When we talk about artificial intelligence (AI) today, we often mean generative AI (genAI), for example chatbots like ChatGPT or Microsoft Copilot that can be used to generate text or images. As a student at Chalmers, you have access to the chatbot Copilot since it’s included in the Microsoft Office package. A chatbot generates seemingly original content after you have asked it a question or given it instructions (prompts). Sometimes you can upload files into the chatbot and make it generate something based on the contents of the files.
Generative AI is based on Large Language Models (LLMs), a type of deep learning that is based on artificial neural networks which means a set of complex algorithms. By feeding an LLM with large amounts of text and data, the LLM will learn which word sequence is more probable to create semantically correct sentences.
All responses and generated contents from a LLM are based on the statistics it found in the training data. Because of this, you cannot count on the responses from a chatbot to always be correct since they are built to produce probabilities, not truths. You can read more about this phenomenon here: Made-up facts and references.
Always make sure that all facts and claims from a chatbot match another credible source before including it in your work!
Figure 1. Example of a sentence that can be generated by a chatbot. "The sky is blue/grey/red today."
The first step when you interact with a chatbot is to write (or in some cases speak) a prompt. After the language model has interpreted your prompt, a response is produced based on the content of the model’s training data, that is, the statistical models. You can imagine a very advanced form of autocomplete. How likely is it that this word follows this word, and so on…
Figure 1 shows a simplified image of how a language model generates sentences. It is likely that the sentence “The sky is blue today” occurs often in the model’s training data, which means that it is this sentence that will be generated. If the most likely sentence according to the training data is “The sky is grey today” the chatbot will give that answer instead.
If you want to deepen your understanding of these concepts, further reading:
Not all chatbots are the same. You will get different answers when you prompt ChatGPT and Claude even though the prompt is exactly the same. Chatbots work thanks to their language models, sometimes called foundation models. Which model has been used, the content of the training data, and how it has been trained affect how the chatbot performs and what answers it provides. You need to choose the chatbot that best fits your current task.
There are many companies that provide AI services, such as OpenAI, Google, Meta, Anthropic and Microsoft. AI companies that create and train their own language models can incorporate these into their own chatbots and AI tools, for example, the models GPT-4.1, o4mini and o3 are available via ChatGPT today (2025-07-25).
It is not uncommon for the same language model to be found in several different tools if the AI companies have sold the rights, an example of this is that Microsoft Copilot uses GPT4o, among other things. There are also many AI models that can be used freely since the source code is openly available, for example via Hugging Face.