Academic Work and AI

Artificial intelligence (AI) refers to the “automation of intelligent behaviour. It is not clear what ‘intelligent’ means, nor which technology is used” (Kompetenzplattform KI.NRW 2024). Generative AI models are one type of Artificial Intelligence. This includes technologies which are based on comprehensive training and can be used to create new content (e.g., texts and images).

AI tools for text generation are based on Large Language Models, which have been trained with large text corpora. The functioning of a language model can be described as the generation of statistically probable word sequences. Through training, the model is able to calculate which words will most likely occur in which contexts. This calculation includes a random factor in order to prevent the generation of the same texts over and over again.

Language models are neither knowledge models nor search engines. They are optimised to give answers that are as humanly, as correct in terms of language and as useful as possible – but not necessarily factually correct. This means there is no guarantee for factually correct answers. Therefore, please do not simply trust the output, but always check for correctness and appropriateness of the content, e.g., on the basis of specialist literature.

You will find more information in the “Handout on Generative AI Tools for Students” (see download box), chapters 1.1 and 1.2.

The functioning of a language model can be described as the generation of statistically probable word sequences. Language models are optimised to give answers that are as humanly, as correct in terms of language and as useful as possible – but not necessarily factually correct. Therefore, please do not simply trust the output, but always check for correctness and appropriateness of the content, e.g. on the basis of specialist literature.

When working with generative AI tools, the responsibility for the content of the texts always remains with you. In terms of copyright, AI tools cannot be considered as authors or creators of the texts that they generate. You are responsible for the factual correctness of the content in your academic text.

Academic work is one type of examination at HSBI. Generative AI can be useful in different ways in the various phases of academic work. The potentials of generative AI are related mainly to the beginning and the end of a writing process as well as to phrasing and translating. But you should always observe the limitations of the use of AI.

The potentials of generative AI lie in generating ideas, creating a schedule and work plan, suggesting methodologies, summarising literature, explaining difficult passages in the sources, suggesting suitable databases, creating draft outlines, adapting to the style of academic language, adding keywords to texts, further developing text fragments, explaining technical terms, finding synonyms, adapting texts to target groups, assisting in understanding source texts, standardising source references, making stylistic, linguistic and formal corrections, summarising your own texts and providing (criteria-based) feedback.

Generative AI reaches its limits when it comes to justifying the selection of an idea, narrowing down topics, adjusting the time and work schedule, selecting the appropriate method(s), reviewing AI-summarised literature, researching and selecting suitable sources, documenting the literature search, revising the outline, creating a detailed outline, reviewing statements/arguments for truthfulness, logic, persuasiveness, validity and correctness, reviewing corrections, summarising your own texts (taking into account data protection/terms of use and providing comprehensive feedback.

You will find more information in the “Handout on Generative AI Tools for Students” (see download box), chapters 1.1 and 2.2.

First and most important recommendation: talk to your examiners and follow their instructions.

After that, the following recommendations apply: HSBI’s declaration of originality requires you to indicate when you have used generative AI. It also specifies what information you need to provide (product name, access source, information on the software functions used and the scope of use).

Case 1: Content extraction
As a rule, we advise against the first case, i.e., extracting content. On the one hand, contents may be hallucinated, which means that they must not be included in a fact-based academic text. On the other hand, generative AI tools are not original sources (which you should always use for citations). Rather, they remix their training data (the exact composition of which is unknown) to form new contents. The substance of these contents is sometimes doubtful.

However, if you are determined to cite content from generative AI despite our recommendations, you have to specify the sources of the extracted content – just like you do when you extract content from other sources. To do this, use the referencing system of your choice for citation.

Case 2: Inspiring or corrective work
For the second option, using generative AI for inspiration or proofreading, you are also required to document the use in order to fulfil the requirements of HSBI’s declaration of originality and to make the originality of your work verifiable and assessable for your examiners.

An ‘overview of aids used’ can serve as documentation, containing information on the product name, access source, software functions used and scope of use (see requirements of the HSBI declaration of originality), which can be inserted into your academic text after the references.

You will find more information in the “Handout on Generative AI Tools for Students” (see download box), chapters 1.1 and 2.1.

As a general rule, you must indicate the sources for passages taken from other works verbatim or in paraphrased form in your written work. This also applies to AI-generated content (see HSBI’s General Examination Regulations of HSBI , § 13(6)). By submitting the university-wide declaration of originality, which is mandatory for all written work, you assure “that you have completed your work independently and without the help of third parties and that you have identified citations as such and indicated the sources” (HSBI’s Framework Examination Regulations, §13 (6)).

In addition, there are legal proceedings against various AI companies for copyright infringements in the training corpus, particularly in terms of image generation. Thus, you run the risk of committing copyright infringement if you use an AI-generated text one-to-one. Even if it is very unlikely, it cannot be completely ruled out that an output will contain a verbatim copy from a training text once in a while.

You will find more information in the “Handout on Generative AI Tools for Students” (see download box), chapters 2.1 and 1.3.

Writing an academic text, e.g., a bachelor thesis, demonstrates “[...] the ability to work independently on a subject-related issue from an interdisciplinary, academic and subject-specific perspective by a specific due date” (see HSBI’s General Examination Regulations §26). Depending on the task, the requirement for working independently can no longer be met when using AI tools.

This is the case, for example, if you outsource too many tasks to AI tools, so that your own share of the task no longer predominates or is no longer recognisable.

The unauthorised outsourcing of the task to a third party, so-called ghostwriting, also applies to AI-supported offers (see HSBI’s General Examination Regulations, §13(6)).

You will find more information in the “Handout on Generative AI Tools for Students” (see download box), chapters 1.1 and 2.1.

Writing an academic text, e.g., a bachelor thesis, demonstrates “[...] the ability to work independently on a subject-related issue from an interdisciplinary, academic and subject-specific perspective by a specific due date” (see HSBI’s General Examination Regulations §26).

Depending on the task, the requirement for working independently can no longer be met when using AI tools. This is the case, for example, if you outsource too many tasks to AI tools, so that your own share of the task no longer predominates or is no longer recognisable. A lack of independence in your work will affect the assessment of your performance.

Talk to your examiners about the performance requirements and evaluation criteria that apply to your specific assignment.

You will find more information in the “Handout on Generative AI Tools for Students” (see download box), chapters 1.1 and 2.1.

Creating good prompts for generative AI tools can help you get more relevant answers. Here are some general tips for formulating effective prompts:

1. Be specific: precise prompting helps you focus AI’s response on your desired topic: try to keep the prompt short and concise without losing important details, as very long or complex prompts can dilute the clarity of your query.
2. Provide context: if context or background is important, let the AI tool know.
3. Be aware of your objective: specify the objective of the response.
4. Be specific in your questions: ask targeted (follow-up) questions to guide the answer.
5. Be specific about the format: if you prefer a specific response structure or formatting, please specify.
6. Be specific about the style and tone: if you want a certain style (formal, informal, funny etc.), let the AI tool know.
7. Examples: an example can help to clarify the query.
8. Proofread for errors: finally, make sure that the prompt does not contain any typographical or grammatical errors that could lead to misunderstandings.

AI tools like http://www.prompt-creator.ai specialise in generating prompts and can assist you in prompting.

Data protection

Many AI tools require registration, which links your own data to your e-mail address and, if applicable, your mobile phone number. Additionally, some of your entries may be used for further training. It is therefore important that you do not enter personal data or business-sensitive data. When using these tools, data is sent to servers outside the EU (including the USA), where different laws apply with regard to copyright and, in particular, data protection.

In compliance with data protection regulations, you can only work with HSBIKI, but not with the ChatGPT version that you can access directly via the OpenAI website. This is because you only register on the university website with your HSBI ID when using HSBIKI. Due to the connection via an interface, there is no data exchange between your personal user account at the university and the OpenAI provider. Neither the teaching staff nor other persons from the university can access your entries in HSBIKI.

Copyright

Generative AI cannot be an author or an originator. If you enter a simple prompt, the output is initially in the public domain, which means that it is not protected by copyright. However, you can be an author or an originator yourself, if you provide an “intellectual contribution to a significant degree” (Hoeren 2023: 23).

In addition, legal proceedings are pending against various AI companies for copyright infringements in the training corpus, particularly in terms of image generation. Thus, you run the risk of committing copyright infringement if you use an AI-generated text one-on-one. Even if it is very unlikely, it cannot be completely ruled out that an output will contain a verbatim copy from a training text once in a while.

Distortions

The training data may contain distorted perceptions or judgments as well as biases, which may also appear in the output of the AI tool, e.g., ethnic or gender bias (Albrecht 2023: 67-70; HUL 2023: 6; Liang et al. 2023; Oertner 2024). The underlying texts are mostly in English and shaped by Western culture. Prejudices and distorted perceptions present therein are transferred into the model and thus into the output.

 “[The text outputs] can be prejudiced, stereotypical, superstitious, ideologically coloured or politically radical and are based on private opinions, dubious reporting, gossip, advertising, commercially and politically motivated disinformation, propaganda and fake news. They do this, alarmingly, even when they are not presented in the style of an opinion, but in the sober style of a factual assertion.” (Oertner 2024: 280, translation by HSBI)

This matter-of fact style can lead to false confidence in AI and the truthfulness or factual correctness of the answers, in particular when people are not very familiar with how generative AI tools work and with regard to topics in which they do not have own specialist knowledge.

Ecological footprint

From an environmental perspective, text-generating AI tools are problematic, in particular training, but also operation, which means that daily use by a large number of users involves high electricity consumption and carbon emissions (Zandt 2023; Landwehr 2023; Strobl 2023; Ludvigsen 2023; Ludvigsen 2023; Heikkilä 2022). A production of approx. 552 tons of carbon is expected in order to completely train GPT3 once – for comparison, “[t]he average emissions of all passenger cars registered in Germany in 2022 [...] at 10,000 kilometers driven per year and vehicle were around 1.1 tons” (Zandt 2023, translation by HSBI). The energy spent on training amounts to 1,287 megawatt hours, which corresponds to “the energy consumption of 320 four-person households in a year” (Strobl 2023, translation by HSBI, emphasis in the original).

Abusive use

It is easy to create many texts with language models in a short amount of time which are linguistically correct and sound human, but this goes along with potential abusive use in fake news and spam e-mails. As mentioned above, language models are not optimised towards factual correctness, but display statistically probable word orders. Thus, GPT 4 is still prone to misinformation.

Working conditions

In order to filter out discriminatory language when training language models, human feedback is necessary, for which a large number of the cheapest possible workers are usually used. It has been revealed that workers in Kenya were paid less than $2 an hour to read and tag violent, racist, sexist and other traumatic texts for the training of ChatGPT 3.5.

You will find more information in the “Handout on Generative AI Tools for Students” (see download box), chapters 1.3 and 1.4.

The application HSBIKI is provided for students and employees of HSBI to enable communication with a chatbot. In this process, no data exchange takes place between the individual user account (i.e., login information) and the provider. Only HSBI-affiliation, date, time and content of the entry are transmitted. The entries, however, will not be used for further training. Neither the teaching staff nor other persons from the university can access your entries in HSBIKI.

You can find more information on HSBIKI, terms of use, data protection and chat access at: https://www.hsbi.de/dvz/it-services/hsbiki

The tools listed here are only examples to inform users and do not constitute a recommendation.

Writing

You can find more writing tools here: https://www.vkkiwa.de/ki-ressourcen/ki-tools/

Literature review

On the following page, you will find general information on AI-supported literature research and the Connected Papers tool licensed by HSBI: https://www.hsbi.de/bib/workshops/ki-recherche

Machine translation

You will find information on how to deal with AI-based translation tools in your studies in the “Guidelines for the Use of AI-Based Translation Tools for Studies” in the download box.

Transcription

On transcription tools: https://youtu.be/DItJ1MZMxb0

E-mail address for all students with questions on writing with generative AI tools:
ki-wissenschaftlich-arbeiten@hsbi.de

Writing consultations at the faculties for students with questions on the general writing process:

• Faculty of Engineering and Mathematics: Kristina Rzehak kristina.rzehak@hsbi.de
• Faculty of Social Sciences: Anna Surmann anna.surmann@hsbi.de
• Bielefeld School of Business: Uta Kentzler uta.kentzler@hsbi.de
• Faculty of Health: Stephanie Wiens stephanie.wiens1@hsbi.de

Heinrich Heine University Düsseldorf offers an AI self-learning course “AI for everyone 1: introduction to artificial intelligence,” which is open to all students (German-taught): https://ki-campus.org/courses/kifueralle-hhu

You can also attend the follow-up course: “AI for everyone 2: understanding, evaluating, reflecting”, which imparts advanced skills (German-taught): https://ki-campus.org/courses/kifueralle2-hhu

You can filter for English language courses on the KI-Campus website: https://ki-campus.org/

The project KI:edu.nrw offers free basic training courses, advanced courses, self-study courses, information material, introductory videos and advice on the topics of AI competence and learning analytics (in German).

HSBI’s Data Processing centre provides students and employees with a (German-language) online self-study programme from SecAware.nrw-Selbstlernakademie on HSBI’s ILIAS portal. This self-study programme also includes a module on AI: https://www.hsbi.de/elearning/ilias.php?baseClass=ilDashboardGUI&cmd=jumpToSelectedItems