AI tools today often use so-called large language models (LLMs).
8. 1 How it (does not) work
Language models do not operate on fixed rules (they are not “smart calculators”). Instead, they are trained on large text datasets, which may include content from websites, books, encyclopaedias, or discussion forums. During training, the model learns to recognise patterns in language, such as which words and sentences typically follow one another, what style fits a particular situation, or how the argumentative structure of an academic text looks. The model does not store entire texts; instead, it creates an internal statistical representation of language – a probabilistic model of what is most likely to come next in a given context.
Once pre-trained, the model is fine-tuned (e.g., using human feedback) and equipped with a system prompt that defines its role, behaviour, and boundaries. It can also be adapted for specific tasks such as conversation, text summarisation, code generation, or literature search. Some models also have access to the internet, databases, or other tools.
The user then interacts with this “ready-made” model in natural language via user prompts. The model responds by generating text that it predicts to be the most likely continuation in the given context.
8. 2 Bias
The output of a language model is influenced by a combination of several factors: the patterns it learned during pre-training (from text datasets), the way it was fine-tuned, the system prompt, additional instructions from the provider or third parties, and, of course, the user prompt, the current context, and previous conversation history.
Each of these factors, individually and in combination, can lead to bias. The nature of the training data may unintentionally reinforce stereotypes. The system prompt or extra instructions may shape the style of responses or favour certain worldviews or products. And the user interaction itself can create a filter bubble effect.
For the average user, the source of a particular bias is often non-transparent, which increases the need for caution – not only when interpreting AI outputs, but also when deciding what to submit to the AI in the first place.
It is also important to remember that, like many other services, AI tools may store and analyse the inputs you provide. For this reason, the Rules and Recommendations for the Use of Artificial Intelligence at VSB-TUO (TUO_LEG_24_002) explicitly state:
„AI tools should not be used to process sensitive data, such as personal information or data included in contractual agreements.”
8. 3 User Moderation and Critical Review
Almost all AI tools now include some mechanisms to reduce bias, in various forms: from source transparency (links, citations), to feedback options (rating answers), to system-level interventions such as content filters or automatic moderation. However, this does not mean that users can rely on AI-generated outputs without further scrutiny.
AI-generated outputs can be regarded, to some extent, as a form of grey literature – with the important difference that they are produced not by human reasoning, but by statistical computation, whose sources and logic are often opaque. This calls for extra caution when working with such material.
How to use AI tools while avoiding hidden bias?
- Well-thought-out prompts and their variations
The way a task is formulated has a major influence on output quality. The more targeted the prompt, the more accurate (and useful) the response tends to be. On the other hand, even small changes in wording, tone, or structure can produce very different answers, sometimes leading to emergent behaviour that seems surprising or illogical. Helpful strategies include:
- trying different variations of the same query,
- hanging the language (e.g., Czech, English, Spanish),
- developing the conversation further, asking follow-up questions or clarifying.
- Critical evaluation and diversification of sources
Just like with traditional search, the golden rule applies here too: never rely on a single source, single model, or single tool. Always cross-check:
- responses from other models,
- information on the web (web search engines),
- archived versions of websites (web archives),
- scientific perspectives (specialized search engines).
These checks should be supported by critical evaluation, common sense, and awareness of the broader context.
- „Why do I think what I think?“
The rise of AI tools makes it even more important to think critically about our own thinking. In the flood of information (and generated content), it is easy to absorb opinions subconsciously until they quietly become part of our worldview. It is worth pausing now and then to ask:
- „Why do I think what I think?”
- „Do I really think this myself?”
- „Why do I believe this is true?”
- „What can I actually verify?”
- „And how can I even think this at all?”
A Few Useful Tools (Not Only) for Literature Search
Keeping up with the rapidly evolving field of large language models is challenging. Below is a short selection of tools, (re)trained for specific tasks:
Scite: Although not primarily designed for beginners, Scite can help verify the quality of scholarly publicationsby clearly showing citations in context (supporting, disputing, or neutral, along with their position in the sentence). For inspiration, you can also try Scite Assistant. Available at VSB-TUO through our subscription.
Google NotebookLM can summarise the content of uploaded files, turn them into study guides, outlines, or even podcasts, and answer questions about their content. Although VSB-TUO does not currently subscribe to NotebookLM, the free version offers a generous token limit and many features useful for study and research.
Writefull Premium s a tool for language editing of academic texts in English. It helps when writing scholarly articles, theses, or abstracts by suggesting improvements to style, grammar, and scientific expression. It uses a combination of proprietary language models and third-party models. Available at VSB-TUO through our subscription.
GPTZero can estimate the likelihood that a text was generated by AI. Offers up to 10,000 words per month for free. For more frequent use, consider Originality.AI. Both are specialised AI tools (not conversational LLMs) that use algorithms and statistical models to estimate whether a text was AI-generated.