Generative AI tools

The technology behind current text generation systems are so-called Large Language Models (LLMs). These are statistical models of very large text corpora that are able to make plausible continuations for a given text input. LLMs gain this ability through several steps: First, they are "trained" on billions of pieces of human-written text (scanned books, excerpts of the internet, etc.). In a further step, these models are adapted according to an extensive manual process of evaluating the (un)desirability of their outputs. As a more comfortable and familiar form of interaction, the user interface for such an LLM often takes the form of a dialogue-like chat interface, as opposed to simply 'continuing' the text.

The main property of LLMs -- which is underlying both their impressive abilities as well as their limitations and problems -- is that their outputs are statistically similar to the data they have been trained with. Assertions made by LLMs therefore often sound very plausible, but aren't necessarily correct. By the same token, biases present in the training data can be reproduced and thereby reinforced. This can include, for instance, the overrepresentation of certain languages and written cultures, or within academia, certain fields, methods, authors etc. Conversely, potential gaps and omissions in the training data can cause an LLM to produce unreliable outputs.

Because language models have no capacity for metacognition about what they 'know' and say, they are unable to recognise when they don't know an answer: They will always give an answer as if they know, and it's up to the user to say whether that answer was reasonable or not. By "fine-tuning" the models through human judgement as mentioned above, the manufacturers of LLMs also try to prevent the reproduction of toxic (e.g. racist, sexist or otherwise dehumanising) content such as it might be found in training data taken from the internet. This makes such outputs less likely, but is by no means always successful.

Awareness of the data with which LLMs are trained therefore plays a major role in assessing their capabilities and risks. Although nearly all manufacturers of LLMs keep the exact composition of their training data secret, as a rule of thumb it is helpful for users to think about how well or poorly, biased or diverse the subject area on which they are questioning an LLM tends to be represented on the internet as a whole.

When using generative AI tools (including BIKI), there is some legal information to be observed. In particular, we ask you not to enter/upload any personal or copyright-protected data into these systems. When using them in teaching and studying, you may also need to take into account aspects of examination law, for which we refer you to the relevant information page of Justiziariat Studium und Lehre (German).

The above-mentioned statement that the output of generative AI tools must always be checked for correctness and appropriateness is not only pragmatically necessary in order to use these tools sensibly, but is also derived directly from the obligation of our university and its members to good scientific practice. Bielefeld University's guidelines for safeguarding good scientific practice can be found here.

The conscious and unconscious attribution of human attributes to systems that can interact with their users in natural language is a phenomenon that has long been recognised in cognitive science and AI research. With the enhanced linguistic capabilities of current generative AI systems, these effects have become even more pronounced. This is already reflected in the vocabulary used to talk about these systems: They "learn" and "know", "the AI" "thinks" or even "hallucinates". Together with the fluent linguistic interaction typical of a generative AI chatbot, this type of language suggests to many users that it is - at least to some extent - an understanding, thinking agent. This in turn has consequences for the further assessment of these systems: a system that "hallucinates" has temporarily lost its touch with reality, but still has the capacity to be "in touch with reality" in principle. This aspect can even extend to legal issues: for example, a frequently cited justification for the use of copyrighted works in the training of LLMs is that they "read" the texts in a way way similar to how human authors do, without the latter being accused of copyright infringment for their writing being influenced by what they previously read.

Both, the training and operation of generative AI systems are very energy-demanding. For example, a query to ChatGPT consumes around six to ten times as much energy as a Google search. The current share of AI systems in the total energy consumption of data centres is estimated at 10-20%, with an upward trend (see here). Generative AI tools should therefore not be used without careful consideration. In particular, for tasks that can already be performed well and possibly more reliably with other tools (search engines, spell and grammar checking, translation, etc.), we recommend to use these tools instead. In this context, we also refer you to Bielefeld University's sustainability mission statement.

Used in a responsible and judicious manner, generative AI systems can be a versatile, complementary tool for education and teaching.

Depending on your level of expertise and the specifics of your subject, different applications of generative AI tools may or may not be suitable. What remains central is that you must be able to evaluate the content of the output, make your use transparent, and take responsibility for the content of the outputs you wish to continue working on. This is most likely to be possible with forms of use where you can simply discard outputs that you find inappropriate. This could, for instance, include the following:

• Support with idea generation
• Generation of clearly outlined text snippets or program code drafts
• Paraphrasing or rewriting your own texts
• Creating drafts for teaching material, e.g.
  • Repetitive exercise material in variations
  • Plausible incorrect answers for multiple choice tests
• Feedback or corrections for your own texts
• Additional help with literature research (as a source of inspiration, not as a database)
• Raw translations

Regardless of whether you use generative AI tools in teaching yourself, we recommend that lecturers treat generative AI - with due brevity - as a subject matter of their teaching. Especially since Bielefeld University gives lecturers a great deal of freedom in regulating the specific use of generative AI tools, it is important for a successful course to ensure that your students understand the learning objectives of your course and to what extent generative AI tools could help or hinder them in reaching these objectives. Ideally, you should enter into dialogue with your students and give them the opportunity to exchange ideas. Proactively addressing generative AI in your course is the best way to prevent ambiguities and problems with the use of unauthorised tools after the course.

ZLL is happy to support you in addressing generative AI in your teaching with its current range of workshops and other events. Additionally, you are also very welcome to contact us at any time for individual support (zll@uni-biefeld.de).

For an assessment regarding the use of generative AI tools in terms of examination law, please refer to the website of the Justiziariat Studium und Lehre (German).

If you as a lecturer are wondering whether you should adapt your examination format, we recommend that you first test some of your previous exam questions or essay topics using BIKI to get a feel for what LLMs can and cannot do in your specific case. If you realise that the scope of questions that BIKI can handle well is too large, there are several options: The most radical "solution" is to switch to exam forms in which generative AI tools cannot really be used, such as oral exams or on-site closed-book exams. However, especially in writing-intensive fields and courses in which, for example, the independent writing of term papers is one of the essential skills to be taught, this is hardly an option. Instead, consideration should be given to how the existing form of examination can be adapted. If your examination regulations and module description allow this and you are not already working in this way, it may be helpful, for example, to move elements of term paper writing (topic identification, outlining, peer discussions, etc.) into your course sessions during the semester. This will give you an insight into the different stages of your students' work and you will be in a better position to assess whether the final submission is consistent.

To make it easier for you to double-check citations, we also recommend (if this is possible for sources commonly cited in your field) that you insist that students list a DOI for each citation. This makes it much easier for you to check (via copy & paste or by clicking) whether the citation exists and whether its content corresponds to what is claimed.