Ungated link here: tinyurl.com/337jsm4e

Happy Labor Day!

inally, this paper is co-authored with two fantastic graduate students, one of which (@karenxni.bsky.social) will be on the job market this year! We plan to make all data in this paper publicly available, and hope this is the beginning of a rich research agenda in this area. 🧵/🧵

Overall, this early look at natn'l job training efforts to accelerate worker adjustment to AI offers some optimism, but returns are higher when workers avoid AI skills altogether. Future work will need to disentangle if effects may differ for on-the-job training w/in firms. 12/🧵

The distribution of training participants is also occupationally representative of the nation. This + the large scale of the job training data allow us to cautiously infer to the national population of CPS unemployed workers. 11/🧵

Although WIOA training participants are mostly low income, they are highly AI exposed. The modal participant displaces from the top quintile of occups in AI-exposure ('5' on the x-axis). Many workers in this quintile were cashiers or customer service reps before training. 10/🧵

More descriptives: high earnings returns are concentrated in the most recent years when labor mrkts were exceptionally tight ➡️ training may carry stronger signal value when firms have to reach deeper into the skill market. Or it may reflect changes in AI—an open question! 9/🧵

We find that 25-40% of occupations are “AI retrainable” (high sahre!). Some occupations (e.g. paralegals) rank higher b/c workers earn more when moving to higher AI-content work, and others (e.g. customer service reps), rank lower because workers are forced to move down in AI-content. 8/🧵

Using our matched sample, we construct an AI Retrainability (AIR) index ranking occups by the share of retrained workers earning ⬆️ wages despite moving into AI-intensive roles. We then ask whether AIR is driven by earnings gains holding AI skills constant, or AI upskilling. 7/🧵

Our main finding is that while workers leaving AI‑exposed occups see strong earnings returns from training (~$1500/qtr, large estimates!), AI‑specific retraining delivers 29% lower returns vs. general training ➡️ frictions in acquiring skills used by AI-intensive occups. 6/🧵

We assemble a new dataset of 1.6m+ job training spells from the Workforce Innovation & Opportunity Act (WIOA) linked to AI exposure measures by @erikbryn.bsky.social , @danielrock.bsky.social , & co-authors. We then analyze earnings returns to training by pre-separation AI exposure. 4/🧵

This paper shines early light on the effectiveness of job training for workers in occs. exposed to AI before job loss, and for those who target AI-intensive occs. in their next jobs (presumably by acquiring AI-compatible skills that safeguard against future job loss). See ⬇️ 3/🧵

While much attention has been paid to the potential impact of AI on headcounts, there is next to no evidence on the adjustment margin firms most commonly pursue to accommodate AI technology: reskilling workers for AI. (See e.g., firm survey results w/ Fed colleagues below). 2/🧵