There are two different types of AI in wide use today. Recent developments have focused on data-driven machine learning, but in the last decades, most AI applications in education (AIEd) have been based on representational / knowledge-based AI.
Data-driven AI uses a programming paradigm that is new to most computing professionals. It requires competences which are different from traditional programming and computational thinking. It opens up new ways to use computing and digital devices. But the development of state-of-the-art AI is now starting to exceed the computational capacity of the largest AI developers. The recent rapid developments in data-driven AI may not be sustainable.
The impact of AI in education will depend on how learning and competence needs change, as AI will be widely used in the society and economy. AIEd should be used to help schools and educational institutions in transforming learning for the future.
Many AIEd systems have been developed over the years, but few of these have shown clear scientific impact on learning. Evidence is lacking partly because the contexts of teaching and learning vary across classrooms, schools, educational systems, and countries. Local knowledge and capacity is critical for effective adoption and shaping of AIEd, and new scaling models are needed. Co-design of AIEd with teachers is a possible way to advance new scaling models.
AI has a great potential in compensating learning difficulties and supporting teachers. The Union/ the EU needs a “clearing house” that helps teachers and policy-makers make sense of the fast developments in this area.
The big picture
In her 2019 Mobile Learning Week opening keynote, Director-General of UNESCO, Audrey Azoulay stated that AI was the biggest innovation in the human history since the paleolithic time. This may well be the case—if AI someday is invented.
Despite the common error of misplaced concreteness, AI is not a thing. It is a domain of research with many sub-disciplines, each with their own histories, domains of expertise, and developmental dynamics. This is important to understand, when assessing the potential impact of AI in education and policy.
There are three essentially different approaches to develop AI systems. Symbolic computation was a key driver in the emergence of AI research in the 1950s. The key idea was that computers are logical machines that can process bits of knowledge instead of only calculating numbers. This led to highly optimistic declarations that as soon as the generic logic of human reasoning could be programmed, computers would gain the essential qualities of human intelligence.
It was soon realised, however, that intelligent action requires extensive amounts of domain-specific knowledge. Towards the 1980s, this led to rapid growth in AI systems that relied on the manipulation of knowledge representations. In particular, representational AI—now often called “good-old-fashioned-AI” or GOFAI—focused on how the cognitive structures of expert decision-makers could be automatically processed. Since the 1980s, many such “expert systems” have been developed and deployed in large companies. Many programming techniques developed in the GOFAI research are now routinely used in all software development. This has led to the adage “When it works, it is not AI anymore.”
In pedagogic uses, the representational approach to AI has been dominant since the 1980s. Intelligent tutoring systems (ITS) typically contain representations of student’s current knowledge, a domain model that describes the knowledge to be learned, and a pedagogic model that steers the learner towards the learning objectives.
The recent interest in AI has its roots in a third approach: artificial neural networks. The first mathematical models of biological neural networks were developed in the 1930s. They became highly influential when it was shown that “universal logical machines” could be constructed from the simplest possible models of neurons as digital on-off elements. For many scientists influenced by logical positivism, this suggested that all rational thinking could be modelled with such networks.
Since the 1950s, many different models of artificial neural networks have been created. With inspiration from studies on biological neural networks, learning in these networks has typically been modelled as the strengthening of connections between simultaneously active neurons. This is known as Hebbian learning. For several decades, a major challenge in these network models, however, was the difficulty of using this simple associative learning rule in larger networks that contained many layers of artificial neurons.
The current AI revolution, to a large extent, results from the fact that it has now become possible to program computers with this simple learning rule. This represents a new paradigm for using computers. This approach is commonly called data-driven AI.
Where do the AI breakthroughs come from?
Data-driven AI has generated major breakthroughs in the last nine years. To put these in a context, it is important to understand the root causes of progress.
Three key technical developments underpin recent advances in data-driven AI. First, in the last 15 years, the rapid expansion of social media, internet use, and smart phones have generated vast amounts of data, text, voice, and images. Second, data-driven AI uses very simple computations that can be done using hardware that was originally developed for graphics processing in computer games. Using these specialised processor architectures, very high computational power can be achieved at low cost for the types of processing that is needed for developing and training data-driven AI models. Third, the Internet has enabled the low-cost distribution of human work at a massive scale. Many of the advances in data-driven AI are based on the availability of data collections that have been processed and labelled by humans.
In the last decade, these three technical trends have converged in a very special innovation dynamic. Internet platform firms, who have access to data and real-time connectivity, have become the dominant users and developers of data-driven AI and major developers of AI research knowledge, software platforms, and processor hardware. As the effective use of real-time big data is impossible without automatic data processing, machine learning and data-driven AI have become a necessity for these companies. At the same time, hundreds of millions of end-users on these platforms constantly classify and categorise data, making separate labelling and categorising redundant. This has led to what can perhaps now be called google-sized natural monopolies of the Internet.
This dynamic is not necessarily a sustainable one. Extrapolations from the extraordinary developments of the last decade may have little predictive power. The number of computations required to generate state-of-the-art models has doubled every 3.4 months since the deep-learning breakthrough in 2012. In June 2019, Jérôme Presenti, Vice-President of AI at Facebook, said that Google and Facebook were now quickly running out of compute power. Data-driven AI can solve some difficult practical problems, but it is probably the most wasteful computational approach invented in the human history. If the brute-force approach to data-driven AI continues, AI may, indeed, rapidly become an important source of global warming.
Digitisation is often considered to be immaterial. It is therefore important to note that, currently, the mining of one Bitcoin requires fossil energy equivalent of 750 tons of concrete, or 60 barrels of oil. Effective policy development for AI in education therefore requires understanding also the technical drivers of AI, as well as the future of education in a world where AI technologies are widely used.
Skills and competences in an AI-enabled world
To understand the potential impact of AI, it is useful to reconsider the EU competence frameworks. In these frameworks, “competence” is understood as a combination of expertise and attitude. Expertise, in turn, is viewed as a combination of knowledge, skill and experience.
A practical interpretation of competence is that it is a capability to get things done. This requires epistemic components, such as domain knowledge, accumulated experience, and skill. Epistemic components of competence are, however, not enough. This is what the EU competence frameworks conceptually capture with “attitude.” More broadly, however, epistemic components of competence need to be complemented with non-epistemic components, such as creative problem-solving, meta-cognitive capabilities, including self-reflection and emotional control, and the capacity to mobilise social resources and knowledge.
Skill, understood as an epistemic component of competence, is commonly associated with specific tools and techniques. In this sense, a car creates a car-mechanic, a computer creates a software programmer, and an anvil and a forge create a blacksmith. “Skill,” therefore, is conceptually a mirror image of current technology. When technology changes, skills become obsolete.
In the current digital transformation, technologies and tools used for work are rapidly changing. Epistemic components of competence are rapidly becoming obsolete. Education, therefore, is shifting its emphasis from epistemic content-related components of competence that were central in the last two centuries towards generic technology-independent “soft skills.” Social skills and capabilities to mobilise networked resources are becoming increasingly important as the Internet enables new forms of access and collaboration.
It is in this “post-Kondratiev” innovation dynamic, where the long-term impact of AI can best be understood. When the technical context of competence is losing its stability, experience, skill and domain-specific knowledge become less important. Generic non-epistemic components of competence, such as creative problem-solving and meta-cognitive learning skills, in turn, become increasingly important. It becomes less relevant what you know, and more important whose knowledge you can mobilise and what you can learn. Social and cultural skills that are necessary to effectively operate in the global networks of production and communication, become increasingly important.
In this setting, AI becomes a general-purpose technology that can perform tasks that previously required human knowledge and skill. Data-driven AI becomes necessary when the world becomes connected in real time, and when constant adaptation is needed to optimise activities in complex global networks that link actors across time and space. This also drives rapid change in skill and knowledge demand. As productive activities become increasingly automated in these real-time networks, human intervention, however, can become difficult. The future of work and demand for skills and education, therefore, cannot properly be understood just by focusing on AI systems themselves. It is necessary to understand the broader drivers that make AI systems economically interesting and socially important. These same drivers also generate important tensions in current educational systems. As a result, AI is often viewed as a way to reduce tensions between the institutions of the past and the needs of the present. For example, AI is commonly viewed as a tool that can provide individual personalised teaching of course material or as a way to automate repetitive teacher tasks. Another way of perceiving the potential of AI in education is to see it as a tool that could allow the transformation of education towards the needs of the future.
AI and digital skills
The rapidly rising visibility of AI has led many educational institutions to expand the provision of AI-related content. The ‘Elements of AI’ online course, developed by the University of Helsinki and Reaktor, has been a very successful effort to provide introductory-level knowledge about AI for broad audiences. Its main objective has been to “demystify AI,” and now over 350,000 people from 170 countries have signed up to this free online course.
A particularly interesting aspect of ‘Elements of AI’ is that about 40 per cent of the learners have been women. This is more than double the average for computer science courses. In general, women now represent less than one fifth of AI researchers. If job growth in the future increasingly occurs in tasks that require AI-related skills, as many experts claim, this may generate important gender imbalances in the labour market.
The very high media visibility of AI and the stories about AI experts being hired with seven-figure annual salaries have created what could be called an AI gold rush. This has important implications for AI skill development. In particular, it raises the question whether AI skill gaps will be filled without policy interventions.
To understand the dynamics of AI competence development, it is useful to distinguish use, modification, development, and creation skills. The ‘Elements of AI’ course addresses basic usage skills, creating awareness and helping the learners to make sense of the essential AI-related concepts and claims. This level of learning is necessary for effective and appropriate use of existing AI-systems. To achieve this level of competence requires less than a week of effort.
The majority of current AI systems are now created by relatively novice developers who rely on tools, frameworks, code, and learning material openly distributed by large companies, such as Google, Facebook and Microsoft. To be able to take into use existing tools and to modify them for specific purposes, some programming skills are needed.
On a very basic level, secondary school students can now build basic chatbots and machine learning systems in a few hours using this approach. There are also development interfaces specifically intended for children that reduce the need to know programming languages. For example, Machine Learning for Kids (https://machinelearningforkids.co.uk/) is now used in many schools and coding clubs. It provides simple programming interfaces to the IBM Watson services and allows children to develop programs in Scratch, Python and APP Inventor. A prominent initiative in this area has been AI4k12.org that has a very active mailing list for teachers who implement AI-projects in the classroom.
More generally, many systems used in service and manufacturing industries are modifications of freely available AI systems. For a relatively competent computer programmer, it takes some months to learn state-of-the-art AI development platforms and to modify existing code for business purposes.
The development of new AI systems and skilful modification of existing computational architectures using state-of-the-art development approaches requires substantially more effort. In general, graduate-level theoretical knowledge, support from competent peers or experts, and access to open source tools and commercial hardware platforms provided by the leading AI firms is necessary. Many universities have expanded their educational offers at this level of competence. Due to the high visibility and economic attractiveness of AI, the number of competent people at this level is increasing very rapidly. In particular, existing education in statistics, mathematics, computer science and physics can relatively easily be converted to AI-specific skills at this level.
The creation of new state-of-the-art AI models requires advanced theoretical knowledge and practical skills. Until recently, there have been only very few researchers with competences required to create new breakthroughs in AI and machine learning. Most of these have been employed by universities. As AI has become a strategic issue for many large corporations, this top-level talent has been in high demand. The largest data-driven AI conference, NeurIPS had some 13,500 attendees in 2019. This represents about 8-fold increase from 2012 and 41 per cent increase from 2018. Based on these numbers, one might roughly estimate that the number of highly competent AI researchers, able to create new AI architectures, is now perhaps 20,000 globally. These are the people that move the current technological frontier.
Radical breakthroughs in AI may, however, require broad and trans-disciplinary skills and knowledge. As was pointed out above, the current deep-learning gold rush is at least partly inspired by historical trends that may be running out of steam and electricity. It is possible that qualitatively new types of processing and compute architectures will be needed to create environmentally viable AI systems. Whereas continuous training of state-of-the-art machine learning systems now requires megawatts of electricity, the human brain works well with about 20 watts. Neuromorphic computing and new non-digital hardware architectures may become increasingly interesting in the future, and many of the hard-learned skills and knowledge of current-day AI experts may have only limited lifetime. It is therefore not clear that formal education in AI-specific knowledge and skills will be able to generate competences that will be relevant in the future. At present, most estimates of AI competence needs are based on rather straightforward extrapolations of the past, and there do not seem to be ongoing attempts to develop more informed scenarios about AI-related competence needs and their potential development paths.
A specific characteristic of AI skill development is also that state-of-the-art knowledge and tools can only be accessed through the Internet. A similar dynamic of competence creation characterises open source communities. For example, in the 1990s, the Linux development community was able to create very rapidly high-level software and computer architecture competences outside the formal systems of education. Internet-enabled peer-to-peer learning is also important in the AI domain, and is now explicitly supported by some of the largest global companies. As very high economic incentives now drain top talent from universities, the universities may face considerable challenges to be able to provide state-of-the-art knowledge in meaningful ways. These challenges are exacerbated by the fact that at present the development of breakthrough AI systems requires access to big data and compute platforms that are, to a large extent, controlled by commercial actors.
The use of AI in education
One common classification of AI in education is based on the main user of the system. A recent NESTA report distinguishes student-, teacher- and system-facing AI. Holmes, Bialik and Fadel (2019) further divide the student-facing AI systems in systems that aim at teaching students, usually based on instructivist pedagogy, and systems that aim at supporting learning, often building on more constructivist pedagogic approaches. The following table shows examples of such systems.
A recent review of peer-reviewed academic AIEd articles found that extant research has covered four main areas of AI in higher education:
• adaptive systems and personalisation
• assessment and evaluation
• profiling and prediction
• intelligent tutoring systems
The different uses of AI can also be categorised based on the student life cycle. In the above dataset, 63 percent of the academic articles described systems for academic support services, 33 percent described administrative and institutional services, and 4 percent covered both. Academic support services included systems for teaching and learning (e.g. assessment, feedback, tutoring). Administrative and institutional systems included systems such as admission, counselling, and library services. A further coding of the articles generated four main areas of AI application, shown in Table 2.
Zawacki-Richter et al. (2019) note that a large majority of the papers that they analysed in detail, were authored by computer scientists and authors from STEM departments. Only nine of the 146 first article authors were from education departments. Almost four fifths of all the articles applied quantitative methods. This, however, probably reflects the selection criteria for journals that were included in the study. In particular, in the U.S. academic setting, publishable articles tend to require quantitative methods, and this often leads to quantitative studies where computer-science researchers use their classrooms as research laboratories. In Europe, much of the AIEd research has been published in conferences that were not covered by this study.
There is relatively scarce evidence about the benefits of AI-based systems in education. In some specific areas, such as mathematics and physics, intelligent tutoring systems have been shown to improve learning, but it is also clear that learning benefits cannot be achieved simply by introducing new tools in a classroom. A recent review of critiques on ITS by Benedict du Boulay argued that a key to educational impact is teacher training that helps the teacher to orchestrate technology use. In other words, the learning outcomes do not depend on technology. It depends on how the teachers can use technology in pedagogically meaningful ways. An appropriate approach, therefore, is to co-design the uses of technology with teachers. This approach has been the starting point in the EU-funded New Era of Learning -project, where the largest Finnish cities have provided opportunities for rapid AIEd experiments and co-design with technology developers, teachers and students. A similar approach was also used in the Joint Research Centre (JRC) of the European Commission “AI Handbook with and for Teachers” -pilot project that finished in December 2019. In the project, teachers knowledgeable of AI and AI developers knowledgeable of teaching jointly developed a prototype model for EU-level co-creation network that would produce an AI handbook that would help teachers and education developers to deploy and use AI in appropriate ways. Such an approach aims to move beyond conventional technology-push and demand-pull models of innovation, adopting a middle-up-down diffusion model. As technology is advancing rapidly and many products will become available, teachers and education administrators will need high-quality information that will help them make sense of this rapidly changing landscape. In the U.S., the Department of Education has invested in the “What Works Clearinghouse” that consolidates scientific evidence on educational products and policies, and there have been many similar initiatives in the Member States. There seems to be clear potential for coordinating such initiatives at the EU level.