What is artificial intelligence?
Artificial intelligence refers to computer-based systems that technically mimic certain human cognitive and linguistic abilities. They can generate text, images or other content, but do not understand it themselves and do not generate their own knowledge. The use of AI systems also carries risks.
When it comes to studying and academic work, this means that AI can provide support, but it must not replace independent thinking, critical evaluation or personal responsibility.
Two basic types of AI models
In artificial intelligence, to put it simply, there are two basic types of model:
- Discriminative AI recognises patterns in existing data and classifies new inputs into known categories or values. Typical applications include spam filters and image recognition. These systems learn to recognise distinguishing features and make decisions based on existing classes.
- Generative AI goes one step further: it generates new content by combining patterns, structures and probabilities drawn from training data. This results in texts, images, music or videos that can appear to be human-created content.
Generative AI models have been ‘trained’ using large amounts of data. This means that they have identified patterns, structures and statistical correlations between words from vast amounts of data, which they then use to generate new data. Depending on the type of data that generative models produce, a distinction is made, for example, between language models, image models, audio models and video models.
Large Language Models (LLMs)
A widely used form of generative AI are large language models (LLMs) such as ChatGPT, which specialise in processing human language. They can generate text by calculating which word or phrase is most likely to follow next in a given context.
This means that an LLM has no consciousness and no genuine understanding of content. In essence, it merely answers this question: ‘Based on the probabilities I have learnt, what is the most likely continuation of this text?’
The model therefore does not access ‘meaning’ in the human sense, but only:
- statistical correlations between words
- Typical sentence structures
- Typical patterns of argumentation
- Typical answers to typical questions
So the model does not actually understand; rather, the patterns and probabilities it has learnt are so good that the result appears as if it were understanding.
No new knowledge is generated in the process: LLMs combine existing linguistic patterns to produce new texts without understanding the content, reflecting on it critically or conducting their own research. Their output is based solely on statistical correlations between words.
-
Bias and misinterpretations
AI models learn the probabilities on which their answers are based from their training data. If, in this data:
- certain perspectives are over-represented
- depict a particular group stereotype
- certain terms frequently appear in the same context
… then the model ‘adopts’ these patterns without realising that they are problematic.
Example:
If the training data often has 'professor' linked to 'male', the model is more likely to connect 'professor' with male forms or examples.
This means the results might not be reliable. AI-generated texts can seem realistic when it comes to language, but they might have factual errors or show biases from the training data. Language models can't tell if their statements are correct or biased because they can't understand content and don't have a body of knowledge to back them up. They've just learned patterns from a lot of text during training.
-
Hallucinations
Sometimes generative AI models produce text that sounds very plausible but is factually incorrect. It hallucinates. This happens because the model
- does not have a genuine, internally stored database
- does not distinguish between ‘sounds plausible’ and ‘is verifiably true’
It simply optimises: ‘What is the most linguistically probable response to this prompt?’
Example:
If you ask for a source or a quotation, and the model has seen many responses during training following the pattern: "The study by XY (2021) shows that …", then it will generate something along those lines, even if the specific study has never actually existed.
What does this mean for the use of AI in the context of university studies?
- Tools such as ChatGPT are neither research tools nor reliable sources of knowledge
- They can't substitute specialist literature and may also provide incorrect or inaccurate information
- They can continue, rephrase or summarise texts, but they can't guarantee that the content is factually accurate.
- There is a risk of plagiarism because AI tools can produce text that is almost identical to existing content. This means that the content may be reproduced without any changes or with only minor alterations, and the source of the content may not be cited.
Anything produced with the aid of AI must therefore be checked and backed up by suitable sources. Students are always personally responsible for their academic work – that is, for its content, arguments and findings.