AI took your job — can retraining help?

Many people worry that AI is going to take their job. But a recent survey conducted by the Federal Reserve Bank of New York found that rather than laying off workers, many AI-adopting firms are retraining their workforces to use the new technology. Yet there’s little research into whether existing job-training programs are helping workers successfully adapt to an evolving labor market.

A new working paper starts to fill that gap. A team of researchers, including doctoral candidate Karen Ni of the Harvard Kennedy School, analyzed worker outcomes after they participated in job-training programs through the U.S. government’s Workforce Innovation and Opportunity Act. Researchers looked at administrative earnings records spanning the quarters before and after workers completed training. Then they analyzed workers’ earnings when transitioning from or into an occupation that was highly “AI-exposed” — a term that refers to the extent of tasks that have the potential to be automated, both in the traditional computerization sense and through generative AI technology.

Across the board, the training programs demonstrated a positive impact, with displaced workers seeing increased earnings after entering a new occupation. Still, those earnings were less for someone who targeted a high AI-exposed occupation than someone who targeted a low AI-exposed occupation.

In this edited conversation, Ni explains the role that job-training programs play as AI use is transforming the labor market.

With all the discussion around job displacement and AI, what led you to focus on retraining in particular?

When thinking about the disruptions that a new large-scale technology might have for the labor market, it’s important to understand whether it’s possible for us to help workers who might be displaced by these technologies to transition into other work. So we homed in on, OK, we know that some of these workers are being displaced. Now, what can job training services do for them? Can they improve their job prospects? Can they help them move up in terms of earnings? Is it possible to retrain some of these workers for highly AI-exposed roles?

We wanted to help document the transition and adaptability for these displaced workers, especially those who are lower income. Because then we can think about how we can support these workers, whether it be better investing in these kinds of workforce-development programs or training programs, or adapting those programs to the evolving labor market landscape.

What can we learn by looking at data from government workforce development programs?

One of the big advantages of using these trainees is that it’s nationwide, and so it’s nationally representative. That allows us to take a broad look at trainees across the entire country and capture a fair bit of heterogeneity in terms of their occupations and backgrounds. For the large part, our sample captures displaced workers who tend to be lower income, making an average of $40,000 a year. Some are making big transitions from one occupation to a completely different one. We also see a fair number of people who end up going into the same types of jobs that they had before. We think those workers are likely trying to develop new skills or credentials that might be helpful to enter back into a similar occupation. Some of these people might be displaced from their occupation because of AI. But the job displacement in this sample could be for any reason, like a regional office shutting down.

Can you provide some examples of highAI-exposed careers versus low AI-exposed careers?

AI exposure refers to the extent of tasks within an occupation that could potentially be completed by a machine or a large language model. Among our sample of job trainees, some of the most common high AI-exposed occupations were customer service representatives, cashiers, office clerks. On the other end of the spectrum, the lowest AI-exposed workers tended to be manual laborers, such as movers, industrial truck drivers, or packagers.

AI retrainability by occupation

What were your main findings?

We first looked at the split before entering job training: if they were displaced from a low AI-exposed or high AI-exposed occupation. After training, we find pretty positive earnings returns across the board. However, workers who are coming from high AI-exposed jobs have, on average, 25 percent lower earnings returns after training compared to workers initially coming from low AI-exposed occupations.

Then we looked at the split after job training, if they were targeting high AI-exposed jobs or low AI-exposed jobs. If you break it down that way, we find that workers generally are better off targeting jobs that are lower AI-exposed compared to the workers who are targeting jobs that are more highly AI-exposed. Those who are targeting the high AI-exposed fields tend to face a penalty of 29 percent in terms of earnings, relative to workers who target more general skills training.

Are there any recommendations that displaced workers could take away from those findings?

I would cautiously say our findings seem to suggest that, for these AI-exposed workers going through job-training programs, going for jobs that are less AI-exposed tends to give them a better outcome. That said, the fact that we do see positive returns for all of these groups suggests that there’s probably other factors that need to be considered. For instance, what are the specific types of training that they’re receiving? What kinds of skills are they targeting? There’s an immense heterogeneity across the different job-training centers throughout the country, in terms of the quality, intensity, and even the types of occupations that they can offer services for. There’s a lot of potential for future work to consider how those factors might affect outcomes.

Also, in this case, the training program is predominantly serving displaced workers from lower parts of the income distribution. So I don’t think we can generalize across the board and say, “everyone should go do a job-training program.” We were focused on this specific population. 

You also created an AI Retrainability Index to rank occupations that both prepare workers well for jobs that are more AI-exposed and also earn more than their past occupation. What did the index reveal about which occupations are most “retrainable”?

We wanted to have a way of measuring by occupation how retrainable workers are if they were to be displaced. Our index ranking shows that, depending on where they’re starting from, you might have more or less capability of being retrained for highly AI-exposed roles. The only three occupational categories that had a positive index value — meaning that we consider these to be occupations that are highly AI-retrainable — were legal, computation and mathematics, and arts, design, and media. So someone coming from a legal profession is more retrainable for high-paying, high AI-exposed roles than someone coming from, say, a customer service job.

Overall, we found that 25 to 40 percent of occupations are AI retrainable, which, to us, is surprisingly high. You might think that if someone is coming from a lower-wage job, it might be really hard to retrain them for a job that has more AI exposure. But what we found is that there may actually be a large potential for retraining.