WP121 Web · Page 13

10
  ILO Working Paper 121
country variation (Erumban et al., 2024), the overall productivity growth has been almost zero in 
LAC ever since the start of the global productivity slowdown of the last 10 years (Dieppe, 2021). 
Compared to other regions, barriers to innovation and technology adoption have been particu­
larly salient factors limiting productivity growth in LAC. 
Could GenAI help unlock this productivity impasse? Recent empirical studies focused on the use 
of GenAI in particular occupational settings suggest that the positive impacts on productivity can 
be large. For example, Peng et al. (2023) implemented a controlled experiment among profes­
sional programmers and found that access to a GenAI assistant reduced the time to complete 
programming tasks by 56 percent. Brynjolfsson et al. (2023) find that access to GenAI increases 
productivity among customer support workers in terms of issues resolved per hour, which is driv­
en mostly by the boost of performance among the novice and low-skill workers. Similarly, Noy 
and Zhang (2023) find that having access to ChatGPT helps improve the productivity of writing 
professionals, by increasing the quality of outputs as well as by reducing the amount of time re­
quired to produce them, with the benefits being the largest for low-ability workers.
While the results of this literature suggest a promising role for GenAI to boost productivity, in the 
context of LAC and emerging economies more broadly, there are important reasons to be cautious.
First, there are good chances that such initial macroeconomic projections are too optimistic and 
based on oversimplified models. As shown by Acemoglu (2024), when the tasks classified by  
Eloundou et al. (2023) as exposed to GenAI in the context of US-based occupations are linked to 
their actual impact on GDP, the average task-level savings and the economic viability AI deploy­
ment (Svanberg et al., 2024), the estimated impact amounts to a modest 0.71 of additional Total 
Factor Productivity at the end of a 10 year period. Accounting for hard-to-learn tasks drops this 
estimate to 0.55 percent of TFP, corresponding to an additional GDP growth due to AI at 0.92 per­
cent over 10 years. In addition, the actual impact on productivity in specific occupations might 
be largely dependent on how these technologies will be implemented at the workplace. For ex­
ample, Doellgast et al., (2023) suggest that productivity benefits might be less consequential if 
the new AI tools are applied mainly for worker control, thereby limiting creativity and the oppor­
tunities for larger value added through innovation in products and services. Acemoglu (2024) 
also demonstrates that the final impact on productivity largely depends on the type of new tasks 
that will emerge due to adoption, and that some of such new tasks might either not contribute 
much new economic value or produce outright “public bads” that can be wrongly accounted as 
part of GDP growth based exclusively on their monetary value.3
Second, the rate of adoption and exposure to GenAI is likely to be slower in developing coun­
tries where fewer workers are using digital technologies than in their richer counterparts. More 
specifically, two individuals with the same occupation could have very different levels of GenAI 
exposure if one of them uses a computer or internet at work, while the other one does not. As 
shown in Figure 3, internet access in LAC countries varies from anywhere below 50 p