Guidelines for AI

• Starting point for the guidelines

  Generative AI are systems that behave like humans and can autonomously produce a new image, text or combine images, text and sound into an autonomous output.

  SDU's guidelines for the use of AI - especially generative AI - are based on the premise that we must still live up to the same high standards of quality and the universities' obligation to seek the truth.

  • In addition, the guidelines are based on already existing guidelines for privacy, consent, marketing, PR and journalistic practices.
  • A human-centric approach, as defined in the EU AI Act (December 2023).
  • Workshop with the communications organisation at SDU on 18 December 2023.
  • Corporate Visual Identity (CVI) of the University of Southern Denmark.
  • CO2 emissions, as some AI systems emit the equivalent of one smartphone charge per prompt.

  AI logic systems are implemented as assists in various programmes today, including image and text programmes. These systems do not produce new things, but optimise. They can and may be used for general editing, such as creating effects (toning, blurring), enlarging images, removing background/isolating subjects.

• Know your responsibilities and read the small print

  Since Generative AI systems create images and text from pre-existing original images and text, you need to pay special attention to the following:

  • Original texts or images may have been collected illegally through surveillance
  • Original texts or images may be protected by copyright
  • Motives or wording may be heavily biased, manipulative, fake or otherwise contain unvalidated statements
  • Original texts, including research articles, contain knowledge the author would not necessarily have in Generative AI systems
  • Generative AI systems can be hacked like other systems, and content protected by the GDPR regulation is particularly vulnerable.
  • Consider who is actually behind the solution you are using. Generative AI systems often reside in a cloud, so don't share everything with everyone.
  • Also, remember to read the small print. For example, are you giving up your own input (data) as payment to get something in return?

  You are responsible for ensuring that text, images and speech comply with copyright laws and do not discriminate, manipulate or distort the truth.

  In other words, this means that when you as an employee publish or redistribute AI-generated content, you are liable for the outcome if the content violates laws and directives or otherwise compromises SDU.

• Good use of AI

  AI is suitable for

  • optimising automated processes.
  • versions of the same basic concept, e.g. advert texts.
  • short, generic texts, e.g. advert texts.
  • illustrations and images when they comply with SDU's visual identity and prior concept work has taken place enriching website metadata (keywords, description) and teaser text for search engine optimisation.
  • idea development as a creative assistant and sparring partner.

  It must be declared for longer texts, audio and images if they are primarily developed using artificial intelligence.

• Where the line is drawn

  We don't use AI for

  • Creative outputs in sound, images or text where there is uncertainty about the copyright of the original images used or where people or context have been manipulated without consent.
  • Photographs of people and locations that purport to be documentation of events that  never happened.
  • Texts, including press releases, op-eds, profiles, longer social media opinion pieces, or entire campaigns without checking copyrights.
  • Images that go against our intention to be authentic and have the appearance of being a collage of stock photos.
• AI for translation

  Machine translation is very useful for smaller texts, such as emails and messages, but think twice about long texts such as articles and news.

  The machine-translated text is still your responsibility, so make sure you read it through and quality check.

  Five tips for those who use machine translation:

  1. Remember that linguistic errors in the source text, such as forgotten words or syntax errors, may be reflected in the machine translation. You should therefore quality assure both the source text and the translation.
  2. Be on the lookout for errors such as idioms or figurative language. Language models translate words, not opinions.
  3. Always tell the recipient if your text is machine translated. As the producer, you are responsible for your text - even when it is machine translated.
  4. Double-check SDU's glossary. SDU has its own glossary of approved SDU terms. Language models don't recognise it, so you need to find the list yourself and align it with the terms used by SDU. Find SDU's glossary of terms.
  5. Take care of confidential data, research results, sensitive personal data or other GDPR considerations. You must comply with current legislation, so be aware of confidential content when uploading text to translation programmes.

   

  Feel free to use SDU's own translation tool, it takes SDU's term list into account. The other advice and recommendations still apply.

  Find SDU's translation tool.