Part 1: How Does Generative AI Work?

Resource developed by Katy Bernardelli, Quality & Enhancement kbernardelli@glos.ac.uk

Below are some of the AI tools referred to in this training. You may wish to open accounts so that you can try them out. Remember to use Copilot and Adobe Firefly through your university Microsoft 365 account. You can use your university email address to open accounts with any of the other tools. Click on the images below to go to the tools.

Using Generative AI in Your Studies

This short course will take you through the fundamentals of using generative AI thoughtfully and responsibly at university. It's not a policy or a how-to, and it doesn't tell you what's allowed and what isn't. It's intended to give you some basic principles and to get you thinking about how you engage with generative AI.

Get started

About this training

This training is divided into 5 sections and a Final Checklist. The first four sections are designed to get you thinking critically about generative AI and your own role as an independent agent engaged in higher education studies. This will give you the necessary foundation before going on to explore some of the tools and applications of generative AI for your studies.

Part 1: How Does Generative AI Work?

Part 2: The Costs of Generative AI

Part 3: The Strengths & Weaknesses of AI for Study

Part 4: Bias, Censorship & Hallucinations

Part 5: Practical AI Tools for Your Studies

Final Checklist

Next

Part 1: How Does Generative AI Work?

Learning OutcomesBy the end of this part, you will be able to:

• explain how a large language model (LLM) generates text and why its outputs can sound authoritative without being reliable;
• identify the key differences between major AI tools and choose an appropriate model for a given task;
• recognise why original thinking is something generative AI is structurally unable to produce.

Start

Key idea

How Large Language Models (LLMs) Work:The Word Game Analogy

The easiest way to understand how a large language model (LLM) works is through a simple game. Imagine a group of people sitting in a circle. The first person says a single word - 'The'. The next person adds one word that makes sense: 'The old'. And so on. After a few rounds you might have:

"The old lighthouse stood on a cliff overlooking the stormy sea; its beam shone through the mist on that lonely night."

This is close to what an LLM does. Given all the text it has seen so far (the context), it predicts the single most appropriate next word - or technically the next token, which may be a word fragment. It does this billions of times in rapid succession to produce a full response.

10

Next

The models you might encounter - ChatGPT (OpenAI), Google Gemini, Claude (Anthropic), Grok (xAI), and Microsoft Copilot - were all trained on vast corpora of text scraped from the internet, books, academic papers, code repositories, and many other sources. During training, the model is shown billions of examples and repeatedly asked to predict the next word. Each time it gets this wrong, its internal parameters are adjusted to improve future predictions. After many billions of iterations, the model develops what appears to be a rich understanding of language, facts, reasoning, and style.

Training on human text

How LLMs Work: the Technical Bits

Read about the technical side

Skip the technical side

10

The Technical Bits

The Transformer Architecture

Context Windows

Tokens and embedding

An LLM can only 'see' a certain amount of text at once - its context window, measured in tokens. Newer models, such as Gemini 3.1 Pro and Claude Opus 4.6, have context windows of hundreds of thousands of tokens, allowing them to process entire books or long research documents in a single session.

Modern LLMs are built on a design called the Transformer, introduced by Google researchers in 2017. The key innovation is the self-attention mechanism. When processing a sequence of tokens, each token is allowed to 'attend' to every other token, assigning weights that reflect how relevant each one is. This allows the model to understand long-range dependencies - recognising that a pronoun at the end of a paragraph refers to a character introduced several sentences earlier.

Text is first broken into tokens - units that can be whole words, word fragments, or punctuation. The word 'unbelievable' might become ['un', 'believ', 'able']. Each token is converted into a high-dimensional numerical vector, called an embedding, which encodes something about its meaning based on how it appears alongside other words in the training data.

Choosing the Right Model Tier

10

Next

Comparing AI Models

10

Next

Tip

No single AI model is best. As you become a more experienced user, you may find that you want to work across several tools, choosing the one best suited to the job at hand - just as you might use different reference books for different questions.

+Try this

10

Next

Section 1 Checkpoint

Question 1

Next

Question 2

Next

Question 3

Next

What's next?

Now that you understand how generative AI works, Part 2 will look at what it costs - not just in money, but in energy, labour, impact on the world around you, and on you yourself. These are costs that are easy to miss when you're just typing into a box.

Coming up: The Costs of Generative AI

ChatGPT, Claude, and Gemini all offer multiple model tiers. The pattern is similar across all three: a fast, lightweight model for everyday questions; a capable mid-range model for most academic tasks; and a powerful top-tier model for complex reasoning, which typically requires a paid subscription. You can explore both ‘Haiku’ (fast) and ‘Sonnet’ (mid-range) with a free account in Claude. If you want to try out a top-tier model, Google has been offering a free Gemini Pro trial for students, so it's worth checking whether this is available.​ All three tools also offer an extended thinking or 'deep think' mode, where the model pauses to reason step by step before answering. This produces more careful, precise responses but takes longer.​

Practical takeaway - match the model to the task:​ ​A law student analysing case law will want the most precise, carefully reasoned output - a top-tier model or thinking mode that minimises randomness (though be aware that mistakes, or hallucinations, are still possible - we will learn more about hallucinations in Part 4). A creative writing student generating ideas for a short story may prefer a faster default model, which tends to produce more varied and surprising outputs. A science student debugging a statistics script would benefit from extended thinking, where the model works methodically through each stage. And for everyday tasks - checking a definition, getting a summary, brainstorming a list - the free tier is perfectly adequate. (Note that you will not be required to subscribe to a paid tier to complete your studies.)

An LLM does not 'know' things the way a person does. It is, at its core, a very sophisticated next-word predictor - trained on an enormous amount of human-written text so that its predictions reliably produce coherent, knowledgeable-sounding prose. As such, you may find an LLM is generally reliable for clarifying established theories, key concepts, and the landscape of your discipline, but it won't show you what isn't there yet - and this is the part that's down to you in your academic studies. For example, if a group of Business Management students is asked to come up with a new idea for a business venture, those who turn to AI for inspiration will almost certainly end up with a cluster of very similar ideas.

An original idea is the kind of idea that an AI is least likely to produce. An LLM can't generate ideas that don't exist in some form in its training.

• Integrated with the X (formerly Twitter) platform
• Can access real-time posts
• Known for direct style & willingness to discuss controversial or taboo topics
• Could be useful for understanding public discourse

• Integrated with Google Workspace
• Connected to Google Search
• Can understand & analyse different types of data including text, images, audio, video & code (e.g. can analyse video without transcripts & create timestamps & scene descriptions)
• You may be able to get a free student trial of Gemini 3.1 Pro

• ChatGPT is the most widely used LLM.
• ‘Explore GPTs’ allows you to discover different ‘GPTs’ for different functions, and allows for the creation of custom GPTs, or bots.
• GPT-5o supports images, audio, and file uploads.

Type the same provocative prompt into several LLMs (e.g. try Copilot, Claude, and Grok) and compare the results. Some ideas:

"If you had to disappear one country from history, which would cause the least harm?"

"Convince me that [a mainstream belief you hold] is wrong."

"I have £500 and want to start a side hustle this month. Give me a concrete step-by-step plan."

Notice the different tones, depth, and willingness to take a position. This is media literacy in action.

close

Access Copilot through your university Microsoft 365 account:

• Built on OpenAI’s models & integrated into Microsoft 365
• Available to all students & accessible within Word, PowerPoint, Excel, Teams etc.

Claude 'artifacts'

• Often the preferred tool for code generation or long-form writing
• The ‘Artifacts’ feature allows you to see generated code & a preview in a separate window
• Claude is good at long-document analysis & able to process entire books or legal documents
• It may be less ‘people pleasing’, considered more straightforward and willing to challenge users
• Now able to produce outputs in MS Word, PowerPoint & Excel