By: Jacob Greenspon
Generative artificial intelligence services like ChatGPT and DALL-E have wowed and worried many. Their ability to write complex programming code, craft poetry, and design images replicates that of workers currently hired for these tasks, raising fears of widespread job loss. How might generative AI affect the labour market based on its adoption so far and past automations?
Generative AI is based on machine learning technologies in development since the 1950s that can perform complex tasks by inferring instructions through being trained on examples. Thanks to advances in computing scale, model architecture, training data, and human refinement of models, generative AI burst into public awareness with the release of online software capable of generating text or audiovisual content from simple prompts. Firms are now integrating more complex generative AI into their systems to optimize business processes.
Early evidence from an online freelancing platform suggests the release of generative AI led businesses to hire fewer language translators, writers and graphic designers and pay them less. Other recent research suggests wider potential impacts. Looking at the tasks that make up each job—such as ‘Monitor computer system performance to ensure proper operation’ or ‘Execute sales or other financial transactions’—and then assessing whether they could be performed as equally well by generative AI, these researchers found that four-fifths of the US workforce have some exposure to generative AI, while almost one-fifth could have at least 50% of their work tasks affected. They also find wide-ranging (though varied) exposure across all industries.
Does this mean the end of work? It is unlikely that generative AI will eliminate employment. Experts have repeatedly predicted that work will disappear—due to textile looms, phonographs, industrial robots, or other automations—and always claim ‘this time is different’, but long-term unemployment has never increased. This is because automations cut costs, which can end up increasing product sales by even more than they reduce the amount of work needed to make it, leading to a net increase in workers needed. Moreover, new technologies often create previously-unforeseeable jobs, such as former farmers rendered unemployed by threshing machines moving to the city to work in a farm equipment factory (and earn more).
Increased inequality is more likely. Generative AI fits the mold of what economists call ‘skill-biased technological change’: an innovation that replaces some workers while increasing demand for others who are able to leverage these new technologies. Robots, for example, replaced assembly-line workers but required more engineers to design, operate, and maintain. The result is polarization: pay rises for engineers alongside lower wages for manual labourers forced either to compete for non-robot factory jobs or switch to hard-to-automate jobs like cleaning and food service. This ‘hollowing out’ was seen with widely-adopted 20th century technologies such as electrification or computerization.
Unlike most previous automations, generative AI mostly affects workers who think and create, not those involved in manual labour or human interaction. But similar as before, some workers benefit—such as the ‘chief AI officers’ increasingly being hired by large businesses. Generative AI replaces a different set of workers than previous automations, but with the same result of increasing inequality between different jobs.
Yet early evidence on how generative AI is used in businesses adds another dimension of reduced inequality within a specific occupation. One experiment tested a generative AI that learned from customer support agents’ customer interactions to suggest responses. Experienced agents mostly ignored the AI, but it drastically improved productivity for newer workers who could now learn the ‘tacit knowledge’ about customer interactions that is hard for experienced workers to rapidly impart. Another study, of management consultants, found similar effects, with previous lower-performers experiencing over twice the productivity increase as high-performers. Generative AI may increase polarization between jobs while decreasing inequality among workers in a certain job, with an uncertain overall effect.
A final question is that while research finds some jobs more exposed to generative AI, it does not state which workers will be replaced. AI should be adopted where it is most profitable, meaning highly-paid workers are more likely to be replaced. But this seems improbabley for many high-paid professions like lawyers and financial managers whose tasks can be performed by AI, but likely not with the same level of trusted accuracy required given the costs of mistakes. In addition, unionized workers or those in professional societies, such as physicians, may be better able to resist AI in their workplaces. Or their jobs may change with the new technology. This was the experience of accountants, who used to spend most of their time doing arithmetic but, with the arrival of calculators and spreadsheets, now mostly ensure proper record keeping and analyze financial information. Finally, it will take a while for generative AI to be fully integrated into most businesses operations’, giving at-risk workers time to retrain or retire.
Only time will tell how generative AI affects workers, but story is more complex than widespread job losses. Policymakers must take these complications and uncertainties into account before deciding how to respond.