Jrcb4 The Impact of Artificial Intelligence on Learning final

address this challenge. . 
Recent AI breakthroughs are based on supervised machine learning. A critical success 
factor of these systems is the availability of huge amounts of pre-categorized training 
data. In contrast to logic- and knowledge-based approaches to AI, we therefore 
characterize these as “data-based” AI systems in this report. Many of these “deep-
learning” neural AI systems may well be characterized as “datavores.” At present, the 
most important technical bottleneck of AI, therefore, is the availability of data. 
This is a qualitatively new development in the history of computing and information 
processing. Without access to vast training datasets, it is very difficult to develop 
successful AI systems. In this report, we put forward an argument that EU policies could 
create data platforms that could redefine the competitive landscape for learning- and 
education-oriented AI systems. 

---

4 
As these supervised AI learning algorithms are based on historical data, they 
can only see the world as a repetition of the past. This has deep ethical 
implications. When, for example, students and their achievements are assessed using 
such AI systems, the assessment is necessarily based on criteria that reflect cultural 
biases and historically salient measures of success. Supervised learning algorithms create 
unavoidable biases, and these are currently extensively debated. From a more 
fundamental ethical point of view, however, the expression of human agency requires 
capability to make authentic choices that do not only repeat the past. Although there are 
already AI systems that deal with creative activities, AI systems will have great 
difficulties in dealing with people who are creative, innovative, and not only average 
representations of vast collections of historical examples.
It is often assumed that AI systems enable new levels of personalisation and diversity for 
information systems; much of this, however, results from fine-grained categorization that 
puts users into pre-defined classes. Although these systems may be able to efficiently 
simulate personalisation, they do not necessarily support deeper levels of diversity. At 
present we can say that the use AI systems in educational settings will shape the 
development of human cognition and self-efficacy, but we don’t know how. It is therefore 
important to continuously evaluate, for example, how the use of AI in educational 
contexts constrains and enables human possibilities for responsible and ethical action. AI 
systems can be excellent predictive machines, but this strength may be an important 
weakness in domains where learning and development are important. A contribution of 
this report is to show that different types of AI and machine learning systems operate on 
different layers of human behaviour
4
. Most importantly, the level of meaningful