The art of talking to AI

An AI assistant can be incredibly powerful, but also terribly frustrating. If you ask it to change something small, it takes and updates the entire text. Sometimes when you ask for a source, it turns out that the source you get is completely wrong. But perhaps worst of all is that it is always so overly kind 😅

Örebro University recommends that you as a student use the AI assistant Copilot in your studies. The service is procured by the University and is free of charge for anyone who has an ORU account. When you are logged in to Copilot with your ORU account, your communication is protected through the University's agreement with Microsoft. No information leaves the University environment and your prompts are not used to train the language models that Copilot is based on. Read more about how to use Copilot here! 

These problems are not because these AI services are bad, but it depends on how they work, and if you understand how they work, you can also avoid most of the pitfalls and get much better results. This guide is about just that. How to communicate with your AI assistant in the most efficient and rewarding way possible.

In this resource, we'll look at:

  1. Why does AI do this? Three common problems and why they occur.
  2. So how do you think? A mindset that makes the use of AI more rewarding.
  3. Concrete tips and methods. How to get started and start using AI right away.

Alright, let's go 🙂

Before we get into how to actually use AI, let's look at three situations that you may have already experienced and provide explanations why. This will help us understand how to make the AI behave the way we want it to later on.

"It changes my entire text"

You've written a text and just want to tweak one thing. Maybe you want it to sound a little more academic, or for a paragraph to be shorter. You therefore give the text to an AI assistant and ask:

Can you shorten the second paragraph?

And you get back something completely different. The style of the text, your wording, your structure. Not just on the second paragraph, but on the entire text. This is one of the most common problems. Perhaps the most common frustration among those who use AI.

Why is this happening? A generative AI chatbot is powered by a type of AI model called a language model. It has learned patterns by analyzing huge amounts of text.

What it does basically is surprisingly simple: it tries to figure out what the next appropriate word in a sentence should be. The chatbot then adds a layer on top: it has been instructed to behave like a helpful assistant. So, it generates text that a helpful assistant would type in response to your particular message.

This has a rather interesting consequence: An AI assistant, in its natural state, can only write text from start to finish, based on its principle of typing the next appropriate word. It can't back down, it can't pick out a specific paragraph, but it has to rewrite your entire text. And since it guesses every word it writes, there's always a chance that it will be wrong. If it happens that it gets a word wrong, then there is no turning back, it cannot change its mind and then the rest of the text will be different from how it was before.

However, there is a solution! Today's AI assistants can make use of tools that allow AI to edit pieces of text. I highly recommend looking into this and you can read more about it in my guide on AI features.

Insight: AI is really bad at following requests for precise wording, especially when it comes to longer texts. It can remember concepts and messages, but when it comes to repeating a text word for word, it's basically impossible. However, there are tools in some AI assistants that enable this.

"It makes up sources that don't exist"

You ask AI for academic sources for a task. You get nice citations, authors, years, journal, volume. Everything looks absolutely correct. But when you search for the articles: they don't exist.

Why is this happening? AI generates text that looks like a source. It has learned the pattern of what an academic reference usually looks like, real research names, credible journals, correct format. But it doesn't have access to a database of real articles, all it tries to do is guess what would have been an appropriate thing to refer to according to what it way before.

Insight: AI doesn't know when it's wrong. It delivers everything with the same confidence, whether it's true or not. You must always be the one who reviews. However, there are AI tools that help with this process, but even here you need to be in control.

"It always agrees with me"

You ask the AI if your interpretation of a concept is correct. It says "Absolutely, that's right!" but your interpretation was wrong. Or: you question the AI's own response, and it changes position immediately. "You're absolutely right, I should have said..."

Why is this happening? The AI is trained and instructed to be accommodating. This means that it has a strong tendency to validate what you say, rather than challenge you. If you look under the hood, there is a kind of instruction to the AI assistant, a so-called system prompt, which basically says "You are a helpful assistant". All the help you then receive is based on the perspective that the way it answers you should be helpful. And according to the training data it has seen (basically all text on the internet), it is relatively rare that the "helpful" person is critical or questioning.

Insight: If you want honest feedback, criticism, or questioning, you need to actively ask for it. This marks that you want that kind of help.

All three examples above point to the same thing: AI is like a committed helper who wants to make you happy, but has no understanding of your situation. It doesn't know if it's right, it doesn't know what you really want, and it doesn't know when it's doing it wrong. The only thing it does is generate the most likely continuation.

And that's why the way you work with AI determines whether the outcome will be good or bad.

If AI is the helper, then you are the project manager. It is you who decides what should be done, how it should be done, and whether the result is good enough. AI can perform, but you are in control.

This may sound difficult, but it really comes down to three things:

1. Define what "helpful" means

AI guesses what helpful means if you don't tell it. Imagine a prompt like "Make my text better". In its current formulation, this becomes a guessing game for the AI. However, you can write "Shorten my sentences and make the tone more academic, without changing the content".

💡Defining exactly what you need help with can sometimes be difficult, because sometimes you don't really know yourself. But here you can actually ask the AI for help, e.g.: "I want to make this paragraph better, give suggestions for different concrete types of improvements without changing anything."

2. Give it context

The AI responds based on the information it has. Without context about your course, your material, your level, the answer will be generic. With the right context, it can adapt to your particular situation.

💡This could mean uploading your course materials, telling them what course you're studying, or giving it your notes. And then there's my absolute favorite trick: ask the AI to ask you questions so that it can better understand your situation.

3. Work in steps

AI rarely solves everything perfectly on the first try. Often, you need to divide the work, iterate, give feedback, and steer. A good rule of thumb is:

~70% preparation and coordination → ~10% execution → ~20% refinement

💡It can feel cumbersome - but that's what actually gives results. And the good thing is: you can use AI in the planning phase as well. Ask it to help you come up with a plan.

Now you know why AI behaves the way it does, and what mindset you need. Here are 10 practical ones; concrete tips that show what it looks like in reality.

Tip 1: Let the AI interview you

If you don't know exactly what you need help with – ask the AI to ask questions instead. This is one of the most underrated methods.

✅ I need help studying for my exam in [course name]. Ask questions to understand my situation better before we start.

Instead of you having to figure out everything necessary to inform your AI assistant about yourself, you let the AI map out your needs through a conversation. Imagine it as a first meeting with a private tutor. They need to understand where you are before they can help you.

Tip 2: Break down work into steps

Don't ask AI to do everything at once. Start with a plan, then execute step by step.

❌ Write a complete summary of the entire course with all concepts, priorities, and study recommendations.

✅ We will map out what I need to understand before the exam. Start by listing the overall areas. We then take one area at a time.

AI gets confused if it has to keep track of too much at once. Divide, you coordinate, AI executes.

Tip 3: Ask for perspective instead of "the best answer"

Instead of asking the AI what it thinks, ask it to list multiple perspectives. Then you avoid the yes-saying behavior and get broader material to think from.

❌ What do you think of my argument?

✅ List three different perspectives on my argument; one that strengthens it, one that questions it, and one that suggests an alternative approach.

In the end, it is you who should think and make the decisions. AI gives you material to think with.

Tip 4: Give AI an alternative

When you ask the AI for a task, it will want to complete it. If you ask for a source, it will want to give you a source. If you want feedback, it will give feedback, even if your text is actually really good. You can therefore give AI a way out, an "escape road", to complete the task without difficult consequences.

❌ Find a source

✅ Try to find a source that supports the claim. If you can't find it, let me know.

❌ Give me feedback

✅ Give me feedback, and tell each point how necessary it is to fix.

In this way, we can tap into one of AI's greatest strengths without blindly relying on its responses.

Tip 5: Ask for criticism - not confirmation

If you want feedback, explicitly say that you want critical feedback. AI's default behavior is to be positive and affirmative.

❌ What do you think of my text?

✅ Review my text and give me three things that can be improved. Be honest and concrete.

Tip 6: Use feedback wisely

Feedback should never be seen as truth. It's up to you to decide whether or not to take on board the feedback. This is true regardless of whether it comes from a human or an AI. Therefore, make sure that you agree with the feedback or that the AI has a very strong argument in case you choose to take it.

Tip 7: Specify what AI should answer for each point

By asking the AI to respond to multiple aspects per point, you force it to be more structured and thoughtful.

❌ List key concepts in the course.

✅ List key concepts. For each concept: give a brief explanation, specify the priority (high/medium/low), justify the priority, and indicate which source is most relevant.

This means that you get better answers, and you can guide the way AI reflects - by deciding which aspects are important to think through.

Tip 8: Save important insights in separate documents

When AI generates something valuable, a mapping, a good analysis, a summary, it is important to copy and save it in a separate document. AI forgets, but you don't have to. This way, you can easily start new chat conversations by pasting these summaries or share insights you gained without having to scroll through long conversations. You'll thank yourself if you want to reuse the material a few days later.

Tip 9: Separate instruction and text with quotation marks

When you paste your own text into a prompt – clearly mark where your instruction ends and the text begins. Use one quotation mark (") or better yet, two quotation marks (""), before and after the text that you paste text.

Shorten the following text and make it more academic. Do not modify the content.

"""
[your pasted text here]


Without this marking, AI can confuse what is your instruction and what is material.

Tip 10: Give the AI your material context

One of the powerful things you can do with AI. Upload or paste relevant material, syllabus, notes, old exams, assessment criteria. It makes a huge difference in the quality of the responses you get.

Without context, AI responds based on general knowledge. With your specific material, it can adapt to your particular course and needs.

💡 There are different ways to handle documents and materials. The easiest is to enter directly into the chat, but then the AI will forget about it after a while.

Summary

What I hope you take away from this guide:

1. Understand why AI behaves the way it does: It guesses the most likely answer, it doesn't know what you want, it doesn't know when it's wrong, and it's always trying to be accommodating.

2. Define what "helpful" means: The clearer you are about what you want, the better the results.

3. Give the AI context: Material, background, your level — everything makes a difference.

4. Work in steps: You are the coordinator. Plan, divide, iterate.

5. Be critical: Always review the results. Ask for criticism instead of confirmation. Double-check all the facts and sources.