must show the real Skrylle – not AI-generated content that resembles Skrylle. If, on the other hand, the feature is about forests more generally, AI-generated content may be used, provided that the message of the content remains unchanged.
Editing and post-production
If an image, video or similar content has been edited or post-processed using generative AI in such a way that the content of the original image has been manipulated or altered, this must be clearly indicated. Examples might include padding to make an image wider, or when something is removed from or added to an image or video.
Exceptions apply where AI has been used for standard editing, such as to improve image quality or brightness. In such cases, no labelling is required, provided that a full human review has been carried out.
Images from procured photographers
If images taken by photographers commissioned by the University are processed using AI, the photographer must always be given the opportunity to approve the material before publication. This is set out in the University’s agreement with the photographers.
[Parts of] Image/video/illustration created using AI.
In situations where space is limited:
AI generated image/video/illustration.
When editing visual content:
Image/video/illustration edited using AI.
- Graphic profile: All visual content produced by Lund University must comply with the University’s graphic profile and visual style.
- Authenticity: Do not use AI-generated visual content that conveys a key meaning in communications about real people, places or events. The recipient must not be at risk of being misled and interpreting the content as a true representation of reality.
- Photographs: If images taken by the University’s procured photographers are processed using AI, the photographer must be given the opportunity to approve the material before publication.
- Copyright and terms: Ensure that you have the right to use and publish the material, and comply with the terms of the AI tools used.
More on the use of AI in the creation of visual content in relation to legal considerations and the University’s graphic profile:
Legal requirements
Labelling of visual content must be visible to the recipient upon first exposure to the content and without requiring any interaction (such as a click). If the label is placed within a nearby text, that text must be visible as soon as the content is displayed.
Recommendations
As visual content may be perceived as realistic or as a depiction of real people, events or settings, the labelling should also be placed within the image or video itself, for example by means of specific icons or technical solutions. This also applies in contexts where the content can be accessed or shared without any accompanying labelling. As the University does not currently have such solutions in place, the following placements are recommended.
Recommended label placement
- Images on web pages/in articles: AI-generated images should primarily be labelled in the caption, where one exists. Where it is not possible to add a caption, the label may instead be placed in a paragraph of text that is directly visible beneath the image. Labelling included in the image’s alt text is not considered sufficient.
Tip: If you are working in Drupal, it is a good idea to choose an infobox rather than an imagebox or an image, as this allows you to add text directly alongside the image. - Video on web pages/in articles: When AI-generated videos are embedded on a web page, the labelling should be included in the caption or in other visible information immediately adjacent to the video.
- Social media: When AI-generated visual content is published on the University’s social media channels, it must be labelled in the post text. The label should be positioned so that the user does not need to click on “show more” or similar. If the platform offers features for labelling AI-generated or AI-processed content, these are to be used in accordance with the platform’s guidelines. However, such features do not replace your own labelling.
- PowerPoint: AI-generated visual content in PowerPoint presentations should be labelled in a visible caption or footnote on the slides where it appears. Including annotations in notes or speech scripts does not count as sufficient labelling.
Video subtitles
When subtitles for videos are created using generative AI, the same labelling principles apply as for text and translations.
Recommendations
Where AI is used as a support tool and the generated subtitles are reviewed, corrected and quality-assured by a natural person with relevant expertise, the use of AI does not, as a rule, need to be indicated.
This also applies where subtitles have been translated using AI, provided that the language and terminology have been reviewed by a natural person and that no content has been added, omitted or distorted.
If, on the other hand, subtitles have been generated in a more automated manner – for example, during live broadcasts or conferences, where human review is limited or, in practice, absent – there may be a risk that the content is interpreted in a way that could be misleading or inaccurate. Consequently, these are treated in the same way as AI-generated and translated text and must be clearly labelled for the recipient.
The subtitles have been created using AI and have not undergone full human review.
For subtitles, the same recommendations apply regarding human review of language and terminology as for text and translations.
- Place the subtitle labelling immediately next to the video player, for example in the video’s description, provided that it is immediately visible to the viewer upon their first viewing of the content and does not require any interaction (such as a click).
- If the descriptive text is not immediately visible when the video is played, the labelling should also be included as a subtitle at the start of the video, so that the viewer takes in the information while the video is playing.
Audio
When generative AI is used to create or process audio such as music, sound effects or synthetic voices, there is a risk that the listener may be misled. For example, the listener may believe that a voice belongs to a real person, that the recording depicts a real event, or that the music was composed by humans.
Recommendations and legal requirements
Under the EU Artificial Intelligence Act, AI-generated or manipulated audio that sounds like real or existing people, audio that occurs in specific locations or at specific events, and that could be mistakenly perceived as authentic or truthful, is known as a ‘deepfake’. By law, these must always be labelled, regardless of whether a human review has been carried out or not.
Avoid using AI-generated voices that could be mistaken for a real person, unless the person has given explicit consent and it is justified in the context.
Processing audio using AI
When audio is processed using AI to improve its quality, the content does not need to be labelled, provided that a natural person with relevant expertise has reviewed and quality-assured the editing and the final result. This applies to improvements such as noise reduction, level adjustments or other standardised post-processing.
If, however, the adaptation is likely to affect the recipient’s perception of the content’s authenticity – for example, if the wording or voices are altered – the content must always be clearly labelled as such. The same applies to synthetic reconstructions, for example, if the audio track has been lost from a clip and AI is used to recreate the sound.
Music/background sounds/sound effects/voice-over/narration/voices] created using AI.
Please ensure that:
- you hold the rights and licences to use music and audio, even when using AI.
- the script and factual content have been reviewed, and a natural person with relevant expertise has listened to the final product.
- you have obtained consent and are complying with legal requirements when using AI-generated voices that sound like real people.
In the case of sound, the labelling must be both audible and visible. The label must be positioned where the recipient encounters the sound, without any additional interaction, such as a click.
Recommended label placement
- At the start of an audio or video playback (such as voice-over and/or subtitles), if all or large parts of the audio are AI-generated.
- In subtitles for specific scenes or sections where the sound is AI-generated (sound effects, voices, etc.)
Contact
Corporate Communication
Anna Johnsson
Communications Officer
anna_v [dot] johnsson [at] kommunikation [dot] lu [dot] se (anna_v[dot]johnsson[at]kommunikation[dot]lu[dot]se)
+46 46 222 70 21
Edina Voloder
Communications Officer
edina [dot] voloder [at] kommunikation [dot] lu [dot] se (edina[dot]voloder[at]kommunikation[dot]lu[dot]se)
+46 46 222 67 42
Related information
The recommendations and requirements are based on the EU Artificial Intelligence Act (Article 50) and Lund University’s policy on the use of generative AI.