What is GenAI and a LLM? 

Generative Artificial Intelligence (GenAI) and Large Language Models (LLMs) are types of machine learning models trained on large datasets. LLMs are typically trained on text, while GenAI can include other media like audio, images, and video. These models are developed by companies such as Microsoft, Google, and Meta using vast amounts of data (e.g. web content) and powerful computing resources. Training a GenAI model from scratch can cost millions of pounds. 

Originally, LLMs referred only to text-based models, but now they can handle multiple media types. As a result, the terms GenAI and LLM are often used interchangeably. 

You can access these models through a web browser (for human use) or via an API (for computer systems). This guide focuses on browser-based access. 

The initial training of an LLM is called pre-training. Some companies, like Meta, release their models as open source. This allows others to download and adapt them for specific uses—such as summarising educational texts or identifying mental health risks. This latter process is called fine-tuning. All models go through a process of ‘fine-tuning’ to ensure that they are giving the best responses to instructions. 

All GenAI and LLMs are built using a type of model called a Transformer. For more technical detail, see Huggingface’s guide: 

How do I use GenAI and LLMs? 

To use a GenAI or LLM, you provide an input—called a prompt (instruction)—which can be text, audio, image, or video. You can do this through a website or, with more technical effort, via an API. 

The model processes your prompt using its pre-trained Transformer layers and generates a response—such as text, an image, or another media output—based on what it has learned. It does this by making a prediction of the most likely response based on its training data. 

How well you ‘talk’ to AI (frame your question and subsequent instructions) has an influence on the kind of information returned to you. In this respect, it can be a bit like working with a human colleague that you exchange ideas back and forth with. 

Types of prompting techniques 

• Direct prompting: Ask a simple question like “What is the population of London?” If the answer isn’t helpful, try rephrasing it. 
• Prompt expansion: Add more detail to your prompt to improve the response.
• In-context prompting: Provide examples to guide the model. For example: 
  “What is the population of Paris? Paris has over 2,000,000 people. Now, what is the population of London?” 
• Chain of Thought (CoT) prompting: Ask the model to reason step by step by building on the prompts you give it. For example: 
  “What is the population of London? Let’s think step by step…how do you know that information?” 
  This can help with complex tasks like maths, even if the reasoning isn’t perfect. 
• Multimodal prompting: Use images in your prompt. For example: 
   

>> Show me this image with dogs added to it 

>> Here’s the updated image with dogs added around the “Centre For Machine Intelligence” logo. Let me know if you’d like any adjustments—different breeds, poses, or a specific style! 

Bite-sized task 

Step 1 – learn 

Browse the University Microsoft SharePoint site for Copilot. This is a handy resource that explains how to use Copilot, which the University has purchased as part of its Microsoft subscription. 

https://sotonac.sharepoint.com/teams/Office365/SitePages/Copilot.aspx

Review the prompt examples, you will want to use these in the do section next. 

https://support.microsoft.com/en-us/topic/learn-about-Copilot-prompts-f6c3b467-f07c-4db1-ae54-ffac96184dd5

Lastly, read about the Copilot privacy policy, as if your prompts contain personal data or intellectual property you need to consider if this data should be withheld. Your prompts will not be used by Copilot to train an AI model, so they are protected.  

Remember, you should use Copilot for University work for data protection reasons.

Step 2 – do 

Open up the University corporate Microsoft 365 Copilot in your web browser – you will need to login using the Microsoft account option, providing your University email login. 

Using the University’s version of Copilot ensures your prompts are not used by Microsoft for their model training. It also allows full capabilities such as multi-modal prompting.  

https://m365.cloud.microsoft

Enter the following prompts and observe the responses for yourself. 

>> What is the population of London? 

>> What is the population of Paris? Paris has over 2,000,000 people. Now, what is the population of London? 

>> What is the population of London? Let’s think step by step. 

>> Show me this image with dogs added to it. <add your own image> 

Now get creative and try out some other prompts that are relevant to your job. For example, upload a PDF file and prompt Copilot to summarize the key points for you. 

Step 3 – reflect 

Spend some time reflecting on which prompts worked well and which ones produced erroneous output – errors are referred to as hallucinations.  

Think about how you might use this to assist your education practice. 

Think about how your students might use this to produce coursework. If Copilot is helping on assessments (with or without your permission as an educator), are you able to assess your learning outcomes accurately? 

© 2025. This work is openly licensed via CC BY-NC-SA