This is true. But I’ve been noting that there’s still a timing issue when trying to make it an AI story. At least using my less sophisticated subtraction method.
Looking at Fig 4, the obvious question is: when does the gap actually start? After all, it predates ChatGPT, the slowdown and everything else. What is the issue then? Can we disregard the fact that your small model does not reflect college graduate trends for whatever reason?
Hiring new grads requires a lot more training and no guarantee that training pays off. It makes sense that companies are being more risk averse during today's labor market and economic environment. I worked at a research agency that built their HR strategy around hiring a big class of new grads each year. We would hire them at a very competitive new grad salary and put them through an intense year of training. But experience and training goes a long way in the research industry. After one year of experience, many of our RAs were being hired away with salaries more than $20k+ what they were getting paid. We ended up having to do significant raises every 6 months in an attempts to keep them longer. But it wasn't a sustainable model. Also, culturally it's really tough to be someone's first job out of college right now. There was a big disconnect between what they expected out of a job in terms of hours, work/life balance, culture, etc. and what those of us who had been working for a bit expected. AI might be making it easier to put off hiring new grads, but it has been less attractive to hire new grads across a variety fo industries for a while.
2) The timing matters. Your figure 5 is for Sept 2025. The cuts to research have decimated positions for recent graduates. These do not just affect universities but the whole private-public research ecosystem. And that's where some young graduates would try to get jobs. You can repeat the same logic for cuts to federal grants supporting schools and NGOs because of the fight to DEI and the DOGE mess. Again, these are usually jobs for college graduates. Before invoking AI, at least for 2025, there are many other explanations that come to mind for a cooling labor market for college graduates.
3) Companies know whose side the administration is on in the labor-capital conflict. The labor market was very tight during the recovery from the pandemic and this threw the employers completely off. AI provides a nice excuse for companies to discipline labor, with more layoff or less hiring. This is true even if AI does not replace many jobs: but it is a good excuse to hire fewer graduates and have the senior staff work harder (anecdotally, I have heard quite a few stories like these).
4) I am sure that AI may matter for some jobs, e.g. anything involving coding. But the hiring of CS graduates went through the roof few years ago, and so there is also some sectoral market adjustment going on.
5) A pet peeve of mine (though tangential to your post). As your Figure 3.a shows, the unemployment rate for African Americans also fell "much further than historical patterns would predict" in 2019, during Trump 1.0. In my experience, this fact is not usually mentioned in polite conversations, and so I mention it here.
6) I do not understand the concern about the positive residuals: if the model is trained on data up to 2019, and the projection is for 2025, then the difference between actual and projected need not be (on average) zero. But maybe I am missing something.
7) Small side technical point. The model is estimated in log-log but then you use unemployment predictions. There are different ways to back out E(U|X) from E(log(U)|X), and I guess you use the smearing factor (I admit I do not know R and so I can't figure the code out). This will affect the levels and I am pretty sure the difference too. I do not know if this is a big issue or not, or even how one would address it. But it seemed worth mentioning.
This is true. But I’ve been noting that there’s still a timing issue when trying to make it an AI story. At least using my less sophisticated subtraction method.
It’s especially bad if you’re a black female. 7.3% unemployment rate. 3 years ago it was 4.7%
Looking at Fig 4, the obvious question is: when does the gap actually start? After all, it predates ChatGPT, the slowdown and everything else. What is the issue then? Can we disregard the fact that your small model does not reflect college graduate trends for whatever reason?
Hiring new grads requires a lot more training and no guarantee that training pays off. It makes sense that companies are being more risk averse during today's labor market and economic environment. I worked at a research agency that built their HR strategy around hiring a big class of new grads each year. We would hire them at a very competitive new grad salary and put them through an intense year of training. But experience and training goes a long way in the research industry. After one year of experience, many of our RAs were being hired away with salaries more than $20k+ what they were getting paid. We ended up having to do significant raises every 6 months in an attempts to keep them longer. But it wasn't a sustainable model. Also, culturally it's really tough to be someone's first job out of college right now. There was a big disconnect between what they expected out of a job in terms of hours, work/life balance, culture, etc. and what those of us who had been working for a bit expected. AI might be making it easier to put off hiring new grads, but it has been less attractive to hire new grads across a variety fo industries for a while.
Nice post. Some thoughts:
1) Kudos for posting the code!
2) The timing matters. Your figure 5 is for Sept 2025. The cuts to research have decimated positions for recent graduates. These do not just affect universities but the whole private-public research ecosystem. And that's where some young graduates would try to get jobs. You can repeat the same logic for cuts to federal grants supporting schools and NGOs because of the fight to DEI and the DOGE mess. Again, these are usually jobs for college graduates. Before invoking AI, at least for 2025, there are many other explanations that come to mind for a cooling labor market for college graduates.
3) Companies know whose side the administration is on in the labor-capital conflict. The labor market was very tight during the recovery from the pandemic and this threw the employers completely off. AI provides a nice excuse for companies to discipline labor, with more layoff or less hiring. This is true even if AI does not replace many jobs: but it is a good excuse to hire fewer graduates and have the senior staff work harder (anecdotally, I have heard quite a few stories like these).
4) I am sure that AI may matter for some jobs, e.g. anything involving coding. But the hiring of CS graduates went through the roof few years ago, and so there is also some sectoral market adjustment going on.
5) A pet peeve of mine (though tangential to your post). As your Figure 3.a shows, the unemployment rate for African Americans also fell "much further than historical patterns would predict" in 2019, during Trump 1.0. In my experience, this fact is not usually mentioned in polite conversations, and so I mention it here.
6) I do not understand the concern about the positive residuals: if the model is trained on data up to 2019, and the projection is for 2025, then the difference between actual and projected need not be (on average) zero. But maybe I am missing something.
7) Small side technical point. The model is estimated in log-log but then you use unemployment predictions. There are different ways to back out E(U|X) from E(log(U)|X), and I guess you use the smearing factor (I admit I do not know R and so I can't figure the code out). This will affect the levels and I am pretty sure the difference too. I do not know if this is a big issue or not, or even how one would address it. But it seemed worth mentioning.