Jrcb4 The Impact of Artificial Intelligence on Learning final

3
 
Impact on learning, teaching, and education 
Since the beginning of the 1980s, and until recently, educational applications of AI have 
mainly focused on the knowledge-based approach.
61
The most prominent line of research 
has been concerned with intelligent tutoring systems, or ITS.
62
These systems use a 
knowledge-based architecture. A typical ITS architecture has a domain model that 
describes the area to be learned and a student model that describes the current state of 
student's knowledge and learning. An expert system or pedagogical model manages the 
introduction of learning materials to the student through an adaptive and interactive user 
interface. 
These systems have traditionally used the knowledge-based approach, now commonly 
known as "gofai" (good-old-fashioned-AI). They have been successful mainly in relatively 
limited and unambiguous domains, such as mathematics and physics.
63
As student 
behaviour and learning can also be monitored in ITS environments in great detail, 
intelligent tutoring environments have also been an important source of data for research 
on learning.
64
The difficulty in developing ITS for broad learning domains has also 
switched the focus to the more narrow problem of using AI and machine learning to 
generate teacher interfaces for student and learning monitoring, and learning 
diagnostics. This is commonly known as learning analytics and educational data mining 
(EDM).
65
3.1
 
Current developments 
In special needs education, AI-based approaches have shown potential, for example, in 
the early detection of dyslexia.
66
A well-published example is the Swedish company 
“Lexplore” that has developed a system that quickly scans for students at risk and 
detects dyslexia by tracking reader eye movements. The system uses data-based pattern 
recognition, and the company is now expanding to the US and UK, offering school and 
school-district wide scanning.
67
AI-based systems have also been successfully developed 
for the diagnosis of autism spectrum disorder and attention deficit hyperactivity disorder 
(ADHD). In particular, child-robot interaction seems to enable new forms of diagnostics 
and special needs educational applications.
68
As student testing plays an important role in many educational systems, many projects 
are trying to explore the use of AI for automatic test generation and assessment. Much of 
this work is aimed at automating summative assessment, with a promise of reducing 
teacher workloads. A possible unintended consequence of this work is that high-stakes 
testing will be increasingly displaced by frequent low-stakes formative assessment, as 
the effort and cost required for assessment decreases. Current AI systems are very good 
in combining evidence from complex and varied sources of data and using them for real-
time pattern recognition. For example, student homework can relatively easily be 
checked and diagnosed by an AI system that has data on both individual student history 
and peer responses. Accumulated formative assessments could, therefore, to a large 
extent make high-stakes testing redundant. AI is also beginning to be used to diagnose 
student attention, emotion, and conversation dynamics in computer-supported learning 
61
For an early example, see Sleeman and Brown (1982). 
62
E.g., Woolf (2009). 
63
E.g. Ritter et al. (2007), Graesser et al. (2005). 
64
E.g., Porayska-Pomsta (2015). 
65
For a compact review of some relatively recent developments, see Luckin et al. (2016) and a JRC report on 
Learning Analytics by Ferguson et al. (2016) 
66
See, e.g., Drigas and Ioannidou (2012). 
67
Jakobsson (2017). For English version, see http://www.lexplore.com/ 
68
E.g., Scassellati (2012), Boccanfuso et al. (2016). 

---

28 
environments, for example for course development and management, in an attempt to 
generate optimal groups for collaborative learning tasks, and to recognize patterns that 
predict student drop-out.
69
To do this effectively, large datasets are needed for training 
the systems. As was pointed out above, this is a major technical bottleneck. Student 
behavior also has to be actively monitored to provide feedback for learning. This creates 
technical needs to unobtrusively monitor students, for example, using video processing 
and remote eye tracking, with associated ethical and regulatory challenges. Ethically less 
problematic are systems that use less granular data to provide recommendations. For 
example, at UC Berkeley students can now get course recommendations using a system 
that relies on neural AI technologies originally developed for natural language processing 
and machine translation.
70