1 s2.0 S2666920X22000170 Main

Computers and Education: Artificial Intelligence 3 (2022) 100062
Contents lists available at ScienceDirect
Computers and Education: Artificial Intelligence

journal homepage: www.sciencedirect.com/journal/computers-and-education-artificial-intelligence
AI & learning: A preferred future

Venkat Srinivasan
Innospark Ventures, Boston, MA, USA
A R T I C L E I N F O
Keywords:
AI & learning
Preferred future for learning
Autonomous learner
Learning science
Multi-sensory engagement
Intelligent learning systems
1. Introduction well as those underemployed or seeking career changes, gig-workers,

veterans).
Rapid progress in AI technologies has generated considerable inter Over the last couple of decades, there has also been significant
est in their potential to address challenges in every field and education is advancement in our knowledge of how we learn. Studies have deter
no exception. Improving learning outcomes and providing relevant ed mined that several practices, loosely referred to as learning science, if
ucation to all have been dominant themes universally, both in the systematically implemented, can aid learners in optimizing their
developed and developing world. And they have taken on greater sig learning potential. However, these practices have not yet found their
nificance in the current era of technology driven personalization. way into educational curriculums in a systematic fashion. It has been
Learning outcomes have however stagnated in the U.S. and have largely up to individual teachers to adapt them as they saw the fit based
remained relatively low in the developing world. In their 2020 annual on their experience. .
letter, Bill and Melinda Gates point out that despite spending billions Another macro trend impacting learning in general has been the
they have not seen the kind of progress they expected with learning evolution of the role of the teacher in learning. From being the sole
outcomes in k-12 education in public schools (CNBC, 2020). In India, influence on learning, digital technologies and AI are transitioning the
each year, the Annual Status of Education report reports in India’s public teacher to being more of a facilitator and coach than being the singular
schools, a 5th grader is unable to read 2nd grade text (ASER, 2019). source for learning. In his far sighted foreword, John Seely Brown
Recently, the Central Square Foundation (Central Square Foundation, (Iiyoshi & Vijay Kumar, 2010) recognizes that all learning does not
2020) found that learning outcomes in private schools in India, while happen inside the classroom arguing forcefully for blurring the
better compared to public schools, are still poor - 35% of rural private distinction between formal learning [classrooms] and informal learning
school students in Grade 5 cannot read a grade 2 level paragraph. [outside]. He prophesized three influences as shaping the future of ed
Lack of equity in education has been a contributing factor for poor ucation – the digital age, social learning and technology. The impact of
learning outcomes. In the U.S., a 2013 report by the Equity and Excel these influences are very relevant today.
lence Commission of the U.S. Dept of Education reported that “some The coronavirus pandemic we are living through will likely forever
young Americans – most of them white and affluent – are getting a alter how the various stakeholders in education interact going forward.
world-class education” while those who “attend schools in high poverty Approximately, 1.5 billion learners and 90 million teachers were forced
neighborhoods are getting an education that more closely approximates to leave the physical classroom (UNESCO, 2020) for a protracted period.
schools in developing nations.” Compounding the challenges in Almost overnight, most educational institutions were forced to go online
providing equitable access to educational opportunities is a new gen and have had to remain either in full remote or hybrid learning modes
eration of learners, with highly differentiated educational needs. This for 18 months or so. The forced transition to remote learning has
includes displaced and differently abled learners (such as immigrants as revealed certain benefits, created interesting challenges and caused
E-mail address: Venkat.Srinivasan@Innosparkventures.com.
https://doi.org/10.1016/j.caeai.2022.100062
Received 28 November 2021; Received in revised form 19 February 2022; Accepted 8 March 2022
Available online 18 March 2022
2666-920X/© 2022 Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
V. Srinivasan Computers and Education: Artificial Intelligence 3 (2022) 100062
changes in behavior like greater willingness to adopt AI technologies. learner.

The challenges related to equity in education have only got exacerbated Based on an analysis of over 400 research articles published between
and come into sharper focus. All of a sudden parents have had to become 2000 and 2019 on the application of AI for teaching and learning, they
para educators. The need for innovative solutions for delivering find shifts in research emphasis between the first and second decade.
high-quality education remotely at speed and scale has never been During the period 2000–2009, the emphasis was on learner-oriented
greater. approaches trending away from instructor-oriented approaches. In the
We believe AI technology has the potential to help the world address second decade 2010–2019, they found that there was a noticeable shift
these chronic challenges in education relating to equity, learning out towards modeling learning outcomes/analytics and student profiling.
comes, and real world relevance. In this paper, we describe a preferred Their findings reflect the growing interest in explicit measurement of
future for the role of AI in learning. Our vision of a preferred future is learning outcomes, and predictive AI models including the use of deep
manifest in a comprehensive AI-first solution framework for learning statistical learning.
with the objectives of scaling quality education affordably, optimizing Xie at al (2019) review trends and developments in
the learning potential of individual learners, enabling teachers to be technology-enhanced adaptive/personalized learning (PLL) by review
more effective and parental involvement to have a positive impact. ing related journal articles in the decade 2007–2017. Of particular note
is Chen et al. (2021) who found that learners’ gender, cognitive styles,
2. Literature review and prior knowledge can determine if personalized learning systems will
be effective. Another notable finding is by Tseng et al. (2008) who found
In this section, we review prior research in a number of relevant that the integration of learning styles and learning behaviors can be used
areas. We review specific research related to AI in Education (AIED), the to determine the appropriate learning materials for the learner. Xie et al.
changing role of the teacher, parental involvement, the neurobiology of (2019) identify wearable personal learning technologies, collaborative
learning and the constructivist theory of education. In arriving at a and immersive personalized learning using virtual reality as areas for
preferred future, we believe a holistic view encompassing all these di future applications of adaptive/personalized learning.
mensions is critically important. Based on a review of 220 research articles on personalized language
learning, Chen et al. (2021) find that (1) multimodal videos promote
effective personalized feedback; (2) personalized context-aware ubiq
2.1. AI in learning uitous language learning enables active interaction with the real world
by applying authentic and social knowledge to their surroundings; (3)
Baker and Smith (2019) categorize AIED applications into 3 cate mobile chatbots with Automated Speech Recognition (ASR) provide
gories: (i) learner-oriented AIED; (ii) instructor-oriented AIED; and (iii) human-like interactive learning experiences to practice speaking and
institutional system-oriented AIED. However often applications that pronunciation; (4) collaborative game-based learning with customized
focus on learning cut cross several stakeholders especially in the current gameplay path and interaction and communication features supports
context and impact a common objective - learning. They are likely to social interactivity and learning of various language skills; (5) AI pro
reflect a collaboration among teachers, learners, parents, tutors and motes effective outcome prediction and instruction adaptation for in
peers. We prefer the following modified categorization: dividuals based on massive learner data; (6) Learning Analytics (LA)
dashboards facilitate personalized recommendations through learning
(i) learning-oriented, including learners, teachers, educational in data visualization; and (7) data driven learning allows personalized and
stitutions, tutors, peers, parents and para educators; immediate feedback in real-time practice. Generally, these findings
(ii) institutional operations oriented financial aid, retention, learner validate that AI technology can be beneficial in personalized language
experience; learning.
(iii) policy oriented including governmental interventions, subsidies, In a review of AI supported eLearning (AIeL) research, Tang et al.
competitive assessments like MCAS in the U.S. (2021) found that several of the studies analyzed had reported positive
impact of AI on e-learning participants’ learning performance or per
Our focus in this paper is on learning-oriented applications in AIED. ceptions. They found that a large portion of AIeL applications are in the
We use AIED and ‘AI in Learning’ synonymously in the remainder of the form of intelligent learning systems, which provide personal guidance or
paper. feedback based on individual students’ learning status. They suggest
Guan et al. (2020) summarize the various definitions of AIED in prior that prediction of students’ learning characteristics by collecting and
research: analyzing their online learning behavior and performance was a ripe
Authors Definition area for research. Moreover, from the perspective of the follow-up
Hwang (2003) Intelligent tutoring system that helps to organize studies, it is concluded that developing adaptive learning environ
system knowledge and operational information to ments and providing personalized learning supports based on the stu
enhance operator performance and automatically dents’ learning characteristics are important trends in AIeL research.
determining exercise progression and remediation
Hwang (2014) define the criteria for ‘smart learning’ environments.
during a training session according to past student
performance. Smart learning environments are defined as enabling learners to access
Johnson et al. (2009) Artificially intelligent tutors that construct responses in digital resources and interact with learning systems in any place and at
real-time using its own ability to understand the any time, and also actively provide the necessary learning guidance,
problem and assess student analyses.
hints, supportive tools or learning suggestions to them in the right place,
Popenici and Kerr (2017) Computing systems that are able to engage in human-
like processes such as learning, adapting, synthesizing,
at the right time and in the right form. The study presents a framework
self-correction and use of data for complex processing to address the design and development considerations of such smart
tasks. learning environments to support both online and real-world learning
Chatterjee and Computing systems capable of engaging in human-like activities. While the framework has several valuable elements, it is still
Bhattacharjee (2020) processes such as adapting, learning, synthesizing,
focused on an appropriate context-aware learning environment beyond
correcting and using of various data required for
processing complex tasks. a purely adaptive learning environment. It does not provide an opera
tional technology blueprint for AIED.
In an important contribution, Chen et al. (2020) conducted an review
The first two definitions of AIED above are focused on intelligent of 45 influential articles and also evaluated definitions of AIEd from
tutoring and the last two are slightly broader but still focused on the broad and narrow perspectives and clarified the relationship among
2
AIEd, Educational Data Mining, Computer-Based Education, and Some research studies (Pinkwart, 2016; Rummel et al., 2016; Schiff,
Learning Analytics. The findings relevant to this paper are that little 2020) have specifically focused on predicting a future for AIED. Pink
work had been conducted to bring deep learning technologies into wart (2016) presents two possible futures - a utopian vision and a
educational contexts; natural language processing (NLP) technologies dystopian vision. Comparing the two extreme scenarios, the author
were commonly adopted in educational contexts, while more advanced summarizes seven challenges AIED might have to face in the future:
techniques were rarely adopted and there was a lack of studies that both cultural context, practical impact, privacy, interaction methods,
employ AI technologies and engage deeply with educational theories. collaboration at scale, effectiveness in multiple domains, and the need
Their recommendations are to seek the potential of applying AI in for greater importance for the role of AIED in educational technology. In
physical classroom settings, spare efforts to recognize detailed entail our opinion, all seven are valid issues to be addressed if AIED is to
ment relationships between learners’ answers and the desired concep succeed at scale. The author does not present any systemic solution to
tual understanding within intelligent tutoring systems; pay more address the challenges.
attention to the adoption of advanced deep learning algorithms such as Rummel et al. (2016) similarly present areas of research they
generative adversarial network and deep neural network; seek the po consider important to avoid the pitfalls of a Dystopian view of the future
tential of NLP in promoting precision or personalized education; of adaptive collaborative learning support (ACLS). Specifically, they
combine biomedical detection and imaging technologies; and closely advocate the need to work towards a comprehensive instructional
incorporate the application of AI technologies with educational theories. framework building on educational theory in order to provide nuanced
We agree with many of their recommendations but have to issue a word and flexible (i.e. intelligent) ACLS to collaborative learners. We agree
of caution about the one suggesting greater adoption of deep learning with the basic premise of their recommendation – the need for nuanced
technologies. and flexible personalized learning support based on educational theory
The above reviews provide valuable summaries of the progress AIED and findings. In fact, the framework proposed in this paper reflects a
has made, especially in personalized and adaptive learning. Broadly the similar underlying rationale.
findings support the hypothesis that personalization makes a difference Schiff (2020) reviews the status of AIED, specifically intelligent
when learner characteristics and differences are carefully considered. tutoring systems and educational agents, and attempts to provide
The trend towards more explicit measurement of learning outcomes is development pathways for AIED in the future. The study notes in
healthy and much needed. In incorporating learning styles, extant particular that MOOCs didn’t succeed widely because of a mistaken
research with the exception of Srinivasan and Murthy (2021) does not assumption of technological linearity and their inability to reproduce
consider the beneficial impact of engaging multiple sensory channels. the in person teacher-student interaction. Overall, he points out that the
The extant research also does not explicitly propose a generalized previous round of educational technologies were not as successful
framework for integrating findings from learning science. Granular because of quality gaps in their implementation. Schiff (2020) presents
personalization governed by learning styles including the influence of two competing futures for AIED – one rooted in efficiency through
different sensory modes is integral to a more holistic AI-led learning automation and the second rooted in robust differentiation at a student
framework. level. In our opinion, we do not see them as competing. AI technologies
Another area that has attracted significant attention is essay grading/ can enable us to achieve both efficiency and individualized learning.
grading of open text responses. (Janda et al., 2019). However, most of In addition to the challenges identified by Pinkwart (2016), AIED
the solutions, if not all are black boxes and not considered viable or also has to address broader challenges with AI – explainability, tracta
effective. By relying on word based statistical models, none of the bility, reliability of data, and full context awareness, in adopting a
‘Automated Essay Evaluation’ systems capture the real meaning of the purely data driven approach to intelligence acquisition (Marcus, 2020;
essay and its relevance to the related learning content. They suffer from D’Amour et al., 2020; Srinivasan, 2017). No amount of data will ever be
other serious limitations – the need for large amount of data for training, complete (Srinivasan, 2017) and the resulting models will contain
the resulting bias, the lack of traceability and the lack of tractability. whatever biases the data has. Explainability will be critical for many
Since the model training is entirely dependent on the data set, they have AIED applications,. A third challenge is tractability. The only way to
been criticized for not being able to recognize gibberish essays [Kolo improve models purely dependent on data is to add more data and hope
wich, 2014] and essays that reflect divergent thinking. They are also not it will improve the model. Revised models with additional data may
configurable for course specific essay assignments since the rubric may change previous results which in some cases creates issues of repro
vary from course to course and there will not be enough data for training ducibility. Finally, in some areas like essay grading and natural language
a statistical model in such cases. Unfortunately, this has resulted in the processing, current technologies do not have the ability to retain full
use of multiple choice assessments and the use of ‘peer grading’ for open context. Neural language models like BERT (Bidirectional Encoder
text responses in online learning. Given the importance of ‘writing’ to Representation from Transformers) for example are word pattern based
reflect complete understanding, AIED needs to focus on automated and do not understand the text despite an attempt to incorporate word
evaluation of text responses with full understanding of the learning level distributional semantics using embeddings, see e.g., Bender and
objectives/rubric and learning material. Koller (2020).
An area of significant concern in developing and implementing AI Barring the issue of bias (Yang et. al., 2019), much of the AIED
applications is the potential misuse of AI, specifically the issue of ‘bias’ literature while extensive has not yet really recognized the implications
from reliance on data and the need for governance. Bias from data is a of these fundamental challenges in the use and adoption of AI for
significant issue which inherently makes all statistical deep learning learning. Some authors even suggest AIED research should adopt more
models unusable. Digital Promise [www.digitalpromise.org] has sophisticated methods like deep neural network models. In our view,
recently introduced a certification product which prioritizes racial eq blind focus on methods without rationale on why and how that would
uity by proactively identifying and minimizing racial bias in a product’s benefit learning will not lead to advances in the use of AI for learning.
algorithms and design. Yang et al. (2019) propose that AI can evolve into
human centered AI (HAI). In particular, they draw attention to how AI 2.2. The changing role of the teacher
can inhibit the human condition and advocate for an in-depth dialog
between technology and human-based researchers to improve under Since the days of the industrial revolution, the widely prevalent
standing of HAI from various perspectives. Yang et. al. (2019) however teaching/learning model used to center around a human teacher. As
do not recognize or provide any solutions to address the root causes of shown in Fig. 1, a good teacher had to be a subject matter expert and be
data driven bias. In our opinion, AIED needs to adopt explainable and current all the time, be an expert in the pedagogy relevant to that subject
tractable models of intelligence acquisition. and had to be an expert at effectively imparting the subject using the
3
digital environment including the internet and digital technology.

Content on every conceivable subject is growing exponentially on the
internet. The teacher and the text book are no longer the only source of
content. Teachers cannot possibly keep up with all the new content and
knowledge by themselves. Further exacerbating the issue, today
autonomous learners are likely to be aware of new content before
teachers. AI can play a significant role in assisting teachers to stay
abreast of new content and knowledge.
In terms of pedagogy, partly reflective of the digitally connected
environment we are in today, in the flipped classroom model, learners
complete assignments in the classroom and study the learning content at
home through video lectures and on-screen tutorials. In the digitally
connected world, there are multiple modes of delivery which all inter
sect and interact with one another. Learners can learn autonomously
anywhere including home and the classroom, can interact with peers,
mentors and parents, and belong to communities.
Today’s teacher is therefore, much more of a coach and facilitator. Of
course, the teacher is also expected to be an expert in the subject matter
but in a more foundational way where they are well versed in the
foundational aspects of the subject.
Further, Covid19 forced all institutions to move online almost
overnight irrespective of their prior beliefs about online education. Even
though the immediate online shift has largely been a ‘lift and shift’ of
materials and instruction, many teachers have been forced to use a va
riety of digital tools online which they were not familiar with before. As
they experience these tools, we can envision they will want to modify
pedagogy and materials to leverage the capabilities offered by these
tools. Besides, the emerging consensus seems to be that the future is
likely to be a blended model in most institutions. As a result, we can
expect to see further changes in the role of the teacher in the future.
The AIED literature does not explicitly recognize the disaggregation
of the teacher’s role, the emergence of the autonomous learner and the
resultant implications for learning. The real opportunity before us is to
use AI to scale this disaggregated model more effectively to improve
learning, teaching and to do so in an equitable manner.
2.3. The role of parents as educators
The pandemic we are living through has forced parents to play a

greater role in the education of their children at least in k-12 settings.
With schools going remote at least partially, parents have had to teach
and/or monitor their children’s learning at home. This has led to a huge
need to support parents in their roles. In many cases, parents may not
have had the necessary education themselves or certainly not trained
pedagogically. We believe this will lead to an increased need for
tutoring.
Fig. 1. The conventional role of a teacher.
Is the greater role for parents temporary? Even as we return normal
functioning of schools, we speculate that parental involvement may
desired pedagogy. In fact, many studies (e.g., Beteille & Evans, 2019) continue at an elevated level.
have found that high-quality teaching can have an outsized impact on The research on the relationship between parent involvement and
learning. In the United States, Popova et al., [2016] found that the dif academic achievement is mixed. McNeal Jr. (2014) finds that while
ferences between a good and a bad teacher can equate to a full year of much research supports the claim that parent involvement leads to
learning for a student. improved academic achievement (e.g. Burcu & Sungur, 2009; Coleman,
There are substantial challenges in scaling good teachers if you 1991; Epstein, 1991; Henderson, 1991; Lee & Bowen, 2006; Patel,
define them as shown in Fig. 1. Many elements of quality professional 2006), other research indicates that parent involvement is associated
development of teachers that lead to impact at a small-scale can be very with lower levels of achievement (e.g. Brookover et al., 1979; Domina,
hard to preserve when implemented at large scale. Among the more 2005) or has no effect on achievement (e.g. Brookover et al., 1979;
commonly reported issues in scaling quality professional development Domina, 2005; El Nokali et al., 2010). Additionally, parent in
of teachers have been cost, maintenance of quality, identification of high volvement’s effect on academic achievement has been found to vary by
quality trainers, and contextualizing the training programs to reflect the minority and/or social status of the student (e.g. Hill et al., 2004; Lee
local context. & Bowen, 2006), by gender (Muller, 1998), and by immigrant status
With the advent of the internet and other technological advances, (Kao, 2004). Finally, many studies find positive, negative, and/or no
learners have started to become more autonomous and the role of the associations between parent involvement and academic achievement
teacher has been changing, especially in the online context [see Fig. 2]. within the same study (e.g. Domina, 2005). Surprisingly, the contra
Fig. 2 shows the disaggregation of the 3 dimensions we defined in Fig. 1 dictory findings are remarkably consistent and cut across grade level,
as characteristic of a good teacher. It reflects the impact of the new measure of academic achievement, and time (spanning the middle
4
Fig. 2. The disaggregation of the Teacher’s role.
1970s to the late 2000s).

To our knowledge, there have been no studies on how AIED can play
a role in enabling parents to be para educators. Intelligent learning as
sistants appropriately personalized and integrated with learning science
at a granular level can make parents and teachers much more effective.
2.4. The neurobiology of learning
Over the last few decades, there has been considerable advancement
in our understanding of how our brains construct and retain knowledge,
i.e., how we learn. A systematic integration of these findings offer a
tremendous opportunity to maximize the learning potential of every
learner.
Our brains are believed to store knowledge in structures commonly
referred to as schemas. It has also been found that our brains encode new
information that fits prior knowledge more efficiently than when in
formation is novel [van Kesteren et al., 2012]. Recent findings across
neuroscience [St Jacques et al., 2013], psychology [Roediger and
Mcdermott, 1995 and educational science [Castro Sotos et al., 2007]
suggest that when a schema becomes too strong it can end up facilitating
the formation of false memories and misconceptions. False memories
and misconceptions obviously signify a major burden for learning. Fig. 3. Knowledge construction in the brain.
In terms of the process of knowledge construction, our brains have
been found to follow a set of sequential processes to construct knowl timeline t, this episodic detail semanticizes where we start losing the rich
edge [Paller & Wagner, 2002]. First, new information becomes encoded detail. At some point, we will barely remember the specifics except for
into a memory. Right after encoding, a memory is usually episodically some aspect of the event if it made a very deep impression on us. The
detailed, containing details about time, place and context. Over time, process of episodic memory gradually being absorbed into long term
encoded memories are consolidated during sleep (and presumably also memory depends a lot on our pre-existing knowledge schema in our
during rest). It is generally believed that consolidation semanticizes brain. As we try to recall this episodic event over time, the degree of
memories such that episodic details fade away and what remains is a specificity of our recall will vary based on when along time we try our
more abstract version of the encoded memory. This is illustrated in recall. This is illustrated by the arrow from the timeline t to ‘3 Recall’.
Fig. 3. Generally, when memories are retrieved after consolidation, they
Fig. 3 shows an episode getting captured initially at the time we have a different neural architecture than when they were first encoded
witness it as episodic memory in rich detail. Over time, illustrated by the
5
[Dudai et al., 2015]. Moreover, they do not always contain the same oral texts ranging from 40,000 to over 100,000 words, numerous regions
details, showing that memories are altered over time beyond our in the brains of the pandits were dramatically larger than those of
conscious awareness and are not retrieved as a solid entity, but rather controls, with over 10 percent more grey matter across both cerebral
reconstructed depending on present cues. The act of retrieval is gener hemispheres, and substantial increases in cortical thickness (“the
ally thought to alter a memory again, updating it with previously and famous Sanskrit Effect”) [Hartzwell, 2018]. Textual memorization is
currently learned or retrieved information. Memories are then recon standard in India’s ancient learning methods and it is widely believed
solidated into existing schemas, and in this way, schemas are thought to that exactly memorizing and reciting the ancient scripts, enhances both
be continuously adjusted to optimize our understanding of the world memory and thinking. There are other studies [Gao et al., 2019; Kamal
around us and to allow prediction of future occurrences [Benoit et al., et al., 2013] which report a significant deactivation of certain areas of
2014]. the brain as a result of religious chanting indicating relaxation and a
Also, it is generally accepted that our brains have evolved to predict meditative impact.
what will happen next. Our brain is continuously absorbing information
from the environment to expand its world model and optimize its in
ternal predictive model. Conversely, detailed episodic memories are also 2.5. The constructivist theory of education
valuable as they can help to update the model when the world happens
to change. In terms of learning, the learner’s pre-existing world model as Educational psychologists have also attempted to recognize how
it relates to the material being learned will have a huge influence on how cognitive processes work and have advocated a theory of learning called
effectively they learn. Active learning [Bonwell & Eison, 1991] is widely the Constructivist Theory of Education [Pass, 2004, p. 74]. It refers to
regarded as a seminal attempt to reflect this process in learning. the idea that learners construct knowledge for themselves and such
Based on the above review of how our brains structure knowledge knowledge construction is dependent on what their existing knowledge
and incorporate new information, it would appear that the way to schema is. Such thinking has led to the idea of ‘Active Learning’.
optimize the learning potential for a learner will be to create the best Bonwell and Eison [1991] defined active learning as “instructional ac
possible memory with both semantic and episodic details. van Kesteren tivities involving students in doing things and thinking about what they
and Meeter [2020] summarize specific actions during encoding, are doing”. Approaches that promote active learning require that stu
consolidation and retrieval that can contribute to the best possible dents do something—read, discuss, write—that requires higher-order
memory for a learner. Many of these are summarized in Fig. 4. thinking. They also tend to place some emphasis on students’ explora
In a related vein, Ghazanfar and Schroeder (2006) also find that the tions of their own attitudes and values. Approaches that promote active
brain has evolved to operate optimally in multisensory environments. If learning often explicitly ask students to make connections between new
a classroom exercise stimulates more than one of the learner’s senses at information and their current mental models, extending their under
any given time, multiple channels of the brain work to process the same standing. Fig. 6 lists various techniques to achieve ‘Active Learning’ in
information, and the brain devotes more cognitive resources to that the physical class room.
information. In addition, the brain creates some interactivity between Techniques to replace some lecturing as part of active learning have
the two modes, essentially storing the information cross-referenced be also been suggested including case-based learning, concept maps
tween the channels [Clark et al., 1991]. Thus, multisensory engagement [Novak & Canas, 2008], student-generated test questions [Angelo &
may accelerate efficient knowledge construction in the brain. Cross, 1993]. There are also other active learning pedagogies many of
Researchers at the University of Waterloo [Forrin & MacLeod, 2017] which have strong following including team-based learning,
found that speaking text aloud helps to get words into long-term process-oriented guided inquiry learning, peer-led team learning, flip
memory. The study determined that it is the dual action of speaking ped classroom model and problem-based learning.
and hearing oneself that has the most beneficial impact on memory. Several of the techniques listed in Fig. 5 signal that writing is the
Very related is the practice of rote memorization where a learner repeats most effective way to reflect understanding. It is a very critical part of all
long passages of text from memory. While its standalone impact on three phases of knowledge construction – encoding, consolidation and
learning is not proven, its general beneficial impact in improving grey retrieval. In fact, complete understanding can only be assessed based on
matter is well established. In a study of India’s Vedic Sanskrit pandits essay and open text responses. Yet, in today’s instructional settings, few
who train for years to orally memorize and exactly recite 3000-year old courses incorporate open text assessments extensively; open text ques
tions have been replaced with multiple choice questions. This is because
Fig. 4. Methods for best possible memory creation.
6
Fig. 5. Some active learning techniques.
teachers and institutions do not have the time to assess open text re learning science. These have not been explicitly reflected in the studies
sponses. Clearly, AI can play a significant role in this context. envisioning a future evolution of AIED or the AIED taxonomies proposed
Many of the Active Learning techniques are variations on findings by in the literature. While there are many successful instances in the real
neuroscience researchers. world where these findings have been used, it is largely left to the
teacher to incorporate them as they see fit. There has been no systematic
effort at the institutional level to integrate learning science techniques
2.6. Literature review summary into their curricula. One reason for this could be that it is difficult and
cumbersome to implement these techniques manually. AI holds out a
We have reviewed the literature in 4 areas that we believe are highly significant possibility to address this challenge.
relevant in the context of a holistic, integrated framework for a preferred In summary, while prior AIED research has made many valuable
AI-led future for learning. We summarize our overall findings here. contributions, we believe there is a need for conceptualizing a holistic
The AIED research reviewed in 2.1 has focused on applications of AI actionable framework reflecting the role of AI in learning.
to specific issues in learning and education. We have presented overall
findings of many of these studies as reported by several meta studies. 3. A preferred future for AI-Enabled learning
More important, we have highlighted throughout the review the limi
tations and challenges of AI applications in education. Some authors In this section, we present a preferred future for the role AI can play
have proposed taxonomies for organizing AIED research meaningfully in achieving effective learning at scale equitably. Our preferred future is
and a few have proposed their vision of a future for educational tech presented in the form of a comprehensive solution framework.
nology. Research on future pathways for AIED is partial or high level The framework approaches learning holistically and attempts to
without any specific actionable technology solution blueprint. integrate foundational learning, learning science and personalization. In
In section 2.2, research related to the changing role of teachers the process, it also addresses the gaps in the AIED literature identified in
validates their evolution as coach and facilitator but prior work in that the previous section. The framework comprises several components
field has not really integrated how AI technology can facilitate their each of which can scale specific aspects of learning irrespective of the
evolving role and scale quality education at scale. Further, there does subject. It comprises foundational components for learning combined
not appear to be any specific research related to how AI technology can with learning science optimized functional content for any topic or
address the knowledge gap created by the continuous flow of new subject or field. The framework brings together all the individual com
information. ponents through a Learning Orchestrator personalized to the individual
In section 2.3, we summarize the literature related to parental learner. Together, it provides a robust foundation that can act as a
involvement in student learning. Contrary to popular belief, the litera framework for any AIED application. The framework is also equally
ture does not seem uniformly conclusive about the beneficial impact of relevant to other stakeholders in the learning process – teachers, par
parental involvement. We did not find any direct study on how AI ents, peers, mentors, educational institution and other learning envi
technology can help parents be effective and potentially even obviate ronments such as corporations interested in re-skilling or up-skilling
the need for parents to play the role of teachers especially if they are not employees.
proficient with the subject. For the purposes of the framework, we define AI to be synonymous
Finally, section 2.4 reviews literature on knowledge construction and
7
with intelligent machines which can perform tasks that humans do. In compared algorithmically against a gold standard version. Discrepancies
telligence acquisition for these machines can be from data, from experts are analyzed presented at the syllable and phoneme level. Learners in
or from both. For most real world usage, intelligence acquisition will lower grades can use this feedback to improve their reading skills on
likely be a combination of expert intelligence and intelligence acquired their own. An adaptive personalizer presents appropriate practice pas
from data as shown in Fig. 6. We refer to this as ‘computational intel sages or words based on an analysis of areas where the learner is having
ligence’. Hence, we define AIED applications as those which incorporate difficulty consistently.
computational intelligence. Further, we propose that in the framework Another set of learners who will find this level of functionality
all AI models be explainable except in cases where explainability is not valuable are non-native (English) learners. For learners in higher grades,
needed. the Reading Assistant serves as a way of accelerating long-term memory
Explainable AI is where the specifics of the AI model, like its features, formation.
is clearly visible and understandable by subject matter experts. Deep
learning using neural networks produce black boxes which are not 3.2. Comprehension assistant
currently explainable. Methods like random forest, CART, natural lan
guage processing technologies like Gyan [www.gyanAI.com] which do The Comprehension Assistant (Fig. 8) is another foundational
not rely on statistical language models e.g., BERT and GPT3 (Generative component applicable to all content regardless of grade. The goal of this
Pre-trained Transfomer 3), are explainable. Deep neural networks may component is to accelerate understanding a word or concept, in context.
be acceptable in areas where it is not essential for the user or modeler to The Comprehension Assistant presents a ‘Meaning Graph’, which is a
understand model specifics or validate causality. network of concepts related to the word along with how they are related.
We also believe the framework must reduce reliance on the need for We believe presenting such a network of related concepts enables
large amounts of data by increasing reliance on expert knowledge where stronger memory formation.
available. A hybrid approach to knowledge acquisition can reduce the The Comprehension Assistant also allows the learner to become
dependence on large amounts of data and minimize data specific issues familiar with all the senses in which the word can be used and their
like bias. Tractability is another preferred aspect. If the only way to associated learning graphs or semantic network. Activity based learning
improve the model is to add more data without understanding whether methods and assessments can be based off of the full range of the word’s
it will improve or worsen the outcomes, it is less preferable. Again, if we semantic networks.
relied on expert knowledge for a priori feature engineering and used
explainable methods, the resulting models will be more tractable.
3.3. Cognitive learning optimizer
The framework components are briefly described below:
We envision the AI-enabled framework will integrate relevant

3.1. Reading assistant learning science algorithms seamlessly into the learning process/
curricula. We can categorize this optimizer into three buckets:
The Reading Assistant is a core component applicable to all content
regardless of grade or level from elementary school to higher ed to adult 3.3.1. Reorganize learning content & curriculum
learners. It enables the learner to read aloud any content. The Reading The framework should enable effective and systematic integration of
Assistant helps both learning to read and reading to learn. Depending on the findings summarized in Fig. 4 with course content. Specifically, the
the level of the learner, the reading assistant can assist in improving curriculum should be designed to incorporate aspects like interleaving,
pronunciation, reading fluency and in higher grades and working spacing, repetition, multi-sensory engagement, active participation and
learners, it can help with faster knowledge construction as described in worked examples. Without AI technology, this can be extremely
the previous section. The Reading Assistant can be unassisted or used in cumbersome and difficult to achieve. AI algorithms can assist signifi
a guided mode. cantly in the reconstruction of the learning content by automatically
Fig. 7 displays a screen shot from Reading Assistant which is suggesting strategies for interleaving, spacing, and repetition.
currently being launched in the k-12 schools in the U.S. The screen There can be many different spacing, interleaving and practice pat
shows assessments of the early learner’s reading – pronunciation, terns for a learning pathway or course. Fig. 9 is illustrative. We envision
reading fluency. Learners can practice on their own. Their reading is a one-time effort to create ‘learning optimized’ version of existing
Fig. 6. Machine intelligence acquisition.
8
Fig. 7. Reading assistant.
curriculum including materials and then an ongoing effort to integrate individual learner.
new material for existing courses and new courses. Another interesting area of future development is the possible
In a pioneering effort, Digital Promise has created Learner Models identification of the preferred learning mode for an individual learner by
rooted in research-based factors and strategies that honor the whole analyzing signals in brain waves through brain computer interfaces
child and adult learners to help instructors and edtech product de (BCI). Currently, BCI relies mostly on electroencephalography (EEG) for
velopers reach each learner in every learning environment (https://di monitoring the electrical activity of the brain. Using AI and non-invasive
gitalpromise.org/initiative/learner-variability-project/about/). Their sensors in wearable devices such as headbands and earbuds, it is now
models can be a starting point in the creation of automated AI algo becoming possible to analyze and extract relevant brain patterns (Rashid
rithms to optimize learning content. Of course, the framework needs to et al., 2020). Analysis of brain signals can be used for detecting
go beyond such learning models and adapt such models for each engagement levels and whether someone is focused or distracted. It can
9
Fig. 8. Comprehension assistant – learning graph.
Fig. 9. Structure learning content to maximize memory and understanding.
be used to alert the onset of fatigue. Neurable [www.neurable.com], a software.

boston based startup, has developed BCI technology to measure brain Based on 20 years of experiments with his colleagues, Mayer [2010]
activity to generate simple, real time insights. presents evidence in the context of medical education that the retention
The Personalized Learning Orchestrator described later should and transfer test performance of people who learn from a lesson with
incorporate all the above aspects to maximize the learning potential of multi-sensory engagement was consistently superior to an identical
the individual learner. lesson learnt without multisensory engagement.
3.3.2. Multi-sensory engagement 3.3.3. Optimize cognitive load

Multi sensory engagement implies the content has to be ideally We envision the framework to be able to optimize the cognitive load
enabled on multiple mediums – auditory, visual and text (Fig. 10). This for every learner. This is largely related to the way the learning material
will be extremely cumbersome to achieve without automation. AI al is presented to the learner and linked to multi-sensory engagement.
gorithms can help create the same content in different sensory modes. Graphical presentation of the content can reduce the cognitive load
Auditory can be in the form of reading aloud in the case of text associated with learning and understanding written text. Chih-Hsiang
[Schramma, 2016], visual can be in the form of learning graphs, and et al. (2013) find that mind maps based collaborative learning approach
simulations and videos as described in Mayer [2010]. This successful significantly enhanced students innovative performance in a
large scale deployment reported by Srinivasan and Murthy (2021) in project-based learning task on a university management course.
dicates that multisensory technology can be beneficial even in the face of Knowledge graphs have been suggested as a way of representing
poor classroom infrastructure, and short amounts of time spent using the knowledge for statistical machine learning; for a review of this body of
10
findings.
GyanAI [www.gyanAI.com] has created a graphical representation
of content for the purpose of learning, ‘Learning Graphs’ (Fig. 11). We
illustrated a learning graph earlier in the context of the Comprehension
Assistant. Learning Graphs contain the essence of the document in a
graphical form in the form of a network of concepts and concept re
lationships. It is presented in layers to promote easy understanding. The
central theme of the document is displayed first along with other major
concepts in the document. The user can then expand the graph along the
path they want to learn or explore. For e.g., they allow the learner to
traverse a book in a non-linear fashion if desired. The idea is that some
learners are comfortable with reading long written text and others prefer
visualization of the ideas in the text. Learning Graphs are a variation of
Concept or Knowledge Graphs with a focus on helping the learner un
derstand and comprehend chunks of material at a time driven by
learner’s interest.
Learning graphs can also be aggregated across multiple documents to
create a synthesized learning graph. Fig. 11 illustrates a synthesized
learning graph across three documents. The combined learning graph
can enable a reader to understand the contents of multiple documents
Fig. 10. Multi-sensory engagement. without having to read through redundant concepts in each document.
knowledge, see Nickel et al. (2016). Knowledge graphs are however 3.4. Automated open text assessment/assisted writing
largely limited to definitional extensions of knowledge, encapsulating
relationships like ‘also known as’, ‘is a type of’ etc. Automated grading of essays/open text responses and assisted
Mayer [2010] present the results of their detailed experiments on writing are more or less two sides of the same coin. Effective writing is
ways of optimizing cognitive load by reducing extraneous processing. not only about communicating effectively to an audience but also about
For example, people can learn better from black-and-white line draw cementing memory formation and knowledge in the writer’s brain.
ings than from color photographs, when interesting but irrelevant stories Natural language processing technologies (BERT, GPT3) provide a less
are deleted, or when stunning but irrelevant video is deleted. A second than full context ability to generate text. We see them deployed in many
example is that a lesson on how the heart works is more effective when it email and short text writing environments where they guess the next
contains headings that correspond to each of the main steps in a process. word or a reply. But as Marcus (2020) points out they can also generate
Another example, in a figure depicting the human heart, is to place text nonsensical text since they do not really understand the full context and
describing each chamber next to the corresponding chamber. In gener lack the ability to do what we might refer to as ‘common sense infer
ating learning optimized versions of learning material, we envision the encing’. Besides these language models, there are many grammar
framework to be able to suggest changes to the content to reflect Mayer’s checkers, which attempt to identify grammatical errors in one’s writing
and in many cases, suggest corrections. A more complete, fully
Fig. 11. Gyan learning graphs.
11
contextual writing assistant will be able to guide the writer with the help Grader’s assessment. This grader can also be used as a Writing
of ‘Learning Graphs’ for any concept reflecting global knowledge about Assistant.
that concept and commonly accepted rhetorical rules about effective
writing. We envision such a Writing Assistant to be part of the 3.5. Personalized learning orchestrator
framework.
The same component can be used for open text assessment. Forma As reviewed earlier, there has been substantial AIED research on the
tive and ongoing assessments are an integral part of assessing the effi benefits of personalization. An adaptive personalization component is
cacy of the educational technology, pedagogy and content. Nowadays integral to any AIED framework. It should identify the learning prefer
many courses/curricula rely mostly on multiple choice questions and ences of every learner over time and attempt to personalize their
minimize open text questions. However, complete understanding of a learning experience and optimize their learning potential [Fig. 13]. The
concept or learning unit is only possible through open text responses. It Personalized Learning Orchestrator includes a personalizer and an
is widely recognized that the inability to do open text assessment at orchestrator with the former continually personalizing the learning
scale is holding back the growth of online learning [e.g., Haldi, 2020]. It content and the latter performing the orchestration with the learner. The
is also well understood that human grading of essays is fraught with component will also integrate multisensory engagement and a cognitive
inconsistencies and other human limitations such as recency bias. load optimizer. This can be done by creating an initial set of granular
Different graders of the same essay are quite likely to grade it differently. learning pathways for the course or learning unit, identifying the com
On top of the limitations of human grading, online learning has adopted bination of learning science mechanisms that work best for a learner and
‘peer grading’. being capable of multisensory engagement with the learner including
The open text assessment component should be configurable easily at contextually rich interactive conversation. The Personalizer should also
a course or learning unit level. It should also be able to analyze the be capable of accommodating personalization at a very granular level in
response in the context of the complete learning content the learner was the learning unit, say at an individual concept level.
exposed to. It should be completely explainable and not need large We envision the creation of a meta taxonomy of learning types over
amounts of data for training. time where each atomic learning item can be categorized into a more
We report results from an explainable, course-configurable Essay abstract learning type. This might allow us to understand learning
Grader at a prominent, globally respected higher educational institution preferences at a more abstract learning type level.
which currently uses peer grading in its online courses for open text At a more macro level, the Personalizer can also use machine
response assessment. The Essay Grader was tested on a sample question learning to see if there are systemic reasons for success in specific atomic
from an engineering course. The Grader was configured to reflect the learning items or in abstract learning types across a learning cohort like
professor’s course-level rubric for grading, and considered all the a class, school or a collection of schools. The Personalizer will track
reading materials relevant for the assignment. Fig. 12 displays a com where a learner is having difficulty and provide an appropriate inter
parison of the Grader’s automated score, peer grades, and grading by 3 vention. The intervention will require the framework to assess if the
teaching assistants (TAs) who graded the same essays manually. difficulty is due to lack of prerequisite knowledge and/or difficulty
As we can see, the peer graded scores were perfect for all 10 students. understanding the material. For the former, the framework should allow
The TA scores had a wide dispersion for the same essay. The automated for the learner to access prerequisite materials and become more pro
grader had a tighter dispersion of scores for the 10 students. For each of ficient where possible. For the latter, the framework should present
the automated assessments, a detailed reasoning report was provided to additional examples, practice problems to provide the learner an op
the professor/institution explaining how the Grader arrived at the final portunity for repetition. It can also incorporate these findings in the
score/grade. In all the cases, the professor and TAs agreed with the interleaving and spacing strategy for the remainder of the course. The
Fig. 12. The explainable essay grader example (Max score = 15).
12
Fig. 13. Personalizer.
course pathway becomes personalized for the learner with timely developments in their field.
remedial interventions. The framework envisions a continuous knowledge acquirer which
would discover new relevant knowledge dynamically and present it to
3.6. Parents/para educators the learner/teacher (Fig. 14). It will access relevant content from the
internet [or other sources] on any topic, auto-curate, auto-synthesize
Parent involvement has been a major cornerstone of several imme continuously on its own. The knowledge discovery feature can be
diate past Presidents educational initiatives - President Obama’s “Race applied to any concept or topic of interest. Depending on context, the
to the Top” initiative, former President Bush’s No Child Left Behind newly discovered content can be staged for the teacher/author to review
initiative, former President Clinton’s 1996 Elementary and Secondary before presenting to the learner.
Act, and former President Reagan’s Goals 2000: Educate America Act, GyanAI [www.gyanai.com] has developed the first on-demand
and is actively promoted in national programs and initiatives (e.g. Head continuous knowledge acquisition engine, Gyan which can provide
Start). However, previous research shows inconsistent relationships auto-curated, auto-synthesized knowledge on any topic any time. Fig. 15
between parent involvement and academic achievement. shows a snapshot from Gyan has automatically found the relevant arti
While we can debate the impact of parental involvement on chil cles for the user’s topic of interest, created an appropriate table of
dren’s education, the pandemic has forced increased involvement of contents, categorized relevant articles and also created a synthesized
parents. This has resulted in a need to train parents on how to be learning graph on the right.
effective in teaching and monitoring learning at home. The AI-enabled Gyan can continuously keep this knowledge updated for any or all
framework presented in the next section eases parental burden and en concepts in the learning graph.
ables effective learning even in a completely remote learning environ
ment. The framework envisions a personalized learning orchestrator 3.8. The complete framework
which can work in an interactive fashion with the learner. Parents do not
have to take on the role of a teacher unless they want to. The framework Fig. 16 combines all the components described in this section into a
also envisions peers and learning communities as part of the learning complete framework. We envision the following components:
network. This can also ease the burden on parents. Reading Assistant – enables read aloud of any content with the
ability for the learner to control their reading.
3.7. Continuous knowledge acquirer Comprehension Assistant – facilitates understanding of specific
concepts with learning graphs.
As pointed out earlier, knowledge is continuously growing on the Automated Grader – assesses open text response in a traceable
internet on every conceivable topic. While foundational concepts and manner providing the assessor/learner complete visibility into assess
facts may not change in most cases, it will be important to have access to ment reasoning. The grader can be configured to course/assessment
new knowledge on a continuous basis. This will serve three purposes. specific rubric.
First, additional material may be useful to enhance a learner’s under Dynamic Learning Optimized Content – integrates new content/
standing of the topic/concept as they are going through a learning knowledge with existing course content/curriculum using learning sci
program. Second, this is a way for the learner to keep learning contin ence algorithms and the human teacher’s input.
uously. Third, a teacher needs an easy way to be up to date with new Learning Science Algorithms/Learning Optimizer – converts content
13
Fig. 14. Continuous knowledge acquisition.
Fig. 15. Continuous knowledge acquisition.
to be optimized for learning reflecting learning science. Continuous Knowledge – acquires relevant knowledge continuously
Learning Lab Certification – educational institutions may need to on any set of topics from any collection of sources, auto synthesizes with
create a common competency to ensure that all their offerings are existing content in a variety of forms.
properly and consistently optimized for learning. There are other important actors including peers and learning com
Personalized Learning Orchestrator – orchestrates the learner’s munities whose role can also be facilitated by the framework
learning according to the learning objectives of the course leveraging the appropriately.
other components and personalizes to the learner using machine
learning based on the learner’s preferences, performance as the course
progresses.
14
Fig. 16. An AI-Enabled framework for learning.
4. Educational institution centric view existing content repositories and the public internet. We envision that in
the future, institutions will be substantially driven by a job market
The previous section has presented our vision of an AI-enabled future which is increasingly competency based. The institution will need to
for learning from a student centric point of view. An equivalent frame rapidly design learning content in response to such market demand at an
work can also be envisioned for the educational institution and for a industry or even a company specific level. The framework envisions the
corporation interested in up-skilling or re-skilling its employees or in use of AI in rapidly designing the required learning content.
continuous talent optimization. Fig. 17 presents a dynamic learning The second aspect is that of re-skilling or up-skilling or merely life
framework from an educational institution/corporation perspective. long learning. Here a learner needs to qualify for a specific job or role.
There are two aspects to note here. The first is the need for a dynamic Fig. 17 shows how the framework can support automated identification
course authoring framework which can find relevant content from of the learner’s current skill from their resume and the normalization of
Fig. 17. Institution centric view.
15
such extracted skills with reference to a skills taxonomy. It also imagines Fourth, the intervention must be adapted to fit in the local envi
the determination of personalized learning pathways for the learner to ronment as suggested in the literature. Language, culture, practices in
achieve the competencies needed. Not all learners are at the same place. the local environment must be considered in the implementation plan.
The personalized learning pathways are determined by identifying the Finally, the framework can enable as much privacy as desired. Sensitive
learner’s skills from their resume, and matching it with the skills data like learner performance and learning style, can be protected where
required for the learning content and identifying a specific learning needed.
pathway. We have also illustrated implementations of several framework
The same framework is applicable for the corporation aiming to components individually and in combination - reading assistant,
optimize its talent pool or an individual wanting to re-skill or up-skill. comprehension assistant, multisensory engagement, cognitive load
optimization, continuous knowledge discovery and dynamic content
5. Summary creation.
In an important contribution, Hwang (2020) outlines a vision and 6. Summary

challenges for AIED in the context of setting a preferred research agenda.
The study proposes categorization of AIED studies into four roles: AIED holds enormous promise to address chronic challenges of
intelligent tutor, intelligent tutee, intelligent learning tool or partner equitable learning at scale. In this paper, we have proposed a preferred
and policy-making adviser. It will be useful to compare the framework future for AI in Education with the twin objectives of scaling quality
proposed in this paper with Hwang (2020). education affordably and optimizing the learning potential of individual
First, the primary purpose of our preferred future framework is not learners. Various components in the framework can scale different parts
for a research agenda even though it could be used for that purpose. Our of the teaching/learning model while allowing the human teacher to
framework is actionable and meant to guide AIED application efforts. It continue to evolve as a facilitator/coach in addition to being a subject
can be implemented in part or whole by one or more application matter expert. The framework strongly advocates the use of explainable
builders. Second, the framework in this paper directly maps to three of AI, the integration of scientific findings on knowledge formation and
the four roles described by Hwang (2020): intelligent tutor, intelligent learning, and the appropriate integration of humans to overcome
tutee and intelligent learning tool. Further, it can be easily extended to various previously observed challenges. The framework applies uni
envelop the policy making advisor role but we leave that for the future. versally to all learners including life long learners. The framework was
The preferred future presented in the paper reflects all the ten research also shown to extend to include the perspective of an educational
areas identified by Hwang (2020). Areas or dimensions in our preferred institution or a corporation.
future framework which have not been explicitly reflected in Hwang
(2020) include the use of AI for dynamic course authoring, continuous
knowledge acquisition, optimizing learning content to reflect learning Declaration of competing interest
science, optimizing cognitive load and the intelligent creation and
management of learning communities. We see the framework proposed None.
in this paper as a superset of the vision outlined in Hwang (2020) albeit
our primary objective is different. References
The preferred future for AIED proposed in this paper addresses the
challenges with AI identified earlier. First, it is based on an inclusive Angelo, T. A., & Cross, K. P. (1993). Classroom assessment techniques: A handbook for
definition of machine intelligence. We specify machine intelligence to be college teachers. San Francisco: Jossey-Bass.
Annual State of Education Report, India. (2019). http://www.asercentre.org//p/359.
a combination of expert knowledge and knowledge acquired from data. html.
This definition addresses issues relating to bias and the need for large Baker, T., & Smith, L. (2019). Educ-AI-tion rebooted? Exploring the future of artificial
amounts of data. If an AI modeling effort starts with expert knowledge intelligence in schools and colleges. Nesta Foundation. Retrieved from https://media.
nesta.org.uk/documents/Future_of_AI_and_education_v5_WEB.pdf (2019) Google
which is an accumulation of many years of experience, it cuts down the Scholar.
need to discover all intelligence from data and will reduce the amount of Bender, E. M., & Koller, A. (2020). Climbing towards NLU: On meaning, form, and
data needed. It will also dramatically reduce the bias in the AI model. understanding in the age of data. In Proceedings of the 58th annual meeting of the
association for computational linguistics (pp. 5185–5198). July 5 - 10, 2020.
Second, we specify that all AI models used must be explainable. This
Benoit, R. G., Szpunar, K. K., & Schacter, D. L. (2014). Ventromedial prefrontal cortex
addresses the issue of black box recommendations and findings which supports affective future simulation by integrating distributed knowledge.
will face significant user adoption issues. There may be areas where Proceedings of the National Academy of Sciences of the United States of America, 111,
explainable models may not be needed and in these cases, the modelers 16550–16555, 2014.
Beteille, T., & Evans, D. K. (2019). Successful teachers, successful students: Recruiting
are obviously free to choose any approach that they find effective. We and supporting the world’s most crucial profession. World bank policy approach to
must stress of course that the requirement for explainability does not teachers. http://documents1.worldbank.org/curated/en/235831548858735497/
imply a compromise for model performance. In our experience, we have Successful-Teachers-Successful-Students-Recruiting-and-Supporting-Society-s-Most-
Crucial-Profession.pdf.
always been able to build explainable models that performed as well or Bonwell, C. C., & Eison, J. A. (1991). Active learning: Creating excitement in the classroom.
better compared to black box approaches. ASH#-ERIC higher education report No. 1. Washington, D.C.: The George Washington
In the case of natural language processing, e.g., open text assessment, University, School of Education and Human Development.
Brookover, W. C., et al. (1979). School social systems and student achievement: Schools can
our framework envisions an explainable natural language understand make A difference. NY: Praeger.
ing platform which can retain full context and truly attempt to under Burcu, S., & Sungur, S. (2009). Parental influences on students’ self-concept, task value
stand the discourse. We briefly reported the results from implementing beliefs, and achievement in science. Spanish Journal of Psychology, 12, 106–117.
Castro Sotos, A. E., et al. (2007). Students’ misconceptions of statistical inference: A
an explainable open text grader which can be configured at a course review of the empirical evidence from research on statistics education. Educational
level and can also consider relevant learning content in assessing Research Review, 2, 98–113, 2007.
responses. Central Square Foundation. (2020). State of the sector report on private schools in India.
Chatterjee, S., & Bhattacharjee, K. K. (2020). Adoption of artificial intelligence in higher
Third, the framework needs to be implemented in phases based on
education: A quantitative analysis using structural equation modelling (in press) in.
the sophistication of the target environment. Srinivasan and Murthy Education and Information Technologies, 2020.
(2021) describe a successful phased implementation where they intro Chen, H., et al. (2020). Application and theory gaps during the rise of artificial intelligence in
duced a multi-sensory intervention first and after its large scale adop education, computers & education (Vol. 1). Artificial Intelligence, Article 100002,
2020.
tion, they have started a more sophisticated version of their intervention Chen, H., et al. (2021). Twenty years of personalized language learning: Topic modeling
for personalized learning. and knowledge mapping. Educational Technology & Society, 24(1), 205–222.
16
Chih-Hsiang, W., et al. (2013). A mindtool-based collaborative learning approach to Mayer, R. E. (2010). Applying the science of learning to medical education. Medical
enhancing students’ innovative performance in management courses. Australasian Education, 44, 543–549, 2010.
Journal of Educational Technology, 29(1). McNeal, R. E., Jr. (2014). Parent involvement, academic achievement and the role of
Clark, J., et al. (1991). Dual coding theory and education. Educational Psychology Review, student attitudes and behaviors as mediators. Universal Journal of Educational
3, 149–210. Research, 2(8), 564–576. https://doi.org/10.13189/ujer.2014, 2014 http://www.
CNBC. (2020). https://www.cnbc.com/2020/02/12/bill-and-melinda-gates-say-educat hrpub.org.
ion-philanthropy-is-not-having-impact.html. Muller, C. (1998). Gender differences in parental involvement and adolescents’
Coleman, J. (1991). Parent involvement in education. Policy perspective. Office of educational mathematics achievement. Sociology of Education, 71, 336–356.
research and improvement. Washington, D.C.: U.S. Department of Education. Nickel, M., et al. (2016). A review of relational machine learning for knowledge graphs.
D’Amour, et al. (2020). Underspecification presents challenges for credibility in modern Proceedings of the IEEE, 104(1), 11–33. Jan. 2016.
machine learning. https://arxiv.org/abs/2011.03395. arXiv.org. Novak, J. D., & Canas, A. J. (2008). The theory underlying concept maps and how to
Domina, T. (2005). Leveling the home advantage: Assessing the effectiveness of parent construct and use them. Technical Report IHMC.
involvement in elementary school. Sociology of Education, 78, 233–249. Paller, K. A., & Wagner, A. D. (2002). Observing the transformation of experience into
Dudai, Y., Karni, A., & Born, J. (2015). The consolidation and transformation of memory. memory. Trends in Cognitive Sciences, 6, 93–102, 2002.
Neuron, 88, 20–32. Pass, S. (2004). Parallel paths to constructivism: Jean Piaget and Lev Vygotsky. Information
El Nokali, N., Bachman, N., & Votruba-Drzal, E. (2010). Parent involvement and Age Publishing, ISBN 1593111452.
children’s academic and social development in elementary school. Child Patel, N. (2006). Perceptions of student ability: Effects on parent involvement in middle
Development, 81(3), 988–1005. school. Dissertation Abstracts International: Section A. Humanities and Social Sciences,
Epstein, J. (1991). Effects on student achievement of teachers’ practices of parent 67(3-A), 838.
involvement. Advances in Reading, 5, 261–276. Pinkwart, N. (2016). Another 25 years of AIED? Challenges and opportunities for
Forrin, N. D., & MacLeod, C. M. (2017). This time it’s personal: The memory benefit of intelligent educational technologies of the future. International Journal of Artificial
hearing oneself, in memory. https://doi.org/10.1080/09658211.2017.1383434. Intelligence in Education, 26(2), 771–783. https://doi.org/10.1007/s40593-016-
October 2 2017. 0099-7, 2016.
Gao, J., Leung, et al. (2019). The neurophysiological correlates of religious chanting. Popenici, S. A. D., & Kerr, S. (2017). Exploring the impact of artificial intelligence on
Scientific Reports, 9, 4262. https://doi.org/10.1038/s41598-019-40200-w, 2019. teaching and learning in higher education. Research and Practice in Technology
Ghazanfar, A. A., & Schroeder, C. E. (2006). Is neocortex essentially multisensory? Trends Enhanced Learning, 12(22), 1–13, 2017.
in Cognitive Sciences, 10(6), 278–285. Popova, A., Evans, D., & Arancibia, V. (2016). Training teachers on the job : What works
Guan, C., et al. (2020). Artificial intelligence innovation in education: A twenty-year and how to measure it (English). World Bank Group. Policy Research working paper
data-driven historical analysis. International Journal of Innovation Studeies, 4(4), no. WPS 7834;WDR 2018 background paper Washington, D.C http://documents.wor
134–147, 2020. ldbank.org/curated/en/122031474915232609/Training-teachers-on-the-job-what
Haldi, T. C. (2020). Online learning’s grading challenge. Blog post. https://intelligentmach -works-and-how-to-measure-it.
ineslab.com/online-learnings-grading-challenge/. (Accessed 18 August 2020). Rashid, M., Sulaiman, N., et al. (2020). Current challenges, and possible solutions of
Hartzwell, J. (2018). A neuroscientist explores the "Sanskrit effect. Scientific American. Jan EEG-based brain-computer interface: A comprehensive review. Frontiers in
2, 2018. Neurorobotics. https://doi.org/10.3389/fnbot.2020.00025, 03 June 2020.
Henderson, A. (1991). The evidence continues to grow: Parent involvement improves student Roediger, H. L., & Mcdermott, K. B. (1995). Creating false memories—remembering
achievement. Washington, D.C: National Committee for Citizens in Education. words not presented in lists. Journal of Experimental Psychology Learning, 21,
Hill, N., et al. (2004). Parent academic involvement as related to school behavior, 803–814, 1995.
achievement, and aspirations: Demographic variation across adolescence. Child Rummel, N., Walker, E., & Aleven, V. (2016). Different futures of adaptive collaborative
Development, 75(5), 1491–1509. learning support. International Journal of Artificial Intelligence in Education, 26(2),
Hwang, G. (2003). A conceptual map model for developing Intelligent Tutoring Systems. 784–795. https://doi.org/10.1007/s40593-016-0102-3, 2016.
Computer Education, 40(3), 217–235, 2003 https://doi: 10.1016/S0360-1315(02) Schiff, D. (2020). Out of the laboratory and into the classroom: The future of artificial
00121-5. intelligence in education. AI & Society, 1–18, 2020 Aug 9.
Hwang, G. (2014). Definition, framework and research issues of smart learning Schramma, E. (2016). Multisensory learning and technology in the acquisition of English as a
environments – a context-aware learning perspective. Smart Learning Environments, 1 second language, working paper. EnglishHelper Inc.. Jan 2016.
(4). Nov 2104. Srinivasan, V., & Murthy, H. (2021). Improving reading and comprehension in K-12:
Hwang, G. (2020). Vision, challenges, roles and research of artificial intelligence in Results for a large scale AI intervention in India. Computers & Education: Artificial
education. Computers & Education: Artificial Intelligence, 1, 100001, 2020. Intelligence, 2(2021), 100019. https://doi.org/10.1016/j.caeai.2021.100019
Iiyoshi, T., & Vijay Kumar, M. S. (2010). In Iiyoshi, & V. Kumar (Eds.), Opening up Srinivasan, V. (2017). The Intelligent Enterprise in the Era of Big Data. NY: John wiley &
education: The collective advancement of eucation through open technology, open content sons.
and open knowledge. MIT Press, 2010. St Jacques, P. L., Olm, C., & Schacter, D. L. (2013). Neural mechanisms of reactivation
Janda, H. K., Pawar, A., Du, S., & Mago, V. (2019). Syntactic, semantic and sentiment induced updating that enhance and distort memory. Proceedings of the National
analysis: The joint effect on automated essay evaluation. IEEE Access, 7, Academy of Sciences of the United States of America, 110, 19671–19678, 2013.
108486–108503, 2019. Tseng, J. C. R., et al. (2008). Development of an adaptive learning system with two
Johnson, B. G., Phillips, F., & Chase, L. G. (2009). An intelligent tutoring system for the sources of personalization information. Computers and Education, 51(2), 776–786.
accounting cycle: Enhancing textbook homework with artificial intelligence. Journal https://en.unesco.org/news/covid-19-school-closures-around-world-will-hit-girls-hard
of Accounting Education, 27, 30–39, 2009. est, (2020).
Kamal, N. F., Mahmood, N. H., & Zakaria, N. A. (2013). Modeling brain activities during van Kesteren, M. T. R., et al. (2012). How schema and novelty augment memory
reading working memory task : Comparison between reciting Quran and reading formation. Trends in Neurosciences, 35, 211–219.
book. Procedia - Social and Behavioral Sciences, 97, 83–89, 2013. van Kesteren, M. T. R., & Meeter, M. (2020). How to optimize knowledge construction in
Kao, G. (2004). Parental influences on the educational outcomes of immigrant youth. the brain”, npj Science of Learning (2020) 5:5 https://doi.org/10.1038/s41539-020
International Migration Review, 38, 427–450. -0064-y.
Kolowich, S. (2014 April 28). Writing instructor, skeptical of automated grading, pits Xie, H., et al. (2019). Trends and development in technology-enhanced adaptive/
machine vs. machine. The Chronicle of Higher Education. personalized learning: A systematic review of journal publications from 2007 to
Lee, J., & Bowen, N. (2006). Parent involvement, cultural capital, and the achievement 2017. Computers & Education, 140, 103599, 2019.
gap among elementary school children. American Educational Research Journal, 43 Yang, et al. (2019). Trends in artificial intelligence-supported elearning: A systematic
(2), 193–218. review and co-citation network analysis (1998–2019). Interactive Learning
Marcus, G. (2020). The next decade in AI: Four steps towards robust artificial Environments. https://doi.org/10.1080/10494820.2021.1875001
intelligence. https://arxiv.org/abs/2002.06177. arXiv.org.
17

1 s2.0 S2666920X22000170 Main

Uploaded by

Copyright:

Available Formats

1 s2.0 S2666920X22000170 Main

Uploaded by

Document Information

Original Title

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

1 s2.0 S2666920X22000170 Main

Uploaded by

Copyright:

Available Formats

Computers and Education: Artificial Intelligence 3 (2022) 100062

Contents lists available at ScienceDirect

Computers and Education: Artificial Intelligence

AI & learning: A preferred future

1. Introduction well as those underemployed or seeking career changes, gig-workers,

E-mail address: Venkat.Srinivasan@Innosparkventures.com.

changes in behavior like greater willingness to adopt AI technologies. learner.

digital environment including the internet and digital technology.

2.3. The role of parents as educators

The pandemic we are living through has forced parents to play a

Fig. 2. The disaggregation of the Teacher’s role.

1970s to the late 2000s).

2.4. The neurobiology of learning

Fig. 4. Methods for best possible memory creation.

Fig. 5. Some active learning techniques.

We envision the AI-enabled framework will integrate relevant

Fig. 6. Machine intelligence acquisition.

Fig. 7. Reading assistant.

Fig. 8. Comprehension assistant – learning graph.

Fig. 9. Structure learning content to maximize memory and understanding.

be used to alert the onset of fatigue. Neurable [www.neurable.com], a software.

3.3.2. Multi-sensory engagement 3.3.3. Optimize cognitive load

Fig. 11. Gyan learning graphs.

Fig. 13. Personalizer.

Fig. 14. Continuous knowledge acquisition.

Fig. 15. Continuous knowledge acquisition.

Fig. 16. An AI-Enabled framework for learning.

Fig. 17. Institution centric view.

In an important contribution, Hwang (2020) outlines a vision and 6. Summary

You might also like