Intelligence Unleashed Publication

Intelligence
Unleashed
An argument for AI in Education
Rose Luckin
Wayne Holmes
UCL Knowledge Lab,
University College London
Mark Griffiths
Laurie B. Forcier
Pearson
Open Ideas at Pearson

Sharing independent insights on the big,
unanswered questions in education
ABOUT
Open Ideas at Pearson

Pearson’s goal is to help people make progress
in their lives through learning. This means we are
always learning too.
This series of publications, Open Ideas, is one of

the ways in which we do this. We work with some
of the best minds in education – from teachers and
technologists, to researchers and big thinkers – to
bring their independent ideas and insights to a
wider audience.
How do we learn, and what keeps us motivated to

do so? What is the body of knowledge and skills
that learners need as we move into the second
half of the 21st century? How can smart digital
technologies be best deployed to realise the goal of
a more personalised education? How can we build
education systems that provide high quality learning
opportunities to all?
These questions are too important for the best ideas

to stay only in the lecture theatre, on the bookshelf,
or alone in one classroom. Instead they need to be
found and supported, shared and debated, adopted
and refined.
Our hope is that Open Ideas helps with this task,

and that you will join the conversation.
OPEN IDEAS AT PEARSON
Pearson UCL Knowledge Lab

Pearson is the world’s learning The UCL Knowledge Lab (previously
company, with expertise in known as the London Knowledge
educational courseware and Lab) is an interdisciplinary research
assessment, and a range of centre at the UCL Institute of
teaching and learning services Education, University College London.
powered by technology. Our mission is both to understand
and to design ways in which new
Our mission is to help people make digital technologies can support
progress through access to better and transform learning and teaching
learning. We believe that learning throughout the life course, at home
opens up opportunities, creating and at work.
fulfilling careers and better lives.
We start from the belief that new
technologies, when we fully exploit
their possibilities, will change not
only the ways we learn, but what
we learn, as well as how we work,
how we collaborate, and how we
communicate. Based on research
and evidence, we are devising new
pedagogies, implementing innovative
digital systems, developing new areas
of knowledge, and informing policy-
makers and educational stakeholders.
Creative Commons
This work is licensed under the Creative Commons Attribution
4.0 International Licence. To view a copy of this licence, visit
http://creativecommons.org/licenses/by/4.0 or send a letter
to Creative Commons, PO Box 1866, Mountain View, CA 94042, USA.
Suggested reference: Luckin, R., Holmes, W., Griffiths, M. & Forcier,

L. B. (2016). Intelligence Unleashed. An argument for AI in Education.
London: Pearson.
Copyright 2016
The contents of this paper and the opinions expressed herein are those
of the Authors alone.
ISBN: 9780992424886
ABOUT
The Authors
Rose Luckin
Rose Luckin is Professor of Learner Centred Design at the UCL Knowledge
Lab, University College London. She has been developing and writing about
artificial intelligence in education (AIEd) for over 20 years, and has been an
active member of the International AIEd society since its inception. Her research
explores how to increase participation by teachers and learners in the design
and use of technologies. In addition to over 50 peer-reviewed articles and two
edited volumes, Prof. Luckin is the author of Re-Designing Learning Contexts
(Routledge, 2010), and lead author of the influential Decoding Learning report
(Nesta, 2012).
Wayne Holmes
Wayne Holmes is a Researcher at the UCL Knowledge Lab, University College
London and he teaches about education and technology at the Graduate
School of Education, University of Bristol. He has been involved in education
and education research for more than 20 years, receiving his PhD in Education
(Learning and Technology) from the University of Oxford, and co-founding an
ed-tech web platform on which students answered 300+ million questions.
His research interests are in education, the learning sciences, and artificial
intelligence in education (AIEd).
Mark Griffiths
Mark Griffiths is Director of Research within the Office of the Chief Education
Advisor at Pearson. He leads on the office’s efforts to use world-class research
to influence and inform Pearson’s products and services. Prior to Pearson
he worked at Nesta – the UK’s innovation charity – where he invested in
over a dozen organisations that use technology or social action to improve
school-aged learning.
Laurie B. Forcier
Laurie Forcier leads the Open Ideas thought leadership series within the
Office of the Chief Education Advisor at Pearson. She has over 15 years’
experience in the education sector, covering research, evaluation, policy, and
administration. She was a member of the research team that produced Land
of Plenty, the final report of the Congressional Commission on the Advancement
of Women and Minorities in Science, Engineering, and Technology Development,
and is co-editor of Improving the Odds for America’s Children (Harvard Education
Press, 2014).
Acknowledgements
The authors wish to thank
Michael Barber and Amar Kumar at
Pearson for giving us the space and
encouragement that has allowed
us to form our own collaborative
learning group around AIEd. We
are also very grateful to the many
Pearson colleagues who have
provided helpful guidance to this
project or comments on this paper,
including John Behrens, Kristen
DiCerbo, Erin Farber, Josh Fleming,
José González-Brenes, Denis Hurley,
Johann Larusson, Nathan Martin,
Janine Mathó, Ashley Peterson-
Deluca, David Porcaro, and
Vikki Weston.
We also want to thank Mutlu

Cukurova, Beate Grawemeyer,
Manolis Mavrikis, and Kaska
Porayska-Pomsta of the UCL
Knowledge Lab, for their support
and colleagueship.
Two people deserve a special

acknowledgement: Ben du Boulay
for being a supportive and critical
friend of this project from the start,
and Joshua Underwood, co-author
of a 2011 paper that provided the
inspiration for this one.
INTELLIGENCE UNLEASHED
Contents
8 Foreword by Sir Michael Barber
11 Introduction
15
13 What is Artificial Intelligence (AI)? Will AI take over
from humans?
17 A brief introduction to Artificial 31
Intelligence in Education (AIEd) Teachers and AIEd
39
23 What AIEd can offer learning right now The ethics of AI
and AIEd
32 The next phase of AIEd 40
AIEd and the
41 Taking it to the next level: How AIEd Physical World
can help us respond to the biggest 52
unsolved issues in education Learning from the
approach that
jump-started
45 Bringing it all together: The continuing driverless cars
race between education and technology 53
Learning from
49 Recommendations to help us DARPA
Unleash Intelligence
56 References
8 OPEN IDEAS AT PEARSON
Foreword by
Sir Michael Barber
For thirty years I have attended conferences where
speakers have spoken to slides comparing images of an
early 20th century classroom with one from today, and have
pointedly asked: ‘why so little change?’ The modern variant
goes something like this: smart technologies have already
transformed so many parts of our lives – from how we date
to how we book a taxi. It would seem that there is no doubt
that AI will also significantly influence what we teach and
learn, as well as how we do it. And yet...
Adopting a puzzled stance as to why things have not

changed more has some value. It prompts us to examine our
assumptions, our habits, and our routines. It only takes us
so far, though. More is needed.
What we need – what we should demand – is an explanation

of why and how things could be different. First, we need to be
empowered by an understanding of what artificial intelligence
in education (AIEd) is, what it delivers, and how it goes about
doing that.
Second, we need a clear explanation of how the field of

artificial intelligence can connect to the core of teaching
and learning, so that we can avoid general-purpose
technologies being used in ways that do not deliver the
step changes in learner outcomes we seek. For example,
smart technologies that adapt to what is liked, rather than
what is learnt, or that deliver more efficient administration,
but not more efficient learning.
Third, we need concrete options that will allow us to make

the potential of AIEd real at the system level – that is, at the
scale that will allow it to support the teaching profession
broadly and impact positively on the learning experience
of each and every student. And fourth, we need to ask and
answer some profound ethical questions – for example,
about the acceptable uses that can be made of the data
that AIEd collects.
In other words, what we need is a degree of specificity about

AIEd that allows us to assess, invest, plan, deliver, and test.
This is what this paper offers – a useful primer on AIEd and
a compelling argument about what it can offer learning.
INTELLIGENCE UNLEASHED 9
From what AI is and how AIEd-driven learning systems are

built, onto its potential role in addressing the profound issue
of robots and machines taking over more and more current
jobs, it covers a vital range of topics with ease and elegance.
It is also a good read, with entertaining references from Pac-
Man and Stephen Hawking, sci-fi and ancient philosophy. And,
yes, it is understandable to a non-technical reader!
To make my own case for reading this paper, let me move

to a more local, anecdotal, level. Recently a member of my
Pearson team talked to me about a phonics learning app
he had bought for his young son. We could easily identify
the affordances that the technology brought – perfect
pronunciation of 42 phonics sounds, infinite patience, and
a healthy spillover of engagement from the software
to learning.
Yet, it was equally easy to identify ways in which some basic

AIEd techniques could have made the app so much better.
Content was re-presented even after it had been mastered,
which led to boredom. Other content was accessible even
though it was much too difficult, leading to frustration.
And there were no speech recognition capabilities present
to verify the learner’s pronunciation, or blending of sounds.
Asking for these features is not asking for science fiction.

Instead, it is asking us to incorporate findings from fields
like the learning sciences into AIEd tools so that these insights
are realised in cheaper, more effective ways. This paper offers
a long-list of where we should look for this combination of
learning insights and technology – for example, collaborative
learning, meta-cognition (or knowing about one’s own
thinking), useful feedback, and student motivation.
Funders and founders, policy makers and philanthropists

– in fact, anyone who takes seriously the urgent need to
embark on the next stage of education system reform –
should read and debate this paper. Only then will we (finally)
make good on the promise of smarter technologies for
learning (and, as a side effect, get rid of those boring slides).
Introduction
We wrote this short paper on artificial intelligence in
education (AIEd) with two aims in mind. The first was to
explain to a non-specialist, interested reader what AIEd is:
its goals, how it is built, and how it works. After all, only by
securing a certain degree of understanding can we move
beyond the science-fiction imagery of AI, and the associated
fears. The second aim was to set out the argument for what
AIEd can offer learning, both now and in the future, with an
eye towards improving learning and life outcomes for all.
Throughout, our approach has been to start with teaching

and learning – and then describe how well designed and
thoughtful AIEd can usefully contribute. Crucially we do not
see a future in which AIEd replaces teachers. What we do
see is a future in which the role of the teacher continues
to evolve and is eventually transformed; one where their
time is used more effectively and efficiently, and where their
expertise is better deployed, leveraged, and augmented.
Although some might find the concept of AIEd alienating,

the algorithms and models that comprise AIEd form the
basis of an essentially human endeavour. AIEd offers the
possibility of learning that is more personalised, flexible,
inclusive, and engaging. It can provide teachers and learners
with the tools that allow us to respond not only to what is
being learnt, but also to how it is being learnt, and how the
student feels. It can help learners develop the knowledge
and skills that employers are seeking, and it can help teachers
create more sophisticated learning environments than would
otherwise be possible. For example, AIEd that can enable
collaborative learning, a difficult task for one teacher to do
alone, by making sure that the right group is formed for the
task-at-hand, or by providing targeted support at just the
right time.
We look towards a future when extraordinary AIEd

tools will support teachers in meeting the needs of all
learners. Drawing on the power of both human and artificial
intelligence, we will lessen achievement gaps, address
teacher retention and development, and equip parents
to better support their children’s (and their own) learning.
Importantly, doing this will require much more than
borrowing the language of AI – we need to go deep,
harnessing the power of genuine AIEd, and then working
to apply it in real-life contexts at scale.
True progress will require the development of an AIEd

infrastructure. This will not, however, be a single monolithic
AIEd system. Instead, it will resemble the marketplace
that has developed for smartphone apps: hundreds and
then thousands of individual AIEd components, developed
in collaboration with educators, conformed to uniform
international data standards, and shared with researchers
and developers worldwide. These standards will enable
system-level data collation and analysis that help us learn
much more about learning itself and how to improve it.
If we are ultimately successful, AIEd will also contribute

a proportionate response to the most significant social
challenge that AI has already brought – the steady
replacement of jobs and occupations with clever algorithms
and robots. It is our view that this phenomena provides a new
innovation imperative in education, which can be expressed
simply: as humans live and work alongside increasingly smart
machines, our education systems will need to achieve at
levels that none have managed to date.
Our response, we argue, should be to take on the role

of metaphorical judo masters. That is, we should harness
the power and strength of AI itself. In that way we can help
teachers to equip learners – whatever their age – with the
knowledge and flexible skills that will allow them to
unleash their human intelligence and thrive in this
re-shaped workforce.
To be candid, the impetus for this paper arose from our

impatience with the status quo. Despite nearly three
decades of work, AIEd is in many ways still a cottage industry,
and the benefits and enormous potential of the field remain
mostly unrealised. Sadly, many of the best ideas in AIEd
currently make it no further than the lab, or perhaps a lecture
hall. AIEd is hampered by a funding system that encourages
siloed research, and that shies away from dealing with the
essential messiness of education contexts. We believe this
needs to change.
This is our attempt at contributing to that change, through

explaining, arguing, and putting forward some evocative, and
perhaps provocative, views of the future. It is our hope that
this paper will provide a deeper understanding of AIEd, and
stimulate a much-needed debate.
Let us start by introducing AI.

What is Artificial
Intelligence (AI)?
It can be difficult to define artificial intelligence (AI), even for ALGORITHM

A defined list of
experts. One reason is that what AI includes is constantly steps for solving
shifting. As Nick Bostrom, a leading AI expert from Oxford a problem.
A computer program
University, explains: “[a] lot of cutting edge AI has filtered into
can be viewed as an
general applications, often without being called AI because elaborate algorithm.
once something becomes useful enough and common In AI, an algorithm is
usually a small pro-
enough it is not labeled AI anymore.”1 Instead, it is considered cedure that solves a
a computer program, or an algorithm, or an app, but not AI. recurrent problem.
MACHINE LEARNING
Another reason for the difficulty in defining AI is the Computer systems
that learn from data,
interdisciplinary nature of the field. Anthropologists, enabling them to
biologists, computer scientists, linguists, philosophers, make increasingly
psychologists, and neuroscientists all contribute to the better predictions.
field of AI, and each group brings their own perspective DECISION THEORY
The mathematical
and terminology. study of strategies
for optimal decision-
For our purposes, we define AI as computer systems that making between
options involving
have been designed to interact with the world through different risks or
capabilities (for example, visual perception and speech expectations of gain
or loss depending on
recognition) and intelligent behaviours (for example,
the outcome.
assessing the available information and then taking the
most sensible action to achieve a stated goal) that we
would think of as essentially human.2
The use of AI in our day-to-day life is increasing ever more

rapidly. For example, AI scientists are currently building on
new approaches in machine learning, computer modelling,
and probability statistics to improve financial decision
making3, and are using decision theory and neuroscience to
drive the development of more effective medical diagnostics.4
And with the recent launch of OpenAI, a non-profit artificial
intelligence research company with an initial investment
of $1B, we expect this acceleration to continue apace –
including, we predict, in the area of AIEd.5
Will AI take over

from humans?
Some in the scientific community General AI is very different from the
worry that AI is a Pandora’s box with “domain specific AI” most of us are
dangerous consequences. As far familiar with. These domain specific
back as 1993, the computer scientist AIs focus on one thing – for example,
Vernon Vinge popularised the notion mastering chess (Deep Blue or Giraffe),
of the singularity, the point at which or Go (Google’s Deep Mind), driving
an AI-powered computer or robot a car (Google’s self-driving cars), or
becomes capable of redesigning and recognising a passport photograph as
improving itself or of designing AI a representation of a particular person.
more advanced than itself. Inevitably, Even for leading AI advocates, the
it is argued, this will lead to AI that singularity appears to be due to arrive
far exceeds human intelligence, at some ever-receding future date,
understanding, and control, and to usually around thirty years from the
what Vinge describes as the end of the time they are writing!8
human era.6 More recently, Stephen
Hawking and other leading scientists However, AI is getting more
including Stuart Russell, Max Tegmark, sophisticated, and it is already having
and Frank Wilczek have also warned a profound impact on our economy.
us about the potential downsides of AI In a 2013 study, the economists Frey
becoming too clever.7 and Osborne used AI techniques
themselves to begin the task of
This worrying idea has fed Hollywood identifying the effects of automation
films for decades, from 2001: A Space on the jobs market. According to
Odyssey in the 60s, the Terminator their estimates about 47 percent of
series of the 80s, and the more recent current US jobs are at high-risk of
Transcendence – all of which depict being carried out by machines in the
one version or another of a dystopian next decade or two.9 To date, it is
world dominated by out-of-control AI. middle-income jobs that have been
most affected, reflecting the tasks
Before we get too worried, however, that are currently most susceptible
it is worth making a note about to automation.10
the current capabilities of artificial
intelligence. Major advances Past waves of profound changes in
in ‘general AI’, AI that that could the economy (for example, the move
successfully perform any intellectual from an agricultural driven economy
task that a human being could, would to an industrial one) have been
be necessary for any singularity to accommodated by changing the
occur. And right now, general AI does reach and the nature of education
not exist. and learning.
The understandable guiding thought that we could not have previously

has been that if “workers have predicted or imagined. Others,
flexible skills and if the educational like Martin Ford, believe that as
infrastructure expands sufficiently, automation takes over more and more
[then] … the race between technology sophisticated tasks, the number of
and education will not be won by jobs will be simply too small to sustain
either side and prosperity will be current employment rates.12
widely shared.”11
We take an optimistic stance that
Whether this will hold true in the face aligns with the historical view. We also
of rapidly developing AI is contested. believe that AIEd has a major role to
An historical view provides evidence play in helping us prepare for the new
that technological change has always roles that the economy will create – a
brought with it a host of new roles topic we return to later in this paper.
A brief
introduction
to Artificial
Intelligence
in Education
(AIEd)
The application of artificial intelligence to education (AIEd) has ADAPTIVE LEARNING

ENVIRONMENTS
been the subject of academic research for more than 30 years. A digital learning
The field investigates learning wherever it occurs, in traditional environment that
adapts teaching and
classrooms or in workplaces, in order to support formal education
learning approaches
as well as lifelong learning. It brings together AI, which is itself and materials to
interdisciplinary, and the learning sciences (education, psychology, the capabilities and
needs of individual
neuroscience, linguistics, sociology, and anthropology) to promote learners.
the development of adaptive learning environments and other
MODELS
AIEd tools that are flexible, inclusive, personalised, engaging, These represent
and effective. something from
the real world in a
computer system
At the heart of AIEd is the scientific goal to “make computationally or process, to assist
precise and explicit forms of educational, psychological and social calculations and
predictions.
knowledge which are often left implicit.” 13 In other words, in addition to
being the engine behind much ‘smart’ ed tech, AIEd is also a powerful
tool to open up what is sometimes called the ‘black box of learning,’
giving us deeper, and more fine-grained understandings of how
learning actually happens (for example, how it is influenced by the
learner’s socio-economic and physical context, or by technology).
These understandings may then be applied to the development of

future AIEd software and, importantly, can also inform approaches
to learning that do not involve technology. For example, AIEd can
help us see and understand the micro-steps that learners go through
in learning physics, or the common misconceptions that arise.14
These understandings can then be used to good effect by
classroom teachers.
As we have said, AI involves computer software that has been

programmed to interact with the world in ways normally requiring
human intelligence. This means that AI depends both on knowledge
about the world, and algorithms to intelligently process
that knowledge.
This knowledge about the world is represented in so called ‘models’.

There are three key models at the heart of AIEd: the pedagogical
model, the domain model, and the learner model.
Take the example of an AIEd system that is designed to provide

appropriate individualised feedback to a student. Achieving this
requires that the AIEd system knows something about:
• E
ffective approaches to teaching (which is represented in a
pedagogical model)
• The subject being learned (represented in the domain model)
• The student (represented in the learner model)
Examples of the specific knowledge that might be integrated into

each of these models is displayed opposite.
AIEd models What the model represents Examples of specific knowledge

represented in AIEd models
Pedagogical The knowledge and expertise ‘Productive failure’ (allowing

model of teaching students to explore a concept
and make mistakes before being
shown the ‘right’ answer)
Feedback (questions, hints, or
haptics), triggered by student actions,
which is designed to help the
student improve their learning
Assessment to inform and
measure learning
Domain model Knowledge of the subject being How to add, subtract, or multiply
learned (domain expertise) two fractions
Newton’s second law (forces)
Causes of World War I
How to structure an argument
Different approaches to reading
a text (e.g. for sense or for detail)
Learner Knowledge of the learner The student’s previous

model achievements and difficulties
The student’s emotional state
he student’s engagement in
T
learning (for example: time-
on-task)
To delve deeper into just one of these examples,

learner models are ways of representing the interactions
that happen between the computer and the learner.
The interactions represented in the model (such as the
student’s current activities, previous achievements,
emotional state, and whether or not they followed
feedback) can then be used by the domain and pedagogy HAPTICS
components of an AIEd programme to infer the success Any form of
of the learner (and teacher). The domain and pedagogy interaction involving
touch. In the case
models also use this information to determine the next of AIEd, this could
most appropriate interaction (learning materials or be something like
learning activities). Importantly, the learner’s activities are a smartphone
vibration in
continually fed back into the learner model, making the response to a
model richer and more complete, and the system ‘smarter’. correct answer.
So, what would a piece of education technology driven by

AIEd look like? Figure 1 (shown opposite) is a simplified picture
of a typical model-based adaptive tutor. It is based on the
three core models as described above: the learner model
(knowledge of the individual learner), the pedagogy model
(knowledge of teaching), and the domain model (knowledge
of the subject being learned and the relationships between
the different parts of that subject matter). AIEd algorithms
(implemented in the system’s computer code) process
that knowledge to select the most appropriate content to
be delivered to the learner, according to their individual
capabilities and needs.
While this content (which might take the form of text, sound,
activity, video, or animation) is being delivered to the learner,
continuous analysis of the learner’s interactions (for example,
their current actions and answers, their past achievements,
and their current affective state) informs the delivery of
feedback (for example, hints and guidance), to help them
progress through the content they are learning.
Deep analysis of the student’s interactions is also used to

update the learner model; more accurate estimates of the
student’s current state (their understanding and motivation,
for example) ensures that each student’s learning experience
is tailored to their capabilities and needs, and effectively
supports their learning.
Some systems include so-called Open Learner Models, which

present the outcomes of the analysis back to the learners and
teachers. These outcomes might include valuable information
about the learner’s achievements, their affective state, or
any misconceptions that they held. This can help teachers
understand their students’ approach to learning, and allows
them to shape future learning experiences appropriately.
For the learners, Open Learner Models can help motivate
them by enabling them to track their own progress, and
can also encourage them to reflect on their learning.
One of the advantages of adaptive AIEd systems is that they

typically gather large amounts of data, which, in a virtuous
circle, can then be computed to dynamically improve the
pedagogy and domain models. This process helps inform
new ways to provide more efficient, personalised, and
contextualised support, while also testing and refining our
understanding of the processes of teaching and learning.
Figure 1
AIEd system showing a simplified picture of a typical
model-based adaptive tutor.
Domain
Model
Algorithms
Pedagogy Learner
Processing the knowledge
Model Model
represented in the models
Learner Interface
Adaptive Content
Learning content (e.g. text or video) adapted to
the needs and capabilities of the individual learner
Data Capture
Feedback
The learner’s interactions, achievements,

affect (emotion), speech...
Data Analysis
AI techniques (such as machine
learning and pattern recognition)
Open Learner Model
Making the learning explicit,

for teachers & learners to see
In addition to the learner, pedagogical, and domain models,

META-COGNITIVE
AIEd researchers have also developed models that represent Metacognition is
the social, emotional, and meta-cognitive aspects of learning. sometimes defined
simply as ‘knowing
This allows AIEd systems to accommodate the full range of about one’s thinking’.
factors that influence learning. It has two elements:
being aware of
thinking and being
Taken together, this set of increasingly rich AIEd models might able to control or
become the field’s greatest contribution to learning. regulate it.
What AIEd can

offer learning
right now
A multitude of AIEd-driven applications are already in use EDUCATIONAL

DATA MINING
in our schools and universities. Many incorporate AIEd and The development
educational data mining (EDM) techniques to ‘track’ the and use of methods
to analyse and
behaviours of students – for example, collecting data on
interpret the ‘big
class attendance and assignment submission in order data’ that comes
to identify (and provide support) to students at risk of from computer-
based learning
abandoning their studies. systems and from
school, college,
Other AI researchers are exploring novel user interfaces, or university
administrative
such as natural language processing, speech and gesture and management
recognition, eye-tracking, and other physiological systems.
sensors, which could be used to augment both AIEd
and non-AIEd software.
Here, however, we focus on three categories of AIEd

software applications that have been designed to support
learning most directly: personal tutors for every learner,
intelligent support for collaborative learning, and intelligent
virtual reality.
AIEd can provide an intelligent,

personal tutor for every learner
One-to-one human tutoring has long been thought to
be the most effective approach to teaching and learning
(since at least Aristotle’s tutoring of Alexander the Great!)
Unfortunately, one-to-one tutoring is untenable for all
students. Not only will there never be enough human
tutors; it would also never be affordable. All of this begs the
question: how can we make the positive impact of one-to-one
tutoring available to all learners across all subjects?
This is where Intelligent Tutoring Systems (ITS) come in. ITS

use AI techniques to simulate one-to-one human tutoring,
delivering learning activities best matched to a learner’s
cognitive needs and providing targeted and timely feedback,
all without an individual teacher having to be present. Some
ITS put the learner in control of their own learning in order
to help students develop self-regulation skills; others use
pedagogical strategies to scaffold learning so that the learner SCAFFOLD
is appropriately challenged and supported. In the context
of education,
scaffolding is a
The 1970s brought some of the first AI systems to offer teaching method
individualised and adaptive instruction. For example that enables a
student to solve a
BUGGY15, a ground-breaking system designed to teach basic problem, carry out
addition and subtraction, used a model of the possible a task, or achieve
misconceptions that learners might exhibit in their procedural a goal through the
gradual scaling
arithmetic. This ‘bug library’, effectively the system’s domain back of outside
model, was used to diagnose each error that a student made assistance.
so that appropriate tutoring could be offered. Initially, it was NEURAL NETWORKS

Networks of
limited by the bugs that it could recognise, those that had interconnected
been included in the original code. Over time, additional data sets, based on
a vastly simplified
misconceptions were found and added to the library.
understanding
of brain neural
Instead of models, many recent ITS use machine learning networks.
techniques, self-training algorithms based on large data sets,

and neural networks, to enable them to make appropriate
decisions about what learning content to provide to the
learner. However, with this approach, it can be difficult to
make the rationale for those decisions explicit.
Modern model-based adaptive systems can be far more

flexible. They enable the rationale for each decision taken
by the system to be made explicit and understandable
by humans (and thus potentially applicable to classroom
teaching). Throughout the last decade, increasingly
sophisticated learner, pedagogy, and domain models have
been introduced in numerous adaptive tutors to support
the individualisation of learning.
For example, the iTalk2Learn system16, designed to help

young students learn about fractions, used a learner model
that included information about the learner’s mathematics
knowledge, their cognitive needs, their affective (emotional)
state, the feedback they had received and their responses to
that feedback.
Model-based adaptive tutors can include a range of AIEd

tools that:
•M
odel learners’ cognitive and affective states17
• Use dialogue to engage the student in Socratic learning
experiences, that is learning experiences that involve
enquiry and discussion, questioning and answering18
• Include open learner models to promote reflection and
self-awareness19
• Adopt meta-cognitive scaffolding (for example, by providing
dynamic help or using a narrative framework) to increase
learner motivation and engagement20
• Use social simulation models – for example, to enable
language learning students to engage more successfully
with speakers of their target language by understanding
cultural and social norms21
AIEd can provide intelligent

support for collaborative learning
Research over decades has suggested that collaboration,
whether between a pair of students undertaking a project
together or a community of students participating in an
online course, can foster higher learning outcomes than
learning alone.22 Collaborative learning is effective because
it encourages participants to articulate and justify their
thinking, to reflect on other explanations, to resolve
differences through constructive dialogue, and to build
shared knowledge and meaning. Collaborative learning can
also enhance motivation; if students care about the group,
they become more engaged with the task and achieve better
learning outcomes.23
However, research also suggests that collaboration between

learners does not happen spontaneously.24 For example,
group members might not have the social interaction skills
needed to collaborate effectively. This can be especially
difficult in the context of online collaborations, where
participants rarely meet in person.
This is where AIEd can contribute. Several approaches have

been investigated and here we focus on four: adaptive group
formation, expert facilitation, virtual agents, and intelligent
moderation.
Adaptive group formation

This uses AI techniques and knowledge about individual
participants, most often represented in learner models,
to form a group best suited for a particular collaborative
task. The aim might be to design a grouping of students all
at a similar cognitive level and of similar interests, or one
where the participants bring different but complementary
knowledge and skills.25
Expert facilitation
Here models of effective collaboration – known as
‘collaboration patterns’ – are used to provide interactive
support to the collaborating students.26 These patterns
are either provided by the system authors or mined from
previous collaborations. For example, AI techniques such as
machine learning or Markov modelling have been used to
MARKOV
identify effective collaborative problem-solving strategies. MODELLING
These can then be used to train systems to recognise when An approach used
students are having trouble understanding the concepts in probability
theory to represent
that they are sharing with each other, or to provide targeted randomly changing
support of the right form at the right time. systems.
They can also show students (and their teachers)

how well an individual is contributing to group work,
an historically difficult activity to parse and assess.27
Intelligent virtual agents

A third approach involves intelligent virtual agents
that are introduced into the collaborative process.28
These AI agents might mediate online student
interaction, or simply contribute to the dialogues,
by acting as:
• A
n expert participant (a coach or a tutor)
• A
virtual peer (an artificial student at a similar
cognitive level to the learner, but one who is
capable of introducing novel ideas)
• S
omeone the participants might themselves teach
– for example, the artificial student might hold
deliberate misconceptions, or provide alternative
points of view to stimulate productive argument
or reflection29
Intelligent moderation
With large student numbers working in multiple
collaborative groups, it can be impossible for a
person to make any sense of the large volume
of data that the participants are generating in
their discussions.
Intelligent moderation uses AI techniques such as

machine learning and shallow text processing to
analyse and summarise the discussions to enable a
human tutor to guide the students towards fruitful
SHALLOW TEXT
collaboration. For example, the system might PROCESSING
provide alerts to human tutors to inform them of A method of text
significant events (such as students going off topic analysis that
identifies – but does
or repeating misconceptions) that may require not ‘understand’ –
their intervention or support.30 particular words.
Intelligent virtual reality to support

learning in authentic environments
Artificial intelligence first appeared in a digital game in 1979,
when the developers of Pac-Man used a technique known
as state machine (transitioning between states depending
on conditions) to control whether or not a ghost ran towards
or away from a player. The AI in most modern digital games
builds on this simple approach. As the game-based story
unfolds, autonomous non-player characters (agents) take
information from both the game and the player and, based
on that information, use AI algorithms to determine the most
appropriate actions to take.
Virtual reality for learning works in a similar way. It provides

authentic immersive experiences (the subjective impression
that one is participating in a realistic experience) that simulate
some aspect of the real world to which the user would not
otherwise have access (such as dangerous environments
or somewhere geographically or historically inaccessible).
Research has shown that giving opportunities for students to
explore, interact with, and manipulate aspects of a simulated
world, perhaps investigating ‘what if’ scenarios, (such as, ‘what
if there is a drought?’), enables them to transfer what they
have learnt to the real world.31
For example, a virtual submarine might allow

the user to shrink to a microscopic level to
investigate natural processes that occur under
the surface of a rock pool, or the student might
be able to explore a nuclear power plant, Ancient
Rome or the outer planets.
Virtual reality becomes ‘intelligent’ when it is

augmented with artificial intelligence. AI might be
used simply to enhance the virtual world, giving it
the ability to interact with and respond to the user’s
actions in ways that feel more natural. Or, drawing
on Intelligent Tutoring Systems, AI might also be
integrated to provide on-going intelligent support
and guidance to ensure that the learner engages
properly with the intended learning objectives
without becoming confused or overwhelmed.
Virtual pedagogical agents might also be included,

acting as teachers, learning facilitators, or student
peers in collaborative learning ‘quests’. These
agents might provide alternative perspectives,
ask questions, and give feedback, all based on a
properly specified pedagogical model.
Many studies have demonstrated that immersion

in intelligent virtual reality can enhance educational
outcomes, enabling students to construct their
own individual understanding of the world being
explored.32 Some have also been shown to have
the potential to release what Chris Dede, a leading
learning scientist, calls ‘trapped intelligence’ – that
is, they allow low-achieving students to build their
self confidence by shifting their self image from
being a poor academic performer to, for example,
a successful virtual scientist.33
In addition, intelligent synthetic characters in virtual

worlds can play roles in settings that are too dangerous or
unpleasant for learners. For example, FearNot is a school-
based intelligent virtual environment that presents bullying
incidents in the form of a virtual drama. Learners, who have
been victims of bullying, play the role of an invisible friend to
a character in the drama who is bullied. The learner offers
the character advice about how to behave between episodes
in the drama and, in so doing, explores bullying issues and
effective coping strategies.34
Intelligent virtual reality can also be used for intelligent team

training during which virtual humans are able to reason about
individual events, carry out actions, and negotiate options,
with the aim of guiding human trainees to make similar
assessments – for example in peacekeeping scenarios.35
...
Taken together, these three types of applications have been

used to create learning environments that are not just more
personalised, but also more inclusive and engaging. For
example, they can provide additional help for learners with
special educational needs, motivate learners who cannot
attend school, and support disadvantaged populations.36
AIEd applications can also be more flexible than the

alternatives. Many are deployed online, meaning that they
can be available on personal and portable devices within,
and beyond, formal educational settings. AIEd researchers
are also exploring the use of mobile devices to deliver
adaptive materials for anytime, anywhere, social and
collaborative learning (while still monitoring and providing
intelligent support as needed).37
AIEd has made great progress, but has barely scratched the
surface. There is exciting promise as the existing technologies
develop, mature, and scale. Yet, the AIEd horizon includes
much more than simply ‘more of the same’. AIEd developers
are getting better at recognising how to blend human and
machine intelligence effectively, which means that future
AIEd is poised to make significant strides in a number of
critical areas.
Let’s now turn our attention to this horizon.

Teachers
and AIEd
We are in no doubt that teachers As this transformation takes place,
need to be central agents in the teachers will need to develop new
next phase of AIEd. In one sense skills (maybe through professional
this is obvious – it is teachers who development delivered through
will be the orchestrators of when, an AIEd system). Specifically they
and how, to use these AIEd tools. will need:
In turn, the AIEd tools, and the data
driven insights that these tools • A
sophisticated understanding
provide, will empower teachers to of what AIEd systems can do to
decide how best to marshal the enable them to evaluate and make
various resources at their disposal. sound value judgements about
new AIEd products
More than this, though, teachers • T
o develop research skills to
– alongside learners and parents – allow them to interpret the data
should be central to the design of provided by AIEd technologies, to
AIEd tools, and the ways in which they ask the most useful questions of the
are used. This participatory design data, and to walk students through
methodology will ensure that the what the data analysis is telling
messiness of real classrooms is taken them (for instance, using Open
into account and that the tools deliver Learner models)
the support that educators need –
not the support that technologists or • N
ew teamworking and management
designers think they need. Teachers skills as each teacher will have AI
who take part in these processes will assistants in addition to their usual
gain increased technological literacy, human teaching assistants, and they
new design skills, and a greater will be responsible for combining
understanding of what AIEd systems and managing these resources
can offer. most effectively
As mentioned earlier, we predict that Most excitingly, with the evolution

the increased introduction of AI- of the teacher’s role will also come
powered tools will serve as a catalyst the evolution of the classroom, as
for the transformation of the role of AIEd tools allow us to realise what it
the teacher. AIEd is well placed to take is unrealistic to expect any teacher
on some of the tasks that we currently or lecturer to do alone. For example,
expect teachers to do – marking and making the positive impact of one-to-
record keeping, for example. one tutoring available to every child,
or realising effective collaborative
Freedom from routine, time- learning (a difficult activity to keep on
consuming tasks will allow teachers track without some form of additional
to devote more of their energies support).
to the creative and very human acts
that provide the ingenuity and
empathy needed to take learning
to the next level.
The next
phase of AIEd
The future of AIEd is inextricably linked to the future of AI.

The increasing consumerisation of AI technologies brings
with it a massive increase in the number of people who are
developing AI. The pace of innovation and development in
general is at its fastest rate ever38 and the current popularity
of AI should mean that innovation in AIEd is a focus of
attention for an increasing number of businesses.
In this section, we look to the edges of theory and practice

to consider some of the advances we expect to see through
the continued growth of AIEd, as well as some of the new
technologies we expect to be developed. Some of these will
happen sooner than others, and all can be seen as both
opportunities and challenges. Having said this, we see these
developments in a positive light, and have here focused on
conveying the exciting potential that AIEd has to improve
education for all.
AIEd will help learners gain 21st century skills

There is an increasing recognition that so-called 21st century
skills are essential for current and future work environments,
with many groups advancing lists (some short and others
long!) of the skills people will need to fully engage in
employment and society.
To take one example, the World Economic Forum have

proposed 16 skills, split across three categories:39
Category Helps students Related skills

approach...
Foundational ...everyday Literacy Technology literacy

Literacies tasks
Numeracy Financial literacy
Scientific literacy ultural and civic
C
literacy
Information
communication
Competencies ...complex ritical thinking and

C Communication
challenges problem-solving
Collaboration
Character ...changing Curiosity Leadership

Qualities environments
Initiative ocial and cultural
S
awareness
Persistence/grit
Adaptability
We agree with the common wisdom that skills like

these are – and always have been – important, and
that they should be part of any approach to lifelong
learning. There are, however, at least two salient
challenges that need to be addressed if we are to
realise this agenda:
1 W
e must develop reliable and valid indicators
that will allow us to track learner progress on all
the skills and capabilities needed to thrive in the
current century – at the level of the individual,
the district, and the country. This will need to
include difficult to measure characteristics such
as creativity and curiosity.
2 W
e need a better understanding of the most
effective teaching approaches and the learning
contexts that allow these skills to be developed.
AIEd can help with both.
First, AIEd has the tools and techniques to conduct

the fine-grained analysis that allows us to track
each learner’s development of skills and capabilities
as they interact and learn over time. This tracking
of individual learners can then be collated and
interpreted as required to provide knowledge about
progress at the school, district, and country level.
The increasing range of data capture devices –

such as biological data, voice recognition, and eye
tracking – will enable AIEd systems to provide new
types of evidence for currently difficult to assess
skills. For example, a practice-based learning
experience that incorporates elements of problem
solving or collaboration might be assessed using
a combination of data sources including voice
recognition (to identify who is doing and saying what
in a team activity) and eye tracking (to explore which
learner is focusing on which learning resources at
any particular moment in time).
Second, the increasing use of AIEd systems PRACTICE-BASED

will enable the collection of mass data about LEARNING
Practice-based
which teaching and learning practices work best. learning is learning
This data will enable us to track learner progress undertaken,
typically in teams,
against different teaching approaches and, in turn, in real-life contexts
will allow us to develop a dynamic catalogue of the (e.g. learning
best teaching practices suited to the development undertaken
by healthcare
of different skills and capabilities, in particular the professionals in
21st century skills, across a range of environments. working hospitals).
Importantly, in investigating these practices, we will also

be able to relate learner progress to the contexts in which
learning has taken place, and then build context models into
our AIEd systems. Already there are fledgling methodologies
that take into account contextual elements such as the
physical or virtual space, the people who are available to
help, and the learning tools available such as the curriculum,
technology, or books.40
Over time, these models will enable us to identify the best

teaching approaches for different contexts. And, they
will help us identify how contextual factors (such as the
combinations of technology, teachers and the environment)
can be adjusted to improve the efficacy of particular teaching,
approaches – insights that will help students gain 21st
century skills as well as other types of knowledge.
AIEd will support a

Renaissance in Assessment
We echo the assertion made by Peter Hill and Michael Barber
in Preparing for a Renaissance in Assessment, that of the three
core components of learning (curriculum, learning and
teaching, and assessment), it is in many ways assessment that
is holding us back.41 We also agree with the assertion that
technology holds part of the solution. In the near future,
we predict that AIEd will contribute to improving assessment
in three key ways.
AIEd will provide just-in-time

assessments to shape learning
The continued and growing use of technologies in education
will allow increasing amounts of data to be collected about
teachers and learners. This so called ‘big data’ is already being
studied using learning analytics to recognise data patterns
of potential educational interest. For example, analytics have
been used with high levels of accuracy to predict when a
student is likely to fail an assessment or ‘drop-out’ from an
online course.
LEARNING
Soon the sophistication of these learning analytics will be ANALYTICS
complemented by AI techniques to provide just-in-time Learning analytics
are used to find
information about learner successes, challenges, and needs patterns in large
that can then be used to shape the learning experience itself. data sets, like those
For example, AIEd will enable learning analytics to identify generated by online
learning systems,
changes in learner confidence and motivation while learning to enable modelling
a foreign language, say, or a tricky equation. and prediction.
This information can then be used to provide timely

interventions to help students – which could be in the form
of technology-assisted support, individual attention from a
teacher, or some combination of the two.
AIEd will provide new insights into how learning

is progressing
In addition to timeliness, the data gleaned from digital
teaching and learning experiences will give us new insights
that cannot be ascertained from existing assessments.
For example, as well as identifying whether or not a learner
gave the correct answer, datasets could be analysed to help
teachers understand how the learner arrived at their answer.
The data might also help us better understand cognitive
processes such as remembering and forgetting, and the
fundamental impact that these have on learning and student
outcomes. AIEd analysis might also identify if and when a
student is confused, bored, or frustrated, to help teachers
understand and enhance a learner’s emotional readiness
for learning.
AIEd will help us move beyond ‘stop-and-test’

As documented by Kristen DiCerbo and John Behrens in
Impacts of the Digital Ocean on Education, the models and
techniques developed by AIEd researchers over the last 25
years have resulted in an ever rising ocean of digital data
on learning and teaching, telling us much about the data we
need to collect in order to assess students while they learn.42
With ongoing AIEd analysis of a student’s learning

activities, there will be no need for the stop-and-test
approach that characterises many current assessments.
Instead of traditional assessments that rely upon
evaluating small samples of what a student has been
taught, AIEd-driven assessments will be built into
meaningful learning activities, perhaps a game or a
collaborative project, and will assess all of the learning
(and teaching) that takes place, as it happens.43
AIEd will embody new insights

from the learning sciences
AI and AIEd have always been interdisciplinary fields. Moving
forward, AIEd will continue to leverage new insights in
disciplines such as psychology and educational neuroscience
to better understand the learning process, and so build
more accurate models that are better able to predict – and
influence – a learner’s progress, motivation, and perseverance.
An example from Education Neuroscience

One example of the way that neuroscience can inform
education and the design of AIEd systems can be found in
the work of Paul Howard-Jones, Professor of Neuroscience
and Education at the University of Bristol. His work suggests
that learning can be improved when it is linked to uncertain
rewards44 – that is, situations in which a learner knows that
a reward may be given upon their completion of a task,
but there is no certainty that the reward will appear on
every occasion. This is counter-intuitive to typical education
practices where rewards are consistently related to success.
The use of uncertain rewards is much more common in the

world of computer games, hence the current interest in the
design of educational games that use the motivational impact
of uncertain rewards to engage learners and to enhance
their learning. The addition of AIEd techniques to the design
of these educational games would enable, for example, the
provision of uncertain rewards to be calibrated to a learner’s
individual reaction to a given level of uncertainty.
An example from Psychology

For several years now psychologists, most notably Carol
Dweck of Stanford University, have been exploring the role
of ‘mindsets’ in learning.45 They make a distinction between
learners who believe that intelligence does not change over
time (a ‘fixed mindset’) and those who believe that their
abilities can be developed (a ‘growth mindset’). Learners with
a growth mindset see challenges as things to be overcome;
they persist and value effort more, which leads to them
enjoying more success as learners. There is increasing
evidence that a growth mindset can be taught and that
changing students’ mindsets can have a substantial impact
on their grades and achievement test scores.46
There is already a role for technology when it comes to

helping learners to develop a growth mindset; indeed
Carol Dweck’s team have developed Brainology, a piece
of software to provide support and content for promoting
a growth mindset.47 The addition of AI to the technology
would bring greater possibilities. For example, with AI

the system could adapt to a learner’s goal orientation or
mindset, or scaffold learners towards a growth mindset.48
More sophisticated learner models would be able to capture
learners’ mindsets, including how these change over time,
and adapt teaching accordingly. This might include providing
targeted feedback to teachers to enable them to support
each learner to develop a growth mindset in the most
effective way.
AIEd will give us lifelong learning partners

It is said that in ancient China each royal prince studied
with a companion as well as a royal teacher. Perhaps the
Chinese emperors knew that their children would learn
more effectively with another; it is certainly the belief in
much contemporary psychology.49
Early AIEd research in the 1980s brought this ancient story

to life through the development of Learning Companion
Systems. These systems provided each learner with a
collaborative computer-based learning companion (or
companions). It was the role of the companion to use
collaboration and competition to stimulate student
learning. The companion could also act as a student for
the human learner to tutor, and in so doing the student
learns by teaching. The computer-based teacher offered
examples and guidance to human and computer student
alike and determined the order and content of the topics
to be tackled.50
The next generation of learning companions will offer

huge potential for future teaching and learning. There are
no technical barriers to the development of learning
companions that can accompany and support individual
learners throughout their studies – in and beyond school.
These lifelong learning companions could be based in the
cloud, accessible via a multiplicity of devices, and be
operated offline as needed.
Rather than teaching all subject areas, the learning

companion might call in specialist AIEd systems or humans
with expertise in the particular subject area required by the
learner. In addition, the companion could focus on helping
learners to become better at learning through developing a
growth mindset or an impressive array of 21st century skills.
And because this type of system can help all learners to
access learning resources that are optimal for their needs, it
will be suitable for struggling learners as well as those who
are high achieving.
intellectual property concerns. Indeed,

The ethics of AI the growing volume and diversity of
data generated by AIEd systems only
and AIEd serves to double-down on the already
existing ethical concerns about what
AI development is accelerating, happens to education data. What
permeating every aspect of our lives. are the implications of the methods,
The question is, are we prepared technologies, and ideologies that
to let that happen without proper underpin the generation, analysis,
debate or control? The ethics of interpretation, and use of AIEd system
AI, as written about extensively data? Who owns the data, who can
by the Oxford philosopher Nick use it, for what purposes, and who is
Bostrum, need especially careful ultimately accountable?
attention: “responsibility, transparency,
auditability, incorruptibility, Another consideration is the way
predictability (...); all criteria that in which AIEd systems aim to effect
must be considered in an algorithm lasting behavioural change on
intended to replace human their users. For example, a system
judgement of social functions”.51 may make recommendations, use
persuasion, or offer feedback to
For example, what happens if AI ‘goes engender personal relationships
wrong’ (see, for example, the role of between humans and machines.
algorithms in the 2010 financial ‘flash Behaviour change is certainly
crash’)?52 Who is responsible, the one intervention that can be truly
end-user or the programmer? What transformative, but it is again not
will happen when an autonomous without serious ethical considerations.
vehicle is involved in a traffic accident?
Will it be possible to understand how Other concerns have been raised
it arrived at its decision so that it with regard to learning companions.
might be corrected to prevent future Although they are intended to support
problems? Already this can be difficult learners throughout their lives, there
if the AI uses neural networks. And are fears that a companion that
are AIs open to manipulation? We are ‘followed’ you would instead result
all too aware of the consequences of in the perpetual recording of
computer hacking; what might happen learner failures to the detriment
if an AI were developed or modified for of future progress.
criminal purposes?
Similarly, the concept of an AIEd
For AIEd, these ethical questions teaching assistant raises worries
are equally, if not more, acute, and that the technology will be used
questions need to be identified and as a classroom spy to record and
addressed. For example, we know that report any perceived suboptimal
the sharing of data is essential to the performance by the teacher.
integration of AIEd systems, and that
sharing of anonymised data has the Finally, we have a new responsibility
potential to move the field forward to ensure that society as a whole has
by leaps and bounds by cutting back sufficient AIEd literacy – that is, enough
on wasteful duplicative efforts. But to ensure that we use these new
this type of sharing introduces a host technologies appropriately, effectively,
of problems and questions, from and ethically.
individual privacy to proprietary
AIEd will connect to the

AIEd and the Internet of Things
The network of objects or ‘things’
Physical World with embedded computing systems,
sensors, and network connectivity
For some, the word ‘artificial’ in AIEd are referred to collectively as the
can give the sense that the technology ‘Internet of Things’ (IoT). IoT permits
is somehow removed from our real, any network-enabled objects to
physical lives. We have made the be interconnected with any other
argument here that AIEd is at its network-enabled object or machine.
core, a very real, very human This opens up new possibilities for
endeavour. Moving forward, AIEd AIEd systems, for example to support
will increasingly draw on our physical learners developing motor skills
environments and our physical beings that need consistent and extended
– and so make these integral to the practice, such as dancing, playing a
learning process. musical instrument, or even learning
surgical procedures.
AIEd will augment our
physical landscape AIEd will be attuned to how we
Augmented reality systems (AR) will feel and how we move
go one step beyond intelligent virtual Recent research supports the idea
reality systems by enabling learners that learning is significantly influenced
and teachers to experience and by how we feel (our affect), and how
interact differently with the physical we move.53 These insights suggest
world around them. AR technology that learning technologies can be
can display an overlay of information improved by taking these additional
about a person’s environment, inputs into account. Already, learner
allowing formal classroom content to models are no longer limited solely to
overlay the learner’s physical reality. representing and recording learners’
For example, the age, architecture academic progress,54 and sensors that
style, or heat efficiency of the buildings can be worn in clothing or strapped to
around a learner could be visualised body parts (e.g. the Fitbit) have already
as they move around the world. been developed.
We have already seen how existing AIEd systems of the future will
AIEd systems feature socially and increasingly support the whole
culturally intelligent avatars that learner through sophisticated
guide and support learners in virtual models that also capture data about
environments. The addition of AIEd to a learner’s emotional and physical
AR systems will allow for personalised, state. These enriched models will
adaptive educational experiences further contribute to what is known
with virtual mentors or tutors guiding about how we learn, and will provide
students through field trips, leaving individual teachers with real time
teachers to concentrate on those information about their students’
learners whose needs are greatest. physical and emotional well-being as
well as their cognitive development,
allowing for appropriate and timely
interventions in all the areas that
matter to learning.
Taking it to
the next level:
How AIEd can
help us respond
to the biggest
unsolved issues
in education
Policy makers often call them ‘wicked-issues’ – social problems that

are complex, connected, and seemingly resistant to intervention.
Sadly, education does not lack for its fair share.
If AIEd is to attract the attention and investment that we believe it deserves,

then it is only right to ask how AIEd can be realistically applied to address
these unsolved issues. Here we take two big issues – achievement gaps,
and teacher development, retention and shortages – and show how AIEd
can provide a response.
Tackling achievement gaps

Currently, we are failing to meet the needs of all learners. The gap between
those who achieve the most and those who achieve the least is a challenge
that teachers, school leaders, administrators, and government officials face
every day, in every country. Globally, students from poorer backgrounds
perform worse than students from richer backgrounds.55 The results of this
achievement gap impacts upon a country’s economy as well as the social
well-being of their population.56 The reasons behind the achievement gaps
in different countries vary, but the fact remains that not all learners are
achieving their potential at school.
We take it as essential that all children should have at least basic skills
(reading, writing, and mathematics), and yet, across the world, we are not
there. For example, in the UK, nine million working age adults have low basic
skills in either literacy, numeracy, or both. To make this real, this means that
these adults will struggle with simple everyday tasks such as assessing how
much fuel is left in their vehicle by looking at the gauge, or understanding
the instructions on over–the-counter medications.57
We have already shown some of the ways in which AIEd can offer a new
set of tools for addressing this challenge. For example, students who need
extra help can be offered one-to-one tutoring from adaptive AIEd tutors,
both at school and at home, to improve their levels of success. Increased
collaboration between education neuroscience and AIEd developers will
provide technologies that can offer better information, and support specific
learning difficulties that might be standing in the way of a child’s progress.
Moreover, and important to addressing the socioeconomic gap, these
AIEd systems will scale broadly as the reduction in their cost makes them
increasingly affordable to schools and school systems.
AIEd could also offer needed support before a learner begins formal
education, perhaps even before they are born. There is strong evidence
that the first five years of a child’s life have a large influence on that child’s
educational attainment.58 Unfortunately, we see evidence of poor school
readiness for many students, particularly children from low-income
families. This means they enter school at a significant disadvantage to their
wealthier counterparts in areas including language, early maths and science
understanding, physical well being and motor development, and social and
emotional development. This can mean that a child may enter school unable
to identify numbers, interact with peers, or use the bathroom on their own.59
Low-income parents may also have had limited education

opportunities, meaning they may face serious challenges in providing
at-home learning support to their children. AIEd systems can provide
tailored support to parents in the same way that they can for teachers
and students, improving education and outcomes for both parents
and their children.
Imagine, for example, providing parents with AIEd assistants that could
advise them about strategies for talking to their child, sharing songs,
and enjoying books. This could enable all parents to provide the right
sort of support in those all important early years. For parents who
have problems with numeracy or literacy, the AIEd assistant could help
boost these skills too.
To avoid what’s known as the Matthew Effect60 – the all too common
situation in which learners who are already privileged gain the most
from new resources, further exacerbating existing inequalities –
AIEd assistants should be available for all parents, with additional
support provided to those parents who need it most. This will
help ensure that all parents are well informed, supported, and
engaged in their child’s education.
Developing teacher expertise, addressing teacher retention,

and providing respite where teacher shortages are acute
Teacher expertise is key to learner attainment, but high-quality
continuous teacher development has significant costs, both in terms
of time and money. In the same way that AIEd systems can offer
one-to-one or group tutoring to students, so it can do the same on an
ongoing basis for teachers. This training can be designed to meet the
specific needs of the teacher, be completed wherever and whenever
they like, and can be used to access a community of like-minded
professionals who give advice and guidance.
AIEd could also help teachers find and share the best teaching
resources. Imagine, for example, navigating a popular tool like TES61
or Teachers Pay Teachers62 with your own AIEd assistant who knows
the resources you have found useful in the past, the details of your
students, and the teaching schemes and curricula used in your
institution. Your AIEd assistant could accurately predict the resources
that would work best for you and your students as well as uploading
the resources you have created and successfully used.
Intelligent support for teachers could also help address the issue of
teacher retention where we see many skilled professionals leaving
the profession due to ‘burnout’.63 Now that a cloud-based intelligent
assistant for every teacher is a realistic possibility, we can provide
support to reduce teacher stress and workload. The teacher’s AIEd
assistant will be available through any and all of the teacher’s devices
so that it can be deployed as needed wherever the teacher is working.
Outside the classroom, the assistant could greatly reduce

the amount of teacher time needed for grading. Inside the
classroom, the teacher could give the assistant the task
of offering one-to-one tutoring to a group of children who
are struggling to understand fractions. The assistant would
maintain a learner model for each student and would use
this to identify suitable teaching materials. This would free
up the teacher to turn their attention to an individual student,
or work with a group on a different topic.
In many parts of the world acute teacher shortages present

enormous challenges. For example, 33 countries do not
currently, and will not have, enough teachers to provide
every child with a primary education by 2030. In fact, the
world will need to recruit 25.8 million schoolteachers to
achieve this goal.64
Although the most effective implementations of AIEd will

deploy it alongside the expertise and empathy that is
peculiarly human, in some instances this simply will not be
possible, at least in the short-term. This means we will need
to rely on technology to make available high-quality learning
experiences to places where this is currently lacking.
One illustrative example is the work of Sugatra Mitra, who

famously provided children in an Indian slum with free access
to a computer placed in a wall between his office and the
public space. He christened this experiment the ‘Hole in the
Wall’, the goal of which was to understand if children could
form effective self-organised learning groups, in this case
around learning how to use a computer.
This work then led to his creation of School in the Cloud, “a

creative online space where children from all over the world
can gather to answer ‘big questions’, share knowledge and
benefit from help and guidance from online educators.”65
Now, imagine if we could add AIEd to technologies like these.

For example, AIEd could provide the intelligent support to aid
learners in their collaboration, it could provide Open Learner
Models to help volunteer online educators understand the
right support that a given learner might need, or it could
intelligently pair the most appropriate volunteer with the
right student.
This is not to disguise the urgent need to make high-quality

educator expertise available to all. Far from it. However,
where there is none, there is a moral urgency to act and
we would be remiss to not leverage AIEd.
Bringing it
all together:
The continuing
race between
education and
technology
66
We do not lack for predictions of how the existing mix of jobs in the
economy will be upended by the steady rise of the robots, and ever
smarter algorithms deployed on ever bigger data-sets. However, the
implications of this for learning has received relatively little sustained
and serious attention.
This is not surprising given that the debate, to date, is understandably

being led by economists rather than educators. It also reflects gaps in
existing quantitative research which focus largely on job categories,
rather than skills, and on the roles likely to be automated, rather than
those likely to be created.
Throughout this paper we have set out the AIEd pieces that could
– with further development and smart real-world testing – offer a
proportionate response to the new innovation imperative in education.
Simply stated the imperative is this: as humans live and work alongside
increasingly smart machines, our education systems will need to
achieve at levels that none have managed to date.
In our minds, this imperative trumps even the significant impacts

of globalisation, and brings existing issues in education, such as
achievement gaps, into ever sharper relief.
To summarise the argument so far we thought it would be useful to

do two things. First, map the catalogue of AIEd tools that can help
us address this enormous challenge by supporting the next phase
of education system reform. And, second, set out the ways in which
AIEd can be deployed to help us understand how successfully we are
realising this reform agenda.
If you like, just as learners need timely and actionable feedback, so

too do our education systems as they prepare learners for the
future economy.
Before we go on to this, however, it is important to remind ourselves

that the ‘purposes’ of education are wider than getting a job. For
example, a list would include discovering your passions, experiencing
the flow and satisfaction of good work, and being a moral person
with the capacity and will to affect positive change in your family,
community, country, and the world.
Having said this, getting a good job is consistent with the list above.
Indeed, it is one of the central reasons why governments invest
in education.
The table opposite shows our mapping of the tools of AIEd against the
likely requirements of the jobs market in 15 years’ time.
Using AIEd to effect education system reform
In 15 years The implication for How AIEd can help...

time... learning...
Many of the new Students will need to Give every learner their own
jobs created learn as efficiently and personal tutor, in every subject
will be much effectively as possible
Provide every teacher with their
more cognitively
own AI teaching assistant
demanding than
those currently AIEd to deliver timely, smarter,
available teacher professional development
AIEd tools that help every parent
support their child’s learning
We will need to seriously AIEd tools that embody new insights

attend to the non- from neuroscience or psychology
cognitive factors that
Making available new insights into how
influence learning –
learning is going for an individual and the
grit, tenacity and
factors that make it more likely to occur
perseverance; affect;
‘mindset’ In light of that, providing the right support,
at the right time, to keep learning on track
Students will need to Intelligent Virtual Reality to allow learners

achieve higher-order to be supported to learn in authentic
skills – e.g. problem environments – and to transfer that
solving – alongside learning back to the real world
‘knowing what’
Social skills will be Students need to be Intelligent support for collaborative learning
where humans effective collaborative
continue to excel problem solvers and
makers, able to build on
others’ ideas and extend
and sensitively critique
an argument
The ability to get on with IEd techniques to help us understand

A
others, to empathise better how to deliver a wider variety
and create a human of attributes, and how well a learner
connection, will continue is acquiring them
to be valued
We will need to We will need new ways of AIEd tools that support learners to become
re-skill large parts equipping adult learners effective, self-regulated learners for lifelong
of the current with new skills – more learning
workforce – in frequently, quickly,
Lifelong learning companions to advise,
essence, creating and effectively
recommend, and track learning
a learning society
More flexible learning environments,
allowing learners to learn at a time
and a place that works best for them
We will need to do all this without a significant uplift in the MOORE’S LAW
A computing
current investment we make in learning. With the steady term, established
application of Moore’s law, alongside wise investment, there is by Gordon Moore
around 1970,
every reason to believe that the cost of AIEd applications will
which states
diminish overtime, allowing this potential to be realised at a that processor
price that is affordable within current spending parameters. speeds, or overall
processing power
for computers,
Using AIEd to measure education system reform will double every
Once we put the tools of AIEd in place as described above, two years.
we will have new and powerful ways to measure system-

level achievement. Through the implementation of sensible
common data standards and data sharing requirements,
AIEd will be able to provide analysis about teaching and
learning at every level, whether that is a particular subject,
class, college, district, or country. This will mean that
evidence about country performance will be available from
AIEd analysis, calling into question the need for international
testing such as PISA and TIMMS, at least in their current
form.67
With this information available, system leaders and strategists

will need to develop fresh skills to probe the data and
establish the potential causes of any underperformance,
and the most likely solutions. For example, AIEd could
produce the school level data analysis that will indicate in
real-time when a school is experiencing problems. A team
of experts could then be called in to determine how these
problems can be quickly resolved.
Education systems will need to be nimble to take advantage

of the rich real-time systems level analysis that will be
continuously available. Synergistically, it may be AIEd systems
that provide the scaffolding to enable leaders and policy
makers to develop these new skills and abilities.
...
The view we have sketched out, that AIEd will play a critical
role in the next phase of education system reform, will not
happen by chance. This takes us to the final section of this
paper: the practical things that need to be done now for the
intelligence of AIEd to be unleashed.
Recommendations
to help us Unleash
Intelligence
In Alive in the Swamp68, Michael Fullan and Katelyn Donnelly

describe three powerful forces that must be combined if
we are to deliver on the promise of technology to catapult
learning dramatically forward. One is pedagogy, or the
science of how we teach and learn; the second is technology
itself, which we have said a great deal about already; and the
final component is system change, or our understanding of
how to deliver change so that it has a positive impact on each
and every learner.
The future ability of AIEd to tackle real-life challenges in

education depends on how we attend to each of these three
dimensions – that is: (i) we need intelligent technologies that
embody what we know about great teaching and learning in
(ii) enticing consumer grade products, which (iii) are then used
effectively in real-life settings that combine the best of human
and machine.
How does AIEd currently fare against these dimensions?

And, more importantly, what needs to be done to unleash
the full intelligence of AIEd?
Pedagogy
Learner,
Parent,
Teacher
System
Technology
Change
Pedagogy Technology
AIEd research has, to date, mainly AIEd is currently something of a cottage
tackled the low-hanging fruit of industry – research and development
education – for example, learning in takes place in small pockets and at
highly structured domains such as modest scale, mostly by researchers
introductory mathematics or physics, with limited funding and without
or applying AI techniques on highly commercial partnerships. The result is
structured datasets such as university that many of the applications that are
administration systems. developed never move beyond the
prototype stage, at which point much
These gains are essential but they of what has been learnt is lost.
are not enough. If we are to bring about
a step-change in the breadth and The solution is not to funnel money into
quality of learning for all learners, the development of a single monolithic
if we are to tackle the persistent and AIEd system that tackles every subject,
unsolved challenges of learning in the and every possible learning scenario.
21st century, funders and researchers
need to go deeper and wider. Instead, success will lie in the
development of a multitude of individual
In short, AIEd needs to begin with the AIEd components that specialise in
pedagogy and be more ambitious! a particular expertise: for example, a
subject area or a specific learner need.
Recommendations
• D
o not get seduced by the To realise this means putting in
technology, start with the learning. place the structures, incentives, and
funding that will allow an ecosystem
• F
ocus existing AIEd funding on the of innovation and collaboration to be
areas that are likely to deliver the created around AIEd.
step-changes in learning that will
make a real difference. Recommendations
• M
ove beyond the disjointed, un- • D
evelop the infrastructure that
prioritised and siloed approaches enables iterative innovation, and
that characterise the current AIEd less re-invention, in AIEd (for example,
funding landscape. APIs, shared data standards, and
• S
cope out a series of ambitious shared learner models).
challenge prizes that begin with • C
reate smart demand for AIEd
insights from the learning sciences technologies. For example,
and educational practice (see governments and philanthropists
overleaf). could guarantee a market for AIEd
solutions that have been shown
to work in real life settings. This
would unlock needed collaborations
between AIEd researchers and
commercial entities.
• F
ound a DARPA for education that
will accelerate the transition of AIEd
tools from the lab into real-world use
(see overleaf).
Similarly, we believe there would be

Learning from value in a suite of well-funded, global
challenge prizes that pose complex
the approach that learning problems, and then reward
those who provide the most exciting
jump-started and effective AIEd solutions.
driverless cars It’s not difficult to create a list of grand

challenges that could form the basis
In 2004, the authors Levy and for such an approach – for example,
Murnane famously wrote that gaps in student achievement, 21st
“executing a left turn against century skills or even preparing
oncoming traffic involves so many students for jobs that do not yet exist.
factors that it is hard to imagine
discovering the set of rules that can Although the XPrize Foundation
replicate a driver’s behavior”.69 are currently running two ‘Grand
Challenges’ in learning – a $7M
In the same year, the US Defense competition to help adult learners
Advanced Research Projects Agency who struggle with literacy, and a $15M
(DARPA) offered $1M for the team that competition to empower children to
developed a self-driving car that could take control of their own learning –
navigate a 142-mile route. For this first neither are specifically encouraging
prize, no team was successful. the development of AIEd based
solutions. In short, there is a big gap
However, just one year later (when the to be filled.
prize money was doubled) five vehicles
completed the course. The winning Moreover, sensible challenge prizes
team was led by Stanford University’s in this area should be as much about
Sebastian Thrun, who went on to consciousness-raising and movement
lead Google’s autonomous vehicles building as they are about the
team and, when there, began resulting solutions. After all, there
‘hoovering up’ the best engineers is still a lot of work to be done to
from the DARPA challenges. convey the usefulness of AIEd to
parents, policy-makers, educators,
and learners.
Just such an approach has been taken

by the UK innovation agency Nesta,
who created huge media interest
around their £10M Longitude Prize
which (as a result of a public vote)
was eventually focused on the urgent
problem of global antibiotic resistance.
These, in turn, would enable deep

Learning from learning techniques to extend and
refine the learner model so that it is
DARPA useful for an increasing number of
learner types and contexts.
In the US Defense Advanced Research
Projects Agency (DARPA) model the On a related note, it is striking that
best talents in diverse fields conduct some of the most ‘successful’ uses
basic research in order to solve of AIEd have been developed by the
real-world problems that are both US military. For example, America’s
ambitious and relevant. The goals Army is an intelligent virtual learning
are clear, but how you get there is environment which uses an AI-
not yet known. driven, first-person shooter, digital-
game format to allow young people
Creating similar centres of to experience virtual soldiering. It
independent interdisciplinary expertise has been used successfully since
in AIEd, funded long term and focused 2002 both for recruiting and as a
on delivering real-world capabilities, pedagogically-robust and immersive
would allow us to cross the chasm of training tool.
basic to applied research, and provide
the long-term funding and ambition Similarly, DARPA have developed a
that is needed. digital tutor to allow navy IT workers
to develop the skills to solve complex
A fundamental goal for these agencies IT problems. It has been reported
should be creating the technical to be more effective than traditional
infrastructure that allows for the classroom learning.70
development of an ecosystem of
AIEd innovation. This could include Imagine what we might see if we
standards for interoperability that were to put the same effort into
enable researchers and developers improving our schools, universities,
to share and build upon each and community colleges with properly
others’ work. researched and comprehensively
evaluated AIEd.
They should also look at developing
capabilities that can be re-applied
in multiple AIEd applications. For
instance, at the moment each and
every AIEd application has to develop
its own learner model, which can take
up to a year of effort. Developing a
learner model that can be called on
by separate applications would head-
off duplicated effort and allow the
creation of bigger data sets.
System change
AIEd will have to function in blended learning
spaces where digital technologies and traditional
classroom activities complement each other.
Realising this means addressing the ‘messiness’
of real classrooms, universities, or workplace-
learning environments, and involving teachers
and learners in the co-design of AIEd so that
our diagram instead looks like this:
Pedagogy
Learner,
Parent,
Teacher
System
Technology
Change
However, too little attention has been given to

designing and describing how AIEd concretely fits
within the lived experience of real learners and
educators.
There has been even less attention given to

providing the right professional support to allow
educators to realise these re-designed models
of learning.
We also need much better evidence of what works

in AIEd when it is implemented in real classrooms,
and universities – after all, how can we ask AIEd to
tackle the big problems in education, or system
owners to take AIEd seriously, if we do not also
provide the means to allow us to establish
whether the proposed solutions work?
Recommendations
• Involve teachers, students, and parents to ensure
that future AIEd systems meet their needs (a
participatory design process that will lead to better
AIEd products, to teachers more knowledgeable
about the processes of learning, and to more
successful learners).
• T
ake the next step to iterate and intelligently
evaluate AIEd applications in real world contexts.
• D
evelop data standards that prioritise both the
sharing of data and the ethics underlying data use.
...
We do not underestimate the new-thinking,

inevitable wrong-turns, and effort required to
realise these recommendations. However, if we
are to properly unleash the intelligence of AIEd, we
must do things differently - via new collaborations,
sensible funding, and (always) a keen eye on the
pedagogy. The potential prize is too great to act
otherwise.
References
1 (2006, July 26). AI set to exceed human brain power. 13 Self, J. (1999). “The Defining Characteristics of
Retrieved from http://www.CNN.com Intelligent Tutoring Systems Research: ITSs
Care, Precisely”. International Journal of Artificial
2 (2005). ODE: The Oxford dictionary of English Intelligence in Education (IJAIEd). 10, 350–364.
(Oxford dictionaries online). Oxford: Oxford
University Press. AND Russell, S. J., Norvig, P., 14 Vanlehn, K., Lynch, C., Schulze, K., Shapiro, J. A.,
& Davis, E. Artificial intelligence: A modern approach. Shelby, R., Taylor, L., ... & Wintersgill, M. (2005).
Upper Saddle River: Prentice Hall. “The Andes Physics Tutoring System: Lessons
Learned”. International Journal of Artificial Intelligence
3 With some systems, known as ‘Robo Advisors’, in Education. 15(3), 147-204.
algorithms create an investment portfolio suited
to an investor’s individual profile (including, for 15 Brown, J. S. & Burton, R. R. (1978). “Diagnostic
example, their investment aims and their personal Models for Procedural Bugs in Basic Mathematical
attitude to risk) and then can manage that portfolio, Skills”. Cognitive Science. 2, 155-191.
making independent decisions, on a daily basis.
Retrieved from http://www.forbes.com/sites/ 16 Retrieved from http://www.italk2learn.eu
robertberger/2015/02/05/7-robo-advisors-that-
make-investing-effortless/#1e029bcf7e48 17 Grawemeyer, B., Mavrikis, M., Holmes, W.,
& Gutiérrez-Santos, S. (2015). Adapting feedback
4 Bennett, C. C., & Hauser, K. (2013). “Artificial types according to students’ affective states.
Intelligence Framework for Simulating Clinical In Conati, C., Heffernan, N., Mitrovic, A., & Verdejo,
Decision-Making: A Markov Decision Process M. F. (Eds.). Artificial Intelligence in Education 17th
Approach”. Artificial Intelligence in Medicine. 57(1),9– International Conference, AIEd 2015. Madrid, Spain,
19. June 22-26, 2015 Proceedings (Vol. 9112). Madrid,
Spain: Springer International Publishing.
5 Retrieved from https://openai.com/blog/
introducing-openai/ 18 Litman, D. (2009). Language processing in AIEd:
Successes and challenges. Presented at the Panel
6 Vinge, V. (1993). Vernor Vinge on the singularity. on the Evolution of AIEd @ AIEd09. Brighton, UK.
Presented at the VISION-21 Symposium sponsored
by NASA Lewis Research Center and the Ohio 19 Dimitrova, V., Mccalla, G., Bull, S. (2007). Preface:
Aerospace Institute. “Open Learner Models: Future Research Directions”.
International Journal of Artificial Intelligence in
7 Hawking, S., Russell, S., Tegmark, M., & Wilczek, Education. Special Issue of the IJAIEd (Part 2).
F. (2014). Transcendence looks at the implications
of artificial intelligence - but are we taking AI 20 Du Boulay, B., Rebolledo-Mendez, G., Luckin,
seriously enough? Retrieved from http://www. R., Martínez-Mirón, E., & Harris, A. (2007).
independent.co.uk/news/science/stephen-hawking- “Motivationally Intelligent Systems: Diagnosis and
transcendence-looks-at-the-implications-of- Feedback”. In: AIEd. 563–565.
artificial-intelligence--but-are-we-taking-ai-seriously-
enough-9313474.html 21 Johnson, W. L., Valente, A. (2009). “Tactical Language
and Culture Training Systems: Using AI to Teach
8 Vinge, V. (1993). Vernor Vinge on the singularity. Foreign Languages and Cultures”. AI Magazine.
Presented at the VISION-21 Symposium sponsored 30(2), 72.
by NASA Lewis Research Center and the Ohio
Aerospace Institute. 22 Dillenbourg, P., Baker, M. J., Blaye, A., & O’Malley,
C. (1995). The evolution of research on collaborative
9 Frey, C. B., Osborne, M. A. (2013). The future learning. In Reimann, P. & Spada, H. (Eds.) Learning
of employment: How susceptible are jobs to in humans and machine: Towards an interdisciplinary
computerisation. learning science (pp. 189–211). Bingley: Emerald.
10 Autor, D. and Dorn, D. (2013). “The Growth of 23 Slavin, R. E. (2010). Co-operative learning: what
Low-Skill Service Jobs and the Polarization of makes group-work work. In Hanna, D., David, I.,
the US Labor Market.” American Economic Review & Francisco, B. (Eds.), The nature of learning: Using
103(5): 1553–1597. http://dx.doi.org/10.1257/ research to inspire practice (pp. 161-178). Chicago:
aer.103.5.1553 OECD Publishing.
11 Goldin, C. & Katz, L. F. (2010). The race between 24 Ibid.

education and technology. Cambridge: Harvard
University Press. 25 Muehlenbrock, M. (2006). “Learning Group
Formation Based on Learner Profile and Context”.
12 Ford, M. (2015). Rise of the robots: Technology and the International Journal on elearning. 5(1), 19.
threat of a jobless future. Basic Books.
26 McLaren, B. M., Scheuer, O., & Mikšátko, J. 38 Kaminska, I. (2015, June 25). Innovating fast or
(2010). “Supporting Collaborative Learning slow? Gates vs Wolf edition. Retrieved from
and e-Discussions Using Artificial Intelligence http://ftalphaville.ft.com/tag/technology/
Techniques”. International Journal of Artificial
Intelligence in Education. 20(1), 1–46. 39 Retrieved from http://www3.weforum.org/docs/
WEFUSA_NewVisionforEducation_Report2015.pdf
27 Upton, K., & Kay, J. (2009). Narcissus: group and
individual models to support small group work. 40 Luckin, R. (2010). Re-designing learning contexts:
In Houben, G., McCalla, G., Pianesi, F., & Zancanaro, Technology-rich, learner-centred ecologies. London:
M. User modeling, adaptation, and personalization Routledge.
(pp. 54-65). Berlin Heidelberg: Springer.
41 Hill, P. & Barber, M. (2014). Preparing for a
28 Vizcaíno, A. (2005). “A Simulated Student Can renaissance in assessment. London: Pearson.
Improve Collaborative Learning”. International
Journal of Artificial Intelligence in Education. 15(1), 42 DiCerbo, K. E. & Behrens, J. T. (2014). Impacts of the
3–40. digital ocean on education. London: Pearson.
29 One example of this type of virtual agent can be 43 Hill & Barber. (2014). DiCerbo, K. (2014, December
found in Betty’s Brain (http://www.teachableagents. 17). Why an assessment renaissance means fewer tests.
org/research/bettysbrain.php), a computer-based Retrieved from http://researchnetwork.pearson.
learning environment developed at Vanderbilt com/digital-data-analytics-and-adaptivelearning/
University. assessment-renaissance-means-fewertests
30 De Laat, M., Chamrada, M., & Wegerif, R. (2008). 44 Howard-Jones, P., Holmes, W., Demetriou, S., Jones,
Facilitate the facilitator: Awareness tools to support C., Tanimoto, E., Morgan, O. & Davies, N. (2014).
the moderator to facilitate online discussions “Neuroeducational Research in the Design and
for networked learning. In: Proceedings of the 6th Use of a Learning Technology. Learning, Media and
International Conference on Networked Learning Technology. 40(2), 1–20.
(pp. 80–86).
45 See, for example, Dweck, C. S., & Leggett, E. L.
31 Barab, S. A., Gresalfi, M., & Ingram-Goble, A. (1988). “A Social-Cognitive Approach to Motivation
(2010). “Transformational Play: Using Games and Personality”. Psychological Review. 95(2), 256-73
to Position Person, Content, and Context”. or Dweck, C. S. (2006). Mindset: The new psychology
Educational Researcher. 39(7), 525–536. of success. New York: Random House.
32 Hassani, K., Nahvi, A., & Ahmadi, A. (2013). Design 46 Dweck, C. S. (2010). “Even Geniuses Work Hard”.
and implementation of an intelligent virtual Educational Leadership. 68(1), 16-20.
environment for improving speaking and listening
skills. Interactive Learning Environments (pp. 1–20). 47 Retrieved from http://www.mindsetworks.com/
brainology/
33 Dede, C. (2009). “Immersive Interfaces For
Engagement and Learning”. Science, 323(5910), 48 Harris, A., Bonnett, V., Luckin, R., Yuill, N., &
66-69. Avramides, K. (2009). Scaffolding effective
helpseeking behaviour in mastery and performance
34 Vannini, N., Enz, S., Sapouna, M., Wolke, D., Watson, oriented learners. In Dimitrova, V., Mizoguchi,
S., Woods, S., ... & Aylett, R. (2011). ““FearNot!”: R., Du Boulay, B., & Graesser, A. C. (Eds.) AIED 2009,
A Computer-Based Anti-Bullying-Programme Frontiers in artificial intelligence and applications
Designed to Foster Peer Intervention. European (pp. 425-432). 200, IOS Press.
Journal of Psychology of Education. 26(1), 21-44.
49 Cole, M. (1996). Cultural psychology: A once and
35 See, for example, Traum, D., Rickel, J., Gratch, J. & future discipline. Harvard University Press.
Marsella, S. (2003). Negotiation over tasks in hybrid
human-agent teams for simulation-based training. 50 See, for example, Chan, T. W. (1991). Integration-
In Proceedings of the Second International Joint kid: A learning companion system. In Mylopolous,
Conference on Autonomous Agents and Multiagent J. & Reiter, R. (Eds.) Proceedings of the 12th
Systems. 441-448. ACM. International Conference on Artificial Intelligence.
2,1094-1099. Australia, Morgan: Kaufmann
36 Sarkis, H. (2004). Cognitive Tutor Algebra 1 program Publishers, Inc.
evaluation, Miami-Dade County Public Schools.
Lighthouse Point, FL: The Reliability Group. 51 Bostrom, N., Yudkowsky, E. (2013). The ethics
of artificial intelligence. Cambridge: Cambridge
37 Upton and Kay, 2009. University Press.
References
52 (2010, October 1). What caused the flash crash? 67 Mayer-Schönberger, V., & Cukier, K. (2013). Big data:
One big, bad trade. The Economist Online. Retrieved A revolution that will transform how we live, work, and
from http://www.economist.com. think. New York: Houghton Mifflin Harcourt.
53 Lindgren, R., & Johnson-Glenberg, M. 68 Fullan, M., & Donnelly, K. (2013). Alive in the swamp:
(2013).“Emboldened by Embodiment Six Precepts Assessing digital innovations in education. London:
for Research on Embodied Learning and Mixed Nesta.
Reality. Educational Researcher. 42(8), 445-452.
69 Levy, F., & Murnane, R. J. (2004.) The New Division
54 Graesser, A. C., Chipman, P., King, B., McDaniel, of Labor: How Computers Are Creating the Next Job
B., & D’Mello, S. (2007). Emotions and learning Market. Princeton University Press.
with autotutor. In Luckin, L. Koedinger,
K. & Greer, J. (Eds.) Frontiers in artificial intelligence 70 Fletcher, J. D. & Morrison, J. E. (2012). DARPA digital
and Applications. 158, 569-571. IOS Press. tutor: Assessment data. Alexandria, VA: Institute for
Defense Analyses. Retrived from http://www.acuitus.
55 Conroy, M.& Rothstein, R. (2013, January 15). com/web/pdf/D4686-DF.pdf
International test show achievement gaps in all
countries, with big gains for U.S. disadvantaged
students. Retrieved from http://www.epi.org
56 Hanushek, E. A. & Woessmann, L. (2010). The high

cost of low educational performance: The long-run
economic impact of improving PISA outcomes. Paris:
OECD.
57 Kuczera, M., Field, S. & Windisch, H. C. (2016).

Building skills for all: A review of england. Paris: OECD.
58 Bertram, T. & Pascal, C. (2014). Early years literature

review. Retrieved from https://www.early-education.
org.uk/sites/default/files/CREC%20Early%20
Years%20Lit%20Review202014%20for%20EE.pdf
59 Paton, G. (2014, October 16). Four-in-10 children ‘not

ready for school’ at the age of five. The Telegraph.
Retrieved from http://www.telegraph.co.uk/
60 A term commonly used in psychology and sociology

that makes reference to a verse from the Gospel
according to Matthew (XXV, 29), “For unto every
one that hath shall be given, and he shall have
abundance: but from him that hath not shall be
taken away even that which he hath.”
61 Retrieved from https://www.tes.com/teaching-

resources
62 Retrieved from https://www.teacherspayteachers.

com/
63 House of Commons Library. (2015). Teachers: Social

Indicators. SN/SG/2626
64 UNESCO. (2015). Sustainable development goal for

education cannot advance without more teachers.
UIS Fact Sheet No.33. Retrieved from http://www.
uis.unesco.org/Education/Documents/fs33-2015-
teachers.pdf
65 Retrieved from https://www.theschoolinthecloud.

org/
66 Goldin and Katz, 2010

INTELLIGENCE UNLEASHED
DESIGN BOND & COYNE

PRINT DAYFOLD
Pearson
80 Strand
London
WC2R 0RL
www.pearson.com
Join the conversation

@Pearson
#Openideas

Intelligence Unleashed Publication

Uploaded by

Copyright:

Available Formats

Intelligence Unleashed Publication

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Intelligence Unleashed Publication

Uploaded by

Copyright:

Available Formats

Intelligence

Open Ideas at Pearson

Open Ideas at Pearson

This series of publications, Open Ideas, is one of

How do we learn, and what keeps us motivated to

These questions are too important for the best ideas

Our hope is that Open Ideas helps with this task,

Pearson UCL Knowledge Lab

Suggested reference: Luckin, R., Holmes, W., Griffiths, M. & Forcier,

We also want to thank Mutlu

Two people deserve a special

Adopting a puzzled stance as to why things have not

What we need – what we should demand – is an explanation

Second, we need a clear explanation of how the field of

Third, we need concrete options that will allow us to make

In other words, what we need is a degree of specificity about

From what AI is and how AIEd-driven learning systems are

To make my own case for reading this paper, let me move

Yet, it was equally easy to identify ways in which some basic

Asking for these features is not asking for science fiction.

Funders and founders, policy makers and philanthropists

Throughout, our approach has been to start with teaching

Although some might find the concept of AIEd alienating,

We look towards a future when extraordinary AIEd

True progress will require the development of an AIEd

If we are ultimately successful, AIEd will also contribute

Our response, we argue, should be to take on the role

To be candid, the impetus for this paper arose from our

This is our attempt at contributing to that change, through

Let us start by introducing AI.

It can be difficult to define artificial intelligence (AI), even for ALGORITHM

The use of AI in our day-to-day life is increasing ever more

Will AI take over

The understandable guiding thought that we could not have previously

The application of artificial intelligence to education (AIEd) has ADAPTIVE LEARNING

These understandings may then be applied to the development of

As we have said, AI involves computer software that has been

This knowledge about the world is represented in so called ‘models’.

Take the example of an AIEd system that is designed to provide

Examples of the specific knowledge that might be integrated into

AIEd models What the model represents Examples of specific knowledge

Pedagogical The knowledge and expertise ‘Productive failure’ (allowing

Learner Knowledge of the learner The student’s previous

To delve deeper into just one of these examples,

So, what would a piece of education technology driven by

Deep analysis of the student’s interactions is also used to

Some systems include so-called Open Learner Models, which

One of the advantages of adaptive AIEd systems is that they

The learner’s interactions, achievements,

Open Learner Model

Making the learning explicit,

In addition to the learner, pedagogical, and domain models,

What AIEd can

A multitude of AIEd-driven applications are already in use EDUCATIONAL

Here, however, we focus on three categories of AIEd

AIEd can provide an intelligent,

This is where Intelligent Tutoring Systems (ITS) come in. ITS

so that appropriate tutoring could be offered. Initially, it was NEURAL NETWORKS

techniques, self-training algorithms based on large data sets,

Modern model-based adaptive systems can be far more

Foundational ...everyday Literacy Technology literacy

Competencies ...complex ritical thinking and