AI in Academic Writing Processes: information for students

• Science relies on verifiability. However, outputs from GenAI do not have the status of a source that is readable and verifiable by third parties, as the output for a given prompt is only valid for that specific prompt in combination with the corresponding version of the tool, license, and potentially even the time, region of access, etc.

To explain:

• AI-generated texts use Large Language Models (LLM), which are trained to recognize patterns in natural language and reproduce them accordingly. These Large Language Models (e.g. GPT-3.5, GPT-4  for OpenAI ChatGPT and CoPilot, Claude from Anthropic, Gemini Nano from Google, etc.) generate text outputs based on complex neural networks and statistical probabilities.
• They are trained on very large amounts of data. However, the dataset used for training a particular application is either a trade secret (e.g. Claude) or consists mainly of a Common Crawl, a specially compiled database of web texts, English Wikipedia, and freely available books (see for GPT-3 OpenAI July 2020: https://arxiv.org/pdf/2005.14165).

Texts created by GenAI have remarkable linguistic quality in most cases. However, high linguistic quality does not necessarily mean high content quality or accuracy. Even if texts are formulated in a complex way and sound credible, GenAI has a problem with factual accuracy.

Because GenAI does not generate content based on specific sources (such as databases, books, etc.), it instead relies on statistical patterns to predict the most likely next element (character, word) in a response. This process creates linguistically coherent texts, but it does not guarantee factual accuracy.
GenAI systems may sometimes "hallucinate" — that is, generate plausible-sounding answers that are not factually correct. Additionally, repeated queries may produce varying responses each time, as the output is influenced by probabilistic modeling.

The verification of accuracy, i.e. factual accuracy and critical questioning of GenAI-generated texts, lies with us humans.

Being the author of a text—whether a scientific publication or an academic thesis—means taking full responsibility for its content and ensuring its accuracy. As an author, you are accountable for the substance, structure, and proper citation of all sources. It is essential to clearly distinguish your own original ideas from those of others and to acknowledge all sources and supporting evidence transparently.

When using text-generating AI systems (GenAI) for more than basic proofreading or stylistic editing, you must: a) critically assess the accuracy and reliability of the content produced by the GenAI, and b) make clear how the use of GenAI should be disclosed in each specific context. Disclosure requirements may vary depending on the type of publication, so it is important to familiarize yourself with the relevant guidelines for your particular case.

A responsible use of GenAI also includes complying with legal regulations. Details on what this entails can be found in this checklist provided by KI:connect:NRW. For example, copyright-protected materials may not simply be uploaded into a GenAI system for further use.

In the current AI Declaration of independence of Bremen University of Applied Sciences, two options are available: -

• Option 1 – Use of AI technologies without a disclosure requirement 
• Option 2 – Use of AI technologies with a disclosure requirement 

In both cases, coordination between students and examiners is required.

• ZLL AI Inspirations held on 24 June 2025 on the topic:  "AI examination regulations & AI disclosure and documentation" (PDF) (the slides are a condensed version of the most important information on AI labelling).
• You can find further AI Inspirations here. 
   

The term ‘artificial intelligence’ (AI) encompasses a wide range of different systems, which can be categorised in a variety of ways. Each of these categories represents a specific perspective within the field of IT. One possible distinction, for example, is the differentiation between expert systems and generative AI. An expert system is a programme that derives recommendations for action from a comprehensive, verified knowledge base, thereby assisting people in solving complex problems in the same way an expert would. They do not, therefore, generate new content, but generally reproduce results that already exist. Generative AI tools, on the other hand, are explicitly designed to create new content. Generative AI is also trained using data and information from which it identifies patterns and structures. However, the AI uses these patterns to produce new content such as text, audio, images or videos. Whilst other forms of AI can also generate content, this is merely a side effect and not their primary function. In generative AI, the primary task is the generation of new content. The EU AI Regulation (EU AI Act) distinguishes between general-purpose AI (GPAI), such as ChatGPT, and specialised AI software.
Large Language Models (LLMs) represent a particularly powerful form of generative AI and are primarily used for text-to-text generation in popular chatbots. They are based on deep neural networks and are capable of generating and processing human-like text in various languages. By training on extensive text corpora, LLMs can recognise and apply a wide range of linguistic patterns, making them versatile tools for tasks such as writing, summarising or translating texts.
Other forms of generative AI include, for example, text-to-image models (image generation). These are also models based on neural networks which, having been trained on large image corpora tagged by humans, are able to generate new images.
Conventional chatbots typically combine various models that run in the background and handle individual tasks.
The way language models work means that their output is based on statistical calculations. Consequently, the output is not necessarily factually correct and is not verified before being displayed. This results in statements that are statistically probable but may not be factually accurate. Furthermore, common chatbots tend to reinforce users’ existing ideas and opinions. The reliability of information from language models should therefore always be checked.
 

At Bremen City University of Applied Sciences, the Academic Cloud and HSBrain (based on HAWKI) services are available. HSB members can log in using their HSB login details and make use of various services, such as the ChatAI chatbot (AcademicCloud), which offers a choice of different language models.
If you have any questions or require advice, or are unsure which models to use, please contact the ZLL-Support-Team.
 

Chat AI provides a selection of various Open-Weight Large-Language-Models (LLMs) that are hosted on the GWDG (Gesellschaft für wissenschaftliche Datenverarbeitung mbH Göttingen) platform in compliance with data protection regulations. The data sent to these models when using ChatAI, including prompts and responses, is not stored in GWDG’s systems. However, ChatAI also offers a number of externally hosted models, such as OpenAI’s GPT-5, GPT-4o and GPT-3. With these external models, data is sent to the respective providers.
The GWGD regularly updates the list of available models as newer and more powerful versions are released. The GWDG also provides a comparison list of the models, which is likewise updated on an ongoing basis (Link).
 

The responsible use of GenAI technologies in an academic context raises questions to which there are often no definitive answers yet. Here, we provide information on the use of language models, as well as their labelling and documentation.

• Please discuss any specific questions regarding the use of language models (and any other models) in your studies with the relevant lecturers or examiners for your module. You can use the recommendations and templates for labelling and documentation provided on these pages.
• Do you have any specific requests or information regarding GenAI that go beyond your modules? Please contact StudiumPlus – we are happy to assist with your needs.
• If you have any questions about GenAI in relation to academic writing, please contact the Writing Lab.
• For more technical questions about AI technologies or our in-house GenAI, please contact the ZLL-Support-Team.
   

The EU AI Regulation establishes a legal framework designed to ensure that the use of artificial intelligence in the EU is safe and ethical. It has been in force in principle since August 2024, with the individual provisions being implemented gradually. The first bans on unacceptable risks have been in force since February 2025, whilst most of the regulations for high-risk AI systems will not come into full effect until 2026 and 2027.
For Bremen City University of Applied Sciences, as an educational and research institution, the classifications set out in the EU AI Regulation are crucial: AI systems posing an unacceptable risk, such as social scoring systems, are prohibited. High-risk systems, such as those used for assessing learning outcomes or exam supervision, are subject to strict requirements, including documentation, monitoring and training. Language models such as ChatGPT also fall under the category of General Purpose AI (GPAI) and are strictly regulated depending on their use.
The HSB is committed to the responsible use of AI within the framework of these guidelines. Lecturers play a key role by not only teaching students the technical aspects of AI, but also raising their awareness of ethical and legal issues. The aim is to enable students to use AI competently, thoughtfully and responsibly.
In consultation with legal experts, Bremen City University of Applied Sciences is working to ensure that teaching staff and students have legally compliant access to generative AI tools in accordance with the EU AI Regulation. If you have any concerns or questions, please contact the ZLL-Support-Team.
 

A degree course is intended to prepare you for your future career, and it is reasonable to assume that GenAI models will become part of professional practice in academia and other fields. It therefore makes sense for you to familiarise yourself with language and other models and learn to use them (critically) whilst you are still studying. 
Your lecturers will assess whether and how GenAI models can and should play a role in their modules and will inform you accordingly.
There are various ways to regulate the use of AI in higher education: 

1. Unrestricted use of general AI
2. Selective use of specific general AI models
3. Targeted integration of general AI with clear rules of use
4. No use of general AI applications permitted.

If you have any questions or are unsure about anything, please always ask your lecturers directly.
Bremen City University of Applied Sciences’ AI Declaration of independent preparation of work  currently allows for the use of GenAI models in teaching and examinations, with or without a documentation requirement. If you have not received any information about the Declaration of independent preparation of work from your lecturers, please always ask them directly which version you should use.
 

If you are considering making the use of GenAI models compulsory in teaching, various aspects must be taken into account, including, amongst other things, the principle of data minimisation, students’ informed consent to the processing of their data by specific services and/or organisations, associated costs and billing options, as well as the terms of service for these services, which specifically govern the handling of usage data. In particular, the terms of use vary depending on the software, model or platform.
If a solution that complies with data protection regulations (under the GDPR) cannot be guaranteed, the use of such services should be viewed critically and permitted only on a voluntary basis.) Academic Cloud und HSBrain (based on HAWKI) offer access to language models and other services that complies with data protection regulations. 
Furthermore, it is generally the case that the copyright of uploaded documents must be respected, particularly in the case of proprietary GenAI models.
 

Academic writing serves a variety of purposes. It not only serves to present and communicate research findings, but also acts as a tool for thinking and learning. By writing and reading academic texts, students develop an understanding of the subject matter as well as the typical structures of thought and argumentation within the discipline. Writers should be made aware of these functions so that they can decide which tasks to tackle independently and which to automate.

At first glance, current research in this field appears contradictory and ambivalent. For example, neuroscientists point out that using language models when writing means fewer cognitive resources are utilised and that brain activity shifts (Kosmyna et al, 2025: your Brain on ChatGPT). There is evidence to suggest that the uncontrolled use of proprietary language models leads to learners learning less (Bastani et al, 2024: Generative AI Can Harm Learning), that such use can lead to mental laziness (Fan et al, 2024: Beware of metacognitive laziness), or that students rely too heavily on generative AI systems (overreliance) and fail to question the outputs even within their own field of study (Krupp et al, 2024: Unreflected Acceptance). Other studies suggest that specially trained GenAI systems (Kestin et al, 2025: AI tutoring outperforms in-class active learning) or GenKI systems with RAG functions (Ma et al, 2024: Integrating AI tutors in a Programming Course) lead to better learning outcomes among students. For students studying in a language other than their first language, the use of GenAI models can lead to improved linguistic quality (Nguyen Minh, 2024: Leveraging ChatGPT for Enhancing English Writing Skills and Critical Thinking in University Freshmen).
The extent to which the use of GenAI fundamentally hinders learning therefore depends on how the models are used. However, given the widespread availability of proprietary GenAI models, there is a risk that they will be used in an uncontrolled manner outside of research experiments. As language models can quickly produce error-free and natural-sounding text, their use is tempting. This can lead to students questioning their own abilities, losing them (Deskilling, PDF in German) or failing to develop them in the first place (Skill Skipping, PDF in German). In the context of programming, this is referred to as ‘Vibe Coding’, i.e. people with no programming knowledge rely on AI programmes to generate code for them (which poses significant security risks). Similar practices are now also emerging in the field of design (Vibe Designing).
For teaching at the university, this means above all that students must be given critical guidance on their use of GenAI. It is not enough simply to allow the use of these models. In addition, issues relating to academic and professional ethics should also be addressed.
Students should learn to engage actively with topics. Critical analysis, interpretation and evaluation should not be left to AI; they are key steps in the thinking and learning process that students should master during their studies. To this end, they must be given the opportunity to do so within their courses.
 

There is no standard policy at HSB governing the use of GenAI.

• It is therefore up to the relevant degree programme and/or lecturer to establish rules regarding this use and to communicate them clearly to students at the start of the relevant seminar.
• Depending on the learning objectives and assessment requirements, the use of GenAI may be permitted in full, restricted, or not permitted at all. 
• The HSB provides two sets of guidelines for written work: The Declaration of independent preparation of work sets out the rules governing the permissible use of GenAI in the relevant piece of work. This may need to be adapted to your specific use case. 
• In addition, there are guidelines for staff and students on three different ways of indicating the use of GenAI in written work.
• Separate arrangements must be made for examination formats such as presentations.
   

The EU AI Regulation (EU AI Act) requires the use of AI systems to be disclosed in certain cases. This applies, among other things, to deepfakes and the publication of texts intended to inform the public about matters of public interest (EU AI Act, § 50(4)). In such cases, lecturers are obliged to disclose the use of generative AI models in their teaching. Research activities are not usually affected by this.
In addition, many lecturers use GenAI models, for example to prepare their lessons. They can lead by example by indicating when teaching materials, exam questions, etc. have been generated using GenAI models. In doing so, lecturers can use the same labelling methods as students or follow other formats, such as the AI Use Cards (see the white paper on this).
 

If the use of GenAI models was not explicitly part of the assignment, the decision to use or not to use them should not, in itself, affect the assessment. If the use of GenAI models is part of the assignment, steps must be taken to ensure that the use of the relevant models 1) is taught and integrated into the teaching, and 2) is carried out in compliance with data protection regulations.
Furthermore, it is relevant to both staff and students that individuals can be excluded by digital service providers, as happened, for example, with Karim Khan, the Chief Prosecutor of the ICC (Webpage in German), by Microsoft.
 

As GenAI outputs are not considered citable sources, the term ‘citation’ is not used in this context. Instead, in accordance with good academic practice, any use of GenAI should be disclosed and documented. Please refer to the guidelines on labelling and the AI-Declaration of independent preparation of work.
Background: Academic citation serves to ensure transparency and traceability. The purpose of referencing a source is to enable others to follow the reasoning behind a researcher’s conclusions. This is not possible with all language models when it comes to text generated by them: The outputs are not readable or traceable by third parties, as each output is generated solely through a combination of the prompt, the user, the AI model, the model context, etc., and cannot be reproduced by third parties. Furthermore, the issue of output traceability is currently the subject of widespread discussion within the IT community; further information on this can be found, for example, at Joshi et al, 2026: Stochastic CHAOS.
 

The use of GenAI models in examinations must be explicitly permitted by the relevant lecturers and clearly communicated to all students on the relevant module. 
The General Examination Regulations state: “The use of artificial intelligence that has not been expressly permitted constitutes the use of an unauthorised aid” (§ 16(2), AT-BPO, AT-MPO, revised version of October 2023). Unless explicit permission is granted by teaching staff, the use of GenAI models is considered to be the use of an unauthorised aid and therefore an attempt at deception.
 

If GenAI models are used in examinations without authorisation, this constitutes the use of unauthorised aids and may be regarded as an attempted deception (see §16 AT-BPO & AT-MPO). The procedure is the same as that for determining other attempts at deception. 
The regulations for the respective examination, including the use of GenAI models, should be established in advance and communicated to students at an early stage.
 

Neither specialised AI detectors nor general-purpose plagiarism software can reliably detect the use of generative AI models. These tools classify automatically generated texts as having been written by humans, and vice versa. Students are particularly affected by the latter, especially those who do not sit written exams in their first language or who come from non-academic backgrounds (Gostmann und Hildermeier, 2026, Webpage in German). As a result, such detectors reproduce existing inequalities.

An attempt at deception requires intentional conduct. Negligent conduct, on the other hand, does not constitute an attempt at deception. It must therefore first be established that the conduct was intentional, usually on the basis of circumstantial evidence (Baresel u. a., 2025, Webpage in German).
As it is not possible to reliably detect the use of GenAI via detectors, examiners must themselves demonstrate any unauthorised use, for example by comparing different passages of text, comparing the work of different students, or identifying typical expressions found in GenAI models (Baresel u. a., 2025, Webpage in German). Examples of typical AI-generated text can be found (in English) on the Wikipedia Project AI Cleanup, amongst other places.
The procedure for identifying an attempt at deception is based on § 16(1) and (2) of the AT-BPO and AT-MPO.
 

Examination papers that go beyond highly structured formats, such as multiple-choice exams, are subject to copyright. Submitting such an examination paper without the prior explicit and informed consent of the students constitutes unauthorised reproduction and an infringement of copyright.
Furthermore, according to the EU AI Regulation, the automated assessment of student performance by AI systems constitutes a high-risk application. Where AI systems are used in this way, the higher education institution and the relevant examiners must fulfil specific obligations (Baresel u. a., 2025, Webpage in German). Until the Bremen City University of applied Sciences has established relevant regulations and introduced appropriate measures, the use of AI systems for high-risk applications is prohibited. Conventional practices, such as the automated marking of multiple-choice exams, in which exam results are determined automatically in a deterministic and rule-based manner, are not affected by this.