Jrcb4 The Impact of Artificial Intelligence on Learning final

2.4.5
 
Neural AI as data-biased technological change 
A recent study by Nedelkoska and Quintini
50
at the OECD provides a good review of 
econometric research on the impact of automation, and extends the Frey and Osborne 
study using the results of the OECD Survey of Adult Skills (PIAAC). Nedelkoska and 
Quintini matched the technical bottlenecks from Frey and Osborne to PIAAC variables on 
job tasks, such as frequency of complex problem solving and advising or teaching others. 
The variables used by Nedelkoska and Quintini are shown in Table 2. For the overall 
sample of 32 countries, they found that the median job had a 48 per cent probability of 
being automated, with large variations across countries. 
Table 2: Technical bottlenecks for automation
Engineering 
bottlenecks 
Variable in PIAAC 
Description 
Perception manipulation 
Fingers (dexterity) 
How often - using skill or accuracy with your hands or 
fingers? 
Creative intelligence 
Problem solving, simple 
How often - relatively simple problems that take no 
more than 5 minutes to find a good solution? 
Problem solving, complex 
Problem solving - complex problems that take at least 
30 minutes thinking time to find a good solution? 
Social intelligence 
Teaching 
How often - instructing, training or teaching people, 
individually or in groups? 
Advise 
How often - advising people? 
Plan for others 
How often - planning the activities of others? 
Communication 
How often - sharing work-related information with co-
workers? 
Negotiate 
How often - negotiating with people either inside or 
outside your firm or organization? 
Influence 
How often - persuading or influencing people? 
Sell 
How often - selling a product or selling a service? 
Source: Adapted from Nedelkoska & Quintini, 2018 
 
Economists have used both skill-biased and task-biased models to study the impact of 
automation, computers and AI. Neural AI and machine learning, however, do not fit 
these models well. The critical bottleneck is not whether a task is routine or non-routine, 
or whether it requires complex problem solving; instead, it is whether the task can be 
learned by a computer. This, in turn, depends on whether there are data that can be 
used for learning. The impact of AI on occupations can, therefore, best be understood in 
a “data-biased” model. If data are available and history repeats itself, current machine 
learning algorithms can at least in principle simulate the past. To the extent that 
learning, innovation and knowledge creation is about combining existing pieces of 
knowledge, machines may also be able to do that. From a technical point of view, such 
operations are purely syntactic. There are good reasons to expect that social, economic, 
and cognitive processes, as well as other systems that can be called living, cannot be 
simulated using such an approach.
51