Yleinen 
Artificial intelligence is perhaps the hottest topic of discussion around the future of learning. It is associated with many types of threats, but it can also have interesting opportunities to offer for learning. How could artificial intelligence be used to support learning and education? In this episode, we dive into artificial intelligence and the future of learning with Headai’s founder, Cognitive Scientist Harri Ketamo, and Haaga-Helia’s Principal Lecturer Lili Aunimo. The host of the podcast is Programme Director Hanna Nordlund.
The language of the podcast is Finnish.
Listen to the episode here:
Transcription of the episode
Hanna Nordlund:
Welcome to the podcast of Digivisio 2030 project, where we talk about what kind of future of learning we are currently building in Finland. I am Hanna Nordlund, and there are different specialist guests sharing the microphone with me. This is where Oppimisen Seuraava Luku [the next chapter of learning] begins.
[Music]
Hanna Nordlund:
Welcome to the fifth and final episode of this season of the Oppimisen Seuraava Luku podcast. Today, we are talking about artificial intelligence and how teaching can be supported by AI now and in the future. Here with me sharing a microphone today are Lili Aunimo, the principal lecturer of Haaga-Helia, a computer scientist and Master of Philosophy, and Harri Ketamo, the founder of Hedai, a cognitive scientist and Master of Philosophy. Great that you could join me. Welcome.
Harri Ketamo:
Thank you.
Lili Aunimo:
Thank you.
Hanna Nordlund:
Lili, you teach about methods of data analysis and the development of AI applications. Do you think it’s easy to motivate students concerning this topic?
Lili Aunimo:
Well, actually it’s very easy to motivate them. They are already very motivated when they come to the course. Often we have less available courses on these topics than we have interested students, and because AI promises an increase in efficiency on many fields, it interests students. Then on the other hand, data analytics interests them because due to digitalisation, we have much more data about everything, and that should be utilised in business.
Hanna Nordlund:
Welcome to you as well, Harri. You have done projects with UN, Unesco and top universities in Europe and US for 20 years. How did you originally become interested in this subject?
Harri Ketamo:
Well, I cannot name one specific thing, but first I got interested in computational modelling of behaviour through sports, and through that non-player character computer games, which means controlling a machine as modelling the realistic behaviour of the characters. Of course my own background in sports played a part in how to model behaviour. Then slowly language got involved. How can we make algorithms that can construe language like a human. And then you could say a broader neural computation. All of these grew together. When I graduated from my master’s, this current direction or package was ready.
Hanna Nordlund:
Today, we are talking about a very interesting subject. Lili, from the perspective of a teacher, how well do you think we can utilise AI in education today?
Lili Aunimo:
I think that there are differences, but we do utilise AI in teaching. Teachers and students utilise it, and then it’s also utilised on the organisational level. But when we think about individual students and teachers, there are differences. On one end, we have the typical early technology adopters, and they are typically computer science students or teachers here at Haaga-Helia. Then on the other end, we have the students and teachers of other subjects. Of course this is a rough generalization, but we do utilise it quite a lot on those three levels I mentioned. The organisation, students and teachers all utilise AI daily.
Hanna Nordlund:
So we are already up to speed, and probably due to ChatGPT, matters related to AI have reached a new level in coffee table conversations. The subject has also been discussed in higher education institutes. Harri, what kind of possibilities do you see AI having for teaching and higher education institutes in the future?
Harri Ketamo:
Yes. The arrival of ChatGPT has been a broader scale opening for discussion that actually has an incredibly long history. It’s been more than 50 years of different themes of computational intelligence. If we think of all the adaptive learning solutions from the 80s and 90s, they are related. But we are in the early stages in the sense that until now, we have made more precise algorithms for precise problems that have been studied for a long time, but the computational power that we have now enables new kinds of approaches. Of course we always have to remember that when we do science, we cannot do random things or black boxes. The Emperor’s New Clothes story is older than 50 years, and we have to keep that in mind.
Hanna Nordlund:
Lili, what kind of possibilities do you think AI will bring to individual learners in the future?
Lili Aunimo:
It will bring many possibilities, and it is already in use. But for example, the students have the possibility to utilise ChatGPT, and our computer science students can generate code. The code also includes comments, and that is what we are already doing on courses. They don’t always have to ask the teacher for help of how a certain code would be, but when they know how to ask it in the right way, they get a ready-made code. In the future, our students could also use ready-made evaluation tools. Then they could get an evaluation of the results of their software development project and their code but also from the documentation. They could also get a ready-made evaluation of their dissertations as well, which would be generated by AI. They could also get interim feedback 24 hours a day and in the middle of summer if needed. Then another thing is personalised study paths and how they could find the courses and events relevant to their goals. We have a lot of very good material and content and events on offer related to different professions, so there is many possibilities in getting those to the right students at the right time. Then it could help the students to network among themselves. We could find the students who are interested in the same subjects and are in the same stage of studies at that moment, and that might help them to find peer support and sparring between students. AI would have a role in this as well. Of course there is a lot of data when we work in digital environments. Then we could analyse all of this data and generate predictive models with machine learning methods, and the students could improve their ways of studying based on the results from learning analytics and improve their networking as well. There are endless possibilities now that we are studying in a digital environment and data is accumulating from that activity. Here were just some examples to begin with.
Hanna Nordlund:
Thank you. These were very good and concrete examples. Actually, the individualised study paths and more flexible learning that you mentioned are also in the core of the Digivisio programme. This is a very important subject. We have already stated that there are many possibilities, and in the future, we can use AI as an aid in many things, but let’s talk about the other side for a moment. For example, the European Parliament has stated that AI can affect and guide people’s purchase decisions, data can be collected and used in unethical ways, and for example in a competitive sense, there might be a situation where companies who have better resources to collect data and utilise it can have huge competitive advances. Then it has also been brought up that the current lack of legislation is a big problem. Harri, what do you think about these threats in the context of learning? Should we be worried?
Harri Ketamo:
Let’s start with a quick answer. Yes, we should. Then let’s talk about what we should be worried about. If we start with the fact that there is no precise regulation related to AI, but we do have many kind of legislation that protects our privacy. GDPR, for example, was a wonderful clarification that reduced the grey area around data significantly. Then on the other hand, we have ethics as a branch of science, and it’s about 3000 years old. We have a back rest we can use, but it would be useful to have a regulation discussion. I am now cutting corners quite a lot. The experts can comment on it on Twitter. But this discussion is circling around regulation. The AI Act of EU aims to limit methods and areas of application that might be harmful for humans. The aim of the AI Bill of Rights of US is that we should be able to verify and assess the algorithms and data sets, and we should be able to do something about them. Transparency and explainability are probably the most descriptive words for it. Of course these are on the European side. Then the third main thing, which I don’t know the name of, but in China, their aim is that AI cannot give statements that are unwanted. But I would say that there is no absolute truth around these things. It’s the same with Europe’s approach of possible applications that are harmful for people. How do we know that, and how do we know that a non-harmful application which is used for booking something suddenly cannot book something unexpectedly. On the other hand the question with USA’s approach is that who validates the data, with what qualifications and how much. There is no clear policy. That’s why I would say that it’s a bit exaggerated to make this a matter of AI because it’s our general lifestyle, whether it’s European, American or Chinese. But we have tools. If we go back to the fact that in science, we have a long tradition of legislation and research and publication of classical ethics. We have to work for it. The discussion around whether we need to stop development for X months is probably a good way to awaken people in its popularity, but we need to talk about a fact which we cannot escape, which is that the transparency and explainability dimension is essential. Then the other side is to understand how applications can be used for good and bad. We have to be worried about it, but burying our heads in the sand and banning everything for safety’s sake doesn’t lead to anything good.
Hanna Nordlund:
So could we say that we have to be worried and alert?
Harri Ketamo:
Well that is a good summary.
Hanna Nordlund:
Lili, do you want to continue here and broaden the subject from the perspective of higher education institutes? What do you think we need to concentrate right now concerning this subject?
Lili Aunimo:
I agree with Harri that we have to be alert. Explainable AI is in a very significant role, especially in higher education institutes. If we utilise AI in counselling, evaluation and in data collection, we have to have a requirement for the AI that we can see the judgement and sources the solution is based on. We also have to have an understanding of how the AI works. It cannot be a black box. When we make procurements, for example, there needs to be open source code so we can know what algorithms it uses, and that it also uses explanatory algorithms when possible and it doesn’t always need to use the most efficient and computationally heavy algorithms of deep learning, even if it might be the most precise in that situation. We can also use a so-called white box algorithm that is explanatory in its nature. It might give slightly worse results, but we know what the decision and judgement is based on. We can also give an explanation for it after the model. So we can use deep learning algorithms, but we need to give it an explanation with certain methods. It is important to require this when we are making procurements. Like Harri said, I also think it’s very important that we utilise the possibilities of AI in education. In working life, the students get to utilise the possibilities of AI because it has the promise of efficiency. But higher education institutes are definitely the places where we can take into account ethical matters because it’s a part of our nature. We cannot expect companies to do that for us, because their goals are different to ours. I think that at higher education institutes, we need to be alert and have this discussion.
Hanna Nordlund:
It looks like Harri wants to say something here in the studio.
Harri Ketamo:
I would strictly disagree with the requirement of open source. We know that open source is an easy word to say to make it safe. But is it. If we think about the biggest recent data security risks that have been transmitted into systems, they have all swum in through open source. Open source is a strategic choice, but it requires for us to read the code precisely. There is a huge amount of different loggers that swim in or ports open. Data can leak. I am now telling you the threats. Of course the listeners know that we are an entrepreneur in the field, and we don’t have open source but products, but then we are in charge of those products and knowing what happens. Bu using random open source components, nobody knows where that data is, but when we are using data from actors who can lend their services to validation, whatever that means, the responsibility is then on the service provider that the data is where they said it was. Then the third dimension. We are talking about open source now, and we are forgetting the biggest factor, language models. If we now speak in ChatGPT terms, the bias and distortion in language models is in a key position. For example, try asking the version based on ChatGPT-3 anything about Russia’s war of aggression against Ukraine. It goes around quite many things, because they are hidden. The fourth version is better, but it’s also quite careful. When we are creating more data than ever with these kind of tools, we have a very big risk of biasing our entire next set of teaching data.
Hanna Nordlund:
Yes. These are interesting perspectives. Lili, do you want to comment on this before we move on?
Lili Aunimo:
Yes, I definitely want to comment. Harri had good points there. But I would defend the open source softwares by saying that of course procurement requires know-how. So I’m also not for random open source code like Harri said. So we need to know what the software has absorbed and what it consists of. We need expertise. But I think that higher education institutes have quite good expertise from the field of computer science. Then we just need the right people to participate in procurement and review at the code that we are procuring. Then I also think that that enables us to procure these services in a more flexible way in the future when there are open interfaces, and we don’t end up in a vendor lock so easily. Plus, if needed, we can also develop something new there ourselves or try something new with the software code. And concerning language models, I also agree that it is very important what language models we use. Right now EU has an initiative that has language models developed in Europe. Of course when we use these applications, we have to look at what language model to use and how it has been made and who made it and so on. But we have to be alert. We agree with Harri completely that this all requires work and vigilance.
Hanna Nordlund:
Yes. Indeed. We have to be alert. What I got from your answers was that we need to know what we are procuring and on the other hand, we have to understand what actors are behind the solutions, and in that way also look at the accountability of the actors. These are very good points. I think these are also quite big challenges in the sense that we all know the complexities of procurement processes, and when we are producing, we need to know what we are doing. It’s much harder to fix it afterwards. This actually takes us to what I wanted to talk about with you next. Data has a very big role in the development of AI. We know that the quality of the data has a great significance in how the AI works. Like you already brought up, different biases in the data can lead to very bad results. Harri, what do we know about this based on research? The quality of data and different biases. What do we understand about it at the moment?
Harri Ketamo:
Let’s stay on the GPT theme since it’s topical and familiar to everyone. When we are building a language model, the data that it is based on is in a key position. We can build a language model based on Wikipedia. It’s a common solution. GPT-3 and 4 are based on much broader things, where half of the internet is in there. But if we think of a language model based on Wikipedia, it represents quite well the perspective of the community that has written it. We can curate things out of there, but then we know that things in Wikipedia are translated directly from one language to another, and the cultural manifestations may include surprising features. For example, if you compare translated pages on Wikipedia about football to a genuinely written page, they are very different. So in a sense you can check how that affects things. Not to mention on what basis we rank some things out. In the earlier versions of GPT, they had taken out all matters related to the Second World War just in case. So these kind of restrictions that someone probably made with good intentions might prevent some theme being discussed. Or who actually thinks that we shouldn’t talk about genocide. If we sweep those under a rug in language models, we know that in the next language model where we need ten times more data, and 80 percent of it is probably generated by the previous GPT, that data becomes even more biased. The same phenomenon happens all around us. This is not an AI problem. It’s a much bigger problem. If we think about the classic recruiting example, where a company made a deep learning model of who they should recruit as a leader based on previous recruitments. It probably doesn’t surprise anyone what the machine recommended, and that is not the fault of the AI. This is a real design mistake that shows that the designer was fully unable to understand the algorithm and the data.
Lili Aunimo:
Yes. We also need to be worried about what kind of topic-setter AI will be from now on. For example, if teachers here start generating sources and get a new subject to teach and then ask GPT what is included in this subject and what sources should I include, you of course get some good suggestions, but critical thinking is in a big role there for the teacher. They need to see if it’s all that you need to say about the topic and if it highlights some topic too much. So the teacher still has to look for the suggestions from many different sources, and then reflect themselves based on their own education what is actually important in the subject and what would be good to teach to the students. The thinking cannot be outsourced to AI. The human has to think for themselves.
Hanna Nordlund:
That is wonderfully said, and that highlights how fast we move to basic questions of higher education and learning from AI. Harri, do you want to…
Harri Ketamo:
I would continue that that is in the heart of what learning is and what conceptual change is. What learning is, how it happens and what knowledge construction is. And now that we are having the discussion of should ChatGPT be banned or not, I have a strict opinion that under no circumstances. We have to question what learning is if it’s on the level of higher education institutes and universities and what kind of development in thinking it should lead to. This is the heart of it. In addition to critical thinking, we also need good general knowledge. I would say that the value of good general knowledge is even bigger than before. Critical thinking and a broad general knowledge are increasing their value.
Hanna Nordlund:
Us Finns are not alone in our excitement and interest of utilising AI. It’s a big phenomenon everywhere at the moment. Let’s talk a bit about what kind of good examples there are in the world at the moment that you think we should look at in Finland and even implement here.
Lili Aunimo:
The first thing that comes to my mind is that for example in the US, in Purdue University, they are using a Signal system that automatically sends a message to the students when it notices that based on learning analytics, it seems like they might be falling behind. So already in the early phases of the course, the student might get an automatic message on their phone that they should do this and this to reach good learning results from this course. So that utilises automatization in counselling and guiding students. This is related to AI in the sense that often the data analysis is done with the help of predictive models. This is not very new. It has been in use for a while. Then some other examples. A while ago I had a discussion that in Sweden, the consortium of higher education institutes has started using a chatbot in student counselling, and they have done it together. The information technology units have made a joint procurement and thought about it. It also looked like quite a practical and interesting solution, and that might help the work of the student counsellor so they aren’t so busy at the beginning of semesters when there are a lot of questions. Of course there are many practical issues to solve, but they have a good case of solving them, and it’s close to Finland. When I speak with my colleagues from abroad, they sometimes bring up that they utilise AI in monitoring exams and making sure that the exam or essay is done by the student who says they have done it. We have to develop new methods for monitoring in this digital world because students are no longer necessarily under the watchful eye of the teacher. These kind of proctoring systems seem to be in use, and they also utilise AI there.
Hanna Nordlund:
That was quite many good examples. Harri, what would you like to highlight from the world?
Harri Ketamo:
I would highlight a tradition of more than a hundred years, adaptive learning. The basic mechanism of adaptive learning is that we measure what the student already knows, then we compare it to the goals of the moment and then give them suggestions of what to read or do. This is a theoretical framework that is over a hundred years old. It has been developed for many generations from rule-based machines to machines that utilise different ontologies. Adaptive learning becomes a trend every ten years. Now we are in an interesting situation because we have something new again. We, for example, have created a group of algorithms called Compass. There is a language model in the background that can evaluate what the person currently knows. So it’s not that we measure different documents like completed courses that are backtracked to the description of the course, and then we get the know-how the student should have. We understand that if you get a one from a course, you don’t have all the needed know-how, and if you get a five, you have more of it. But that’s better than nothing. Now that we know what the student knows, we can compare the know-how the student has acquired to the know-how required in their field’s labour market. We don’t have the time to go into this, but when we normalise the language in a way where the same words mean the same things, we can compare where the gap is between the labour market need and the know-how acquired, and then give more suggestions such as complete these courses during your last year and you are closer. So we could call it macro-level suggestions. In the old adaptive learning applications, they would take corrective actions during the course, but now there are suggestions when you have completed the course. I think the next big interesting thing is European Data Spaces in Europe. They are trying to flow data between different silos in a commensurate way. One of these data spaces is Data Space for Skills. When we proceed with this and actually get the data to flow between educational institutes, recruiting companies, employers, different governmental actors, the employment agency and ministry of education, and European-wide, not only in Finland, then we can make quite interesting suggestions for learning and regional selection optimisation. Because now it may be that we in Finland have a need that they have an answer to in France, and then we don’t have the problem anymore of being able to recruit a teacher in a specific location, if we can suggest courses. The aim in Europe is to get to more compatible teaching. This is not only related to degree programme education, but especially to continuous learning. At what point do we need to update the degree of a mechanical engineer, since we are moving from universal joints to electric motors in cars. This has its own challenges, and this can be simulated in advance, and we could give people suggestions about it even today. There are solutions to this that are in use in Europe. As an example, I can tell you that we are working together with Metropolia, and if they have a job application for construction with a skills profile, they can be recommended a course from Metropolia that helps them to move along on their career toward their goals. This is already common place, and that is why the data space will take it to new heights, so that we can hopefully start doing this across Europe very quickly.
Hanna Nordlund:
We have often talked about the know-how of teachers in Digivisio and about how the need for continuous learning is of course related to teachers, like any other professions. The fast development of technology, constantly growing possibilities of technology and new tools mean that we have to manage new things. We have spoken about a specific thing today. There are teachers who are getting a bit nervous about AI and another huge thing that creates a lot of possibilities that they need to manage and that they have to stay alert. What kind of encouraging words do you have for teachers related to this?
Lili Aunimo:
I can start. I would say that they should try things in practice and have maybe just one thing per semester that they try. For example, if they haven’t created an account on ChatGPT, they could start with that and give the exercises of their course to ChatGPT and see how it handles it. That could be something they could start with. Then they should share their experiences with their colleagues. We at Haaga-Helia are organising a lot of events for teachers related to AI and utilising it in teaching, so they should attend those actively as well.
Hanna Nordlund:
What are your encouraging words, Harri?
Harri Ketamo:
The possibilities are great, but I would start from the basics. If you haven’t yet, go make your own skills-based CV. And I would argue that every teacher that has completed a higher education degree actually have knowledge about the basics of statistics, basics of doing research and a lot of material that is actually a hundred percent the same skills that AI prediction does. It’s one-to-one with statistical interference. There are a bit more variables. If we were to draw it on paper, we would need a couple of truck-loads of it. But none of the principles change. So if you comprehend this and start with the fact that actually every AI can be returned into IF logic or transistor logic. This has to be kept in mind. Every teacher has a wonderful basis, if they just think about what skills they already have from that field. After that they can think about what they are missing and go forward from there. But the most important thing is to understand that no teacher is lost here, because you cannot qualify for a teacher without completing specific critical basics. This would be my starting point. You have the skills, now we just do things.
Hanna Nordlund:
Your advice would be that they already have the skills and that they shouldn’t be afraid to try things. Thank you, Lili and Harri, for your visit. I could have talked with you about so many more things. I would summarise our discussion by saying, that like Harri said, this is something that has existed for a long time, and now it has emerged to broader consciousness. These are endless possibilities, but at the same time we need to be alert, preserve critical thinking, but then also understand what we are doing when we work with AI. I want to end on Lili’s wonderful idea of how thinking can never be externalised. This was the last episode of Oppimisen Seuraava Luku podcast for now, but you can continue the discussion about this and other subjects related to the future of learning with the hashtag Digivisio2030. I am Hanna Nordlund, and this was Oppimisen Seuraava Luku.