Thank you very much for this opportunity to share my thoughts with all of you.
Generative AI renews concerns for job stability, education and the future of work, because generative AI is capable of things that were unimaginable from AI systems just 10 years ago. The conventional wisdom from labour economics recognizes that technology does not automate occupations wholesale, but instead automates specific activities within a job.
The challenge is that workplace activities and AI applications vary across the entire economy. Therefore, efforts to predict automation and job stability need to rely on simplifying heuristics. Cognitive, creative and white-collar workers are assumed to be safe from automation, for example, because creativity is difficult to assess objectively and because the creative process is difficult to describe algorithmically.
However, generative AI, including tools like large language models like ChatGPT and image generators like Midjourney are doing creative work when they write essays, poetry or computer programming code, or when they generate novel images from just a prompt. This means that today's AI shatters the conventional wisdom that has been used to inform economic policy and economic research.
For example, unlike past automation studies, a recent report from OpenAI and the University of Pennsylvania found that U.S. occupations with the most exposure to large language models tended to be the occupations requiring the most education and earning the highest wages. Departing from a heuristic-based approach to predicting automation will require some new data that reflects the more direct implications of generative AI.
However, just like past technologies, generative AI performs specific workplace activities, which means that AI's most direct impact on occupations is through a shift in workers' skills and activities towards other skills that would complement AI. However, if workers fail to adapt then a job separation can occur. These separations include workers quitting or being fired by their employer. Job separations will lead workers to seek new employment, but if they struggle to find a job, then they may receive unemployment benefits to support them while they continue job seeking.
This lays out a pipeline of AI impact that identifies the most and least direct implications from AI and highlights that data that better reflects shifting skill demands, job separations by region or industry or occupation, and even the data on the unemployment risk experienced by occupations across the economy, will improve efforts to predict AI's impact on workers.
There are some emerging data sources, including job postings, workers' resumés and data from unemployment insurance offices, that offer some new options for describing these details in the labour market but are often missed by traditional government labour statistics.
Finally, because shifts in skills are the most direct consequence from exposure to generative AI, prudent policy should focus on the mechanisms for skill acquisition. If generative AI will mostly impact white-collar jobs, then we should focus on the skills taught during a college education since a college education is the typical mechanism for getting students into those white-collar jobs.
While labour statistics abound, insight into college skills are more difficult to find. If college skills are quantified, then just as we study generative AI in the workforce, we can also assess the colleges, students and major areas of study with the greatest exposure to AI. However, educational exposure to generative AI should not be shied away from. Recent case studies find that generative AI tools do not out-compete or significantly improve the performance of experts, but they do make a big difference in raising the performance of non-experts to be more comparable to that of the experts in those applications.
If this observation holds across contexts, then incorporating generative AI into learning curricula has the potential to improve learning objectives, especially for underperforming students, and therefore could strengthen educational programs.
In summary, generative AI is new and exciting and will impact the workforce in new ways from previous technology. In fact, generative AI shatters the conventional wisdom used to predict automation from AI in the past, because it does the work of occupations that were previously thought to be immune to automation.
A better path forward would focus on the data and insights reflecting what AI can actually do from the perspective of workplace skills and activities as well as the sources of those skills among workers in the workforce.
With that, thank you.