Does better AI make teleworkers more productive, or more replaceable?

Does better AI make teleworkers more productive, or more replaceable?

Factful Friday (based on VoxEU.org column joint with Toshihiro Okubo)

29 March 2024

By Richard Baldwin 瑞士洛桑国际管理发展学院 (IMD) - 商学院

Readers of Factful Fridays will have almost surely, within that last 48 hours, read a new story or blog, or listened to a podcast on how AI is changing the nature of work (Brynjolfsson et al. 2023). And most of you will have also engaged with information about how telework is changing the nature of work (Barrero et al. 2023, Palmou 2021, Drenik 2022).

But have you considered how these two are related? Probably not, and that’s strange. Both trends are driven by digital technology and so advancing at an eruptive pace. And both trends are affecting the same sorts of jobs – white collar and professional jobs. So, why isn’t there more writing on whether the trends are complements or substitutes when it comes to their impact on jobs?

  • Are AI and RI substitutes or complements?

Of course, some of us wrote a whole book on it in 2019, cough, The Globotics Upheaval: Globalisation, Robotics, and the Future of work, cough, but even those 300 pages did not get directly to the substitutes/complements question.

New data.

Back in 2018, when I was finishing the manuscript, Toshi Okubo (a professor of economics at Keio University) and I started to gather data on the use of AI in Japanese workplaces, and remote work in Japanese workplaces. And when I say ‘we’, I mean Toshi was the driving force, and I came along for the ride.

The results are in! They are preliminary. No econometrics (we are looking for tooled-up labour economists to jointly exploit the data), but the unconditional results are suggestive. In some cases, Japanese firms who did more AI investment also did more telework.

Just as an aside, I’m writing this Factful Friday from a remote local (on my way to Seoul) and I’m using Chatty to help me rewrite the Vox column into this form. I must say that I find the AI is definitely complementary to the RI since it is much easier to write stuff without reference material when I have Chatty as my wingman.

RI and AI: How do they interact in the data?

If we conceptualize telework as 'Remote Intelligence', or RI for short, we can phrase the core research question is a snappy way: Are AI and RI complements or substitutes. However, the discussion does not encompass the entirety of AI's domain. Hence, we refer to the segment of AI that mirrors the capabilities traditionally ascribed to white-collar workers as "white-collar robots", setting them apart from their industrial counterparts—the so-called blue-collar robots—that populate manufacturing environments.

The apparent chasm between the discourses on AI and RI is intriguing, if not bewildering. It might seem quixotic to speculate on this yet, this oversight becomes all the more perplexing when reflecting upon how academia approached the intersection of technology and the economy in the past, particularly through the prism of Information and Communication Technology (ICT). The seminal work by Feenstra and Hanson (2001), for instance, meticulously dissected the dual impacts of ICT-induced wage disparity (attributable to skill-biased technological advancements) and the intensifying competition borne out of globalization (manifested through offshoring).

The survey data.

To study AI and RI interactions, Okubo and NIRA (various) set up a volley of surveys across Japan (see graphic below for the survey waves and Covid events and cases in Japan). This endeavour collected, in multiple rounds, data from approximately 10,000 workers. The timing of our initial survey phase could not have been more fortuitous, commencing just as the prelude to the COVID-19 pandemic in Japan and extending through to the latter part of 2022.

The advent and the eventual ebbing of the pandemic acted as an exogenous shock, providing us with a unique vantage point from which to examine the interconnections between AI and RI. In our recent study, detailed in Baldwin and Okubo (2023), we aim to empirically discern whether AI and RI function more as complements or substitutes within the service sector. Our preliminary findings lean towards a complementary relationship between the two, indicating a synergistic rather than a competitive interaction. But that’s getting ahead of the story.

?COVID-19 restrictions and deaths in Japan March 2020 to July 2022.

Source: OWID, https://ourworldindata.org/coronavirus/country/japan. Black line is the number of Covid cases in Japan.

The Globotics Quadrant.

The first set of stylised facts that are useful in addressing the complementary/substitutability question relates to the teleworkability and the automatability of various jobs. Not all jobs are amenable to both.

  • AI can’t be a substitute for RI when it comes to jobs that aren’t teleworkable.

Certain occupations necessitate physical co-presence, rendering them resistant to telework, while others are characterized by tasks that cannot yet be replicated by what we term 'white-collar robots', hence are deemed non-automatable. Within this framework, the conversation around whether telework and automation act as complements or substitutes gains relevance primarily for those occupations that are amenable to both telework and automation. The critical inquiry then becomes: what proportion of occupations fall within this category?

To systematically explore this query, we introduce a scatter plot named the Globotics Quadrant (Figure 1), serving as a visual representation of the intersection between teleworkability and automatability across various occupations. The methodology for assessing teleworkability draws upon Dingle and Neiman (2020), which provides a vertical scale for this dimension, while the automatability of each occupation is gauged using the framework established by Frey and Osborne (2013). The plot is bisected horizontally and vertically at the mean values for all surveyed occupations, creating a quadrant system wherein occupations situated to the left of the vertical demarcation are deemed less automatable than average, and those below the horizontal threshold are considered less teleworkable than average.

This visualization, which we have dubbed the Globotics Quadrant, is showcased with data pertaining to the United States within our analysis. For insights into how these dynamics play out within the context of Japanese occupations, we invite readers to consult the detailed discussion presented in our paper. Through this analytical lens, we aim to elucidate the nuanced interplay between automation and teleworkability across the occupational landscape.

Figure 1: The US globotics quadrant: occupations by automatability and teleworkability.

Source: Authors’ elaboration of data from Dingel-Neiman (teleworkability), and Frey-Osbourne (automatability). Note: Each point represents an occupation; x-axis shows the automatability score (from 0 to 1), and y-axis is teleworkable Score (0 to 1). Occupations grouped into Japan’s NIRA38 aggregates. The point labels refer to the Japanese occupation categories. 1, Administrative and managerial workers 2, Researchers 6, Data processing and communication engineers 12, Management, finance and insurance professionals 19, General clerical workers 20, Accountancy clerks 21, Production-related clerical workers 25, Office appliance operators 3, Agriculture, forestry, and fishery engineers 4, Manufacturing engineers 5, Architects, civil engineers and surveyor 7, Doctors, dentists, veterinarians, and pharmacists 9, Medical Technology and Healthcare Professionals 11, Legal Professionals 13, Management and business consultants 14, Teachers ?15, Workers in religion 16, Authors, journalists, editors 17, Artists, designers, photographers, film operators ?18, Other specialist professionals ?8, Public health nurses, midwives, and nurses 10, Professional social welfare workers 22, Sales clerks 23, Outdoor service workers 24, Transport and post clerical workers 26, Sales workers ?27, Workers in Family Life Support and Care Service 28, Occupational health and hygiene service workers 29, Food and drink cooking, staff serving customers 30, Manager of residential facilities and buildings 31, Other service workers 32, Security workers 33, Agriculture, forestry and fishery workers 34, Manufacturing process workers 35, Transport and machine operation workers 36, Construction and mining workers 37, Carrying, cleaning, packaging, and related workers 38, Other .


The insights gleaned from the Globotics Quadrant are straightforward, but critical. They underscore the multifaceted impact of digital technology across the occupational spectrum, which leads to several key observations:

1.?????? The distribution of occupations across all four quadrants of the Globotics Quadrant illustrates the diverse ways in which advancing digital technologies, such as AI and RI, will influence different jobs.

This dispersion effectively negates any possibility of a one-size-fits-all answer to whether AI and RI function as complements or substitutes across the board. Each occupation's unique positioning within the quadrant underscores the need for a tailored analysis to understand the specific implications of digital advancements.

2.?????? An analysis of the occupations situated within the Northeast quadrant reveals that approximately 12 million workers, accounting for about 10% of the occupations we classified, are found in this segment (detailed in Table 1 of the Annex).

This subset of the workforce operates in jobs that are both highly teleworkable and automatable, highlighting a specific segment that could be at the forefront of experiencing the dual forces of AI and RI.

3.?????? Furthermore, our examination does not uncover a definitive correlation between an occupation's teleworkability and its susceptibility to automation.

This absence of a clear relationship further emphasizes the complexity of the digital transformation landscape and cautions against oversimplified interpretations or predictions.

4.?????? These foundational observations necessitate a nuanced approach to understanding the evolving dynamics at the intersection of AI, RI, and the future of work.

It is clear that the implications of digital advancements are not uniform but rather highly occupation-specific, demanding a nuanced and thoughtful analysis to navigate the ongoing transformations effectively.

Figures 2 and 3 show the variation over time of RI and AI usage in Japan over the period of the survey. It shows that both RI and AI rose with the Covid shock and then stayed up.

Figure 3: Telework usage over time.

Source: Okubo-NIRA dataset.

Figure 4: IT tool usage over time.

Source: Okubo-NIRA dataset.

Evidence from software investments: main results.

To distil the complexity inherent in the diverse responses collected through our survey, we synthesized the data into two composite indices:

  • One reflecting attitudes and experiences towards Artificial Intelligence (AI) - the pro-AI index, and
  • The other towards Remote Intelligence (RI) - the pro-RI index.

For an in-depth explanation of the methodology, see our papers. The insights derived from these indices, as illustrated in Figure 5, offer a suggestive narrative on the adoption and impact of AI and RI technologies across different occupational categories.

The occupational landscape was categorized into three overarching groups for analysis:

  • Office workers,
  • Professionals, and
  • Walking workers.

A notable and robust positive correlation emerged from the data (Figure 5), indicating a significant association between the levels of AI and RI software utilization across these occupational categories. This trend was particularly pronounced among office workers and professionals, where a substantial portion of respondents reported considerable exposure to both AI-enhancing and RI-enhancing technologies. Notably, ICT engineers and researchers stood out as the top users, reflecting the deep penetration of these technologies in fields heavily reliant on digital tools and platforms.

Conversely, the group labelled as walking workers displayed markedly lower levels of engagement with both AI and RI technologies. This discrepancy underscores a discernible digital divide within the workforce, with certain occupations, especially those requiring physical presence and mobility, lagging in the adoption of these advanced digital tools.

This segmentation and the observed correlations not only illuminate the current state of digital technology integration across various occupations but also highlight the potential areas for increased adoption and the challenges that may need to be addressed to facilitate a more inclusive digital transformation in the workplace.

Figure 5: Usage in 2020 of software facilitating telework vs task automation

Source: Authors’ elaboration of data from Okubo-NIRA data set.


Note: the indices are coded to be 1 or 0 at the respondent level and then averaged over workers with that occupation in all industries.

To further investigate this negative correlation, we divided the overall office automation index into two subcomponents:

  • Direct service task automation software (such as robotic process automation) and
  • Automation of office management systems (which are often related to handling information concerning workers in the office.

The findings are displayed in Figure 6.

Figure 6: Growth in usage of pro-telework software versus management automation and task automation software (2020 to 2022).

Source: Authors’ elaboration of data from Okubo-NIRA data set.


Note: Growth in pro-teleworking software usage is on x-axis of both charts; the y-axis for the left chart is the change in usage of office management software while the y-axis of the right chart shows changes in usage of repetitive service task automation like RPA

In both panels, the figure plots the horizontal axis is the change in usage of software that promotes telework (pro-telework software, e.g., zoom, slack, cloud-base file sharing, etc.). The left panel shows the automation software related to managing workers (management automation software) (e.g., HR management tools).

Thus, there is weak evidence for substitution of AI and RI for software which automates some management tasks related to workers. Further research would be needed to establish causality, but the charts are certainly in line with the notation that less management-facilitating software is needed when there are fewer workers in the office.

The left panel of the data shows a mild negative correlation, or no correlation at all, between the usage of task automation software and pro-telework software that supports remote work. This is not unexpected since the need to expand the usage of management software might not be augmented as the share of workers actually in the office falls. The facts on office management automation and telework-facilitating software suggest that, at the occupation level, firms do more of both at the same time.

Evidence from workers’ expectations

The last part of this preliminary look at this very rich data, use questions that probed workers' anticipations regarding the future role of automation technologies—encompassing AI and robots—in their professions.

  • Workers were asked whether their tasks would be supported by automation technologies in the future or whether their roles could potentially be supplanted by such advancements.

The interpretation of the findings illustrated in Figure 7 warrants a nuanced approach, given that the respondents are currently employed individuals. One interpretation, which aligns with the hypothesis that AI and RI could act as substitutes, suggests a scenario where workers who remain employed are likely to engage more extensively in teleworking and leverage AI more significantly. This scenario hints at a possible reduction in workforce numbers as AI enhances the productivity of the remaining employees.

However, despite these complexities, the data presents a compelling narrative at the occupational level. It vividly demonstrates that the occupations deemed most conducive to teleworking are simultaneously those most vulnerable to automation. This correlation provides substantive evidence to inform the ongoing discourse on the interplay between AI, RI, and the future of work. It suggests that as occupations become more adaptable to telework, they also potentially increase their exposure to automation, underscoring the intricate balance between leveraging technological advancements to augment work and the risk of displacement by the same technologies.

Figure 7: Workers’ expectations of telework and task automation in their jobs.

Source: Authors’ elaboration of data from Okubo-NIRA data set.


Note: the indices are coded to be 1 or 0 at the respondent level and then averaged over workers with that occupation in all industries. Ag workers, 33; Carrying, cleaning etc, 37; Construction workers, 36; Food services, 29; Buildings managers, 30; Manuf workers, 34; Occ'l health workers, 28; Other, 38; Other service workers, 31; Outdoor serv workers, 23; Social work prof'l, 10; Nurses etc, 8; Sales clerks, 22; Sales workers, 26; Security workers, 32; Transport workers, 35; Transport workers, 24; Family care workers, 27

Summary and concluding remarks.

Our findings provide tentative, albeit not definitive, evidence from occupational and worker levels data that AI and RI technologies appear to be acting more as complements than substitutes.

  • We observe a positive correlation between the deployment of AI-driven software, designed to automate routine information-processing tasks, and investments in technologies that facilitate telework.

More careful econometric work is needed to move beyond our exploration of the unconditional facts, but we can already speculate about policy implications. One concerns that trend toward international offshoring of office jobs, i.e. international telework, or telemigration. GenAI is replacing some of these foreign workers (e.g. low-end call-centre workers replaced with chatbots), but at the same time offshore workers with excellent GenAI tools can provide better service. Thus, the arrival of ChatGPT may not threaten the trend towards service-led development that relies on back office offshoring.

References

Baldwin, Richard and Toshihiro Okubo (2023). Are Software Automation and Teleworkers Substitutes? Preliminary Evidence from Japan NBER working paper 31627. https://www.nber.org/papers/w31627 .

Baldwin, R. and T. Okubo. Software automation and teleworkers as complements and substitutes,VoxEU.org, 26 Mar 2024.

Barrero, Jose Maria, Nicholas Bloom and Steven Davis (2023). The Evolution of Work from Home. NBER Working Paper 31686, https://www.nber.org/papers/w31686.

Brynjolfsson, Erik, Danielle Li, and Lindsey R. Raymond (2023). Generative AI at Work, NBER Working Paper No. 31161 April 2023, revised November 2023.

Dingel Jonathan and Brent Neiman (2020). How Many Jobs Can Be Done at Home? NBER Working Paper 26948. https://www.nber.org/papers/w26948

Drenik, Andres, Alberto Cavallo, Javier Cravino, Agostina Brinatti (2022). Remote wages in a globalised labour market, VoxEU.org, 19 May 2022.

Feenstra, Robert C., and Hanson, Gordon H., "Global Production Sharing and Rising Inequality: A Survey of Trade and Wages," NBER Working Paper No. 8372, July 2001.

Frey, Carl and Michael Osborne (2013). The Future of Employment: How susceptible are jobs to computerisation? Citi GPS: Global Perspectives & Solutions.

Frey, Carl and Michael Osborne (2015). Computerization and The Future of Employment in Japan, Nomura Research Institute

Okubo, T and NIRA (various), “Report on the results of a questionnaire survey concerning the impact of the use of telework to respond to the spread of the COVID-19 on working styles, lifestyles, and awareness”, Nippon Institute for Research Advancement.

Okubo, T (2022), “Telework in the spread of COVID-19”, Information Economics and Policy 60, 100987.

Palmou, Christina, James Browne, Jeegar Kakkad, and David Britto (2021). Anywhere jobs and the future of work, VoxEU.org, 10 Jul 2021.

Annex

Table 1: US occupations and jobs (millions) in US globotics quadrants.


NICOLAS GOLDSTEIN

Co-founder Talenteum.com ??Change lives by hiring differently. ?? Co-President La French Tech Mauritius

11 个月

Yes AI should increase productivity for sure Richard Baldwin ??

要查看或添加评论,请登录

Richard Baldwin的更多文章

社区洞察

其他会员也浏览了