>The paper notes they did an adjustment for the end of ZIRP. I dont know enough econometrics to understand whether this adjustment was sufficient

Looking at the paper [0], they attempted to do it by regressing the number of jobs y_{c,q,t} at company c, time t, and "AI exposure quintile" q, with separate parameters jointly controlling for company/quintile (a), company/time (b) and quintile/time (g). This is in Equation 4.1, page 15, which I have simplified here:

log(y_{c,q,t}) ~ a_{c,q} + b_{c,t} + g_{q,t}

Any time-dependent effects (e.g. end of ZIRP/Section 174) that would equally affect all jobs at the company irrespective of how much AI exposure they have should be absorbed into b.

They normalized g with respect to October 2022 and quintile 1 (least AI exposure), and plotted the results for each age group and quintile (Figure 9, page 20). There is a pronounced decline that only starts in mid-2024 for quintiles 3, 4, and 5 in the youngest age group. The plots shown in the article are misleading, and are likely primarily a reflection of ZIRP, as you say. The real meat of the paper is Figure 9.

A potential flaw of this method is that ZIRP/Section 174 may have disproportionately affected junior positions with high AI exposure, e.g. software engineers. This would not be accounted for in b and would thus be reflected in g. It would be interesting to repeat this analysis excluding software engineers and other employees subject to Section 174.

[0] https://digitaleconomy.stanford.edu/wp-content/uploads/2025/...