AI - how does it work?

How are the Large Language Models trained and on what sources?

Large language models (LLMs) are trained on vast collections of human-written texts, known as training data. However, it is not always clear which texts have been used. This lack of transparency not only influences what the models are capable of producing, but also raises concerns around copyright and data protection. If you are interested in learning more about the ethical issues and copyright questions related to AI, you can find more information further down on this page.

Different AI tools have been trained on different types of data, and they also have access to different sources. Chatbots like ChatGPT and Copilot generate responses based on a wide range of material, not just scientific texts. In contrast, AI search tools such as Elicit, Consensus, and Scite have been designed to provide answers based on scientific publications, which they also cite. If you are looking for reliable scientific information, it is therefore best to use an AI tool that specializes in searching academic literature. You can read more about how to search for information using AI further down on this page. 

How do LLMs work? 

Large language models are highly skilled at imitating human communication. Their responses often sound confident, which has led to the term “machine splaining” or “AI splaining", when a model explains something that sounds true and confident, even if the information is incorrect. It is important to remember that these models have no concept of what they are producing. They simply generate text based on patterns in language. That is why they are sometimes called 'stochastic parrots' like parrots: they can repeat human-like speech, but without any understanding of its meaning.

Large language models are often connected to databases or search engines using a method called Retrieval-Augmented Generation (RAG). This technology allows them to include references to real sources in their responses. However, even though the sources they cite exist, the quality and reliability of those sources may vary.

AI tools may also summarise sources incorrectly. Since a large language model cannot understand text, it cannot really summarise text; it has no genuine understanding of what the text is about and what the main message is. When a model "summarises" texts, the result is usually just an abbreviation, where you cannot be sure that what has been removed was irrelevant and that what remains is the core of the text, or was even part of the original text.

Further, both the training data itself and the way large language models function may result in AI tools reproducing factual errors, stereotypes, or bias. The answers provided by the tools are also not reproducible; you will get different answers if you use the same prompt in the same tool again.

Below is a list of different AI tools based on the tasks they were intended to perform. Please note that the list is not exhaustive and that you are responsible for using the tools ethically, and in a manner appropriate to your task. You can read more about AI and ethics further down. Also, the tools work in different ways, so you need to familiarize yourself with how they work, how to protect your data, and how to avoid copyright infringement.

General purpose tools (so called chatbots/assistants)

• ChatGPT
• Microsoft Copilot
  At KI, both students and employees have access to Microsoft Copilot, which is a large language model similar to ChatGPT. If you use it when logged in via your KI login, the text you enter is protected.
• Gemini
• DeepSeek

Writing (specifically designed to improve text) 

• Instatext
• Grammarly
• For more tools, see Writing, reading and summarising with AI

Programming 

• Github Copilot

Image generating 

• Dall-E
• Firefly
• Picsart

Audio generating

• Notebook LM
• Soundraw  

Reading

• Notebook LM
• For more tools, see Writing, reading and summarising with AI

Searching for articles 

• Elicit
• SciSpace
• Perplexity
• For more tools, see Searching for information with AI

Transcribing

• Whisper Web
• Onenote
• Klang.ai

Translating  

• DeepL
• Google Translate

Coding

• NVivo 

Generating text, image and audio (so called multimodal models) 

• Notebook LM

Questions to ask yourself when choosing an AI tool: 

• Are you aware of the limitations of the tool? Do you know what the tool is intended to do?
• Who is behind the tool? (Country, company, training data). How might this affect the answers you receive?
• Do you know how the tool manages personal and other types of data?
• Will your data be protected?
• Will your data be used to train future language models?
• Are you using AI in a way that avoids unnecessary resource waste? 
   

Image-generating AI

One type of generative AI creates images. Depending on the training data it has learned from, the tool may be limited to certain kinds of images, such as portraits, or it may be able to produce a wide variety of images. Most tools use text prompts to generate images, but some can also work from a rough sketch. 

AI agents

Another type of generative AI is known as an AI agent or agentic AI. While the term lacks a universally accepted definition, it is commonly used to describe computer programs designed for specific tasks, often featuring a user interface powered by a large language model. These tasks may include booking tickets or conducting advanced searches. 

Classifier

Another type of AI is designed to determine whether a given input belongs to a specific category. For example, in medicine, classifiers can analyze X-ray images to identify those that show signs of lung cancer.  

Vector databases

Vector databases use a relatively simple method to conduct searches, based on a language model that groups words with similar meanings together. This makes the search result less dependent on the exact word used and resolves one frequent problem with search databases: you need to use exactly the right words. For instance, a document containing the word “fire” can only be found using the word “fire” and not its synonym, “blaze.” 

AI detectors

In many settings, it is not permitted to submit content generated by large language models. For example, schools often prohibit AI-written assignments, and scientific journals generally require full disclosure of any AI involvement. However, enforcing these rules is challenging, as it is not always possible to determine whether a text was written by AI. As a result, several companies have developed AI detectors that claim to identify AI-generated content. So far, none of the detectors have proven reliable enough to be considered effective.

Predictive AI

Predictive AI models try to predict real-world outcomes, particularly human behavior. Predictive AI has been used to forecast whether students will drop out of school, or whether convicted criminals will reoffend. However, a consistent problem with these predictive AI models is that they often generate incorrect predictions.