One open question here is whether we’re seeing youth employment decrease because AI is effectively replacing entry level workers in these fields, or because executives wrongly *think* AI can or will soon be able to do so?
I’ve played around with AI in my job, which I’m pretty sure Anthropic would classify as highly exposed to AI (think something similar to accounting). It’s really helpful for a small number of tasks that comprise maybe 10% of my job, but pretty much useless at the rest. If it’s impacting entry level jobs in my field at all, I really think it’s more that those jobs will change a bit than be totally replaced. I suspect that at some point once the AI hype fever breaks, companies will simply reconfigure entry level jobs a bit to incorporate AI and begin hiring again. I could be wrong, but that’s what I’d bet on right now.
I’m not sure why 10% of junior sales and marketing was laid off at a stroke and rehired a year later, or why junior software engineers have a hard inflection at October 2022, but those seem much likelier to be executive decisions — they just don’t fit any reasonable S-shaped adoption curve.
This tracks how I think about the situation. It’s way too early to associate jobs to automation or augmentation.
Targeting AI exposed jobs in layoffs with the hope that efficiencies will follow has been the talk track of C-suite execs for months. If those efficiencies don’t materialize, we might be talking about an AI fueled boom for coders in a year or two, but the reality would be companies cut deeper than they could sustain and equilibrium needed to be reestablished while AI did little to nothing.
And something I worry about, how are we going to get capable workers without their having first done these entry level jobs. There is the perpetual catch-22 of companies who want to hire people with 3 years experience to do an entry level job (Or any job really). But what I think is more crucial is that you can't doo the next job up if you haven't learned the lessons before. You need to be an intern to be a resident, a resident to be an attending etc. How will they know how to do anything? A lot of the frustration with the current generation is the expectation they should just know it. Nobody wants to teach. Teaching isn't a brief walk through of procedures either.
It is logical to me that AI would impact some entry level jobs first. The effectiveness of AI bs humans is still to be determined. If AI does replace many “entry level” jobs, where do the humans receive the training necessary to take over the more experienced roles that AI does not seem (at least yet) to be capable of replacing?
yeah, i think any claims about AI affecting junior software developer jobs needs to address the fact that hiring in the tech industry peaked in mid-2022 and has not even come close to rebounding in the subsequent 3 years, so of course there will be a headcount bias towards older employees since then
Anyone who has tried to use the "new customer service experience" is in for a huge surprise. If AI is being implemented here, it is a huge failure. I tried filing a long-term healthcare claim online with a major US insurer. I could not do it as the website did not include the name of the person the claim was being filed for. I called customer service and went through two or three different menus and was asked at the end of each unsuccessful try if I would like to complete a customer service survey. I finally get a real person with a heavy foreign accent who after five minutes of collecting information said she could not help me and gave me another phone number. I called that number and got through to a person who took the same information again and said I would be contacted within five workdays by a claims manager!!!
This one experience is not unique as I've had other unpleasant experiences in the past couple of years with "customer service" at other major US companies. Sometimes real people cannot be replaced by machines.
Insightful, thoughtful, and well-researched as always. Thank you, Derek! This article along with your recent one on the social implications of AI as a therapist/counselor were so interesting and thought-provoking.
I might have to dig into how they control for the slowdown in tech hiring. I don’t think it can be overstated how much the culture in tech has shifted from 2022 where companies would brag about how much they were growing to boost investor confidence to now where companies are bragging about how lean they are. The headcount of big tech has not grown in 3 years.
I am a senior software engineer. I think AI absolutely has a huge impact, where it can both augment the workflow of certain tasks and automate some entirely (or 95%). But when I see that we are hiring fewer juniors at my company, it seems almost entirely explainable by industry trends and almost entirely unrelated to AI productivity boosts.
The dirty little secret is that coding is the least important part of the job for most software engineers. The other dirty little secret is that even if every software engineer became 50% more productive, there would still likely be more work to do, at least at my company.
New graduate software engineers and new customer service representatives are both quite vulnerable to normal cyclical activity, though. Those trends could both be explained by a general belief among hiring managers that they have over-hired and/or that there is a recessio coming, both of which are quite likely.
I also happen to work in software and I don't see any sign of AI replacement. I do see reduced hiring because the US government is poking at things all over the economy without any clear idea of what they're doing. If there was AI replacement happening, we'd be reducing outreach to new graduates and aggressively pushing mid-level employees to use AI tools. But we're not doing that. Most software companies actually seem very cautious about AI tools because of the risk of IP contamination.
Also, I happen think software engineers are not very replaceable. AI can fill-in-the-blanks in code, but someone still has to set out the requirements and make sure they're filled. In spite of various industry fictions, that is still largely done by the people who write the code.
Where I would be looking for AI replacement is first year lawyers and accountants. Historically what they do is read documents trying to find information their seniors have asked for, because the seniors are trying to build some argument that the partners are trying to make. This is exactly the kind of work LLMs are very good at. The only potential problem is professional liability, but its the partners who take that risk, and the partners are also the ones who stand to profit if they can economize.
The interesting question then is, this messes with the structure of professional services firms, which are generally up-or-out pyramid structures. Some percentage of first years are cut and sent out with good references to become future customers and the rest then graduate to incrementally more responsible work. If you hire few (or no) first years because their job is now to type questions into a chatbot, who does that incrmenetally more responsible work next year? And who goes out into the world and refers your future customers?
As a CS professor at a liberal arts college, I often feel like I'm teaching my students to fly the plane while I'm also learning how to fly it and the controls are constantly being rearranged.
The shape of the line for 22-25 year old software devs (Figure 1) is interesting. It shows the big drop starting in late 2022 and continuing for most of 2023, then a period of relative stability until the second part of 2024, followed by a second steep drop into 2025.
A possible story is that the first drop was primarily due to overall conditions and post-COVID hiring pullback. The early LLMs were not particularly effective at coding. The second drop, though, roughly lines up with the introduction of Claude Sonnet 3.6 in October 2024, which was a big step up in coding ability and kicked off the current wave of agentic programming tools. So the first wave of job losses may have been partially driven by AI hype, but the second is driven by agentic tools increasing the productivity of mid-career devs.
There's some evidence that CS enrollments are softening as students look to less risky majors. I'm still seeing good numbers in my intro course, but lower throughput to later courses. I think we're entering a period of reduced demand for CS as a standalone major, but much higher demand for programming across the curriculum. Using AI to solve problems with code is an emerging skill that's going to transform a lot of disciplines.
Great write up, thanks. It would be useful to consider supply effects here. We are producing more CS grads each year, so at some point you’d expect labor market saturation even without AI. Also, the first group that AI will replace are college students — at least the ones who are willing to be replaced by letting AI do their coursework. Which raises the question about whether recent grads are comparable to past ones. Love to learn more.
Hiring trends bounce around a great deal at the occupational and sectoral levels, sometimes in tandem with the aggregate economy, but sometimes not. While I haven't read the underlying paper, what's presented in this article doesn't seem like a slam-dunk confirmation of AI effects. Basically, as with some of the other commenters, I'd be hesitant to declare the Hand of God at work in what could plausibly be secular changes in hiring rates. But maybe the article undersells the full breadth of evidence presented in the paper?
Indeed, before the TCJA’s enactment, businesses deducted the total amount of R&D expenditures as an expense in the taxable year. Beginning in 2022, all costs related to R&D must now be amortized over five years for US-based companies or 15 years for non-US companies.”
Software jobs are definitely often classified in R&D, and this new rule generated a massive headwind from
2022 until repeal in July 2025 (via the OBBB). Obviously the timing overlaps with end of Covid, so it’s tricky.
Two questions:
1. How does this interact with the age cohort of the software developers?
2. How will hiring into these jobs change now that the OBBB has passed?
I think one potential issue here is looking at data starting in 2021. There was such a huge increase in employment post-covid recovery in 2022. The number of software engineering job postings massively outpaced other sectors in 2022, so it may have become much more saturated. I'm sure the model in the paper controls for time trends, but it's difficult to ascertain when employment was increasing so drastically for all sectors 2021-2022.
I think seeing the trends pre-Covid as well would make this more convincing. Youth unemployment has been increasing relative to overall unemployment since 2011. However, I know there may be data limitations here.
I'm interested in learning about the potential long-term impact of entry-level workers not learning tasks that are easily automated, assuming the data continues to support this and becomes more robust. I suspect that certain tasks which can be automated are actually important. The question then becomes: which entry-level tasks? I have some ideas in mind, but they require more thought before I comment further.
One open question here is whether we’re seeing youth employment decrease because AI is effectively replacing entry level workers in these fields, or because executives wrongly *think* AI can or will soon be able to do so?
I’m not closed to the idea that AI is displacing some young workers, but I also see out of touch executives and investors buying into a lot of AI hype that I’m not seeing reflected on the ground. There was a recent study that showed the vast majority of AI initiatives fail: https://fortune.com/2025/08/21/an-mit-report-that-95-of-ai-pilots-fail-spooked-investors-but-the-reason-why-those-pilots-failed-is-what-should-make-the-c-suite-anxious/
I’ve played around with AI in my job, which I’m pretty sure Anthropic would classify as highly exposed to AI (think something similar to accounting). It’s really helpful for a small number of tasks that comprise maybe 10% of my job, but pretty much useless at the rest. If it’s impacting entry level jobs in my field at all, I really think it’s more that those jobs will change a bit than be totally replaced. I suspect that at some point once the AI hype fever breaks, companies will simply reconfigure entry level jobs a bit to incorporate AI and begin hiring again. I could be wrong, but that’s what I’d bet on right now.
I’m not sure why 10% of junior sales and marketing was laid off at a stroke and rehired a year later, or why junior software engineers have a hard inflection at October 2022, but those seem much likelier to be executive decisions — they just don’t fit any reasonable S-shaped adoption curve.
I think there is some emerging evidence you may be right. Companies were cautious about hiring for a time, but that may be picking back up.
Take a look at the JOLTS job openings for Information for the most recent month currently available (June 2025): https://fred.stlouisfed.org/series/JTU5100JOR
We're at the highest rate of job openings in this industry since 2022!
I am doing a more detailed writeup about how to interpret these numbers, should be published by mid-September.
This tracks how I think about the situation. It’s way too early to associate jobs to automation or augmentation.
Targeting AI exposed jobs in layoffs with the hope that efficiencies will follow has been the talk track of C-suite execs for months. If those efficiencies don’t materialize, we might be talking about an AI fueled boom for coders in a year or two, but the reality would be companies cut deeper than they could sustain and equilibrium needed to be reestablished while AI did little to nothing.
And something I worry about, how are we going to get capable workers without their having first done these entry level jobs. There is the perpetual catch-22 of companies who want to hire people with 3 years experience to do an entry level job (Or any job really). But what I think is more crucial is that you can't doo the next job up if you haven't learned the lessons before. You need to be an intern to be a resident, a resident to be an attending etc. How will they know how to do anything? A lot of the frustration with the current generation is the expectation they should just know it. Nobody wants to teach. Teaching isn't a brief walk through of procedures either.
That was my take as well. Not replaced, but not hired yet because of hype and an overly optimistic executive take on usefulness.
I suspect we will see reversals as with Klarna. See https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ for metrics - AI actually made programmers slower although they thought they were faster.
It is logical to me that AI would impact some entry level jobs first. The effectiveness of AI bs humans is still to be determined. If AI does replace many “entry level” jobs, where do the humans receive the training necessary to take over the more experienced roles that AI does not seem (at least yet) to be capable of replacing?
yeah, i think any claims about AI affecting junior software developer jobs needs to address the fact that hiring in the tech industry peaked in mid-2022 and has not even come close to rebounding in the subsequent 3 years, so of course there will be a headcount bias towards older employees since then
Anyone who has tried to use the "new customer service experience" is in for a huge surprise. If AI is being implemented here, it is a huge failure. I tried filing a long-term healthcare claim online with a major US insurer. I could not do it as the website did not include the name of the person the claim was being filed for. I called customer service and went through two or three different menus and was asked at the end of each unsuccessful try if I would like to complete a customer service survey. I finally get a real person with a heavy foreign accent who after five minutes of collecting information said she could not help me and gave me another phone number. I called that number and got through to a person who took the same information again and said I would be contacted within five workdays by a claims manager!!!
This one experience is not unique as I've had other unpleasant experiences in the past couple of years with "customer service" at other major US companies. Sometimes real people cannot be replaced by machines.
I know a division of folks, writers, laid off and explicitly replaced with AI. Definitely it is taking jobs.
Insightful, thoughtful, and well-researched as always. Thank you, Derek! This article along with your recent one on the social implications of AI as a therapist/counselor were so interesting and thought-provoking.
I might have to dig into how they control for the slowdown in tech hiring. I don’t think it can be overstated how much the culture in tech has shifted from 2022 where companies would brag about how much they were growing to boost investor confidence to now where companies are bragging about how lean they are. The headcount of big tech has not grown in 3 years.
I am a senior software engineer. I think AI absolutely has a huge impact, where it can both augment the workflow of certain tasks and automate some entirely (or 95%). But when I see that we are hiring fewer juniors at my company, it seems almost entirely explainable by industry trends and almost entirely unrelated to AI productivity boosts.
The dirty little secret is that coding is the least important part of the job for most software engineers. The other dirty little secret is that even if every software engineer became 50% more productive, there would still likely be more work to do, at least at my company.
New graduate software engineers and new customer service representatives are both quite vulnerable to normal cyclical activity, though. Those trends could both be explained by a general belief among hiring managers that they have over-hired and/or that there is a recessio coming, both of which are quite likely.
I also happen to work in software and I don't see any sign of AI replacement. I do see reduced hiring because the US government is poking at things all over the economy without any clear idea of what they're doing. If there was AI replacement happening, we'd be reducing outreach to new graduates and aggressively pushing mid-level employees to use AI tools. But we're not doing that. Most software companies actually seem very cautious about AI tools because of the risk of IP contamination.
Also, I happen think software engineers are not very replaceable. AI can fill-in-the-blanks in code, but someone still has to set out the requirements and make sure they're filled. In spite of various industry fictions, that is still largely done by the people who write the code.
Where I would be looking for AI replacement is first year lawyers and accountants. Historically what they do is read documents trying to find information their seniors have asked for, because the seniors are trying to build some argument that the partners are trying to make. This is exactly the kind of work LLMs are very good at. The only potential problem is professional liability, but its the partners who take that risk, and the partners are also the ones who stand to profit if they can economize.
The interesting question then is, this messes with the structure of professional services firms, which are generally up-or-out pyramid structures. Some percentage of first years are cut and sent out with good references to become future customers and the rest then graduate to incrementally more responsible work. If you hire few (or no) first years because their job is now to type questions into a chatbot, who does that incrmenetally more responsible work next year? And who goes out into the world and refers your future customers?
As a CS professor at a liberal arts college, I often feel like I'm teaching my students to fly the plane while I'm also learning how to fly it and the controls are constantly being rearranged.
The shape of the line for 22-25 year old software devs (Figure 1) is interesting. It shows the big drop starting in late 2022 and continuing for most of 2023, then a period of relative stability until the second part of 2024, followed by a second steep drop into 2025.
A possible story is that the first drop was primarily due to overall conditions and post-COVID hiring pullback. The early LLMs were not particularly effective at coding. The second drop, though, roughly lines up with the introduction of Claude Sonnet 3.6 in October 2024, which was a big step up in coding ability and kicked off the current wave of agentic programming tools. So the first wave of job losses may have been partially driven by AI hype, but the second is driven by agentic tools increasing the productivity of mid-career devs.
There's some evidence that CS enrollments are softening as students look to less risky majors. I'm still seeing good numbers in my intro course, but lower throughput to later courses. I think we're entering a period of reduced demand for CS as a standalone major, but much higher demand for programming across the curriculum. Using AI to solve problems with code is an emerging skill that's going to transform a lot of disciplines.
Great write up, thanks. It would be useful to consider supply effects here. We are producing more CS grads each year, so at some point you’d expect labor market saturation even without AI. Also, the first group that AI will replace are college students — at least the ones who are willing to be replaced by letting AI do their coursework. Which raises the question about whether recent grads are comparable to past ones. Love to learn more.
It’s true - no ordinary technology has ever caused temporary job losses.
Hiring trends bounce around a great deal at the occupational and sectoral levels, sometimes in tandem with the aggregate economy, but sometimes not. While I haven't read the underlying paper, what's presented in this article doesn't seem like a slam-dunk confirmation of AI effects. Basically, as with some of the other commenters, I'd be hesitant to declare the Hand of God at work in what could plausibly be secular changes in hiring rates. But maybe the article undersells the full breadth of evidence presented in the paper?
Did you guys look at Section 174 of the TCJA?
https://www.thomsonreuters.com/en-us/posts/tax-and-accounting/section-174-expenditures/
Indeed, before the TCJA’s enactment, businesses deducted the total amount of R&D expenditures as an expense in the taxable year. Beginning in 2022, all costs related to R&D must now be amortized over five years for US-based companies or 15 years for non-US companies.”
Software jobs are definitely often classified in R&D, and this new rule generated a massive headwind from
2022 until repeal in July 2025 (via the OBBB). Obviously the timing overlaps with end of Covid, so it’s tricky.
Two questions:
1. How does this interact with the age cohort of the software developers?
2. How will hiring into these jobs change now that the OBBB has passed?
I think one potential issue here is looking at data starting in 2021. There was such a huge increase in employment post-covid recovery in 2022. The number of software engineering job postings massively outpaced other sectors in 2022, so it may have become much more saturated. I'm sure the model in the paper controls for time trends, but it's difficult to ascertain when employment was increasing so drastically for all sectors 2021-2022.
I think seeing the trends pre-Covid as well would make this more convincing. Youth unemployment has been increasing relative to overall unemployment since 2011. However, I know there may be data limitations here.
I'm interested in learning about the potential long-term impact of entry-level workers not learning tasks that are easily automated, assuming the data continues to support this and becomes more robust. I suspect that certain tasks which can be automated are actually important. The question then becomes: which entry-level tasks? I have some ideas in mind, but they require more thought before I comment further.
Yes. I too thought and what are they not learning that they need to have experienced in the future.
“(M)ost people live in the past, hanging onto stale narratives and outdated models”
Underrated, widely applicable point!