1  Intro to using AI

There are a range of AIs and an even larger number of niche apps and platforms but we’re going to focus on Microsoft’s Copilot signed in via your UofG account. The reason we’ve chosen Copilot is that all University of Glasgow students have access to the same plan and functionality. The UofG Copilot also has enhanced security, which is particularly important when you’re considering using AI to help with data analysis or uploading teaching materials, which are copyright protected.

Since 2024, generative AI tools have become embedded in most productivity software (Word, Excel, PowerPoint, Teams, Google Workspace, Adobe Creative Suite). Whilst this book uses Copilot, you should expect to see AI features in many of the tools you already use.

1.1 A brief introduction

1.1.1 AI model vs AI platforms

There are more qualified people who can explain AI better than I, but a brief but important distinction is between the AI model and the AI platform. The underlying AI model is a Large Language Model (LLM) that essentially does extremely sophisticated predictive text - it’s not thinking, it’s not conscious, it just is very good at deciding what word is likely to come after another word in a way that is human-like.

Different software companies have created different AI models. OpenAI has a series of models it refers to as Generative Pre-trained Transformers (GPTs) and the explosion in public awareness of AI came about when it released ChatGPT in November 2022. Other widely used model families include Claude (Anthropic), Gemini (Google/DeepMind), and Llama (Meta).

Each company now releases a family of models rather than a single one, and the naming is confusing. Within a family there are usually tiers: a large, slow, expensive model at the top and smaller, faster, cheaper ones below it. At the time of writing (September 2026), OpenAI’s top model is GPT-6 Astra, with Sol and Luna as its cheaper tiers, and Anthropic’s line-up runs from Fable 5.1 at the top down through Opus, Sonnet and Haiku. The version numbers go up every few months and each new one is claimed to be significantly better, both at passing standardised tests and at sounding human. Any version number printed in a book will be out of date by the time you read it: at the time of writing, UofG Copilot runs GPT 5.6, which is already a generation behind what ChatGPT offers. Different platforms may also give slightly different results because they are tuned differently.

Separately from which model you’re using, most platforms now let you choose how much “thinking” the model does before it answers. In Copilot this is the Auto, Quick response and Think deeper selector at the top left of the chat. Thinking here means the model generates a lot of intermediate text to work through the problem before it writes its reply. It’s still predictive text, just more of it, but it makes a real difference on anything with several steps, like code. Quick response is fine for a definition; use Think deeper for anything you would want a human to check twice.

These days, most current models are multimodal by default: they can process text, images, and in some cases audio and video. Copilot inherits some of this, for example analysing images or diagrams when pasted in.

In some cases, the name of the model and the name of the platform you use are the same. ChatGPT is the name of the platform, but the underlying model it uses is a GPT. Additionally, other companies and apps can licence use of the models so Microsoft’s Copilot also uses OpenAI’s GPT models, which means it will behave much like ChatGPT with the same model underneath.

It isn’t necessary that you understand all of this in detail. The key thing is that the underlying model is the same for some platforms (so you should expect similar capabilities) and that this model can and will update.

1.1.2 Ethics and AI

It’s beyond the scope of this book (and my expertise) to go into much detail about the ethics of AI but I want to acknowledge that in several ways, they’re hugely problematic.

First, and most importantly, AIs are trained on huge corpora of human language. Humans are full of bias and prejudice and consequently, so are our offspring AIs who have been repeatedly found to encode sexist, racist, and ableist views in their outputs.

• Teaching AI Ethics: Bias and Discrimination

Second, many of the developers that have produced AI models have done so behind closed doors and are very secretive about exactly what data the models have been trained on. One of the reasons for this is because it is very clearly apparent that many AI models have been given access to copyrighted sources (e.g., books and film scripts) and there are an increasing number of lawsuits being filed although it remains unclear what the future holds given how far and how fast the horse has already bolted.

• AI Infringement Case Updates

Third, whilst the use of AI has been touted as a solution to help us manage climate change, it’s also a direct contributor to the problem because of the energy demands.

• Explained: Generative AI’s environmental impact

Fourth, training and moderation often rely on hidden human labour. Low-paid workers, often in the Global South, are employed to filter harmful content or annotate data. This raises ethical questions about labour exploitation.

• How We Investigated the Human Labor Behind AI

Regulation is catching up, slowly. The EU AI Act came into force in stages from 2025 (bans on some uses first, then obligations on general-purpose models), but in July 2026 the EU pushed the deadlines for “high-risk” systems back to December 2027 and August 2028. The UK has no equivalent law and is relying on existing regulators. This means that how we are expected to use AI responsibly may change rapidly. AI Now 2025 Report provides an annual overview of social, ethical, and political issues in AI.

1.2 Criticality and AI

When using AI, it is not enough to know how to operate the tools, you also need to approach them critically. That means asking not just “what can this do for me?” but also “how trustworthy is this output, and what are the wider implications of relying on it?”

Here are some of the things that should always be in the back of your mind when using AI.

1.2.1 Accuracy

AI tools are trained to generate fluent and convincing text, but this does not mean it is correct. They frequently “hallucinate”, producing confident but false information such as incorrect references, fabricated statistics, or misattributed quotes. You should treat outputs as drafts or starting points, not final answers and verify with authoritative sources. References are the clearest case: an AI will produce citations that look real, with plausible authors, journals and years, and some proportion of them will not exist or will not say what it claims. Don’t ask it for references unless you already know how to verify every one of them against the original source (I do use AI for this, but with a tool I built specifically to check each DOI, and I still find errors). If you’re not at that stage, use the reading list and the library.

1.2.2 Authority and Expertise

AI cannot tell the difference between high-quality peer-reviewed research and a random blogpost unless explicitly instructed. It mirrors patterns in its training data rather than exercising judgement and a huge proportion of that training data is Reddit. Always consider who should be the authority on a question (a textbook, journal article, or your lecturer) and cross-check AI outputs accordingly.

1.2.3 Transparency

You cannot see the training data or the “reasoning” steps of the model. That lack of transparency makes it hard to know why it gave a particular answer or whether it is missing key perspectives. Ask yourself what might have been left out or under-represented, and be cautious about treating AI answers as comprehensive.

1.2.4 Epistemic Caution

Epistemic caution means being careful about what you claim to know. AIs generate text that looks authoritative, which can give a false sense of understanding. This risks shallow learning if students accept outputs uncritically rather than grappling with concepts themselves - do you really know it or can you just repeat what the computer told you? Use AI as a “thinking partner,” but make sure you can explain ideas in your own words without relying on its phrasing.

1.2.5 Equity and Access

Even though UofG provides Copilot securely, in wider society access to premium AI tools is uneven. This creates new digital divides between those who can afford subscription services and those who cannot, or between those trained to use AI critically and those left behind. Reflect on whose voices are privileged or excluded in AI-mediated work.

1.2.6 Over-reliance and deskilling

Outsourcing too much to AI can erode your own skills in writing, problem-solving, or coding. This is cognitive offloading. Think of it like going to the gym: getting someone else to carry your kit bag is fine, getting them to lift the weights is not, because lifting the weights was the point. It is not hypothetical. In a study of nearly a thousand secondary school maths students, those given a standard ChatGPT-style tool did better on practice problems whilst they had it and then did worse than students who never had it once it was taken away for the exam. Students given a version that was set up to tutor rather than answer didn’t show that drop (Bastani et al., 2025). The difference wasn’t the model, it was whether the tool was set up to make the student do the thinking. That is what the rest of this book is about: you setting up Copilot so that it behaves like the tutor version rather than the answer version. Use AI strategically, but keep developing your own capacity, especially in core academic skills such as argumentation, critical reading, and data analysis.

There is a psychological cost as well as an academic one. Confidence in your own ability (self-efficacy) is built mainly through mastery experiences: every time you attempt an answer first and then use AI to check and correct it, you accumulate evidence that you can do it. Every time you skip straight to the answer, you don’t. Throughout this book we’re going to ask you to reflect on how your use of AI might affect you psychologically, because it’s not all about academic integrity and cheating; there are serious risks to your long-term success and resilience.

1.2.7 Copilot

Microsoft Copilot is not one single tool but a family of services (e.g., in Word, Excel, Edge, Outlook). In this book, we focus on Copilot Chat accessed via Office365 using your University account, because this is the secure, enterprise version.

Microsoft Copilot can be accessed by signing into Copilot Chat (or via the app launcher in Office365) with your University of Glasgow student account. Whilst ChatGPT is the most widely known and used, the advantage of using your Enterprise account is that the data is processed more securely and what you input won’t be used to train the model.

This is really, really important and if you’re found to have uploaded sensitive data to an unsecure AI you can be accused of academic and/or research misconduct. So if you’re using AI for any University work, use Copilot. The University’s guidance on using AI in your studies and assessments is here.

You can check that you’re in the protected version: there is a green shield icon in the top right of the chat which, if you hover over it, says “Enterprise data protection applies to this chat”, and your name in the bottom left should have a “Work” label next to it. If you don’t see either of those, you’re not signed in with your University account.

WarningThe elephant in the room

Copilot Chat is the tool this book uses because it is the one the University has an agreement with, not because it is the best one. Since 2025 the biggest change in AI has not been the models, it has been tools that work on your computer rather than in a chat window. Claude, ChatGPT and Copilot all now have versions that can read the files in a folder, run code, and write the results back, and for anyone who works with data they are far more useful than pasting into a chat.

That doesn’t change the rule, but it does make clear what the rule is actually about. The question was never “chat window or not”, it is “where does this material go, and who has agreed to what”. Being frustrated that the sanctioned tool is the worst one is a reasonable position. Uploading someone else’s data or teaching materials to a better tool because the sanctioned one is annoying is not.

1.2.8 Teaching materials are copyrighted

The same logic applies to teaching materials. Lecture slides, lecture recordings, lab handouts, reading lists, exam papers, and the books and articles on your reading list are all copyright protected. The University and your lecturers own the teaching materials, and publishers and authors own the readings. When you upload any of these to an AI platform, you are copying someone else’s work into a system you don’t control, and you don’t have the right to do that without their permission. If the platform also uses your uploads for training, the work is absorbed into a commercial product on top. Uploading a lecture recording to a free AI transcription or summarising app is the most common version of this we see and it is exactly the same problem as uploading a dataset.

There are three rules:

1. Only ever use your UofG Copilot for teaching materials. This isn’t because Copilot is magically exempt from copyright; it’s because the University has an agreement with Microsoft that covers University material being processed there, and doesn’t use what you upload to train the model. No such agreement exists with ChatGPT, Gemini, NotebookLM, or any of the many “summarise my lecture” apps, so never use them with University material.
2. Check that you have permission. Some lecturers and courses are happy for you to upload their slides to Copilot, some are not, and some materials (e.g., recordings that include other students, copyrighted textbook chapters) you should never upload anywhere. If your course handbook doesn’t say, ask. If you’re enrolled on Applied Data Skills, you have our permission to upload course materials to UofG Copilot, but not to any other platform.
3. Recordings are for you. Lecture recordings are provided for your own study. Sharing them or uploading them to other platforms is not what they’re for, and that includes running them through a third-party app to get a transcript or summary. If you need transcripts or other alternative formats as part of your access arrangements, use the University’s own tools (the recording platform has captions and a transcript built in) or talk to the Disability Service; that is exactly what those adjustments are for. Check the University’s policy on recorded teaching if you’re unsure.

Whilst we’re on the topic, uploading the reading to get a summary doesn’t just have a copyright problem, it has a learning problem. The summary is the bit your brain was supposed to do.

1.3 Check your understanding

What is the relationship between ChatGPT and UofG Copilot?

They are different platforms that can run the same underlying model, so you should expect similar capabilities but not identical answersCopilot is a cut-down version of ChatGPT that Microsoft has made deliberately less capableThey use completely different models, so what you learn about one doesn't apply to the otherChatGPT is an AI and Copilot is a search engine

An AI gives you a reference with a plausible author, journal and year. What should you assume?

It's real: an AI wouldn't invent an author, journal and year that specificIt's real if you click Regenerate and it gives you the same reference againIt may not exist, or may not say what the AI claims it says, so check it against the original source before you use itIt's real if the DOI looks correctly formatted

How can you tell you’re in the protected version of Copilot?

The model selector says AutoThere is a green shield icon in the top right and a Work label next to your nameThe web address starts with copilot.microsoft.comYou had to enter a password to log in

You upload a lecture recording to a free app that gives you a transcript. Is this ok?

Fine, because the recording is for your own study and you're not sharing itFine, as long as you delete the transcript afterwardsNot fine: the recording is copyrighted teaching material and the app has no agreement with the University covering itFine, because the recording is already online

When should you switch from Auto to Think deeper?

A quick definition of a term you've forgottenAnything with several steps, such as checking your codeEvery prompt, because slower is always betterOnly when Auto gives you an error

1.4 Login

NoteActivity 1

Login to Copilot using your University of Glasgow account.

Once you’re logged in, find the model selector at the top left of the chat window (described above). It defaults to Auto (“Decides how long to think”); you can change this to Quick response or Think deeper, or open the GPT sub-menu to pick a specific model (at the time of writing, September 2026, this is GPT 5.6, in a Think deeper and a Quick response version). For most of the activities in this book Auto is fine. For anything involving your code (chapters 4 to 6) use Think deeper.

Remember that “Copilot” is a family of tools. The link above takes you to Copilot Chat, which is the safest environment for coursework. Copilot in Word/Excel/PowerPoint may look slightly different but uses the same underlying model.

1.5 Prompting

The key difference between using an AI and a regular search engine is that you can shape the output of the AI by “prompt engineering”, which is the phrase used to describe the art of asking the question in a way that gives you the answer you want. It’s easier to show than tell.

NoteActivity 2

Think of a book, TV show, film, or video game that you know really well. I chose Final Fantasy VII (the original PlayStation version). Ask Copilot to give you three different summaries for different audiences.

1. Ask it for a summary without any additional context e.g., “Give me a summary of Final Fantasy VII (original playstation version)”

2. Then, ask it for a summary but give it a steer on the intended audience, e.g., “Give me a summary of Final Fantasy VII (original playstation version) for someone who knows nothing about videogames” or “Give me a summary of Final Fantasy VII (original playstation version) for an expert gamer who has played it many times”

3. Finally, ask it for a summary, but give it a steer on how it should act, e.g., “Act as an expert videogame reviewer for the Guardian. Give me a summary of Final Fantasy VII (original playstation version) for an expert gamer who has played it many times” or “Act as someone who spends too much time on Reddit and thinks they know everything. Give me a summary of Final Fantasy VII (original playstation version) for an expert gamer who has played it many times”.

Reflect on the differences in the information and detail it gives you for each one and how you’re able to evaluate the accuracy of the summaries because of your existing expert knowledge.

1.6 Copilot features

Whilst many AIs use the same underlying model, the platforms can differ in their functionality. The interface gets rearranged every few months so this book deliberately doesn’t use screenshots. Here is where things live as of September 2026; if something has moved, look in the nearest menu.

In the message box

1. The + button on the left of the message box lets you Upload images and files from your computer, or Attach cloud files from your OneDrive. Because you’re signed in using your UofG account it will automatically connect to your OneDrive. Even though Copilot connects securely to OneDrive, check the copyright section above before uploading slides, recordings, readings, datasets, or assignments. The same menu has Add capabilities (currently just image generation) and Chat with an agent (see below).
2. The microphone icon on the right is dictate, which allows you to speak your prompt to the AI. On a laptop you may have to ensure your microphone settings allow access. This is often easier on a phone. It’s worth trying even if it feels odd, because it’s a good tactic: when people speak to an AI they say far more than they would type, explaining the context, what they’ve already tried, and what they actually want. A typed prompt is usually a sentence; a spoken one is usually a paragraph, and the more detail you give it, the closer the answer gets to what you intended.

Under each response

3. Copy copies the response so you can paste it into e.g., a Word document. The … menu next to it has Share response, Edit in Pages and Export to > Word. Any of these are useful if you’re ever asked to provide more information on how you have used AI at University, because you can keep a record of the prompt and the response and be transparent about it. Export to Word is the easiest way to keep a permanent copy.
4. Read aloud (also in the … menu) will get Copilot to read its response to you.
5. Regenerate (the circular arrow) asks for a different answer to the same prompt. It’s worth clicking a couple of times on a factual question, because if you get three different answers, at least two of them are wrong.
6. Suggested follow-up prompts appear as buttons under the response and give you ideas for further prompting (although I have never found these useful, and in chapter 3 we’ll see that they can actively get in the way).
7. Edit in Pages opens an editable page so you can add to and edit the response. You can also then ask an additional prompt and add new responses and share with others. We’re not really going to use this feature but if you want to know more, Microsoft have additional info and training.

In the sidebar

8. New chat, Search (of your previous chats) and Library (files and images you’ve used) are self-explanatory. Every chat you’ve had is listed below them, so you can go back to one and carry on.
9. Agents are versions of Copilot set up for a particular job. At the time of writing there is a Study and Learn agent, which we’ll use in chapters 2 and 3.

Settings

10. The … in the top right opens Chat settings. The Personalisation section is where you set up custom instructions (chapter 2) and control what Copilot remembers about you.

NoteActivity 3

Try out each of the features described above and reflect on which ones you find useful and why. If something isn’t where the book says it is, look for it in the nearest menu before giving up; Microsoft move things around a lot.

1.7 Check your understanding (2)

You ask Copilot a factual question, click Regenerate twice, and get three different answers. What should you conclude?

At least two of them are wrong, so none can be trusted without checkingThe most detailed one is the correct oneThe one it gives most often is the correct oneIt has learned from the earlier attempts, so the last one is the most accurate

Which of these prompts is most likely to get a useful answer?

Explain t-testsExplain t-tests. Keep it short.Act as a statistics tutor for a first year psychology student who has covered the normal distribution but not hypothesis testing. Explain what a t-test is for, in under 200 words, and then ask me a question to check I've understood.What is the best explanation of a t-test?

TipKey takeaways

Reflect on the differences between the summaries - the technicality of the language, the accuracy and nuance of the information, what it chose to focus on, and what it chose to omit. The reason I asked you to create a summary of something you already know well is that you’re aware of where the gaps are. Remember this when you’re asking an AI something you don’t know as well. Just because you can’t see the gaps, inaccuracies, and biases, doesn’t mean they’re not there.

• Tell it how to act
• Tell it who you are
• Give it detail. If you find typing it all out tedious, dictate it
• Always check for gaps, errors, or bias
• Fluency ≠ truth and persuasive text may still be wrong
• Authority matters! Ask yourself “who should be the expert here?”
• Over-reliance risks deskilling. Keep practising your own academic skills