ML technologies such as deep learning, image & speech recognition, natural language processing, and predictive analytics have empowered machines to match, or even surpass, human capabilities of performing certain types of tasks, thereby, engendering economically significant application across diverse sectors and occupations.
While advances in Maching Learning (ML) are significant, and automation is already having significant effects on many parts of the workforce, this raises the question of which tasks will be most affected by ML and which will be relatively unaffected. The survey instrument, which was based on Brynjolfsson & Mitchell (2017), had a task evaluation rubric that comprised 23 questions pertaining to SML.
A survey of approx. 3,100 individuals employed in diverse occupations across India was conducted during June-December 2019 to assess the suitability of machine learning and, in turn, the resultant susceptibility of 106 Indian occupations, as defined by the National Classification of Occupations (NCO 2004).
Occupations such as painters, building structure cleaners, administrative associate professionals among others have high SML scores whereas occupations involving mining, construction, potters, glass makers, domestic and related helpers, cleaners among others score low on the SML Index.
Refer Fig. 1
The sectoral SML index shows high scores for creative, arts, entertainment, architecture and engineering activities, computer programming, among others. Similar to occupations, sectors such as mining, quarrying, building construction, among others score low on the SML sectoral Index scale.
Refer Fig. 2
We use the SML index to understand how the COVID-19 pandemic differentially impacted industrial sectors in India. Our hypothesis, prior to the data analytics, was that sectors with higher SML scores would see less adverse impact of COVID on economic activity.
Extending the methodology used in Eyre et al. (2020), we use publicly available Twitter posts of NIFTY 500 firms as a proxy for economic activity. We assess the impact of sectoral scores of SML on firm-level tweeting activity, including quantum and sentiment of tweets.
For the analysis, we compared the time series of volume and sentiment of Twitter posts of the NIFTY-500 firms after the lockdown with their average posting activity before the lockdown. Using the sample of non-COVID related tweets (Panel B of Fig. 1), we find that firms, on average, decrease their tweeting activity during the lockdown. However, those in high SML sectors see an increase in overall posting frequency.
For both panels in Fig. 1, the dependent variable in columns (1)-(3) is the log(count of tweets) and in columns (4)-(6) is a dummy variable equal to one if a firm tweeted on the focal day and zero otherwise. The results are based on estimating the following equation:
where yif,t is one of the dependent variables. The lockdown dummy, 𝐿𝑜𝑐𝑘𝑑𝑜𝑤𝑛t equals one for the post-lockdown period (March 25, 2020 - April 14, 2020) and zero for the pre-lockdown period (Mar 4, 2020 – March 24, 2020). The results in Panel A of Fig. 1 are based on all tweets, whereas those in Panel B of Fig. 2 are based on non-COVID tweets. The observations for regressions are at Firm x Date level. Clustered standard errors are reported at Firm x lockdown level.
For both panels in Fig. 2, the dependent variable in columns (1)-(3) is a continuous variable which ranges from -1 to 1. The measure represents overall sentiment score as calculated using the algorithm as documented in Hutto & Gilbert (2014). The dependent variable in columns (4)-(6) is a dummy variable equal to one if the average sentiment of combined tweets is positive in week w and zero otherwise. The results are based on estimating the following equation:
where (𝑇𝑤𝑒𝑒𝑡 𝑆𝑒𝑛𝑡𝑖𝑚𝑒𝑛𝑡)𝑓,𝑤 is one of the above described dependent variables. 𝐿𝑜𝑐𝑘𝑑𝑜𝑤𝑛𝑤 is a dummy equal to one if the week was under lockdown, otherwise zero. The lockdown dummy equals one for each of the three-week periods in the post-lockdown period (March 25, 2020 - April 14, 2020) and zero for the weeks in the pre-lockdown period (Mar 4, 2020 - March 24, 2020). 𝑆𝑀Ljf is the sectoral SML index for firm f in sector j. The results in Panel A in Fig. 2 are based on all tweets while in Panel B in Fig. 2 they are based on non-COVID tweets. The Sector FE is at 2-digit NIC level. The observations for regressions are at Firm × Week level. Clustered standard errors are reported at Firm × lockdown level.
Sectors and occupations with greater levels of digitization will be able to counter the impact of COVID-19 pandemic, thereby demonstrating a high degree of digital resilience when compared to their peers with lower levels or no digitization.
With growing advancements in machine learning (ML) and artificial intelligence, various sectors have adopted these technologies to derive wide-ranging benefits. We studied the impact of these emerging technologies in various occupations and sectors of India using suitability of machine learning (SML) scores for these occupations and sectors.
The scores have implications for labor markets. The SML score, when correlated with labor market outcomes and firm performance, can provide valuable and nuanced insights that go beyond job losses to include complementing labor, increasing demand for it by lowering costs, changing demand by changing overall income, changing information flows, or reorganizing work. Much attention has been devoted to substitution of human labor by machines but the biggest effects in the coming years could well be a combination of the above impacts.
We expect our assessment of SML of diverse occupations and sectors will drive future research that will help inform policymakers, researchers, and industry executives about the implications of artificial intelligence and machine learning.