Tag Archives: ai

Learning Materials Shouldn’t Form Filter Bubbles

Snellmanin ala-asteen luokan oppituntiIn printed learning materials bubbles might be formed between the school book cover and the plastic covering it. However, bubbles can emerge also with digital learning materials. These bubbles are completely different since they isolate learners from each other. That’s why it is very important to avoid and burst them.

Last time I wrote about how artificial intelligence is teacher’s friend not enemy. Intelligent digital tools and materials are not about to replace the teacher. Instead they support the teacher, making her job easier.

Now it is time to address another theme: filter bubbles. Filter bubbles are known from social media and politics. People with similar opinions and attitudes talk solely with each other, causing them to form a bubble which drifts away from other bubbles. In the future the same unwanted mechanism can also be affecting learning.

Politicians, education officials, application developers, publishers, and many others: one of the most central objective for parties working with education is to advance personalized learning. This means that learners are encountered as individuals rather than as homogeneous groups. For example my employer Sanoma Pro as an educational publisher aims to offer diverse materials with something for everyone.

In digital environments personalized materials can be composed and distributed automatically. For example in Bingel the pupil is offered easier or more difficult exercises depending on whether he got previous answers right or wrong.

Personalization cannot jeopardize the equality of learning

Personalization and adaptivity can be taken further than what present Bingel has. This is exactly where the danger of filter bubbles looms. In principle the learners could complete exercises and according to the answers be directed in different routes which might not cross in the future. This could replicate the unwanted scenario from social media with learners in their own bubbles, isolated from each other.

In real life the teacher has a central role in guaranteeing that each learner goes through at least particular topics. Also the curriculum mandates this; personalization has to be implemented respecting he curriculum.

What to personalize then, and on what grounds? Some propose learning styles as the answer. It means offering interactive content for kinesthetic learners, videos for visual learners, and so on. However, learning styles and their positive impact to learning has been debunked several times. Not worth to waste time in them.

Personalization for engagement

My proposal is to personalize in order to engage and motivate the learners. One can assume that a learner is more motivated if she gets offered learning materials tailored for her. For example, exercises should be hard enough to keep the learner interested. They cannot be too difficult, however, so that the learner doesn’t get frustrated. It is natural for people to like both routines and change. Learning materials should reflect this by offering both familiar content and surprises.

To conclude, an important note related to personalization and more generally applications utilizing AI: both the learners and the teachers should have the feeling of being in control and understand the tools and content they use. This is why it is worthwhile to offer recommendations rather than completely automatic adaptivity. Let the learners and teachers to select whether to follow recommendations or not.

Even better if the recommendation includes argumentation of why just that particular content is offered to that particular learner. For example like this: “You seem to be practicing for tomorrow’s geometry exam. You have already mastered areas but there is still some work to do with volumes. Here is a 30 minute exercise package for you and there is a three minute video covering the basics. You can also familiarize yourself with volumes with the following containers found in your kitchen….”

When implemented right, personalized learning motivates the learner, eases teacher’s daily work, and does not isolate learners from each other.

This post was first published in Sanoma Pro’s blog in Finnish.


Hey Teacher: AI is Your Friend, Not Enemy


We are told from left and right that artificial intelligence (AI for short) is going to take our jobs: four million jobs in the UK in danger to disappear, 47 percent of US jobs are going away, as much as half of all professions vanishing during the next decade.

Luckily not all news stories are as gloomy. One research dealt with a thousand companies and estimated that AI will create new jobs in 80 percent of the cases. More moderate predictions claim that professions are going to both appear and disappear. A recent study by PricewaterhouseCoopers (PwC) stated that in the years 2017-2037 the UK job market is going to see seven million jobs to go away, but 7,2 million new ones to emerge.

You Cannot Replace a Teacher with a Machine

How about teachers, then? Are they going to be fine? Yes, claims PwC. Teaching is one of those professions which are hard to replace with computers or automation. I couldn’t agree more. Deep knowledge of learners’ personalities, ability to adapt to dynamics in the classroom context, and interacting with the study group accordingly are in the hard core of teaching. No one or nothing knows the class better than the teacher.

Instead of replacing teachers, AI can help them in doing their job better. Teaching involves several routine tasks which could be partly or completely automatized. For example in Bingel, Sanoma Learning’s gamified learning environment for primary education pupils, checking the answers of simple exercises is automated. Assessment has further opportunities for automation and AI, as does for example planning the upcoming weeks, days, or lessons during a semester.

AI can help also the learners, for example by virtue of personalized learning. A machine can learn to provide the best possible exercises and other content to each learner, as long as it has access to enough data of the learning history. The accuracy can be further improved by adding other context attributes into the equation. For example the social dynamics of the study group, time of the study activity, and even the learner’s emotional state could help to select the most appropriate content. Doing all this can increase the motivation of the learner, which in turn can improve the learning results. In other words: the machine learns to make the human learn.

Learning materials are always designed and tested together with the users of the materials, namely teachers and learners. This applies also to intelligent systems which learn. Even the best data scientist or developer cannot implement good pedagogical solutions without the precious know-how of education professionals. The machine learns to make the human learn, but it all starts with the human teaching the machine.

This post was originally published in Finnish on Sanoma Pro’s blog.

Google’s Oppia Turns Programming into a Dialogue

Update, March 5th, 2014: I found the man behind Oppia! A true developer really devoted to what he is doing. He helped me to get a bit deeper insights on some things:

  • Oppia is not Google’s product, so the title of this post is a bit off. Oppia was initiated as the 20% time project, meaning the time employees can use for their own projects. Something I proposed a while ago to be included in education, too. And then found out it is already happening.
  • The fact that it is called Oppia is an accident of sorts. The team has no connections to Finland. They had to come up with a name and were going through different languages searching for translations to the word “learning”. The runner-up was Swahili word “kujifunza” but “oppia” is easier to pronounce.
  • No immediate plans to integrate with Google’s data & algorithms. For the time being Oppia is an independent system and any integration with web content (e.g. to find out that Finnish is linguistically close to Estonian) is up to the community. If someone wants to code an interface to web sources, fine, they are free to do so. But you shouldn’t expect Google itself to implement such, at least for now.
  • Community is at the core. This is the most characterizing feature of Oppia. You could compare it to Wikipedia or Linux in its ways of working. This actually brings a fresh approach to personalized feedback and learning paths. Usually metadata, semantics, and powerful computing is harnessed to perform these tasks, but Oppia relies on the community. The more people contribute to Oppia and its explorations, the better it becomes. If it works for Wikipedia, hey, why not here?


Last week I bumped into two interesting technology launches: Wolfram Language and Google’s Oppia. Both bring us closer to actually discussing with a computer rather than only giving it commands in a one-directional fashion. I managed to discuss Oppia with Google folks yesterday here at SXSWedu.


Google’s search engine has long ago started to resemble the classic vision of artificially intelligent computer, which gives answers to questions humans ask it. I remember this from the comics I read as a kid in the 1970s. There are questions Google can’t answer off the shelf, though, and that’s why it would benefit from having a dialogue with the user.

Oppia is an open-source project Google has initiated. It aims at providing a convenient way for anyone to create and share “bite-sized educational explorations”. What caught my attention first when I heard of the project was its name. Oppia is Finnish and means “to learn”.

This factoid was also used in the demo the Google guys gave me. The demo included a question along the lines of “What language is the word Oppia?” The demonstrator first typed “Spanish”. Oppia gave a negative feedback in a polite way and asked the user to try again.

“Greek”, he wrote. Then Oppia replied “Closer, but not quite, please try again.” I got excited: could this system know about languages and how close they are linguistically? I asked him to type “Estonian” next. I know that Finnish and Estonian are very close to each other, hoping for an answer indicating that.

But no, it did not recognize this aspect. Bad news. The good news is that it could. We could’ve created a rule on the spot stating this fact, setting Finnish and Estonian linguistically very close. The huge news, however, is the potential.

At the moment Oppia is a separate system relying on explicitly states rules. What if it had all Google’s muscles behind when interacting with the user? I am sure there are several sources on the web where the closeness of Finnish and Estonian is mentioned.

While I am waiting for the integration between Oppia and Google’s information & algorithms to happen, I am marveling its dialogue approach of interacting with the users. It is mimicking the communication between a teacher and a student. When the user/student answers incorrectly, Oppia/teacher not only tells that the answer is wrong but tries to encourage and give a push to the right direction. Very nice.

Another phenomenon Oppia contributing to is the lowering of barrier for anyone to start programming. Like the Wolfram Language, it bears the possibility of bringing people and computers closer to understanding each other.

I have yet to locate “the man behind Oppia”, but I’ve heard he is at the conference. If I find him and get deeper insights, I’ll update this post or write a new one.

Artificial Intelligence, Y U Follow Me?

Artificial Intelligence (AI) has the tendency to follow me wherever I go. Maybe I am to blame, since I first went to it when still studying at the universiy. I wrote my Master’s thesis in 1999 about software agents, which at that time were the main manifestation of AI.

These agents were supposed to start roaming the web delivering all sorts of interesting content to their masters, the web users. Then I moved on to researching the Semantic Web, which was intended to provide healthy food for these agents.

Next were context-aware mobile phone apps. The simplest form of context-awareness is based on the user’s location: provide the user with content which is relevant in her current place, e.g. nearby restaurants with good deals.

There are many more attributes describing the user context than mere location, however, and combining those into a meaningful framework is tricky. It’s so tricky that doing it right in my opinion requires quite strong AI. You’d have to know the user’s current goals, social surroundings, mental and physical state, etc. in order to provide her with the most suitable content and services.

I was able to hide from AI for a couple of years, working mainly on market research and social media. Now since more than two years I’ve been in education and AI has yet again found me, this time in the form of personalized learning.

Personalized or adaptive learning is a great goal to pursue! What better for a student than to provide him with the materials and solutions that best fit his skill levels and learning style preferences? By the same token, doing it all the way it is tricky. Very.

Complete understanding of the user context basically requires telepathy. As Nagel pointed out, one person cannot fully understand what it is like to be another person. It cannot be objectively communicated. Similarly for learning: students are unique individuals and facilitating each one best is (at least next to) impossible.

There are many things that can be done, however, and are actually being done already. Adaptive learning solutions are going to enter education the same way intelligence enters the automotive industry: slowly but surely, without us even noticing it.

If you would bring a modern self-parking car with cruise control and windshield wipers adjusting to the weather to say 1940s, locals would treat is as (artificially) intelligent device for sure. However, these innovations have emerged over time and that’s why we are reluctant to claim them as AI.

So keep calm and carry on, adaptive learning solutions are going to pop up here and there. It’s going to be an evolution, not revolution.