Export der Lerninhalte aus dem Moodle-Trainingskurs · zum Lesen ohne Moodle. Interaktive H5P-Elemente, Videos und Feedback-Aktivitäten sind im Original enthalten; hier als Hinweis markiert.
Modul 1 · AI Basics
Module overview
🚀AI Basics
Welcome to this module! 🎉 By the end of this module, you'll understand why you should/ or shouldn't use LLMs, how to interact with them effectively, and what to watch out for when you are evaluating AI tools. You will not only actively work on your AI skills but also work on a capstone project that you can use for your work.
🏆 Capstone Project: Your own AI Evaluation Workflow
Your overarching task for this module is to design a structured process for evaluating AI tools. You will undergo this process several times, so let everything you learn in this module influence your evaluation workflow!
At the end of this module, we ask you to: ✅ Define 3 very small real-world use cases where AI could help (e.g. searching for literature, summarizing a script, creating quiz questions). ✅ Select 2 AI tools for each use case. ✅ Compare & evaluate these tools using your workflow.
We require both the workflow and the results for the live training!
🔍 What will you learn in this module?
1️⃣ What Are LLMs & How Do They Work? 🤖
What AI & LLMs are and how they differ from other technologies.
How LLMs are trained, their core functions, and their limitations.
2️⃣ Prompting - Easy to learn, hard to master 🎯
What prompting is and why you profit from good prompts.
How to structure effective prompts to get the best results.
Adopting the right mindset when working with AI.
3️⃣ Evaluating - Can AI do that? 📊
Benchmarks & Metrics: Understanding which AI benchmarks are useful.
How to realistically evaluate LLMs for your own needs.
A quick and structured approach to deciding if a tool is right for you.
4️⃣ The Limits & Ethical Considerations of AI ⚖️
Understanding what AI should and should not do today.
Key ethical dilemmas and how to form your own stance.
If you are on the parrot track we have an extra task for you.
🎉 Hi there! Here's your chance to really wrap your head around the concept of Artificial Intelligence (or AI).
First off, it's clear that there's a lot of debate about the definition of AI. Some folks are convinced that everything is AI, whereas others argue that only specific use cases fit the bill. 🤷 It isn't very clear, we know! Even experts have a hard time finding a definition everyone can agree on.
This video tutorial aims to clear up that confusion, offering you a firm grasp of AI. In this Chapter, we will try to explore with you what AI really means and how it works. By the end of it, you'll be able to differentiate between different AI types and will be able to tell how they work. Pretty useful, right? 🙌 At least we hope so.
Before diving into the rest of the module, grab a piece of paper and sketch a mind map of everything you currently know about AI. As the videos play into even richer content, continue updating your map, adding or modifying parts where necessary. At the end, you will have made quite a journey and seen what you learned.
What are LLMs?
What is an LLM?
Written Version:
What is AI, really?
Honestly, I don’t have a final answer to that. Basically, we’re trying to give human capabilities to machines (minus the coffee addiction and existential crises 🤖☕). And that covers a lot of ground. In short, AI covers anything that mimics human skills, involves some form of intelligence, and is done by a machine.
A smaller area within this is machine learning. Normally, when we code, we define each step clearly, first this, then that. But with machine learning, we let the machine figure it out. We give it inputs and a situation, like where it is in a room with someone else, and then let it learn how to act, for example, by playing a game like tag 🏃♀️🤖.
The machine learns by trying things out, it sees what works and what doesn’t. If it does well, we give it positive feedback 👍 if not, we don’t 👎. Over time, it learns how to succeed without us programming every step.
A powerful example of this is neural networks. These systems are built to mimic the way a brain works 🧠, with layers of decision-making. Input comes in, decisions are made at each layer, and those choices guide the next steps. It goes on like that until it reaches an output.
From this, we also get transformer models, ones that can pay attention to context and understand things like language. That’s where models like GPT-4 come in. They’re trained to recognize patterns, focus on what matters, and improve through human feedback.
With that in mind, let’s go explore! 🚀
Time for our task:
The task at hand is to discover an AI application that doesn't involve Generative AI. Search for fun stuff, be it videos, websites or papers!
Understanding how a language model (LLM) works may seem incredibly complex, but a few steps are quite clear and worthwhile to understand.
1️⃣ Tokenizing: Whatever text you input, be it a greeting or a question, the LLM breaks it down into smaller parts, or "tokens." These could be whole words, or even parts of words — the idea is to make it manageable for the computer.
2️⃣ Number Assignment: These tokens are transformed into unique numerical IDs as the AI can handle numbers better than text.
3️⃣ Embedding: The numbers then get context by being placed in a higher-dimensional "embedding" space. This is like a description of the word. Think of it as giving the word numerous attributes to better understand its meaning in a particular sentence — like the word "king" might be associated with "royal," "male," "rich," and so on.
4️⃣ Attention: Once the attributes are understood, the LLM assigns importance to them according to their relevance in the context of the sentence — this is called attention. It helps the model focus on what matters most, like understanding that 'Paris' was 'beautiful', not 'he'.
5️⃣ Context Window: The AI has a "context window" that limits how much it can hold in its 'short-term memory'. Depending on the model, this can range from 1k to 128k tokens.
6️⃣ Prediction: Then, the LLM tries to predict the next word based on given input. It doesn't just guess — it calculates probabilities and lists likely words that come after a given phrase or word.
7️⃣ Temperature and Sampling: Instead of always choosing the most probable word, the LLM introduces some randomness via temperature and sampling. Higher temperatures (like 0.8 or 0.9) make the model more creative, while lower ones (around 0) make it more predictable. So you could say it’s a bit random. 🎲
So, with a seemingly simple text input, an LLM runs through an intricate process to give you relevant output! 💬👩💻
Check your understanding:
Check your understanding of how LLM works with this short quiz:
Let's solve a little challenge. You heard that LLMs do quite a few interesting things before answering. So choose one of these challenges and explain why LLMs behave the way they do:
1️⃣ Consistent Answers: AI might struggle to give consistent answers to the same question, such as "How do I look today?". This becomes particularly complicated when used in contexts like exams.
2️⃣ Citing Sources: This section explores why AI often falls short in correctly citing sources. It is an important skill in many academic and professional environments, yet AI systems seem to struggle with it. What's behind this?
3️⃣ Reading Lengthy Texts: Many have noticed that AI's answers tend to deteriorate when it's asked to comprehend and respond to questions about a whole book. Is the length of the text the reason? Why do the responses improve when the text is shorter?
4️⃣ Counting Items: Most AI systems (except a few models like ChatGPT's o1 struggle to correctly count "r"s, such as the ones in "strawberry". Why is this seemingly straightforward task so hard for AI?
In this chapter, we talk about how to treat genAI and how we can ensure that we get the most from our prompts. But before we start: How do you treat AI? Do you see it as a person or just a tool you use? Do you address it as "he" or "she," 🚻 or simply as "it?" Could it have a personality in your perception, or does it seem plain and boring? 💭🤖 How do you picture AI? How to best treat LLMs and how to prompt?
Let's dive into the picture of the 'Parrot Trainee' to explain the concept of how to work with generative AI. The 'Parrot' signifies the AI’s ability to mimic human language perception but without a full understanding — much like a parrot repeating words without comprehending them. The speaker stresses that the AI 'parrot' relies heavily on the quality of the data we feed it; high-quality data will yield better responses, while "feeding it trash will result in trash out". 🦜➡️🗑️
Moving on to the second part, 'Trainee' depicts a zealous worker — quite clever, eager to assist, but who can sometimes overstep bounds due to their enthusiasm and missing knowledge of your problem. In order to get the best out of the trainee, clear instructions and regular evaluations of their output are required. You should be able — and willing — to assess the parrot trainee’s performance. Only then can you be sure to produce a good result.
In conclusion, generative AI is like a 'Parrot Trainee'. It is capable, but requires careful data input, ongoing assessment of its output, and regular feedback for optimum performance. That is our picture of generative AI.
The question is: What is your picture?
Your turn!
We want you to create your own illustration of generative AI. How is your relationship with genAI? Please write a paragraph or two about why you think your picture of AI is helpful and create a picture with whatever tool you deem worth to illustrate your picture. We recommend Firefly, DALL-E, or Midjourney to generate pictures, but there are other generators out there. If you don't feel comfortable, the text is enough.
👋Hello everyone! Get ready to go on a journey to becoming a millionaire! 💰 Just kidding, but we are diving deep into the concept of "prompt engineering" today. It actually can be as simple or as complex as you make it. First:
Create a simple prompt to receive a summary of this text.
Let's explore an easy way to develop effective prompts using the COFEE framework! Broken down, the acronym COFEE stands for Context, Order, Format, Example, and Evaluation.
☕ Context first — if, for instance, you are a professor needing a point-by-point summary of your slides, be sure to mention that. Different situational contexts call for varied approaches in generating content.
☕ Order (clear directions) — Specify what exactly you need. If you're requesting a special kind of summary, say so. If you want to execute multiple steps, like asking for quality criteria first and the text that you want to summarize as a second step: Do it! 🙂
☕ Format — the structure of what you want; it could be a PDF, a Word document, a markdown file, or a specific arrangement. You can also ask for a short summary, then bullet points, and finally, a structured summary.
☕ Example — Define your quality criteria such as: a good summary only covers main points, it's short, trimmed to only necessary words and sentences. Or go the extra mile and search for good examples of summaries you like and place them in the prompt.
☕ Evaluation — It’s crucial to assess the resultant piece. If the response doesn’t fit the bill, fine-tune your prompt based on your assessment.
The main point is to stay flexible in adapting your COFEE blueprint. Iterate, re-evaluate, and tailor your prompt until the generated responses meet your needs optimally. For instance, you can command the AI first to deliver the main points, upon which you give feedback, and then generate the summary.
Perfecting your prompt is about refining it to meet your needs and expectations. So go ahead, sip your COFEE, and get those effective prompts flowing! 🔥👩💻👨💻☕🚀
Check your understanding
Check your understanding of the COFEE framework with the following short quiz:
🧩 Interaktiv: COFEE.h5p?time=1749816963689
Time for the task:
Let's refine your prompt with coffee!
You will create a summary for a university grade presentation on a topic you currently study. Assume that you will do a presentation in two weeks and are low on time. What do you need to create a good presentation? Should your summary have bullet points on your slides or, instead, concise summary texts to guide your understanding? Subsequently, would it help to add relevant examples, images, or any other visual aids that could enhance your presentation? Think about what you want and put it all in the prompt.
Also, ensure that it sounds like you sound. What is your style when doing presentations? Make it fit for you! You want to throw the information and get something you like and can work with.
So we don't want to throw away high quality prompts, but make them reusable. Whether it's a class discussion preparation or creating PowerPoint slides — automation can be a game-changer. But how do we make a prompt reusable and not tailored to one specific scenario?
1️⃣ Input: The starting point of any task is defining inputs and their relevance in the given context. For instance, when creating a text document, you need to feed in your requirements to the AI and expect the machine to act upon it. Let the user (of the prompt) be asked for it 🖥 Example: "Ask me for any relevant text documents before you do anything else!"
2️⃣ Decisions: This phase involves refining the inputs and setting clear guidelines for the AI to follow. If the outcome of the task is mission-critical, take the reins here and make necessary decisions about the task execution process! 📝 Example: "Give me the main points of the summary and let me evaluate them before you proceed!"
3️⃣ Output: This stage is all about defining the final product, be it a filled out Word template or a PowerPoint presentation. Is there a format you have to adhere to? 💼 Example: "The result should look like this:
# Bulletpoints
Write 5 bullet points about the input
# Short summary
Summarize the article in 3 sentences here
# Structural summary
For every headline in the text give a short summary per headline here."
By following these steps, you can create a prompt that you can reuse, saving time and effort. It's highly advantageous for tasks that require similar processes again and again. Now you try it!
Here is an example of how a mega-prompt could be written: "I'm a professor and want to create a presentation for university students in their master's. First, ask me for the input material, then summarize it in a bullet point list. Separate the important parts from the unimportant parts, and let me give feedback. When I have given feedback. Create the summary like this:
# Bulletpoints
Write 5 bullet points about the input
# Short summary
Summarize the article in 3 sentences here
# Structural summary
For every headline in the text, give a short summary of the headline here.
Please export it in a docx file."
Can you mega prompt?
I want you to create your first mega-prompt. We adhere to the example before and you should create a mega-prompt that you can always use to create material for a presentation.
1️⃣ Start simple: Don't feel the need to go big immediately. Writing a single, concise sentence can get a lot accomplished.
2️⃣ Quality Assurance: If you are aiming for better quality, think of coffee! ☕ By ensuring all elements are included from the get-go, you can continually refine and perfect your prompt.
3️⃣ Reusing and Streamlining: Embrace the Input, Decision, and Output routine. Incorporating this structure into your workflow can really enhance your results and working time.
Experiment with your methods, iterate on your results and share them in the forum. We are excited to see your progress! Get into that engineering mindset and try out our tools!
Before you start working with LLMs, it helps to know which one fits your needs best. In this video, Ben walks through the strengths and limitations of popular tools such as ChatGPT, Claude, and Gemini, comparing their features in everyday study and research tasks. Watch the video to get a practical overview and tips for making smart choices when it comes to LLMs. ☺️
Here is an activity for you. Match the correct tool and LLM with certain situations.
Evaluating LLM (Tools)
What is success for you?
👋 Hey there! This video unravels a very intriguing topic - 🧐 "What is Success?".
So, 🤔 what is success? Some might say it's about acing all benchmarks, but are these benchmarks set within any context or reality? Is success merely just going off without any guideposts? Are you more than your exam results? I surely hope so. But we are in a pickle here, especially considering genAI. Therefore, Let's explore how we can judge if an LLM is any good... or better: If AI can help you with any given task.
In this chapter, we discuss 'Why Benchmarks Matter', how we can approximate if LLM can do something and we also walk you through creating your very own benchmarks. Before you dive in, take a moment to think - "What genAI benchmarks do I know? What do they measure? What even is a benchmark?"
Let's face it, LLMs are pretty good at almost every cognitive task. But, when everything is good, how can we distinguish the good from the excellent? This is where 'benchmarks' come in!
Benchmarks are like sophisticated scales that help us evaluate and compare large language models, identify their strengths, and areas for improvement. They help us see if the LLMs improve (or stagnate) and what they can and can't do. These days, we see nearly as many benchmarks as new models — so how do you find benchmarks that actually tell you what you need to know?
Well, Google is your friend — but the interesting part starts after you’ve found a benchmark worth looking at. You should ask yourself:
What does the benchmark focus on, and is this relevant for me? → If it's not relevant, you may be wasting your time or misjudging the LLM’s competencies.
Who are the evaluators — subject matter experts or other language models? → Can you trust that the benchmark is measuring what it claims to?
Is the test data public, and could the models have been trained on it? 🧐 → If so, it’s like handing out the exam two weeks before the test! 😉
These questions help us unveil whether the benchmarks actually suit our needs. They ensure we’re evaluating LLMs fairly and meaningfully.
Let's make an example:
We use HumanEval to explore what a benchmark can look like. Here is a short article if you are interested in exploring this benchmark in-depth: Link
Written Version:
The process begins by setting up different challenges that language models have to overcome — in this case, it's programming exercises. The model then attempts to provide suitable solutions in response to these tests.
The primary purpose is to verify if the solutions produced work. 📝 If they work after a certain number of tries, we consider the LLM capable of fulfilling the job. So if we have pass@5, we try it 5 times, and if one of the solutions is working, we consider it successful. If it's pass@1, we only try once, and it has to work instantly to get a score point.
The tests were created by industry experts who wrote the tests by hand, which is a good thing. However, there's a weak point in this process: the data utilized for these benchmark tests is public, leading to the risk of the models already being comprehensively trained with it. Since we don't know if the model was trained on this data, we have to be a bit more careful when interpreting this test.❗
Now it's your turn:
Now, you have seen a benchmark in action. Find a benchmark yourself and try to answer our questions!
First off, I encourage you to take some time to hunt for an LLM Benchmark that resonates with you. With so many options out there, finding the right one is like searching for a hidden treasure! 🗺️💎 Once you've located "the one", it's time to see if it already exists in our forum. If yes, fantastic! Make it even better by enhancing the article with your fresh insights. If not, don't hesitate to create a new article!
We are really curious about what you will choose! See you soon.📝
👋 We think benchmarks are helpful in general. However, there are a few problems that we need to address so we don't blindly trust them:
The first problem is 🔬 contamination — the data used for tests (like with HumanEval) might already be available, and hence, accessible to the LLM. This makes it a bit too easy for the model, potentially skewing the results.
The next hurdle is that benchmarks can quickly become outdated 💾. (Have you seen one of those in your life?) Benchmarks are not only costly but also time-consuming to create. Researchers usually conduct them once every year or two — but that’s a problem, given how fast AI evolves.
Furthermore, some benchmarks rely on bold evaluation strategy choices 📝. For instance, some contain a massive number of questions that can’t be reviewed in detail, which makes it hard to judge whether the results or the questions themselves are even accurate.
Now, let’s talk about my favorite benchmark — LLM Arena! 🏟 It’s kind of like a chess match leaderboard 🎯. Two models go head-to-head in a blind test. You input a prompt, get two responses, and rate the better one — without knowing which model wrote which. 💭 That way, you're judging only the quality of the output, leading to more unbiased results.
There’s even a leaderboard that tracks model performance. If a “weak” model beats a “strong” one, it gains a lot of Elo points. If it loses, it only loses a few. This makes the competition fair — just like in chess or other ranked games.
But enough talking... You guessed it: Try it out yourself!
Try it yourself:
Have fun, try a few blind prompt comparisons yourself 🎮, and see what the best models for your area could be in the leaderboard.
You can also find the leaderboard in a tab on the same page.
All in all...
Rule of thumb
How to decide on whether to use an LLM: Rule of thumb
Let's keep it short in this chapter. When you use genAI, you should always consider if an LLM is really the right tool for the job. Therefore, we have a short checklist for you. You should be good to go if you can answer all these questions with a "Yes!". If not, you might need to explore other ways to get to the result you wish for.
Checklist
Training data
Is the knowledge required accessible (preferably) on the internet?
LLMs are unlikely to be helpful when working with undigitized literature or content behind paywalls, such as academic research, or legal texts, because they haven’t been trained on this data. This limitation poses a significant challenge when addressing specific, specialized questions.
Is the knowledge required time-stable?
LLMs also struggle with time-sensitive data because they are trained at a fixed point in time and don’t automatically reflect ongoing changes. This poses challenges in rapidly evolving fields like law, software engineering, or startups. However, there are solutions for handling such dynamic contexts.
Task Performance
Was the kind of task performed often (preferably on the internet)?
LLMs can be used for both knowledge-based and competency-based tasks, and it's helpful to distinguish between the two. Even if an LLM lacks specific knowledge in its training data, it can still perform tasks, such as answering questions, summarizing, or writing emails, if the necessary information is provided in the prompt. The key strength of LLMs lies in their ability to carry out cognitive tasks, not just store knowledge.
So, in a nutshell, here is a decision tree for you:
Time to test your knowledge
Let's see if you've got it. Read each case and the sentence below it, then mark whether the sentence is true or false. (Leave the mouse on the image for extra information)
🧩 Interaktiv: true-false-question.h5p
🧩 Interaktiv: True%20or%20false%20Case%203.h5p
🧩 Interaktiv: True%20or%20false%20case%204.h5p
🧩 Interaktiv: true%20or%20false%20case%204.h5p
To sum it up
If you only take one thing with you from this chapter: When you assume (or better tested) that the required knowledge isn't available through the model: You have to deliver itsomehow.
Your turn. Share two examples with us of your field:
This chapter is a cornerstone for your capstone project, so maybe enjoy it twice and think about how you want to incorporate it into your cornerstone project. 😉
🔍 Let us walk you through creating your own benchmark for evaluating AI tools. The focus is on achieving clear, measurable outcomes based on unique use cases.
Use cases
💡 First, let's emphasize the importance of starting with a use case — something you want to achieve — and defining what good results really mean to you. If you don't know what you want, then really nobody can help you. Search for products or results you think are great. What makes them so good? What is likeable about them? It could be the speed at which a task is accomplished, the quality of the output, or simply reducing extra work on your part. Whatever it is, write it down and be clear about what you expect the LLM tool to achieve.
Example: When I want to have a good summary, I want to get the most important content presented in an easily readable way — fast.
Metrics
⚖️ Next, it's about choosing the right measurement for evaluation. It could be quantifiable figures, or “vibes,” referring to your subjective feeling when using a tool.
Example: The most important content: Summarize a text manually and list the key points. Then, test the LLM and compare using 3 metrics:
- Success: Every point the LLM gets right earns a score.
- Correctness: Each nonsensical output is a point (the fewer, the better).
- Focus: Each unimportant bit is a point (again, fewer is better).
Readability: Give it a vibe check — how much do you like the text on a scale from 1–10? Speed: Measure the time it takes to summarize manually vs. the time the LLM takes.
Let’s assume we test every LLM three times and take the average of your metrics. Now we have a solid benchmark for summarizing — one we can trust to help pick the best tool. Maybe some tasks are better done manually, while others benefit from LLM support.
Failure analysis
🧐 Sometimes the LLM output will be garbage... and that's okay! If an LLM doesn't help with your task, analyze the problem. If you’ve tried honestly and it still fails — great! Send us an email, post it in the forums, let us know. We love to document what LLMs can’t do — that’s science! 😄
Dirty work
🔄 Lastly, a prediction: when you benchmark anything, it’ll feel like a “working but dirty” solution. That’s totally fine. Trying something new often means working without a polished outcome. Embrace the mess and give it your best try. And don’t forget — share what feels off with others!
So... ready to roll up those sleeves and have fun benchmarking? 😊💪🎯💻
Are you ready for some benchmarking?
So, let's work on the capstone project. Choose three use cases and two different tools for each of these cases. 🧪📈🛠️
Why is this so important? Well, it's a practical and systematic method to evaluate and compare tool performance. It's a solid foundation on which to make informed decisions about different tools' efficacy. The last thing you want is spending money and time on something that is just not worth it. 🙌📚
When you are done with this task: Great!This exercise was not just an academic task—it elevates your ability to critically assess and validate genAI tools, putting you well ahead of others!👩🔬📝
If you want to go the extra mile: Here is your chance! Update your initial answer what you know about Benchmarks and answer again: What did you learn, were you wrong about something? Did your perspective change somehow?
It's not all sunshine in the world of genAI. We have quite a few legal and ethical dilemmas and problems we have to address. We will explore some of these cases and you will explore your value system.
⚖️Let's discuss! ⚖️
Can you imagine any case where we should not use AI, or LLMs?
Keep in mind, we are no lawyers, but we know quite a few that will help us out. Stay tuned for new videos on the topic of AI law in education! Nevertheless, we have a few topics that are fairly strict and clear. So let's dive into the main issues we'll be covering today: Copyright issues, generative AI-based decisions, and cheating.
Firstly, on the hot seat is copyright 📚! We all know that the internet is a playground for creativity. Nevertheless, you own your work — be it text, pictures, or other forms of content. Well... unless a big AI company decides to ignore your rights. Currently, we see ongoing legal disputes between major companies. It doesn’t seem like users of generative AI are being targeted, but it still raises the question: Is it okay to just use someone else’s work?
On a positive note, Adobe Firefly sets a good example — it was trained only on Adobe's own material. That’s great from a legal perspective, but the results don’t quite match up to tools like Midjourney. So we’re in a bit of a pickle: obey the law, and get a worse product. Copyright in the age of AI remains a complex issue — let’s see how it plays out.
The second key topic is the ethical use of generative AI 🤖. AI should not replace human judgment in decision-making that affects people or communities. Generative AI is excellent for support tasks, but the responsibility and liability must stay human. Relying on AI for general help? Fine. Letting it make critical decisions? Not so much.
Just imagine your academic work being judged by an AI. What if the training data had gender bias? A male psychologist or a female engineer could be unfairly evaluated. That would not only be unjust but also very hard to detect. For this reason, important decisions must always involve human oversight.
Lastly, let’s talk about cheating 😱. Accusations of AI-assisted cheating are serious. You should always check with your supervisors before using AI. The real twist? It's currently almost impossible to reliably detect whether a text was written with generative AI. This makes fairness and enforcement tricky — especially with bachelor's or master's theses. Is it fair to use AI as long as everyone does? Now that’s a conversation to have with your lecturer. 😉
So, in conclusion: No decisions by AI, no cheating, and apparently no copyright for anyone till further notice. Weird times ahead. But that’s just my two cents — now it’s time for a little reflection. 🧠✨
AI-Made Decisions
Try to think of a use case where LLMs or AI is making big decisions. Who would profit from LLM-based decisions, and who would suffer? Share your thought experiment here.
Ethical considerations
Let's talk about our values. Don't worry, the aim here isn't to force any ready-made decisions upon you. In fact, we don't want to make any decisions for you at all. What we do want to encourage is thoughtful reflection, meaningful decision-making, and an ability to articulate those decisions.
We have chosen a few dilemmas and problems we see with AI. Pick one of the three subjects we will provide. After this, your mission is to formulate a stance on the chosen topic. ✍️📚 Go deep, question your viewpoint, and take time to develop your position. There are no wrong answers here. The important thing is to reflect on why you made these decisions 🤔. Try to come up with arguments for all sides and finally justify your stance. Put it all in one page and share it with us all.
Remember, ethics is all about the lens we use to view the world and our decisions. Enjoy exploring yours. We are very excited to read about it!🧭
Student learning
Just picture this - a new student steps into a university, an institution with little guidance and brimming with knowledge. They have likely used AI in their high school years to help with their homework. Now, this student has learned that Generative AI is a credible source of truth that can help them with all tasks. But here's the rub. This reliance on AI might make learning quite easy, and as a result, students might shy away from grappling with the more challenging tasks. This could lead to lesser cognitive strain tolerance, therefore less cognitive strain and learning results.
On the flip side, Generative AI smartly distills key takeaways for learners. It cuts down the noise and pulls out relevant insights that students need to understand their course material better. 🎓
So, here's a question for you: As a student yourself, what advice would you give to a new student? Should they use Generative AI to assist with their courses at all, or should they exercise a bit of caution given its pros and cons? 💭 Maybe they should just embrace it and do everything with AI? How would they know if they are doing it "right"? What do you tell them?
🌍 Let's talk about the intersection of genAI and climate change 🌡️. Most likely, we've already surpassed the critical 1.5 degrees Celsius threshold, hinting at a critical level for our planet. We made huge progress in transitioning to renewable energy and alternative forms of transport... and then comes genAI...
We could reap numerous benefits from Generative AI on one side of the spectrum. Let's think about advancing research in new climate tech or better predication models. On the other side, this technological giant has a massive energy appetite 🍽️. Private companies such as Google and Microsoft are even considering nuclear energy supplies to meet the demands of AI, underlining how energy-hungry these technologies are 🏭.
Should we continue using Generative AI, should we abandon it, or should we advocate for a 'greener' Generative AI with more eco-friendly infrastructure 🌿? If you were a technology leader and could have a say in how we should build and use AI in Germany, what would your goals be? Should we make it more efficient and faster? Should we ban it? Only build it with renewable energies or only use nuclear? Or is it not important for climate change? Research how much of a threat AI is for our climate goals and write your leadership goals. Share them here.
Information as a weapon
Let's talk about mass misinformation in today's digital era, brought about by the lack of gatekeepers and the rise of generative AI. 🌐 This enables anyone, particularly those engaged in conflicts or those looking to destabilize democracies, to churn out misinformation at a dizzying scale. 🔄
Worryingly, this issue extends beyond media to even newer AI technologies. There's a danger of 'bad actors' intentionally introducing biases to the data that feeds our AI systems, subsequently skewing the results in dangerous ways. 📊👥
Let's make a very practical example. Should you use Grok?
Your first, but hopefully not only source, are two (biased) YouTube videos in German
or another one in English
They are a bit different, but the task is the same: What ethical problem could you see when we would give this model to university students? What if people used it en masse? What problems could you see?
Your thoughts help us make this module better for you and future students. We’ve put together a quick, anonymous feedback form that takes less than 5 minutes to complete. Click the button below to share your insights:
As an AI Guide, you will be using AI in educational settings. In order to do this well, you have to know a bit about how education works. What better way to do that, than to take on the role of a teacher.
In this module, you will design a workshop - which is like a lecture with more interaction - on a topic of your choice. You'll walk through the design process, from
the logistics of planning to the self-recflection at the end.
Remember: when you design a workshop using proven pedagogical methods, your participants will both learn more and have more fun!
Note: To "turn the pages" of this book, click on the arrows at the top right and left of each page. They look like this:
First Step - Don't Panic
Familiarize yourself with the logistics of holding a workshop. Considering these factors will also help alleviate your worries about becoming the person at the front of the class!
Now it's your turn. Imagine you are giving a workshop about a topic of your choice. It could be like a tutorial for new students in your field, or an introduction to your hobby. Pick a topic, a location, and all the other details. Describe all of this briefly in the Teaching Basics forum. Check out what other people are doing, too, and tell them what you think.
For the Parrots (optionally for Robins), fill in the details in the form below. On the last page, you can either download a docx file, or copy and paste your answers to a text-editing program of your choice. Save this to submit with a few other documents at the end of this unit.
🧩 Interaktiv: SItuation-Analysis.h5p
Prepare the Content
You can think of a workshop as a type of learning environment, just as a course is another type. The purpose of learning environments is to enable students (or participants) to acquire certain knowledge, or a certain skill.
When designing a learning environment, there are three initial questions that need to be answered. Click on the image below to see them.
🧩 Interaktiv: Constructive_Alignment_triangle.h5p
The association of learning outcomes, teaching and learning methods, and examinations is called Constructive Alignment.
In the following pages, you will learn more about how to set goals (aka learning outcomes), how to design the content of your workshop (aka teaching and learning activities), and how to know if your participants learned anything (aka assessment).
Setting Goals
Set goals for your workshop by completing this sentence: At the end of my workshop, participants will be able to ....
Notice that the next word should be a verb, an action! These goals (also called a "learning goal" or "learning outcome" in education) help focus your planning and communicate to participants what
they will get out of attending your workshop.
Here are some guidelines for creating good goals:
Precise learning goals release concrete images - formulate your goals as precisely as possible.
Write it down; this increases commitment.
Goals are always formulated positively. We have no inner images for a "NO".
A good learning goal challenges, but gives the feeling that it is achievable.
The SMART formula can help you with the formulation of your learning goals.
🧩 Interaktiv: SMART_Goals.h5p
Check your understanding...
🧩 Interaktiv: SMART_Goals_test.h5p
This type of goal setting is great for everyday life. In academia, there is an additional model we often use, called Bloom's Taxonomy. This model breaks down cognitive processes (thinking) into six levels, which can be visualized in the triangle below. The idea behind this hierarchy is that students need to master the base of the triangle before they can master higher cognitive functions.
🧩 Interaktiv: Blooms_Taxonomy.h5p
Now, it's your turn. For your upcoming workshop, write a few goals for your participants. What do you want them to be able to do? What do you want them to learn? Post your goals to the forum and then give feedback to your peers.
For the Parrots(optionally for Robins), after you get some feedback from your peers in the forum write down your revised goals. You can use the form below to create a document, or any other program. This is part 2 of your TB capstone, which you'll submit iat the end of this unit.
🧩 Interaktiv: Workshop-Goals.h5p
Organize the Content
Now, on to the meat of your workshop. What are you actually going to do with your participants?
We recommend using the acronym ARIVA, which stands for:
Arriving Reactivating Informing Processing (Verarbeiten auf Deutsch) Assessing
When you break your workshop up into three main parts, the ARIVA parts fit nicely into each:
Introduction = Arriving
Body = Reactivating + Informing + Processing
Conclusion = Assessing
We'll take a look at each part in turn in the next few pages.
Introduction
The participants are walking through the door. What do I do?!
First things first! Watch the following video for ideas on how to start your workshop off on the right foot.
In short, you have three main things to accomplish in the first 10% of the time you have with your participants:
Welcome the participants
Introduce everyone
Set expectations
Here are a few ice breakers we like. You can find hundreds of other ideas by searching online (or asking your AI partner). Drag the names of the activity to the description below.
🧩 Interaktiv: Ice%20Breakers.h5p
On the next page, we'll get into the meat of your workshop. Turn the page to learn more...
Body
The main part of your workshop consists of the ARIVA parts: Reactivating + Informing + Processing. First, a little detail about each of these parts, then a short video about putting
it all together.
Reactivating
Reactivating means reminding your participants of some knowledge that they already have that's related to your topic. For example,
Ask participants to think about how they already use AI in their daily lives. This also gives you a sense of how experienced your group is.
Ask participants what the biggest benefits and risks to using AI in teaching are. This could prepare them for a later discussion on the ethics of AI.
Depending on the size of your group, how you engage them may differ. For example,
For small groups, have them share with the whole group one at a time.
For larger groups, they can share in pairs or in small groups.
For really large groups, you can use a digital tool to poll them.
Here are a few reactivation activities we like. Drag the names of the activity to the description below.
🧩 Interaktiv: Reactivating.h5p
Informing
Informing is more-or-less what it sounds like. This is the "lecture-like" part of your workshop, where you present new information to your participants. This often takes the form of a PowerPoint presentation, but it could also
be a demonstration of an AI tool. I'm sure you have more ideas! If you use a whiteboard, for this workshop or for teaching, check out the tips below.
🧩 Interaktiv: board-presentations.h5p
Most importantly, keep your informing part short, 10-15 minutes at a time is best. Depending on how long your workshop is, you might have 2, 3, or even 5 different informing parts. Each informing part should be preceded by a brief reactivating activity and followed by some time for processing.
Processing
The Processing part of ARIVA is the fun part! This is when participants get to dive into the topic and really grapple with the new ideas you've presented. There are dozens of ways to get your participants
active and engaged. Here are just a few:
Small (or large) group discussions.
Using an AI tool for a specific task, like making an interactive video.
Using an LLM to research sides of a debate.
Here are a few of our favorite processing activities. Drag the name to the description of the activity.
🧩 Interaktiv: Processing.h5p
Conclusion
At the end of your workshop, you'll want to know if the participants have acheived the goals you created for them. You'll also want to know how you could improve for
next time. This is the last A of ARIVA (assessing).
Likewise, at the end of a course, instructors need to evaluate the students to see if they have acheived the learning goals. They usually do this in some form of exam or assessment.
🧩 Interaktiv: Quality-Criteria.h5p
What does this look like in a workshop?
Now it's your turn. Revisit your goals and develop specific questions that you could ask your participants or tasks for them to complete to evaluate whether they have acheived those goals or not.
For the Parrots(optionally for Robins), once you get some feedback, save your revised assessment and submit at the end of this module.
🧩 Interaktiv: Assess.h5p
Chunking
But how do you know how many activities to do with your participants? A good rule of thumb is 45 minutes to an hour per "unit". This includes a brief reactivation (5 minutes), you informing the participants of something new (10 minutes), then they process
the new information (30 minutes). When you "chunk" your workshop, it makes planning much smoother. I explain more about chunking in the next video.
Remember: A tutorial is very similar to a workshop, because you want to give your students a chance to process (V) the information that they got in lecture. In lecture, there is much more time spent "informing (I)" and a lot less
time for "processing (V)". But it's still useful to include it all.
Now it's your turn! Map out what learning goal and activity you will cover during which part of the workshop.
Give the participants (and yourself) breaks! Use Excel, Word, Miro, or whatever tool you want. Don't forget to ask your friendly LLM for help. Here is a link to a practice Miro board, if you want to check it out. If you like it, you can make a free educational account here.
Post your plan to the forum(image, text or video description, ...). Then give feedback to at least two others.
For the Parrots(optionally for Robins),after revising your plan based on the feedback you got, you'llsubmit this document with the others at the end of this unit.
Afterward
Your workshop is over, now what?
I am sure many of you already know some of these things, so let's just reactivate your knowledge by dragging the words to their correct space in the text:
First of all
There may be some participants who approach you afterward for a private chat. Give them your attention first, as much as possible.
🧩 Interaktiv: drag-the-words%20%281%29.h5p
🧩 Interaktiv: drag-the-words.h5p
Wrapping it all up
In this book, you learned all about pedagogy (didaktik auf Deutsch). In designing a workshop, you:
Created goals for your participants,
Chose or designed activities to help your participants acheive those goals,
Created tasks for your participants to do, to prove they acheived those goals,
Your thoughts help us make this module better for you and future students. We’ve put together a quick, anonymous feedback form that takes less than 5 minutes to complete. Click the button below to share your insights:
Sometimes when thinking in German, it’s easy to put the verb at the end of the sentence. In English, however, the verb should come first, followed by the knowledge. So instead of writing “students should be able to [knowledge] [verb]”, the correct order is “students should be able to [verb] [knowledge].”
It’s interesting that ChatGPT was referred to as “he”. Assigning a gender to an AI can reinforce unconscious biases and anthropomorphize a tool that isn’t human. A more neutral choice is to use “it” or simply “ChatGPT”.
Written Version:
The core of this tutorial is the development of effective learning goals, which form the backbone of all educational tasks, content, and assessments. As budding scholars, we often face the challenge of formulating these learning goals from scratch or tailoring existing material to match our intended outcomes.
To simplify this process, I'll demonstrate a method of creating learning goals based on pre-existing content ('our own stuff') from an article I wrote earlier. Download the article here or use a source of your own. However, the approach I am about to show you isn't a one-size-fits-all solution and should serve as a guide. Please work along, but modify the process to your liking. The goal is to create a prompt that allows me to throw text-based material in and get learning goals out.
We begin by defining the context: "I need learning goals for a seminar I'm conducting as a university lecturer." Specifying the format is crucial. In this case, the format is a list of bullet points with content-specific components for each entry.
Let's dive into the process with an exploration of my earlier article. For demonstration, I used ChatGPT to parse my article and generate associated learning goals. It’s okay if the tool makes mistakes. For example, let's examine the first learning goal proposed by our AI: "Understand the role of gamification in motivation." While this appears comprehensive, it's rather vague. It could use some specificity and context.
With this in mind, I revised the learning goal format to "After the seminar, students should be able to ___", with a blank space for the requisite knowledge and an appropriate verb from the learning taxonomy. Keep in mind that working with AI is a process of iteration. I encountered situations where the AI tool needed explicit instructions in certain areas, like language preference or specification of learning content and goals. But once cleared, it churned out more pinpointed learning goals like: "After the seminar, students should be able to explain the basic principles of gamification."
Once the AI tool generates a list of learning goals, it is the lecturers' task to sift through these, cherry-pick the most relevant ones, and further refine them. For instance, in this case, I classified '...explain how reinforcement mechanisms influence behavior in gamified environments' as a good, specific, and effective learning goal. Remember to iterate and improve your process continuously. Over time, a task that could take upward of 30 minutes only takes about five minutes with an efficient system in place.
In conclusion, we can turn basically anything into learning goals. Do you have a dream job? Collect a few job offers and write a prompt to summarize what learning goals you have to hit to land your dream job! Here is the "rough process" of how we got to our final prompt.
Process steps:
🧩 Interaktiv: sort-paragraphs.h5p
Well done! Last but not least, remember: Always have refinement in mind.
Using genAI tools means: Iterating until you are satisfied
Accept that AI-generated content may still need refinement.
The goal is to save time, not perfection.
Now it's your Turn: When you are finished, tell us about your experience in the forum. We don't need to know your learning goals, just your process for creatingthem. Respond to at least two others. How does your experience compare with that of your peers?
Get Trustworthy Materials
Navigating the Sea of Online Resources
Welcome to this chapter, where we aim to get knowledge concerning a topic FAST!
The key focus is on specific strategies that help to gather reliable data quickly. In the video, I guide you through effective techniques using prominent online tools and services, focusing primarily on their applications within the education field.
Below are some tools for gaining knowledge mentioned in the video, each varying in reliability and purpose. Can you match them correctly?
🧩 Interaktiv: drag-the-words.h5p
Your turn: Now that you've used AI to design some presentation materials, tell us about your experience in the forum. Respond to your peers about what they did, as well.
Slides
Creating Slides with GenAI
Ebru’s Two Cents
As you may notice in the video, the slide deck does not appear on screen, even though Ben is referring to it in his presentation. This is a technical issue, and our team is already working on a resolution.
Written Version:
Welcome! We are diving into the world of PowerPoint presentations, slides, and the complexity of mastering these tools. The crux of the problem lies within the density of information present in many slide-based presentations. Overflowing text, scanty use of imagery, all while trying to listen and grasp the subject, can make listening quite challenging... But whom am I telling that? So how do we not make bad presentations?
I personally use three in my presentations: Pictures, single words, and sometimes my desktop (just as it is now).
But let's consider, for instance, the Gamma app, an AI tool that claims to generate engaging slides on the spot. When we try to create a presentation on AI and climate change risks using this tool, the outcome is a versatile set of editable cards. Not only can you customize these cards, adjusting the topic as you wish, but you can also select your style preference, image AI availability, and model choice for the pictures. Voila, a rapid generation of gorgeous slides! It may be gorgeous, but it could be better.
Though stunning, these slides pose a challenge to readability, cramming extensive information in limited space. This can hinder the primary function of a slide – to prompt what the presenter needs to say alongside aiding the audience in "getting back on track" if they let their thoughts wander for a bit. It isn't that these AI-generated slides can't be worked with. But they might not always suit every presenter's unique style, particularly if the focus should be more on the speaker and the story they are telling.
The crux of the issue of AI-generated slides is the overwhelming text, which, while good-looking, distracts the audience, who end up straddling between reading and listening. But we could definitely just reduce it, so maybe just use the style and do the text yourself?
In conclusion, the process of forming effective slides for a lecture is a bit of a tightrope walk, made seemingly simpler by AI but not necessarily more efficient. With AI, iteration might be faster, but mastery in delivering engaging presentations, it appears, still lies in the human hands of the presenter.
But enough talking, it's time to put your AI tool evaluation skills to proof:
You can use the Gamma App we talked about here, or you can pick a different tool if you prefer. To do this task
1. Pick a topic you know very well and use an AI tool to help you create a presentation about it.
2. Design your evaluation criteria. What should a good presentation on this topic be like? How detailed should it be? What are the main points? What kind of images should it show?
3. Write a good prompt on the tool you chose to evaluate.
4. And finally, evaluate your results and share your evaluation with us here. Did the tool meet your criteria? How much editing would you need to do to improve your presentation? Is this tool worth it overall? Would you use it again?
Activities
Designing Discussion Activities with GenAI
<
Written Version:
"The best way to learn is by struggling against the world and succeeding." - Ben Lenk-Ostendorf Cite it whenever you like; it sounds great, and who needs reliable scientific sources these days?
But on a more serious note, here’s what you need to understand:
The most valuable moments occur when learning is a challenge.
Creating problems that facilitate this learning process can be daunting and time-consuming.
Despite this, obtaining high-quality and stimulating tasks is paramount in fostering effective learning.
The question is: Can we let AI support us with that? We certainly think so. Let's create a workshop to have an intense discussion about a topic. Preparing a discussion might not be particularly challenging for an expert; however, creating a document that initiates meaningful discussions among students is significantly more demanding. Yet, this is where our solution shines - taking often time-consuming preparation and condensing it down to a simple process.
Remember this:
The complexity of creating a prompt is not the main issue; it's the task of carefully crafting the correct format, selecting quality criteria, and defining what a successful activity looks like.
The solution here is to find and utilize prior didactical knowledge and bake it into the prompt
Knowledge is still key, but genAI can execute on it. Try it out and enjoy exploring this innovative approach to learning.
Your turn:Let's put this into practice by creating an activity that supports your learning goals, share your activity in the forum. Remember to give feedback to your peers!
Assessment
How to design tests using GenAI
Written Version:
In this module, we delve into the significance of testing during and at the end of a semester. We aim at assessing the effectiveness of the lectures and whether the students comprehend the information being relayed. It's critical to ensure that tests are a fair measure of knowledge and skills, extending the same chance to every student... but it's also a test for the lecturer: Did you succeed in teaching these students?
One of the ways to achieve this entails basing our test questions on our learning goals. For instance, after a seminar, if the learning goal was for students to critically assess the role of generative AI's energy consumption on climate change, our tests should validate whether this goal has been met.
Here are a few different ideas for generating test questions:
Creating a Quiz Question based on the learning goal.
Framing a comprehensive Essay Question related to our learning goal.
Providing interpretation-centered questions where students analyze datasets pertinent to the topic.
Hosting a Debate for an oral examination.
Assigning a Policy Brief wherein the students must compile a written assignment or a case study analysis.
The confines of the exam situation can greatly influence the approach to testing. For instance, if the exam is being administered in a lecture hall, the options might be more restrictive. Therefore, the exam design should be adaptative to the real-world scenario. The first information we need is the context of the exam. After this restriction, we can aim to verify our learning goals. You created a workshop, so how would you test your learning goals?
Here is a little exercise for you:The quiz below was developed by an LLM, and it contains some of the classic errors LLMs can make when creating a quiz. Your task is to find the errors. After that, share in the forum below how you can prevent these kinds of errors.
🧩 Interaktiv: single-choice.h5p
Your turn: Once you have successfully created an assessment using AI, share your experience in the forum. What went well? What didn't? What would you do differently next time?
Assessment for humans
How to design tests for humans
Written Version:
AI is a Horror for Professors – But Maybe It Shouldn’t Be
Let’s be honest — generative AI has completely disrupted how we design and think about exams. In the old world, students had to do the cognitive heavy lifting: understanding, researching, analyzing. But now, AI can do most of that — often better, faster, and definitely cheaper than a human fresh out of university. That’s a problem.
So what’s left for us to teach? If students learn skills that AI can perform effortlessly, what exactly is their edge when they graduate?
The knowledge itself isn’t the issue — it's widely accessible, just like it’s always been with Wikipedia and academic papers. But now, even processing that information — analyzing, contextualizing, making sense of it — is being handled by AI. So we need to shift our focus.
What I believe we’re seeing is a new type of work. Not just doing a task, but orchestrating it. That means knowing where the information lives, defining the process, using AI to do the grunt work, then coming in to refine and apply judgment.
To help rethink how we teach, I developed a framework called ELITE — it’s not perfect, but it gets the ball rolling. We start with existing exam questions and check: can AI answer this? If it can, maybe it's not testing what we really care about.
Take an example: a multiple-choice question about AI's impact on climate change. AI can answer that. So we ask: what's the real learning goal here? Maybe it’s not just recalling facts, but being able to critically evaluate claims about AI and energy use. That’s a much richer goal — and something we can structure differently.
The ELITE Framework
Experiment – Can AI already solve the question?
Learning Goal – What should students actually learn?
Iterate – Do we need to shift the task or the learning goal?
Test – Try new questions and see if AI can still solve them.
Expose – Show students the task, compare their answers with AI, and see where the human value lies.
Sometimes, the right solution is going back to supervised, in-person, paper-based exams for core knowledge. Other times, we need to redefine tasks to reflect this new orchestrator role — where students are evaluating, adapting, and applying AI-generated content in meaningful ways.
The point is not to resist AI — it’s to embrace it smartly. Let’s not fight generative AI but collaborate with it. We should teach students how to use these tools effectively, thoughtfully, and ethically. That’s how we create value, and frankly, that’s what gets me excited again about teaching in this new world.
So yeah — AI is a horror… but it’s also an opportunity. Let’s build something better.
Module Feedback
We'd love your feedback on this module.
Your thoughts help us make this module better for you and future students. We’ve put together a quick, anonymous feedback form that takes less than 5 minutes to complete. Click the button below to share your insights:
We have crafted a collection of resources for you, including guides, studies, and insights on AI in education. Feel free to browse and share your thoughts, participate in discussions, and suggest additional materials in the resources forum. When participating, please remember to be respectful and considerate of others' perspectives. We hope you enjoy!
Websites
Websites
In this chapter, you will find different websites that may interest you. We've included guides on using AI in education, engaging explorable explanations of machine learning, and interactive essays on complex topics for those who want to dive deeper into the workings of AI and machine learning. If you find any websites you think should be included in our resources, write about them in our resources forum.
This project, developed by metaLAB, an organization located at Harvard University, the Freie Universität Berlin, and the Academy of Arts and Design Basel, aims to introduce AI tools relevant to educational settings while also clarifying misconceptions and cutting through the hype surrounding AI. It also proposes assignments that aim to foster informed discussions on the risks, benefits, and potential impacts of AI tools. The metaLAB also developed an AI code of conduct for AI implementation in education.
This crowdsourced collection book contains ideas from educators in multiple educational scenarios, as well as researchers and student-educators, to use AI in education for various subjects and learning goals. You can download the book for free on the website or righthere.
If you are interested in how AI works, you can explore this visual explanation of how machine learning, a core component of AI tools, uses statistical learning to automatically identify patterns in data.
This explorable exploration takes you on a journey through the world of generative music, from its origins to the current techniques powered by machine learning. Discover various generative music systems and experiment with creating your own tunes through a visually immersive experience.
A neural network is a specific type of machine learning model designed to mimic the human brain by using layers of interconnected "neurons" to process data.This websiteallows you to tinker with a neural network by choosing various statistical settings and controls to adapt to what you want to explore.
This website explores how training sets are created and utilized in building AI models and their role in shaping biases within these systems. Developed by Knowing Machines, a research project dedicated to uncovering the histories, practices, and politics of machine learning training, it offers a deeper understanding of how AI learns to interpret the world.
As machine learning usage grows, it raises key questions about fairness, model predictions, and data privacy. This website offers interactive essays to guide you through these topics.
In this chapter, you will find peer-reviewed and non-peer-reviewed research studies investigating the use of AI in educational contexts, especially in special higher education. Some studies have an optimistic view of AI, whereas others are more skeptical. If you find any studies you think should be included in our resources, make sure to write about it in our resources forum.
While cautious regarding the challenges of AI in education, the following studies present an optimistic view and investigate how it can be used to improve teaching and learning, especially in higher education scenarios.
1. Do AI chatbots improve students' learning outcomes? Evidence from a meta-analysis
AI chatbots are becoming more popular in education, and many studies have explored how they affect students' learning. However, the results of these studies have been inconsistent. This study reviewed 24 randomized experiments to better understand the impact of AI chatbots on learning outcomes. The findings showed that AI chatbots had a strong positive effect, especially in higher education compared to primary and secondary education. Additionally, shorter interventions with chatbots were more effective than longer ones, possibly because the novelty of the technology wore off over time. The study suggests that to improve learning outcomes, AI chatbots should include features like human-like avatars, gamification, and emotional intelligence.
2. Artificial Intelligence and the Transformation of Higher Education Institutions: A Systems Approach
The rapid growth of AI, especially generative tools like ChatGPT, presents new opportunities and challenges for higher education. While much has been written on AI in this context, few studies take a systems approach. This article fills that gap by creating a causal loop diagram to map how AI transforms higher education institutions (HEIs). The model shows how AI investments can improve student learning, research, and administration while addressing challenges like academic integrity and adapting to market changes. The study emphasizes the need for HEI leaders to think systematically to manage AI’s impact and avoid pitfalls and suggests further research into the sustainability of AI transformation in HEIs.
Education is undergoing a transformation with the integration of Artificial Intelligence (AI), which is reshaping teaching strategies. Central to this change is personalized learning, where AI customizes content to suit each learner's needs and pace. This paper explores AI-driven personalized learning, including its impact on e-learning, AI-powered tutors, and the ethical challenges it presents.
4. Revolutionising Distance Learning: A Comparative Study of Learning Progress with AI-Driven Tutoring (not peer-reviewed)
While generative AI is expected to greatly impact education, its potential at the university level has yet to be fully realized. This study provides initial evidence that generative AI can significantly speed up learning. We tested the AI-powered assistant Syntea with hundreds of distance learning students across 40+ courses at IU International University of Applied Sciences. Results show that Syntea reduced study time by about 27% on average after three months, highlighting its potential to accelerate and personalize learning at scale.
The studies below examine the potential negative impacts of AI in education and emphasize the need for a cautious approach when integrating AI into learning environments.
1. The effects of over-reliance on AI dialogue systems on students' cognitive abilities: a systematic review
As AI dialogue systems become more common in education, it's important to understand how they impact students' thinking. This review looks at how over-reliance on AI affects critical cognitive skills like decision-making and critical thinking. We examined 14 studies and found that students often prefer quick AI-generated solutions, which can hinder their ability to think critically and make decisions, even amid ethical concerns.
2. ChatGPT: The cognitive effects on learning and memory
As AI technologies like ChatGPT become more integrated into education, understanding their impact on cognitive functions is crucial. While AI holds great potential for personalized learning and easy access to information, excessive reliance on AI could lead to reduced critical thinking and memory retention. This article reviews these potential risks and emphasizes the need for AI to complement, not replace, human cognitive abilities. The paper calls for further research on the long-term cognitive effects of interacting with advanced AI models.
3. Generative AI Usage and Exam Performance (not peer-reviewed)
This study examines the impact of using generative AI tools like ChatGPT on students' exam performance. By analyzing student essays with GenAI detection systems, we found that students using GenAI scored an average of 6.71 points lower than non-users. While GenAI may enhance learning and engagement, its use is linked to lower exam scores, especially for students with high learning potential. These findings offer crucial insights for the ongoing debate on integrating GenAI in higher education and highlight the need for careful consideration by educators and policymakers.
4. Can We Delegate Learning to Automation?: A Comparative Study of LLM Chatbots, Search Engines, and Books (not peer-reviewed)
With the rise of LLM-based chatbots, students are using them more for learning, which raises concerns among educators. They worry these tools might cause students to skip critical steps in the learning process. In this study, we compared three learning sources with varying automation levels using 92 university students. We found that LLMs help students understand concepts but have limited impact on knowledge retention. Additionally, students with higher academic competence engaged more deeply with content, relying less on search activities.
This forum is open for you if you want to share your thoughts, discuss any topics, or suggest additional resources. When participating, please remember to be respectful and considerate of others' perspectives.
