Artificial Intelligence in Education – Challenges and Possibilities By Hemant S.
Artificial Intelligence (AI) has been a term that has been in use for over six decades now. The laypersons understanding of AI would probably stem from watching science fiction movies where machines are anthropomorphized. Following the coining of the term AI at the Dartmouth conference in 1956 by John McCarthy, AI has had its share of ups and down. However, over the last decade, thanks to an unprecedented amount of data being generated, we now see AI being revived on an unprecedented scale. In this article, we examine some possible scenarios for AI in education – the theory and practical applications based on research and practice. We will also look at policy and ethical aspects in having AI as part of the education process. While the article seeks to provide a broad overview, it looks closely at the potential for personalization of education and what this means. It looks at what frameworks of learning could this new system being employing. Interestingly, an area for future research could be to understand how machines learn and whether an existing framework would be applicable.
What is the buzz about?
When we hear the term ‘intelligence’ we tend to think of advanced intellect. The person that we know who has demonstrated to be highly intelligent in school or career. We do not pause to think about ordinary everyday things like having a good conversation or reading a book in the bus on the way to work. We take these for granted and in most instances group them under the term ‘common sense’. Therefore, if we were to compare ourselves to an AI tool or system, we should begin by knowing that most human beings are orders in magnitude advanced than the AI tool.
The challenge in defining what intelligence is an enormous one. While the general perception of intelligence takes us to the cerebral domain, matters get complex when emotional, social, and spiritual dimensions of intelligence are also to be factored in. For example, if one were to find themselves in an accident in a remote place, one would hope to have a compassionate person pass by rather than a Mac Arthur grant winner for Physics. This range of diversity in human intelligences has made it difficult to come up with one common definition of intelligence. A study carried out by Legg and Hutter (2007) sought out to distil the different definition of intelligence, identify similarities and come up with one definition. They identified everyday definitions, clinical definitions, and AI researcher definitions to come up with the following:
‘Intelligence measures an agent’s ability to achieve goals in a wide range of environments.’ (Legg and Hutter, 2007)
Campbell et.al (2002) in their examination of the Deep Blue chess machine, developed by IBM that beat the reigning chess grandmaster in six games in 1997, found that there were many factors that aided the victory. Some of them are a single-chip chess search engine, a massively parallel system with multiple levels of parallelism, a strong emphasis on search extensions and a complex evaluation function. However, Deep Blue could not do anything beyond playing chess. It could not do many of the things a human baby or even some animals could do. This distinction needs to be made to understand the ‘narrow’ nature of AI. Researchers are working towards an Artificial General Intelligence which is when an AI system can perform a range of tasks. Adam et.al (2012) have stated that the goal of AI research is to create systems that can demonstrate a broad range of general intelligence found in humans. Another reason for making this distinction at the outset is to help us understand what AI can and cannot do in education.
Having outlined some of the existing limitations of AI, let us now look at some of the possibilities offered by AI and why it could be potentially revolutionary. The amount of data produced each day is estimated to be a mind boggling 2.5 quintillion bytes. It is this data that is fueling machine learning which in turn is driving AI. 5G is estimated to increase the internet traffic by 1000 times in the next decade (You, 2018). 5G wireless technology is much faster than the existing technology and has considerably less latency. This makes it ideal for communication between devices (Internet of Things) and for driverless cars. So, in addition to humans creating data, we will move to an era where our devices create data on our behalf. All this would result in more data for machine learning algorithms to train on. To get a better sense of what this looks like, let us look at some crucial distinctions. Machine learning is a subset of AI which in turn is a subset of computer science. Within Machine learning, there is another subset which is driving progress in AI, known as deep learning. Deep refers to the number of layers in an Artificial Neural network and in the instance of Deep Learning, it is more than one and this distinguishes it from other forms of Machine Learning. It also makes use of multiple algorithms and thereby increases its accuracy. Let us look at what an AI algorithm looks like and what it does. The image below is an AI algorithm model that is used to identify cats and dogs.
Below is the result (output)
The importance of displaying a simple algorithm in this article is to illustrate the fact that while the task of creating an efficient algorithm may be an intensive and time-consuming process, it can be shared and replicated in a matter of seconds. Imagine the potential this holds for education.
Perhaps, the most significant contribution to education by AI could possibly lie in the sphere of personalized learning. Personalizing learning was the privy of the royalty and select few in the past. There never will be enough human teachers to achieve this on a mass scale even if the impetus were there. This is where AI has a scalable effect unlike anything we have ever seen. The same algorithm, designed for one student can be used to teach millions of students in which the algorithm model learns and adapts to the needs of the student. However, before we explore this further, we need to have a basic understanding of what we mean by personalized learning. Prain et. Al (2013) in their review of personalized learning observed that there did not seem to be a consensus on what it meant and different institutions adhered to a different outcome.
Sebba et.al (2008) have identified five key components of personalized learning as: assessment for learning (AfL), effective teaching and learning (including grouping and ICT), curriculum entitlement and choice, school organization (e.g., workforce remodeling), and beyond the classroom (e.g., extended schools). Underwoord et. al defined personalized learning as ‘the tailoring of pedagogy, curriculum and learning support to meet the needs and aspirations of individual learners, irrespective of ability, culture or social status, in order to nurture the unique talents of every pupil’. And finally,
However, we need to be mindful that the definitions and outcomes stipulated above were meant for a different system and process. So much so, it is my opinion that all of them put too much pressure in a system that was barely equipped to deal with even any additional demands. Therefore, for the purposes of this article, we look at a very narrow definition of personalized learning – address the unique needs of learners to help them meet (at least) the common minimum standards at each level at a pace that is not general. I feel the need to keep it simple is crucial for the success of such a tool.
With the advent of artificial Intelligence and Machine Learning, a subset of the personalized learning process is ‘adaptive learning’. As Murray and Perez (2015) points out ‘Adaptive learning tools are technology-based artifacts that interact with learners and vary presentation based upon that interaction’. This model would help bring a level of personalization that caters to individual students needs rather than a traditional classroom where a one size fits all model is the norm. While this mode of learning has made in-roads into the continued learning and career space, it is yet to make considerable impact on education at school level. One plausible reason can be found in Murray and Perez’s definition of it being ‘technology-based artifacts’. While learners in higher education and careers can leverage the use of technology to their advantage, this does not currently have an effective place in the school system. When Mitra and Dangwal (2010) sought to look at ways in which children can organize themselves for learning without teachers being around, they presented a model in which a group of students use one computer that is connected to the internet, in an open and collaborative setting. This objective of SOLEs differs from the adaptive learning model which tries to not merely personalize learning but also consolidate it at a foundational level. Mitra and Dangwal had this to say about the nature of learning involved in SOLEs ‘It is necessary, at this point, to mention that while we seem to have found evidence of ‘understanding’, the relative importance of ‘deep learning’ in the context and location of this study needs elaboration.’ It is my understanding that the authors of SOLE were trying to involve students who had no others means of education rather than helping the privileged students get better. Therefore, at the face of it, SOLEs and Adaptive Learning may seem to serve two different purposes but the possibility of it merging as the technology associated with it becomes democraticised should not be overlooked. As Mitra and Dangwal (2010) responded to criticism of the nature of learning associated with SOLEs, they made a very nuanced point about the nature of learning and how it differs in parts of the world – ‘….To berate their efforts with comments about the lack of ‘deep’ learning is a sad reflection on our ‘shallow’ understanding of the problem. We would argue that what children achieve routinely in hundreds of ‘Holes-in-the-Walls’ in some of the remotest places on earth is nothing short of miraculous—a celebration of learning and the power of self-motivation. The strange, non-intuitive and unbelievable emergent phenomenon of self-organizing systems is something that few social scientists are aware of, leave alone the possible ramifications.’
Let us take a brief look at how adaptive learning, using AI, has the capability to deliver personalized learning experiences. The framework that would be most useful for adaptive learning systems to guide a learner from beginner to advanced level would be Bloom’s taxonomy (1956). Woolf et. al (1992) has identified four interdependent models which make up Adaptive learning systems. These are – the student model, the pedagogical module, the domain knowledge module, and the communication module. It is worthwhile to note that Woolf et.al came up with this model in 1992. However, it did not find much success due to a limitation of large data samples, lack of internet access and hardware inadequacy. Woolf et.al referred to these systems as Intelligent Tutoring Systems. A brief description of each of the models outlined by Woolf et.al (1992) can be found below:
Student Model – The student model stores information that is specific to each individual learner. At a minimum, such a model tracks how well a student is performing on the material being taught. A possible addition to this is to also record misconceptions. Since the purpose of the student model is to provide data for the pedagogical module of the system, all of the information gathered should be able to be used by the tutor.
Pedagogical Module – This component provides a model of the teaching process. For example, information about when to review, when to present a new topic, and which topic to present is controlled by the pedagogical module. As mentioned earlier, the student model is used as input to this component, so the pedagogical decisions reflect the differing needs of each student.
Domain Knowledge – This component contains information the tutor is teaching, and is the most important since without it, there would be nothing to teach the student. Generally, it requires significant knowledge engineering to represent a domain so that other parts of the tutor can access it. One related research issue is how to represent knowledge so that it easily scales up to larger domains. Another open question is how to represent domain knowledge other than facts and procedures, such as concepts and mental models.
Communications Module – Interactions with the learner, including the dialogue and the screen layouts, are controlled by this component. How should the material be presented to the student in the most effective way?
To understand this better, I would like to use a frame of reference that some of us maybe already familiar with – Netflix. Netflix uses an algorithm to determine what to serve each user based on their preferences and therefore to personalize entertainment to individual user’s tastes and preferences. This is what Netflix (2019) had to say about the process its algorithms follow:
‘Whenever you access the Netflix service, our recommendations system strives to help you find a show or movie to enjoy with minimal effort. We estimate the likelihood that you will watch a particular title in our catalog based on a number of factors including:
- your interactions with our service (such as your viewing history and how you rated other titles),
- other members with similar tastes and preferences on our service , and
- information about the titles, such as their genre, categories, actors, release year, etc.
In addition to knowing what you have watched on Netflix, to best personalize the recommendations we also look at things like:
- the time of day you watch,
- the devices you are watching Netflix on, and
- how long you watch.’
If we were to extrapolate the factors involved in the entertainment giant’s algorithm to an educational Adaptive learning tool, we will see the following tentative factors:
- your interactions with the learning platform (such as your learning history and how you rated various learning content or courses) – Student, Pedagogy, Domain
- other learners with similar tastes and preferences on the learning platform, learning content that helped them overcome similar barriers and – student, communication
- information about different courses, such as domain area, university, tutors, time required for completion, number of learners enrolled etc. – Domain, Pedagogy
In addition to knowing what you have learned on our platform, to best personalize the learning experience, we also look at things like:
- the time of day you learn – student, communication
- the devices you are using for learning – student, communication
- your interaction with the learning community – student, communication
- your preferred learning style (quiz, reading, videos, lectures, discussions etc.) – student, communication, pedagogy, domain
- how long you learn at a stretch – student, communication
The above is a rudimentary process flow to help us understand what set of rules are followed to create an adaptive learning system, based on Woolf’s models. Unlike in the industrial age or other major revolutions that came before, the likelihood of AI and its allied technologies being democraticised in education is high. This is because it does not involve much costs. The algorithms can be tweaked to different contexts and manufacturing costs are nearly non-existent. As Mahidhar and Davenport (2018) point out ‘such systems will probably add little value to your business if they are completely generic, so time is required to tailor and configure them to your business and the specific knowledge domain within it’. In the context of education, there would still be a business element and therefore the providers who have invested using AI in education such as Knewton, Century Tech, Quizlet, Cogni, EdX, Nuance, Coursera and Sana labs amongst others would stand to gain for being the early adopters. However, unlike in sectors like finance, manufacturing and retail there is the possibility in education that the trickle-down effect will be more visible. This can be seen from the success of Khan Academy. Khan Academy has been experimenting with machine Learning for the past decade and has kept their algorithms open source (Hu, 2011). A study commissioned by the Bill & Melinda Gates foundation (2014) reported the following:
‘Students primarily used Khan Academy problem sets to practice and refine skills individually and with classmates, while getting immediate feedback that resulted in a range of observed outcomes across the research sites including:
- Learning new math skills
- Filling in gaps in learning and shoring up weak spots from past instruction
- Tracking and monitoring school work to hold themselves accountable for their performance
- Spending more time in peer teaching and collaborative work with their classmates
- Receiving more opportunities to direct their own learning, and
- Allowing teachers to spend more time assisting individual students or small groups of students.’
This is where the link to SOLEs become critical. How can we ensure that the maximum number of students, spread across the world have a personalized learning experience without costing a fortune? I do not have an answer and I feel more research into this area is warranted.
While personalized learning has been portrayed as the holy grail of AI in education, let us now look at some other possible areas in which more progress has been made. Using AI systems to grade coursework is becoming more commonplace. According to Ramalingam (2018) ‘The Graduate record examination (GRE) is one of the best results of this method. The students’ writings are graded by both human and automated essay grading system. Then the average is taken’. One in every four schools in China is purported to be using machines powered by AI to grade pupils work and give feedback on how to improve (SCMP, 2018)
Another potential area where AI can help aid teaching and learning is by identifying gaps in course syllabus. It can do this by either alerting the professor about a potential gap based on collating large numbers of student feedback or response to a wrong answer. MOOCs have begun employing this practice where they gather feedback from hundreds of thousands of students on their platform and use that to improve their courses. Is this something new? Probably not! What it does change is the labor involved in collating and analyzing this feedback while also making feedback immediate and actionable.
In some parts of the world, facial recognition software is being introduced in classrooms to track facial expressions, emotions and actions that pinpoint whether a student is puzzled, bored, or interested by the content that is being taught. The makers of the software purports to be zeroing in on identifying in learners Vygotsky’s Zone of Proximal Development (1978) to justify the need for this tool in the classroom. It is also replacing smartcards for an array of purposes in schools. Critics have derided this move and questioned the government’s motives. Issues about privacy violation are already rampant. However, the technology has managed to go through and the company that owns it has catapulted to being the highest valued AI startup after raising $600 million in 2018.
As technology gains strides in areas of ‘computer vision’ – a field of study that seeks to develop techniques to help computers ‘see’ and understand the content of digital images such as photographs and videos (Brownlee, 2019) – and Natural Language processing, where computers are able to analyze and respond to human speech, we will see new systems emerge. However, we should be mindful that technology does not exist in a vacuum. It must factor in the context. A good example of this would the current state of AI affairs in China – China may overtake the US with the best AI research in just two years (MIT TR, 2019). This has become possible because the context in China is vastly different and privacy may not be as big a concern as it is in other parts of the world. This gives China the advantage of having massive amounts of data from the most populous country in the world to build and train their AI systems. Another major concern that one would have to deal with is bias. The humans training these machines have biases and these are being passed on to the machines. According to researchers at IBM, ‘AI systems are only as good as the data we put into them. Bad data can contain implicit racial, gender, or ideological biases. Many AI systems will continue to be trained using bad data, making this an ongoing problem. But we believe that bias can be tamed and that the AI systems that will tackle bias will be the most successful.’
What does the future portend?
While we have briefly looked at how AI could impact education, we should now return to a more fundamental question – what constitutes learning and what needs to be learnt? From the discussion above, it appears that AI systems would be adept at helping learn content. What about skills? A report by McKinsey global institute (2018) placed skills such as critical thinking, creativity, and collaboration as the most sought-after skill in the workforce a decade from now. If the process of AI systems outlined above is any indication, the odds are it will hinder the development of these skill sets in learners. For instance, a hands on learning innovation in Africa is hoping to equip learners with the skills for employment – ‘Educate! tackles youth unemployment by partnering with schools and governments to reform what schools teach and how they teach it, so that students in Africa have the skills to start businesses, get jobs, and drive development in their communities’. How can we bring about synergy between technology and the context on the ground? There is also the question of ethics involved in having robots as teaching assistants in the classroom. This had been briefly looked at in the future of learning module. Sharkley (2016) looked at this very phenomenon and concluded thus – ‘Children could form attachments to robot companions, or robot teachers and this could have a deleterious effect on their social development.’
Therefore, it is my opinion that, AI is not a panacea for education. It may not even make a considerable impact at scale. To ensure that a level of synergy is achieved between technology and non-technology practices, we should first strive to understand what purpose the new technology serves? It is also vital that learners and educators are well versed with digital learning environments before interacting with them. The Digital Citizen standard (ISTE) states the following – Students recognize the rights, responsibilities, and opportunities of living, learning, and working in an interconnected digital world, and they act and model in ways that are safe, legal and ethical. This standard assumes a lot of significance as the pace of technology grows exponentially. According to Kurzweil (2004),’ An analysis of the history of technology shows that technological change is exponential, contrary to the common-sense ‘intuitive linear’ view. So, we won’t experience 100 years of progress in the 21st century — it will be more like 20,000 years of progress (at today’s rate)’. And I personally believe from experience that this is true. Growing up in India in the late 80’s, it was hard for us to get a landline telephone connection. There was a waiting period of two years before we got our connection. Today, technology has acted as a great leveler and enabled anyone to get a mobile phone set up in under two hours. It is my experience of having seen the technological revolutions that has enabled widespread communication including the ability to take online classes from reputed universities in the comfort of my home at a fraction of the cost, that has led me to believe that AI does offer possibilities to improve educational outcomes across the world. How we go about all of this is the key. AI should be about Augmenting our Intelligence rather than replacing it.
Adams, S., Arel, I., Bach, J., Coop, R., Furlan, R., Goertzel, B., Hall, J., Samsonovich, A., Scheutz, M., Schlesinger, M., Shapiro, S. and Sowa, J. (2012). Mapping the Landscape of Human-Level Artificial General Intelligence. AI Magazine, 33(1), p.25.
Adaptive Learning to Traditional Learning. [online] Available at: https://www.informingscience.org/Publications/2165 [Accessed 1 May 2019].
Bloom, B. S. (1956).Taxonomy of Educational Objectives. Vol. 1: Cognitive Domain. New York: McKay.
Brownlee, J. (2019). A Gentle Introduction to Computer Vision. [online] Machine Learning Mastery. Available at: https://machinelearningmastery.com/what-is-computer-vision/ [Accessed 2 May 2019].
Built In. (2018). 7 companies using AI in education to enhance the classroom. [online] Available at: https://builtin.com/artificial-intelligence/ai-in-education [Accessed 1 May 2019].
Campbell, M., Hoane, A. and Hsu, F. (2002). Deep Blue. Artificial Intelligence, 134(1-2), pp.57-83.
Educate! | Preparing youth in Africa with the skills to succeed in today’s economy. (2019). The Educate! Model — Educate! | Preparing youth in Africa with the skills to succeed in today’s economy.. [online] Available at: https://www.experienceeducate.org/the-educate-model [Accessed 5 May 2019].
Forbes.com. (2019). How Much Data Do We Create Every Day? The Mind-Blowing Stats Everyone Should Read. [online] Available at: https://www.forbes.com/sites/bernardmarr/2018/05/21/how-much-data-do-we-create-every-day-the-mind-blowing-stats-everyone-should-read/#4bb8cf3360ba [Accessed 2 May 2019].
Informingscience.org. (2019). InformingSciJ – Informing and Performing: A Study Comparing
Iste.org. (2019). ISTE Standards for Students | ISTE. [online] Available at: https://www.iste.org/standards/for-students [Accessed 6 May 2019].
Kurzweil, R. (2004). The Law of Accelerating Returns. Alan Turing: Life and Legacy of a Great Thinker, pp.381-416.
Mitra, S. and Dangwal, R. (2010). Limits to self-organising systems of learning-the Kalikuppam experiment. British Journal of Educational Technology, 41(5), pp.672-688.
Legg, S. and Hutter, M. (2007). A Collection of Definitions of Intelligence. Proceedings of the 2007 conference on Advances in Artificial General Intelligence: Concepts, Architectures and Algorithms: Proceedings of the AGI Workshop 2006, [online] pp.17-24. Available at: https://dl.acm.org/citation.cfm?id=1565458 [Accessed 19 Apr. 2019].
Mcleod, S. (2019). Zone of Proximal Development and Scaffolding | Simply Psychology. [online] Simplypsychology.org. Available at: https://www.simplypsychology.org/Zone-of-Proximal-Development.html [Accessed 6 May 2019].
Prain, V., Cox, P., Deed, C., Dorman, J., Edwards, D., Farrelly, C., Keeffe, M., Lovejoy, V., Mow, L., Sellings, P., Waldrip, B. and Yager, Z. (2012). Personalised learning: lessons to be learnt. British Educational Research Journal, pp.1-23.
Sebba, J., Brown, N., Steward, S., Galton, M., James, M., Celentano, N. & Boddy, P. (2008) An investigation of personalised leaning approaches used by schools (Falmer, Research Report RR843, University of Sussex).
South China Morning Post. (2018). China’s schools are quietly using AI to mark students’ essays. [online] Available at: https://www.scmp.com/news/china/society/article/2147833/chinas-schools-are-quietly-using-ai-mark-students-essays-do [Accessed 6 May 2019].
Sharkey, A. (2016). Should we welcome robot teachers?. Ethics and Information Technology, 18(4), pp.283-297.
Skymind. (2019). Artificial Intelligence (AI) vs. Machine Learning vs. Deep Learning. [online] Available at: https://skymind.ai/wiki/ai-vs-machine-learning-vs-deep-learning [Accessed 2 May 2019].
Sri.com. (2019). [online] Available at: https://www.sri.com/sites/default/files/publications/2014-03-07_implementation_briefing.pdf [Accessed 1 May 2019].
Underwood, J., Baguley, T., Banyard, P., Coyne, E., Farrington-Flint, L., & Selwood, I. (2007). Impact 2007: Personalised learning with technology: Coventry, England: British Educational Communications and Technology Agency (Becta).
- V. Ramalingam et al 2018 J. Phys.: Conf. Ser. 1000 012030
View of Khan Academy as Supplemental Instruction: A Controlled Study of a Computer-Based Mathematics Intervention | The International Review of Research in Open and Distributed LearningIrrodl.org. (2019). View of Khan Academy as Supplemental Instruction: A Controlled Study of a Computer-Based Mathematics Intervention | The International Review of Research in Open and Distributed Learning . [online] Available at: http://www.irrodl.org/index.php/irrodl/article/view/2984/4221 [Accessed 7 May 2019]. Hu, D. (2019). How Khan Academy is using Machine Learning to Assess Student Mastery | David Hu . [online] David-hu.com. Available at: http://david-hu.com/2011/11/02/how-khan-academy-is-using-machine-learning-to-assess-student-mastery.html [Accessed 4 May 2019].
Woolf, B. 1992. AI in Education. Encyclopedia of Artificial Intelligence, Shapiro, S., ed., John Wiley & Sons, Inc., New York, pp. 434-444.
You, X., Zhang, C., Tan, X., Jin, S. and Wu, H. (2018). AI for 5G: Research Directions and Paradigms. [online] arXiv.org. Available at: https://arxiv.org/abs/1807.08671 [Accessed 2 May 2019].