This post is a summary of my literature review on various effects of AI on work. The economic literature was selected primarily through the recommendations of several researchers in the field - gathered through personal correspondence, the reading list from Anton Korinek’s graduate course, and bibliographies of key papers. This post is intended as a summary of some key findings: for more detailed justifications and discussions, please see the review itself, or else feel free to ask questions below.
By writing a review, a sought to uncover two phenomena. Firstly, I sought to locate critical disagreement, by which I mean the points of contention that explain differences between expert views. A critical disagreement is a crux in a topic, and may be an empirical dispute, or else a core assumption, through which the opinions of economists diverge. Finding such disagreements is enormously helpful in both understanding a field, in forming one’s own views on a set of issues, and in figuring out where future research should be undertaken. Secondly, I sought to uncover key areas of agreement. Expert agreement, such as concerning policy recommendations, is notable, and is further worthwhile when experts reach the point of agreement using independent methods or reasons. I attempted to map some contours of agreement and disagreement over several areas: which jobs are most exposed to the advances of Generative AI? Which workers will be most complemented? What will be the impact on worker-wellbeing and job quality? How will this affect income inequality? Which considerations and values should guide policy choices? Which policies are most commonly debated? Finally, how plausible are “explosive growth” scenarios?
Findings
Several recent papers use different methods to conclude that higher-middle income and higher-middle skill jobs are most exposed to the advances of Generative AI. Routine intellectual tasks are most exposed. (Section III).
Few higher-middle income and higher-middle skill jobs are likely to be fully automated; rather, they may be augmented. Increased worker productivity has been a key result in several studies, with a smaller variance of productivity between workers. (Section V). Impacts on inequality depend heavily on the degree to which exposed workers are automated as opposed to being augmented by AI. (Section IV).
Advances in AI are likely to increase the “digital divide” between advanced economies and developing economies. A further potential inequality-risk associated with AI is damaging effects on democracy. (Section IV).
The effects of complemented human labour with AI on worker wellbeing is currently under-researched. Both a quantitative and qualitative survey agree that tasks become easier, time is saved, yet—importantly, job satisfaction may decrease. A key emphasis is that the impact on well-being is not deterministic, and is the result of managerial and executive decision-making. (Section V).
There is agreement over some guiding principles concerning policy: policies should recognise the central importance of work as a source of human meaning, and so technological development should be steered so as to complement, rather than replace, human labour. Policies should seek to reduce expected channels of inequality. Policies should also be flexible. (Section VI).
Three highly debated policies include Universal Basic Income, raised minimum wages, and the taxation of capital. Major empirical disagreements exist around each of these. (Section VI).
Under many long-run models with increasing reduction in human labour, income inequality and GDP inequality will increase. (Appendix).
Economists proposing explosive growth scenarios forward two arguments: one based on labour becoming accumulable through digital workers, and one based on the automation of the discovery of ideas. More research needs to be done on the critical disagreements here to evaluate these arguments: specifically, on why research productivity is currently decreasing, and on whether there are longterm bottlenecks to human labour being replaced. (Appendix).
AI and Work: Summarising a New Literature Review
Link post
Introduction
This post is a summary of my literature review on various effects of AI on work. The economic literature was selected primarily through the recommendations of several researchers in the field - gathered through personal correspondence, the reading list from Anton Korinek’s graduate course, and bibliographies of key papers. This post is intended as a summary of some key findings: for more detailed justifications and discussions, please see the review itself, or else feel free to ask questions below.
By writing a review, a sought to uncover two phenomena. Firstly, I sought to locate critical disagreement, by which I mean the points of contention that explain differences between expert views. A critical disagreement is a crux in a topic, and may be an empirical dispute, or else a core assumption, through which the opinions of economists diverge. Finding such disagreements is enormously helpful in both understanding a field, in forming one’s own views on a set of issues, and in figuring out where future research should be undertaken. Secondly, I sought to uncover key areas of agreement. Expert agreement, such as concerning policy recommendations, is notable, and is further worthwhile when experts reach the point of agreement using independent methods or reasons. I attempted to map some contours of agreement and disagreement over several areas: which jobs are most exposed to the advances of Generative AI? Which workers will be most complemented? What will be the impact on worker-wellbeing and job quality? How will this affect income inequality? Which considerations and values should guide policy choices? Which policies are most commonly debated? Finally, how plausible are “explosive growth” scenarios?
Findings
Several recent papers use different methods to conclude that higher-middle income and higher-middle skill jobs are most exposed to the advances of Generative AI. Routine intellectual tasks are most exposed. (Section III).
Few higher-middle income and higher-middle skill jobs are likely to be fully automated; rather, they may be augmented. Increased worker productivity has been a key result in several studies, with a smaller variance of productivity between workers. (Section V). Impacts on inequality depend heavily on the degree to which exposed workers are automated as opposed to being augmented by AI. (Section IV).
Advances in AI are likely to increase the “digital divide” between advanced economies and developing economies. A further potential inequality-risk associated with AI is damaging effects on democracy. (Section IV).
The effects of complemented human labour with AI on worker wellbeing is currently under-researched. Both a quantitative and qualitative survey agree that tasks become easier, time is saved, yet—importantly, job satisfaction may decrease. A key emphasis is that the impact on well-being is not deterministic, and is the result of managerial and executive decision-making. (Section V).
There is agreement over some guiding principles concerning policy: policies should recognise the central importance of work as a source of human meaning, and so technological development should be steered so as to complement, rather than replace, human labour. Policies should seek to reduce expected channels of inequality. Policies should also be flexible. (Section VI).
Three highly debated policies include Universal Basic Income, raised minimum wages, and the taxation of capital. Major empirical disagreements exist around each of these. (Section VI).
Under many long-run models with increasing reduction in human labour, income inequality and GDP inequality will increase. (Appendix).
Economists proposing explosive growth scenarios forward two arguments: one based on labour becoming accumulable through digital workers, and one based on the automation of the discovery of ideas. More research needs to be done on the critical disagreements here to evaluate these arguments: specifically, on why research productivity is currently decreasing, and on whether there are longterm bottlenecks to human labour being replaced. (Appendix).