Is it working from home, and not generative AI, that is harming the prospects of young workers?

There is growing evidence that the labour market for young workers is challenging. Graduates are finding it more difficult to get jobs after graduation. Several research papers have noted that generative AI may be to blame (see this post, for example), with one research paper referring to the changes in the labour market as seniority-biased technological change (see this post).

But the challenge with trying to attribute changes in the labour market to the rise of generative AI is that there are other contemporaneous changes affecting the labour market as well. One of those changes is the rise of working from home (as I noted in yesterday's post). Working from home may reduce the prospects for junior workers in part because it costs more to supervise and monitor them when they are working from home. Junior workers also benefit from on-the-job learning when they work with other people, and that on-the-job learning is less effective when they work from home. Combining those two effects, working from home reduces the incentive for employers to hire junior workers.

This new working paper by Peter Lambert (University of Warwick) and Yannick Schindler (Ellison Institute of Technology, Oxford) tries to disentangle the effects of generative AI and working from home on employment of younger workers. They use data from Revelio Labs that is made up of monthly matched employer-employee records collected from résumés (predominantly from LinkedIn) to construct a measure of the junior share of all new hires. They also use data from Lightcast on the near-universe of online job postings across thousands of online job sites and other websites. They use the Lightcast data to construct a measure of the share of job postings that require three or fewer years of experience. Their data from both sources covers the period from 2017 to 2025, and includes four countries: the US, the UK, Canada, and Australia.

Lambert and Schindler then use that data, along with measures of 'exposure to generative AI' and 'exposure to working from home' at the occupation level, in a difference-in-differences strategy. That means that they essentially compare the change in the share of junior job hires (or job postings) between occupations that are more or less exposed to generative AI (or working from home). Their main results are neatly summarised in Figure 3 from the paper:

Panel (a) shows that the junior share of new hires decreases significantly in jobs that are more exposed to working from home, from 2023 onwards (the black line). When they also control for exposure to generative AI (the red line), the effect of working from home barely changes. In contrast, Panel (b) shows that the junior share of new hires also decreases significantly in jobs that are more exposed to generative AI, from 2023 onwards (the black line). However, when they also control for exposure to working from home (the blue line), the effect of generative AI becomes much smaller and statistically insignificant. The results are similar for the share of job postings requiring three or fewer years' experience, as shown in Panels (c) and (d) of the figure.

The size of the effects are quite large too. A one-standard-deviation increase in exposure to working from home reduces the junior share of new hires by about two percentage points, and the share of job postings requiring three or fewer years' experience by 1.5 percentage points.

Lambert and Schindler conclude that, based on their results, working from home is a better predictor of the decline in junior hiring than generative AI. Given potential benefits of working from home, they are reluctant to recommend policies against working from home, instead noting that:

...micro-level adjustments may be required to help firms adapt their organizational practices, so as to enjoy the benefits of WFH [work from home] arrangements while simultaneously managing the development of early-career talent.

Seen alongside the negative mental health impacts of working from home (as noted in yesterday's post), this should give us further pause for thought. However, it is worth noting that even if working from home is a better predictor of reductions in junior hiring than generative AI within their model, that doesn't let generative AI off the hook entirely. Since both trends are happening at the same time, reducing working from home might not eliminate the negative impacts on junior hiring, but instead make generative AI appear more important as an explanation. Lambert and Schindler note early in their paper that it is often the same occupations (white-collar occupations) that are most exposed to both working from home and generative AI. Given that, perhaps Lambert and Schindler's recommendation for micro-level changes in organisational practice may be the best mitigation strategy available to us.

[HT: Marginal Revolution]

Read more:

• ChatGPT and the labour market
• More on ChatGPT and the labour market
• The impact of generative AI on contact centre work
• Some good news for human accountants in the face of generative AI
• Good news, bad news, and students' views about the impact of ChatGPT on their labour market outcomes
• Swiss workers are worried about the risk of automation
• Generative AI and entry-level employment
• Generative AI and seniority-biased technological change

Two new studies on who works from home, and its mental health impacts

The pandemic caused a massive rise in working from home and now, even though lockdowns are long since over and many workers have returned to the workplace, we are beginning to understand working from home (WFH) a lot better. Two new studies have recently added to our understanding.

The first is this article by Cevat Giray Aksoy (European Bank for Reconstruction and Development) and co-authors, published in the AEA Papers and Proceedings (ungated earlier version here). They use data from the monthly US Survey of Working Arrangements and Attitudes, limiting their data to the period from January 2024 to December 2025, and document three facts about WFH. First, employees are more likely to work from home if they work for a younger firm, and peaks among those working for employers that were founded in the height of the pandemic, in 2020.

Second, employees are more likely to work from home if they work at a firm with a younger CEO. Specifically:

Firms led by CEOs under 30 have an average of 1.4 WFH days per week, compared with 1.1 days at firms led by CEOs who are 60 or older.

That doesn't seem like a lot, but an additional 0.3 days per week is a little more than three working weeks per year of WFH for those working for the youngest CEOs compared with those working for the oldest. However, this relationship between CEO age and WFH appears to be partly explained by the fact that younger CEOs are more likely to be leading younger firms. When Aksoy et al. put both CEO age and firm age in the same regression model, only firm age remains statistically significant. It is a similar story for CEO gender, which is initially statistically significant, but since female CEOs tend to be younger and to be CEOs of younger firms, CEO gender isn't statistically significant once those other variables are controlled for.

Third, the self-employed are much more likely to work from home. Specifically:

Self-employed workers report two to three times as many WFH days per week as wage and salary employees, depending on employer size. Compared to wage and salary employees, the self-employed are more than three times as likely to work in a fully remote capacity.

This last result is not entirely surprising, given that the self-employed typically have a lot more flexibility over scheduling. And, the self-employed may be the type of people who most value flexibility as well.

The second new article is this one by Natalia Emanuel (Federal Reserve Bank of New York), Emma Harrington (University of Virginia), and Amanda Pallais (Harvard University), published in the prestigious journal Science (open access). They look at the mental health impacts of WFH, using US data from a variety of sources, and a difference-in-differences approach. This involves comparing occupations that are more or less amenable to WFH, between the time before the pandemic and the time after the pandemic. They refer to the occupations that are more amenable to WFH as 'remotable'.

Emanuel et al. first document the dramatic rise of WFH:

The pandemic led to a large increase in remote work for those in remotable jobs, such that by 2024, workers in remotable jobs spent 31.1% of workdays fully remote, whereas people in nonremotable jobs spent only 8.9% fully remote... Those in remotable jobs experienced a 17.9 percentage point (pp) differential increase in fully remote work...

They then show that this rise is associated with more time spent alone:

Along with spending less time in the office, workers in remotable jobs spent more time working alone after the pandemic, logging 1.2 more work hours alone per day relative to nonremotable workers (58.0% increase; P < 0.0001).

Even for those of us who are introverts, more alone time may not necessarily be a good thing. Emanuel et al. are concerned about how WFH and working alone affects mental health. Their main outcome variable is the Kessler (K-6) Psychological Distress Scale, which is:

...based on how often in the past 30 days the respondent felt worthless, hopeless, restless, nervous, that everything is an effort, or so sad that nothing could cheer them up...

Their main source of data is the Panel Study of Income Dynamics covering the period from 2011 to 2023 (from which they exclude the pandemic years 2020 and 2021). Analysing that data, they find that:

Between the pre-and postpandemic periods, mental distress increased for everyone, but it increased significantly more for those in remotable jobs...

Among those in remotable jobs, there was a 0.3 unit increase in the K-6 distress score relative to an average score of 3.0 before the pandemic (standard deviation change = 0.08; P = 0.063) in the Panel Study of Income Dynamics (PSID). In the National Health Interview Study (NHIS), we found the same 0.3 unit deterioration (P = 0.007). We saw deterioration in each of the six subcomponents of the K-6 distress scale: feeling worthless, hopeless, restless, nervous, that everything is an effort, and so sad that nothing can cheer them up...

Importantly, the deterioration in mental health is concentrated among people living alone, which is consistent with the idea that WFH affects mental health through increasing social isolation. Emanuel et al. also find that people in remotable jobs are more likely to seek help from a mental health practitioner, and take relatively more prescription medications for mental health conditions such as anxiety or depression. These changes aren't simply the result of greater flexibility allowing more time to be devoted to health care generally, as there was no change in visits to the doctor and no change for other prescription medications such as statins.

Finally, Emanuel et al. looked at whether the rise of generative AI, rather than the increase in WFH, might explain the results (an important check, given the paper I will blog about tomorrow). They find that results from the same analysis, but substituting an AI occupational exposure index in place of the 'remotability' index, are not statistically significant.

Now, many workers are very keen on WFH - as noted in this post, about half of Australian workers would be willing to give up some salary in order to work from home. Why would people choose more WFH if it may worsen their mental health? Of course, a rational worker would weigh up the benefits and costs of WFH, and may decide that the mental health costs are more than offset by other benefits. However, Emanuel et al. point to another related possibility, which is:

...that the benefits of remote work (e.g., skipping a daily commute) are immediate and salient, whereas the costs of remote work (e.g., frayed connections with co-workers) take time to materialize.

So, a rational worker may be essentially weighing up benefits that occur today, against uncertain costs that may occur sometime in the future and therefore should be discounted (in the same way that we should discount future cashflows in a financial analysis). In that sort of exercise, where the mental health costs are discounted, it is more likely that workers would choose to work from home. They would be even more likely to do so if they are quasi-rational and heavily discount the future, as I note in the first week of my ECONS102 class. In that case, the mental health costs would be heavily discounted. Finally, maybe workers are simply unaware of the mental health costs of WFH. If that is the case, then an information intervention might be helpful in improving mental health among workers who would otherwise be WFH. In the meantime, this research suggests that the post-pandemic rise in WFH may have contributed to some part of the growing mental health crisis, especially through increased time spent alone.

[HT: Marginal Revolution for the Emanuel et al. article]

Read more:

• Remote work, wages, and compensating differentials
• Willingness-to-pay for working from home

Artificial intelligence and the 'age of leisure'

My ECONS101 class covered constrained optimisation last week, and one of the models we looked at was the labour-leisure trade-off for workers. Now artificial intelligence, and in particular generative AI, is likely to have large impacts on the labour-leisure trade-off. As the Financial Times reported last year (paywalled):

The idea that technological progress can enable people to work fewer hours is not outlandish...

But in order to believe a similar trend is going to take hold again, you have to assume three things. First: that AI will deliver a substantial boost to economic productivity...

Second, you have to assume the economic gains will be widely distributed...

Third, you have to believe workers will “cash in” those proceeds in the form of extra leisure, rather than higher income. But will they? In many developed countries, there has been a slowdown in the reduction in working hours in recent decades...

Far from trading income for leisure, it is the people with the highest salaries who tend to work the longest hours.

Will workers trade off higher productivity for more leisure time? Are we about to enter an 'age of leisure'? The constrained optimisation model for the worker (see also this post) can help us clarify the possibilities. In this model, we'll assume that AI increases productivity, and that the increase in productivity is represented by higher wages for workers. [*] The model will then tell us whether workers might respond by consuming more, or less, leisure.

Our model of the worker's decision is outlined in the diagram below. The worker's decision is constrained by the amount of discretionary time available to them. Let's call this their time endowment, E. If they spent every hour of discretionary time on leisure, they would have E hours of leisure, but zero income. That is one end point of the worker's budget constraint, on the x-axis. The x-axis measures leisure time from left to right, but that means that it also measures work time (from right to left, because each one hour less leisure means one hour more of work). The difference between E and the number of leisure hours is the number of work hours. Next, if the worker spent every hour working, they would have zero leisure, but would have an income equal to W0*E (the wage, W0, multiplied by the whole time endowment, E). That is the other end point of the worker's budget constraint, on the y-axis. The worker's budget constraint joins up those two points, and has a slope that is equal to the wage (more correctly, it is equal to -W0, and it is negative because the budget constraint is downward sloping). The slope of the budget constraint represents the opportunity cost of leisure. Every hour the worker spends on leisure, they give up the wage of W0. Now, we represent the worker's preferences over leisure and consumption by indifference curves. The worker is trying to maximise their utility, which means that they are trying to get to the highest possible indifference curve that they can, while remaining within their budget constraint. The highest indifference curve they can reach on our diagram is I0. The worker's optimum is the bundle of leisure and consumption where their highest indifference curve meets the budget constraint. This is the bundle A, which contains leisure of L0 (and work hours equal to [E-L0]), and consumption of C0.

Now, let's say that the situation shown above is the situation before the advent of AI. After AI is introduced, productivity increases, and so wages increase (from W0 to W1). This causes the budget constraint to pivot outwards and become steeper (since the slope of the budget constraint is equal to the wage, the slope has increased from -W0 to -W1). The worker can now reach a higher indifference curve, and it is the position of that higher indifference curve that determines the worker's response in terms of whether they consume more leisure or not. If they move to the higher indifference curve I1, then the worker's new optimum is the bundle of leisure and consumption B, which contains leisure of L1 (and work hours equal to [E-L1]), and consumption of C1. For this worker (whose response is shown in red on the diagram), leisure hours decrease as a result of the higher wage. On the other hand, if they move to the higher indifference curve I2, then the worker's new optimum is the bundle of leisure and consumption C, which contains leisure of L2 (and work hours equal to [E-L2]), and consumption of C2. For this worker (whose response is shown in blue on the diagram), leisure hours increase as a result of the higher wage. [**]

Either of these possibilities could happen. In fact, both could happen, with some workers increasing leisure time and others decreasing leisure time. By itself, this model doesn't answer the question of what will happen, but shows that both increased leisure and decreased leisure are possible outcomes.

The key difference here comes down to the size of the income effect of the increase in wages. When wages increase, the opportunity cost of leisure increases. That makes leisure relatively more expensive, and workers should respond by consuming less leisure. That is what we call the substitution effect - workers substitute away from leisure as it becomes more expensive. However, increased wages also lead to an income effect. Leisure is a normal good, which means that as the worker's income increases, they would like to consume more leisure. Notice that the substitution effect and the income effect are working in opposite directions here. For workers who overall decrease their leisure, the substitution effect (which says they should consume less leisure) must be bigger than the income effect (which says they should consume more leisure). For workers who overall increase their leisure, the reverse is true - the substitution effect must be smaller than the income effect.

AI may lead us into an age of leisure. But only if productivity gains lead to higher wages, and the income effect of higher wages more than offsets the substitution effect.

*****

[*] The assumption that productivity gains will lead to higher wages is a strong assumption. Indeed, the FT article questions whether this assumption is valid. If productivity gains don't lead to higher wages, then this model doesn't help us evaluate whether we're about to move into an 'age of leisure', and the impacts might be more macroeconomic than microeconomic. That is, we may end up with leisure, but arising through weaker labour demand, reduced hours, or unemployment rather than through workers voluntarily choosing more leisure as wages increase.

[**] Notice that the indifference curves I1 and I2 are crossing, and indifference curves cannot cross. However, those two indifference curves are for different workers, so there is no problem. I could easily have drawn two different diagrams, one for each worker, but I've kept them both on the same diagram for efficiency.

Did employers value an AI-related qualification in 2021?

Many universities are rapidly adapting to education in the age of generative AI by trying to develop AI skills in their students. There is an assumption that employers want graduates with AI skills across all disciplines, but is there evidence to support that? This recent discussion paper by Teo Firpo (Humboldt-Universität zu Berlin), Lukas Niemann (Tanso Technologies), and Anastasia Danilov (Humboldt-Universität zu Berlin) provides an early answer. I say it's an early answer because their data come from 2021, before the wave of generative AI innovation that became ubiquitous following the release of ChatGPT at the end of 2022. The research also focuses on AI-related qualifications, rather than the more general AI skills, but it's a start.

Firpo et al. conduct a correspondence experiment, where they:

...sent 1,185 applications to open vacancies identified on major UK online job platforms... including Indeed.co.uk, Monster.co.uk, and Reed.co.uk. We restrict applications to entry-level positions requiring at most one year of professional experience, and exclude postings that demand rare or highly specialized skills...

Each identified job posting is randomly assigned to one of two experimental conditions: a "treatment group", which receives a résumé that includes additional AI-related qualifications and a "control group", which receives an otherwise identical résumé without mentioning such qualifications.

Correspondence experiments are relatively common in the labour economics literature (see here, for example), and involve the researcher making job applications with CVs (and sometimes cover letters) that differ in known characteristics. In this case, the applications differed by whether the CV included an AI-related qualification or not. Firpo et al. then focus on differences in callback rates, and they differentiate between 'strict callbacks' (invitations to interview), and 'broad callbacks' (any positive employer response, including requests for further information). Comparing callback rates between CVs with and without AI-related qualifications, they find:

...no statistically significant difference between treatment and control groups for either outcome measure...

However, when they disaggregate their results by job function, they find that:

In both Marketing and Engineering, résumés listing AI-related qualifications receive higher callback rates compared to those in the control group. In Marketing, strict callback rates are 16.00% for AI résumés compared to 7.00% for the control group (p-value = 0.075...), while broad callback rates are 24.00% versus 12.00% (p-value = 0.043...). In Engineering, strict callback rates are 10.00% for AI résumés compared to 4.00% for the control group (p-value = 0.163...), while broad callback rates are 20.00% versus 8.00% (p-value = 0.024...).

For the other job functions (Finance, HR, IT, and Logistics) there was no statistically significant effect of AI qualifications on either measure of callback rates. Firpo et al. then estimate a regression model and show that:

...including AI-related qualifications increases the probability of receiving an interview invitation for marketing roles by approximately 9 percentage points and a broader callback by 12 percentage points. Similarly, the interaction between the treatment dummy and the Engineering job function dummy in the LPM models is positive and statistically significant, but only for broad callbacks. AI-related qualifications increase the probability of a broad callback by at least 11 percentage points...

The results from the econometric model are only weakly statistically significant, but they are fairly large in size. However, I wouldn't over-interpret them because of the multiple-comparison problem (around five percent of results would show up as statistically significant just by chance). At best, the evidence that employers valued AI-related qualifications in 2021 is pretty limited, based on this research.

Firpo et al. were worried that employers might not have noticed the AI qualifications in the CVs, so they conducted an online survey of over 700 professionals with hiring experience and domain knowledge, but that survey instead shows that the AI-related qualification was salient and a signal of greater technical skills, but lower social skills. These conflicting signals are interesting, and suggestive that employers are looking for both technical skills and social skills in entry-level applicants. Does this, alongside the earlier results for different job functions, imply that technical skills are weighted more heavily than social skills for Engineering and Marketing jobs? I could believe that for Engineering, but for Marketing I have my doubts, because interpersonal skills are likely to be important in Marketing. Again though, it's probably best not to over-interpret the results.

Firpo et al. conclude that:

...our findings challenge the assumption that AI-related qualifications unambiguously enhance employability in early-career recruitment. While such skills might be valued in abstract or strategic terms, they do not automatically translate into interview opportunities, at least not in the entry-level labor market in job functions such as HR, Finance, Marketing, Engineering, IT and Logistics.

Of course, these results need to be considered in the context of their time. In 2021, AI-related skills might not have been much in demand by employers. That is unlikely to hold true now, given that generative AI use has become so widespread. It would be interesting to see what a more up-to-date correspondence experiment would find.

[HT: Marginal Revolution]

Read more:

ChatGPT and the labour market

More on ChatGPT and the labour market

The impact of generative AI on contact centre work

Some good news for human accountants in the face of generative AI

Good news, bad news, and students' views about the impact of ChatGPT on their labour market outcomes

Swiss workers are worried about the risk of automation

How people use ChatGPT, for work and not

Generative AI and entry-level employment

Survey evidence on the labour market impacts of generative AI

Why the effects of a guaranteed income on income and employment in Texas and Illinois shouldn't surprise us

The idea of a universal basic income (sometimes called an income guarantee) has gathered a lot of interest over recent years, particularly as fears of job losses to artificial intelligence have risen. The underlying idea is simple. Government makes a regular payment to all citizens (so it's universal) large enough to cover their basic needs (so it's a basic income). However, other than a number of pilot projects, no country has yet fully implemented a universal basic income (UBI), and many have apparently changed their minds after a pilot (see here and here). There are a couple of reasons for that. First, obviously, is the cost. A basic income of just $100 per week for all New Zealanders would cost about $26 billion per year. That would increase the government budget by about 14 percent [*]. And $100 is not a basic income, because no one is going to be able to live on such a paltry amount. Second, there are worries about the incentive effects of a universal basic income. When workers can receive money from the government for doing nothing (because it's universal), will they work less, offsetting some (if not all) of the additional income from the UBI?

That brings me to this NBER working paper by Eva Vivalt (University of Toronto) and co-authors. The paper was originally published back in 2024, and received quite a bit of coverage then (for examples from the media, see here and here), but has been revised since (and I read the September 2025 revision). Vivalt et al. evaluate the impact of two large guaranteed income programmes in north central Texas (including Dallas) and northern Illinois (including Chicago), both of which were implemented by local non-profit organisations (with the programmes funded by OpenResearch, founded by OpenAI CEO Sam Altman). These are not quite UBIs of course, because they weren't available to everyone. Nevertheless, they do help us to understand the incentive effects that could apply to a UBI. Like many would hope a UBI would be (ignoring the immense fiscal cost), the programmes were quite generous (for those in the treatment group, at least) and:

...distributed $1,000 per month for three years to 1,000 low-income individuals randomized into the treatment group. 2,000 participants were randomly assigned to receive $50 per month as the control group.

Vivalt et al. look at the impacts on employment and other related outcomes. There is a huge amount of detail in the paper, so I'm just going to look at some of the highlights. In terms of the overall effect, they find that:

...total individual income excluding the transfers fell by about $1,800 per year relative to the control group, with these effects growing over the course of the study.

So, people receiving the UBI received less income (excluding the UBI - their income increased once you consider the UBI plus their other income). In terms of employment:

The program caused a 3.9 percentage point reduction in the extensive margin of labor supply and a 1-2 hours/week reduction in labor hours for participants. The estimates of the effects of cash on income and labor hours represent an approximately 5-6% decline relative to the control group mean.

People responded to receiving a UBI by working less, just as many of those who had concerns about the incentive effects of a UBI feared. However, the negative incentives also extended to others in the household:

Interestingly, partners and other adults in the household seem to change their labor supply by about as much as participants. For every one dollar received, total household income excluding the transfers fell by around 29 cents, and total individual income fell by around 16 cents.

So, although households received $1000 extra per month from the UBI, their income only increased by $710 on average, because the person receiving the UBI, and other adults in the household, worked less on average. What were they doing with their extra time? Vivalt et al. use American Time Use Survey data, and find that:

Treated participants primarily use the time gained through working less to increase leisure, also increasing time spent on driving or other transportation and finances, though the effects are modest in magnitude. We can reject even small changes in several other specific categories of time use that could be important for gauging the policy effects of an unearned cash transfer, such as time spent on childcare, exercising, searching for a job, or time spent on self improvement.

So, people spend more time on leisure. Do they upgrade to better jobs, which is what some people claim would happen (because the UBI would give people the freedom to spend more time searching for a better job match)? Or do they invest in more education, or start their own business? It appears not, as:

...we find no substantive changes in any dimension of quality of employment and can rule out even small improvements, rejecting improvements in the index of more than 0.022 standard deviations and increases in wages of more than 60 cents. We find that those in the treatment group have more interest in entrepreneurial activities and are willing to take more financial risks, but the coefficient on whether a participant started a business is close to 0 and not statistically significant. Using data from the National Student Clearinghouse on post-secondary education, we see no significant impacts overall but some suggestive evidence that younger individuals may pursue more education as a result of the transfers...

Some people have concluded that the results show that a guaranteed income or UBI is a bad policy. However, the guaranteed income did increase incomes (including transfers) overall and therefore makes people on average better off financially. Leisure time is an important component of our wellbeing, so we shouldn't necessarily consider more leisure time a bad outcome for a policy. In fact, Vivalt et al. also find that on average the guaranteed income increases subjective wellbeing on average (but only in the first year, after which subjective wellbeing returns to baseline). 

The results should have surprised anyone. They are consistent with a simple model of the labour-leisure tradeoff that I cover in my ECONS101 class. The model (of the worker's decision) is outlined in the diagram below. The worker's decision is constrained by the amount of discretionary time available to them. Let's call this their time endowment, E. If they spent every hour of discretionary time on leisure, they would have E hours of leisure, but zero income. That is one end point of the worker's budget constraint, on the x-axis. The x-axis measures leisure time from left to right, but that means that it also measures work time (from right to left, because each one hour less leisure means one hour more of work). The difference between E and the number of leisure hours is the number of work hours. Next, if the worker spent every hour working, they would have zero leisure, but would have an income equal to W0*E (the wage, W0, multiplied by the whole time endowment, E). That is the other end point of the worker's budget constraint, on the y-axis. The worker's budget constraint joins up those two points, and has a slope that is equal to the wage (more correctly, it is equal to -W0, and it is negative because the budget constraint is downward sloping). The slope of the budget constraint represents the opportunity cost of leisure. Every hour the worker spends on leisure, they give up the wage of W0. Now, we represent the worker's preferences over leisure and consumption by indifference curves. The worker is trying to maximise their utility, which means that they are trying to get to the highest possible indifference curve that they can, while remaining within their budget constraint. The highest indifference curve they can reach on our diagram is I0. The worker's optimum is the bundle of leisure and consumption where their highest indifference curve meets the budget constraint. This is the bundle A, which contains leisure of L0 (and work hours equal to [E-L0]), and consumption of C0.

Now, consider what happens when the worker receives a UBI. This is shown in the diagram below. At each level of leisure (and work), their income (and therefore consumption) is higher. That shifts the budget constraint up vertically by the amount of the UBI. If the worker spends no time at all working, they now have consumption of U, instead of zero, and if they spend all of their time working (and have no leisure) their consumption would be W0*E+U. The worker can now reach a higher indifference curve (I1). Their new optimal bundle of leisure and consumption is B, which contains leisure of L1 (and work hours equal to [E-L1]), and consumption of C1. Notice that the worker now consumes more leisure and more consumption as well. Because leisure has increased, that means that the number of work hours has decreased. The increase in leisure, decrease in work hours, and increase in income overall (when the UBI is included), are consistent with what Vivalt et al. found.

So, based on a simple model of the labour-leisure tradeoff, the results of this guaranteed income programme are not surprising. We should have expected a reduction in work, and a reduction in labour income, and that's what Vivalt et al. found. The question policymakers are left with is whether a large income transfer like this is worth it for government, if each $1000 transferred increases incomes by just $710 on average.

[HT: Marginal Revolution, back in 2024]

*****

[*] Of course, if other welfare payments were scrapped in favour of a universal basic income, then the net cost would be lower. Nevertheless, the point that the cost is very high still stands.

Immigration and the wages of the native-born population

Restrictions on immigration flows are getting a lot of policy attention of late. The argument is that immigration reduces wages for the native-born population. But, is there evidence for that? As you might expect, there are literally dozens of studies that have looked into this question, and there are now several meta-analyses that combine the results across many studies (including the meta-analysis that I referred to in this 2016 post. That post referred to this 2005 article by Longhi et al., which found that:

Overall, the effect is very small. A 1 percentage point increase in the proportion of immigrants in the labour force lowers wages across the investigated studies by only 0.119%.

Longhi et al. then followed up with another article in 2010, which also found a very small effect of immigration on wages, specifically:

...a 1% point increase in the immigration to population ratio reduces wages by only 0.03%.

A new meta-analysis article by Amandine Aubry (Université de Caen Normandie) and co-authors, published in the journal Labour Economics (open access), picks up those two earlier meta-analyses, and extends the analysis up to 2023. Specifically, their analysis includes:

...88 studies published between 1985 and 2023, encompassing 2,989 reduced-form estimates of the wage effects of immigration.

Many post-2010 studies use shift-share (Bartik) instruments to estimate the causal effect of immigration on wages. These instruments predict regional immigrant inflows by interacting a region’s pre-existing settlement shares by origin with national inflows from those origins. They then use the predicted inflows as an instrument for actual inflows in an instrumental variables framework. This approach helps address the concern that immigrants may sort into destinations with stronger labour markets, which would make immigration and wages correlated for reasons other than a causal effect of immigration on wages.

Now, Aubry et al. are more concerned with investigating the heterogeneity in the estimated effects of immigration on wages, rather than the overall estimate. Nevertheless, I think the overall estimate is interesting and important, and for that they find:

...a 1% rise in the immigrant labour force reduces native wages by about 0.033% on average.

This overall effect is very similar to that from the second meta-analysis by Longhi et al. But it's tiny - a 1 percent larger immigrant labour force would reduce the wages of a native-born worker earning $1000 per week by about 33 cents. And, there is substantial variation around that small overall estimate, which Aubry et al. investigate in some detail. They find that:

...contextual heterogeneity explains part of the variance in the estimates. Estimates for Anglo-Saxon and developing countries are systematically larger than those for other economies, and the historical period covered by a study also affects the results, with later periods being associated with smaller effects. Third, methodological heterogeneity is key... In particular, instrumental variable estimations, which are commonly used to infer causality, yield smaller coefficients than OLS...

More recent studies tend to estimate smaller effects of immigration on wages, as do studies that employ instrumental variables (which also tend to be more recent studies). That accords with the results from the two Longhi et al. meta-analyses, where the second study found a much smaller overall effect than the first study. The shift-share instrument only became established as a method by David Card and others in the early 2000s, so its use only began diffusing from then. Given that these sorts of analyses have become the industry standard now, we can generally expect future studies to find smaller effects than older studies.

The results for developing countries, where the effect of immigration on wages is more positive than for developed countries deserves more exploration. Aubry et al.'s sample includes estimates from only a handful of developing countries (Colombia, Costa Rica, Malaysia, Peru, South Africa, and Thailand). This also suggests that more studies on the effect of immigration on wages in developing country contexts would be useful.

The overall takeaway from this meta-analysis is that immigration on average has a negligible overall effect on the wages of the native-born population on average. Unfortunately, this is one of those cases where the empirical results do not accord with 'folk economics'. Although the average effect is negligible, the wages of some subgroups may be negatively impacted by immigration in some contexts (and Aubry et al.'s results are consistent with the idea that the impacts are negative in some contexts or for some groups). The general public (and policy makers) will tend to focus on those negative impacts. Nevertheless, it should be possible in principle to address those negative impacts through policy (economists refer to this as the compensation principle), so that those who benefit from immigration (including immigrants themselves) can continue to do so.

Read more:

• The (small) effect of immigration on wages
• In the labour market, you can't have the best (or the worst) of both worlds
• Immigrant restrictions and wages for locals
• Immigration, native employment, and political backlash
• Michael Clemens vs. the Roy model of migration

What we learn from Freelancer.com about labour market signalling in the age of generative AI

In yesterday's post, I outlined my case for why generative AI reduces the quality of signalling in education. That is, how good education (qualification, or grades) is as a signal to employers of an applicant's ability. There is evidence to support this case, from two recent papers.

The first paper is this pre-print by Jingyi Cui, Gabriel Dias, and Justin Ye (all Yale University), which looks at the signalling benefit in cover letters. Specifically, they study:

...the introduction of a generative AI cover letter writing tool on Freelancer.com, one of the world’s largest online labor platforms. Freelancer connects international workers and employers to collaborate on short-term, skilled, and mostly remote jobs. On April 19, 2023, Freelancer introduced the “AI Bid Writer,” a tool that automatically generates cover letters tailored to employers’ job descriptions that workers can use or edit. The tool was available to a large subset of workers depending on their membership plans.

Cui et al. use eight months of data on two skill categories (PHP, and Internet Marketing), which covers over five million cover letters submitted to over 100,000 job opportunities. They observe who had access to the tool, as well as who used the tool to generate a cover letter, and how much time they spent refining the AI-generated cover letter.

Cui et al. look at the impact of the availability of the generative AI tool on callback rates, using a difference-in-differences research design. This effectively involves comparing differences in callback rates between applicants with and without access to the tool, before and after the tool was made available. Cui et al. find that:

...access to the generative AI writing tool increased cover-letter tailoring by 0.16 standard deviations, while actual usage raised tailoring by 1.36 standard deviations. Applying the same design to callbacks as the outcome, we find that access to the generative AI tool increased the probability of receiving a callback by 0.43 percentage points, and usage raised it by 3.56 percentage points. The latter represents a 51% increase relative to the pre-rollout average callback rate of 7.02%.

All good so far. Job applicants are made significantly better off (in terms of receiving a callback) by using the tool. However:

Our second finding is that AI substitutes for, rather than complements, workers’ pre-AI cover letter tailoring skills... We find that workers who previously wrote more tailored cover letters experienced smaller gains in cover letter tailoring—indeed, the best writers... experienced 27% smaller gains than the weakest ones. By enabling less skilled writers to produce more tailored cover letters, AI narrows the gap between workers with different initial abilities.

In other words, employers are now less able to distinguish the quality of the worker by using the quality of the writing in the cover letter. The consequence of this is that:

The correlation between cover-letter tailoring and receiving a callback fell by 51% after the launch of the AI tool, and the correlation with receiving an offer fell by 79%. Instead, employers shifted toward other signals less susceptible to AI influence, such as workers’ past work experience. The correlation between callbacks and workers’ review scores—the platform’s proprietary metric summarizing past work experiences on the platform and determining the default ranking of applications—rose by 5%. These patterns suggest that as AI adoption increases, employers substitute away from easily manipulated signals like cover letters toward harder-to-fake indicators of quality.

The total number of interviews and job offers were unchanged during this period. Cui et al. don't directly report whether the number of callbacks changed, but if we infer that from there being no aggregate change in the number of interviews, then this is consistent with the idea that the key difference is in the distribution of who received the jobs (and callbacks). Workers with a strong alternative signal (other than a well-written cover letter) received more callbacks, meaning that workers who lack an alternative signal received fewer callbacks. That has an important distributional consequence. New workers typically lack past review scores, so as employers lean more heavily on reviews, workers who are new to Freelancer.com will be disadvantaged and will find it more difficult to get a callback. Overall, in this case, the impact of the generative AI tool on the quality of signalling is negative.

The second paper is this job market paper by Anaïs Galdin (Dartmouth College) and Jesse Silbert (Princeton), who also use data from Freelancer.com. The difference is that they carefully evaluate employers' willingness-to-pay for workers, using the bid data. They also look at customisation of the text of the whole proposal, not just the cover letter. Another difference is that Galdin and Silbert look at a different job type, coding. Their data covers 2.7 million applications to 61,000 job openings, by 212,000 job applicants. Although Galdin and Silbert's paper is far more technical than the Cui et al. paper, Galdin and Silbert's results are somewhat similar (in terms of what they tell us about signalling):

First, we show that before the mass adoption of LLMs, employers had a significantly higher willingness to pay for workers who sent more customized proposals. Estimating a reduced-form multinomial logit model of employer demand using our measure of signal, we find that, all else equal, workers with a one standard deviation higher signal have the same increased chance of being hired as workers with a $26 lower bid... Second, we provide evidence that before the adoption of LLMs, employers valued workers’ signals because signals were predictive of workers’ effort, which in turn predicted workers’ ability to complete the posted job successfully. Third, we find, however, that after the mass adoption of LLMs, these patterns weaken significantly or disappear completely: employer willingness to pay for workers sending higher signals falls sharply, proposals written with the platform’s native AI-writing tool exhibit a negative correlation between effort and signal, and signals no longer predict successful job completion conditional on being hired.

This is strong evidence that, in this context at least, the introduction of the generative AI tool substantially reduces the quality of the job application signal. Galdin and Silbert then build an economic model calibrated based on their empirical results, and using that model they find that:

Compared to the status quo pre-LLM equilibrium with signaling, our no-signaling counterfactual equilibrium is far less meritocratic. Workers in the bottom quintile of the ability distribution are hired 14% more often, while workers in the top quintile are hired 19% less often.

This suggests an even worse outcome than what Cui et al. find. Galdin and Silbert's results suggest that the distributional changes in who gets offered work make high-quality workers worse off, and low-quality workers better off. That is what we would expect when the quality of signalling is reduced. Galdin and Silbert go on to say that:

These effects are driven by three mechanisms. First, employers previously relied on signals to make hiring decisions, so losing access to them impinges on their ability to discern worker ability. Second, more indirectly, the significant positive correlation between a worker’s ability and cost implies that, when employers lose access to signals and workers are forced to compete more intensely on wages, the prevailing workers with lower bids tend to have lower abilities. Third, since workers’ observable characteristics are poor predictors of their ability, employers have little to no information to distinguish between high and low-ability workers.

These changes to hiring patterns lead to a 5% reduction in average wages, a 1.5% reduction in overall hiring rate per posted job, a 4% reduction in worker surplus, and a small, less than 1%, increase in employer surplus.

The overall takeaway from both papers is that generative AI reduces the quality of signals to employers. They don't speak directly to the quality of education signalling, but we can infer that if the quality of other signals of worker quality are reduced by generative AI, then the quality of the education signal likely is as well. That's because proposals and cover letters on Freelancer.com play much the same signalling role as degrees and grades. In both cases, employers can’t observe ability directly, so they rely on an observable, costly signal. On Freelancer.com, that is the proposal or cover letter, and for education, that is the degree or grade. Generative AI makes it much easier for almost anyone to produce a polished proposal or assessment, so the observable output becomes less tightly linked to ability, weakening the value of both kinds of signal.

Read more:

• The impact of large language models on job interviews, from both sides
• David Deming on generative AI and commitment to learning, and the impact of generative AI on signalling in education

Employers strongly prefer applicants who complete in-person rather than online qualifications

As I've noted before (see this post and the links at the end of it), on average online and blended learning don't appear to make students any better off, or any worse off, in terms of learning (however, that conclusion hides important heterogeneity, with more engaged students doing better with online learning, and less engaged students doing worse). So, on average, the human capital or skills gained from online learning appear similar for both online and in-person learning. If employers cared only about skills, they shouldn’t care whether those skills were acquired online or in-person.

However, human capital development is only part of the benefits of higher education. In his book The Case Against Education (which I reviewed here), Bryan Caplan presented an estimate that the education premium is 20% human capital and 80% signalling. And as I have noted several times (most recently in this post), the signal from online education is much weaker than the signal from in-person education. Putting that all together, we should expect employers to be more skeptical of online qualifications, and to be less willing to hire graduates who have online qualifications than those who have studied in-person.

This 2021 article by Conor Lennon (University of Louisville), published in the journal ILR Review (ungated version here), uses a correspondence experiment to demonstrate exactly that. A correspondence experiment involves the researcher making job applications with CVs (and sometimes cover letters) that differ in known characteristics. The researcher then counts how often CVs with different characteristics receive callbacks (a positive phone message or email, or an invitation to an interview). A very simple regression model can then be used to estimate the effect of each characteristic on the probability of receiving a callback. That is what Lennon did in this experiment, with the key characteristic being whether the applicant studied online or in person. As he explains:

...I examine employer responses to 1,891 job applications using 100 unique fictitious applicant profiles. The fictitious profiles are based on real résumés, gathered from a major online jobs website, and represent recent college graduates in four broad areas: business, engineering, nursing, and accounting. For each real résumé, names, dates, contact information, addresses, and previous employer and education details were anonymized. At random, for 50 of these résumés, the researcher added the word ‘‘online’’ in parentheses next to the name of the listed college or university. The researcher then used these résumés to apply for suitable job openings... Because employers typically left voicemail messages without specifically offering an interview time, any positive personalized contact is considered a ‘‘callback.’’

To avoid the employers detecting that they were being subjected to research, each job opening received only one randomised application. However, this 'unmatched' design is still appropriate, because randomisation and a large sample size mean that, on average, the only systematic difference between the two groups of résumés is whether the degree is listed as online or in-person. Lennon finds that:

The effect of having an online degree is large and negative in all specifications. Specifically, the estimates... suggest a 7.3 percentage-point difference in callback rates between traditional and online degree holders, all else being equal... Given that the mean callback rate for online degree holders is 8.3%, a 7.3 percentage-point difference suggests that a résumé reflecting a traditional degree will receive almost twice as many callbacks for interviews as a résumé reporting an online degree, all else being equal.

That is a huge effect and, because 'online' versus 'in-person' was randomly assigned across otherwise similar résumés, we can interpret the 7.3 percentage-point difference in callbacks as a causal effect of completing an online qualification.. Lennon then tests whether the effect is larger depending on the gender or race of the (fictitious) applicant, or by profession. The results show some differences, but I wouldn't read too much into them because they’re driven by a small proportion of the sample. On the other hand, the effect on online education does make a difference to the effect of GPA on the probability of getting a callback. Specifically:

...GPA matters significantly but only for in-person degree holders. Put another way, if you earn an online degree, even a 4.0 GPA will not help all that much. This estimate is a confirmation of the main takeaway of this article: Employers currently do not appear to trust online education.

This is a clear indication of the difference in the signalling value between an online qualification and an in-person qualification (note that the qualifications that Lennon chose were those that could be completed online or in-person, and were otherwise identical). GPA is also a signal of quality. If GPA makes no difference to callback rates for an online qualification, then employers aren't distinguishing between high-GPA and low-GPA graduates of the online qualification. Employers don't seem to value GPA as a signal of applicant quality, if the applicant completed an online qualification. In contrast, GPA makes a large and statistically significant difference for in-person qualifications, showing that GPA remains a strong signal for employers when students complete an in-person qualification. Lennon concludes that:

Because learning outcomes appear to vary little between in-person and online instruction... fewer callbacks for those with online degrees would support the idea that employers view having a traditional degree as a better signal of employability... Alternatively, employers may be inferring some socioeconomic characteristics, or they may believe that human capital formation is diminished in online programs relative to traditional degrees (even if it is not), that the individual will be less socially adept, or that a traditional college education gives students something more than just grades written on a piece of paper.

Lennon rightly notes that his results apply to new graduates, and may not apply to second-chance learners, who often have more real-world experience prior to beginning (or returning to) higher education studies. This study was also conducted in 2015-2017, and some things have definitely changed. Large language models may actually make online qualifications even less of a quality signal than they did when this research was conducted.

The new graduate market is important to universities. We need to understand how employers view our graduates. Based on this study, we should be very cautious about encouraging students into online-only qualifications, lest we hamper their chances of employment when they graduate.

Survey evidence on the labour market impacts of generative AI

A picture of the labour market impacts of generative AI is slowly emerging. At this stage, there is little consensus on what the impacts will be. I just stumbled across this working paper, by Jonathan Hartley (Stanford University) and co-authors, which I had put aside to read earlier this year. Unlike some of the research I have discussed in recent posts (linked at the end of this post), Hartley et al. make use of a nationally representative survey of US workers.

The survey has had three waves in the US (plus one Canadian wave), and the first US wave had over 4200 respondents (Hartley et al. don't report how many respondents there were for the other waves). The results make for interesting reading. First, in terms of who is using generative AI, they report that:

...LLM adoption at work among U.S. survey respondents above 18 has increased rapidly from 30.1% as of December 2024, to 43.2% as of March/April 2025, and to 45.9% as of June/July 2025...

Conditional on using Generative AI at work, about 33% of workers use Generative AI five days per week at work (every weekday). Roughly 12% of Generative AI users use such tools at work only 1 day at work. About 17% and 18% of Generative AI users use Generative AI tools at work two and three days per week respectively...

That is a lot of people using generative AI for work, and using it often when they do. It is interesting to sit these results alongside those of Chatterji et al. (whose paper I discussed in this post). They found growth in both work-related and non-work-related ChatGPT messages over time.

Who is using generative AI at work, though? Hartley et al. find that:

...Generative AI tools like large language models (LLMs) are most commonly used in the labor force by younger individuals, more highly educated individuals, higher income individuals, and those in particular industries such as customer service, marketing and information technology.

These results are similar to those of Chatterji et al., except that Hartley et al. also report gender differences (with greater use of generative AI by men), whereas Chatterji et al. report that the gender gap that was apparent among early adopters of ChatGPT has closed completely.

Hartley et al. then move on to estimating the productivity gains from generative AI. Given that this is survey-based, and not observational or experimental, we should take these results with a very large grain of salt. Hartley et al. ask their respondents how long it takes then to complete various tasks with and without generative AI. The results are summarised in Figure 12 in the paper:

Notice that every task is reported to take less time with generative AI (the green dots) than without (the blue dots). The productivity gains are different for different tasks. However, I find this figure and the data to be very fishy. How could generative AI create a huge decrease in time on 'Persuasion' tasks? Or 'Repairing' (which has one of the biggest productivity gains). Also, notice how almost every task takes between 25 and 39 minutes with generative AI. I strongly suspect that the research participants are anchoring their responses to this question on 30 minutes with GenAI for some reason. Without seeing the particular questions that are being asked though, it is hard to tell why. [*]

Hartley et al. then try to estimate the impact of generative AI on job postings, employment, and wages, using a difference-in-differences research design. They find no impact on job postings or employment, but significant impacts on wages. However, here things get strange. The coefficients that they report in Tables 6 and 7 of the paper are clearly negative, and yet Hartley et al. write that:

Our estimated coefficients... imply economically meaningful wage effects: a one-standard deviation increase in occupational Generative AI exposure corresponds to a significant increase in median annual wages...

Going back to their regression equations, their 'exposure to generative AI variable' is more positive when exposure is high, so a negative coefficient should imply that more exposure to generative AI is associated with lower wages. I must be missing something?

Given the deficiencies in the data and the regression modelling, I don't think that this paper really adds much to our understanding of the labour market effects of generative AI. Which is disappointing, because survey-based evidence would provide us with a complementary data source that would help us to triangulate with the results from other data sources and methods.

[HT: Marginal Revolution]

*****

[*] On a slightly more technical note, we might expect there to be as much variation (in relative terms) in the 'with GenAI' data as in the 'without GenAI' data. However, the coefficient of variation (the standard deviation expressed as a percentage of the mean) is 0.109 for the 'with GenAI' data, but 0.226 for the 'without GenAI' data. So, there is less than half the variation in the reported task times with GenAI than without. Again, that suggests that this data is fishy.

Read more:

ChatGPT and the labour market

More on ChatGPT and the labour market

The impact of generative AI on contact centre work

Some good news for human accountants in the face of generative AI

Good news, bad news, and students' views about the impact of ChatGPT on their labour market outcomes

Swiss workers are worried about the risk of automation

How people use ChatGPT, for work and not

Generative AI and entry-level employment

Generative AI and seniority-biased technological change

Skills-biased technological change occurs when technology increases the productivity, and hence the value created, by workers with higher skills, compared to those with lower skills. One of the canonical examples is computers, which have made skilled white-collar workers more productive, but automated away the jobs of some lower-skilled workers.

As noted in yesterday's post, something similar may be happening with AI. In this case though, generative AI is making more experienced (senior) workers more productive, while at the same time automating away the jobs of entry-level (junior) workers. Think of this as seniority-biased technological change. At least, that's what this new working paper by Seyed Hosseini and Guy Lichtinger (both Harvard University) calls it. Their explanation goes like this:

In many such [high-skill, white-collar] jobs, workers begin at the bottom of the career ladder performing intellectually mundane tasks, i.e., routine yet cognitively demanding activities such as debugging code or reviewing legal documents, which are likely to be especially exposed to recent advances in GenAI. As these workers gain experience, they typically move up the career ladder to more senior roles that involve more complex problem-solving or managerial responsibilities... If GenAI disproportionately substitutes for entry-level tasks, the lower rungs of these career ladders may be eroding...

Hosseini and Lichtinger use data from Revelio Labs, which is drawn from public LinkedIn profiles. Importantly:

A key feature of the dataset is the standardized seniority level variable for each position, constructed by Revelio through an ensemble modeling approach based on multiple sources of information.

Hosseini and Lichtinger group the standardised positions into juniors (Entry and Junior levels) and seniors (Associate and above). They also use data on job postings from Revelio. The resulting dataset is huge, and:

...covers 284,974 firms that were successfully matched to both employee position data and job postings, and that were actively hiring between January 2021 and March 2025... For these firms, we observe 156,765,776 positions dating back to 2015 and 198,773,384 job postings since 2021...

The raw data shows a pattern that is very similar to the pattern from the Brynjolfsson et al. paper I discussed yesterday. Figure 1 from the Hosseini and Lichtinger paper charts the change in employment compared with January 2015, for juniors and seniors (and overall):

Notice that junior and senior employment follow similar trends until 2020, at which point they diverge, with senior employment continuing to grow, while junior employment does not (and even starts to decline after 2023). Brynjolfsson et al. showed that entry-level employment started declining from 2022, so these results are similar (although the point of departure is somewhat different).

Brynjolfsson et al. weren't able to show definitively that generative AI was the cause of the divergence (although they were able to eliminate general trends in their robustness checks). Hosseini and Lichtinger use the raw description data for each job posting, and identify "GenAI integrator" positions - "those reflecting an active attempt to recruit workers tasked with adopting or implementing GenAI in the firm’s workflows". They then:

...define a firm as a GenAI adopter if it has posted at least one GenAI integrator vacancy. By this criterion, 10,599 firms qualify as adopters. While they make up only 3.72 percent of the 284,974 firms in our sample, adopters are disproportionately large... and account for 17.3 percent of the employment (positions) in our dataset.

Hosseini and Lichtinger then compare employment changes between GenAI adopter firms and 'non-adopters', between the period before the first quarter of 2023 and the period after, in a 'difference-in-differences' analysis. They find that:

...junior employment in adopting firms fell by 7.7 percent relative to controls six quarters after the diffusion of generative AI. By contrast, coefficients for senior workers show a persistent upward trajectory throughout the sample, suggesting that adopting firms expanded senior employment more strongly than non-adopters over the last decade.

Hosseini and Lichtinger then extend their analysis to a triple-difference-in-differences analysis, comparing the difference in employment between juniors and seniors, in GenAI adopter and non-adopter firms, before and after the first quarter of 2023. In this more strenuous analysis, they find that:

Aside from a brief dip in early 2021, the coefficients are essentially flat through 2022Q4. Starting in 2023Q1, however, the coefficients decline sharply, reaching roughly a 10 percent drop after six quarters.

That means that juniors in GenAI adopting firms had a 10 percent greater decrease in employment relative to seniors than juniors in non-adopter firms, between the time before and after the first quarter of 2023 (phew!). This is strong evidence in favour of seniority-biased technological change arising from the adoption of generative AI.

Hosseini and Lichtinger go on to show similar effects using an event study research design, and similar effects comparing juniors in occupations that are more exposed to generative AI (compared with those in less exposed occupations). The latter is similar in nature to the results of Brynjolfsson et al.

Hosseini and Lichtinger then look at whether the change arises from a decrease in hiring of junior employees, an increase in job separation, or a change in promotion, finding that:

...that the sharp contraction in junior employment among adopters is driven primarily by a slowdown in hiring, rather than by increased exits. Specifically, the coefficient on Hiring implies that, relative to non-adopters, GenAI-adopting firms hired on average 5.0 fewer junior workers per quarter after 2023Q1... For senior employees, by contrast, hiring shows little change, while separations rise modestly, leading to a small net decline in senior headcount.

So, again the news is not good for new graduates moving into the workforce. Firms that have adopted generative AI are employing fewer junior employees, and that's because they are hiring fewer junior employees. The effects for new graduates are somewhat heterogeneous though, as Hosseini and Lichtinger also find that:

Juniors from tier-3 and tier-4 universities experienced the steepest relative declines in employment, while juniors from tiers 1, 2 and 5 also saw reductions, but of smaller magnitude.

'Tier-3 universities' are "strong national or regional institutions", while 'tier-4 universities' are "lower-tier but standard institutions". It's easy to see why they might be more affected than 'tier-1 universities' (the Ivy League and elite universities), and 'tier-2 universities' (highly respected international institutions). Signals still matter in education. However, it is hard to see why "tier-5 universities", which are "weak or diploma-mill-type institutions" are less affected. Perhaps students from the lowest quality universities select into occupations that are less likely to be affected by generative AI? Hosseini and Lichtinger don't control for the specific occupation in that analysis, but that might provide an answer.

Unlike Brynjolfsson et al., Hosseini and Lichtinger don't try to end their paper on a positive note. Instead, they conclude that:

GenAI adoption appears to shift work away from entry-level tasks, narrowing the bottom rungs of internal career ladders. Because early-career jobs are central to lifetime wage growth and mobility, such shifts may have lasting consequences for inequality and the college wage premium. Taken together, our evidence suggests that GenAI diffusion constitutes a form of seniority-biased technological change, with far-reaching implications for how careers begin, how firms cultivate talent, and how the gains from new technologies are distributed.

I prefer the more upbeat conclusion of Brynjolfsson et al., which is that the labour market will eventually adjust, and the workers who are disadvantaged now, will end up redeployed into other jobs that open up as a result of generative AI. I guess we will find out as this technological change plays out in real time!

[HT: Marginal Revolution]

Read more:

• ChatGPT and the labour market
• More on ChatGPT and the labour market
• The impact of generative AI on contact centre work
• Some good news for human accountants in the face of generative AI
• Good news, bad news, and students' views about the impact of ChatGPT on their labour market outcomes
• Swiss workers are worried about the risk of automation
• Generative AI and entry-level employment

Generative AI and entry-level employment

Does technological change increase employment, or decrease employment? The answer depends on what you believe about the technological change. If the technological change primarily automates tasks that were previously done by human workers, then it may decrease the demand for labour and reduce employment [*]. On the other hand, if the technological change primarily makes workers more productive, so that they generate more value for their employers, then it may increase the demand for labour and increase employment [**].

Which of those two situations is AI creating? In reality, it's probably a bit of both. However, which effect is stronger? We can get a sense of where things are at from this recent working paper by Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen (all Stanford University). They use data from ADP, the largest payroll processing firm in America, which contains records on between 3.5 and 5 million workers each month between January 2021 and July 2025. Using this data, Brynjolfsson et al. demonstrate six key facts about the labour market over that time:

• First, we find substantial declines in employment for early-career workers in occupations most exposed to AI, such as software development and customer support.

• Second, we show that economy-wide employment continues to grow, but employment growth for young workers has been stagnant.

• Third, entry-level employment has declined in applications of AI that automate work, with muted effects for those that augment it.

• Fourth, these employment declines remain after conditioning on firm-time effects, with a 13% relative employment decline for young workers in the most exposed occupations.

• Fifth, these labor market adjustments are more visible in employment than in compensation.

• Sixth, we find that these patterns hold in occupations unaffected by remote work and across various alternative sample constructions.

The first key fact is demonstrated by looking across all occupations. However, it is most clearly seen in Figure 1 from the paper, which shows how the number of workers (by age group) employed as software developers or customer service representatives (two of the occupations most cited in the media as being affected by AI) have changed compared with October 2022:

Notice how the blue line (employment of early career workers aged 22 to 25 years) trends downwards, starting from 2022, while employment of the most senior workers (especially those aged 35 years and over) continue to trend upwards.

The second key fact builds on this, to show the trend is apparent when you pool all workers together, as demonstrated in Figure 4 from the paper:

Note that the trends are not as wildly different as they are for software developers or customer service representatives, but remember that figure shows the trends across all workers, including those in jobs like nurses, welders, or baristas, whose employment is unlikely to be very impacted by AI (yet!).

Their third key fact relates most closely to the point I raised at the beginning of this post. On this, Brynjolfsson et al. find that:

...occupations with the highest estimated automation shares have experienced declining employment for the youngest workers...

...occupations with the highest estimated augmentation shares have not experienced a similar pattern...

The relevant figures are Figures 7 and 8 from the paper (which are too large for me to reproduce sensibly here). For their fourth key fact, Brynjolfsson et al. use a Poisson regression model within each age group, which allows them to control for firm-time-specific and firm-quintile-specific effects (where the quintiles are quintiles of exposure to AI, drawn from the paper that I discussed in this 2023 post). The firm-time effects will control for firm-specific shocks that affect all of the firm's workers, while the firm-quintile effects will control for different trends affecting a firm's workers that have similar exposure to AI. The analysis won't control for all of the relevant differences, and the analysis remains correlational rather than causal. Nevertheless, Brynjolfsson et al. find that:

For workers aged 22-25, estimates for higher quintiles are large and statistically significant, with a 12 log point decline in relative employment... Estimates for other age groups are generally much smaller in magnitude and not statistically significant.

A 12-log point change is about 11.3 percent, so relative to older workers working in the same firm and with the same exposure to AI, the youngest workers have suffered an 11.3 percent decrease in employment.

Brynjolfsson et al.'s fifth point is simply that the effects show up for employment, but not for wages. And their sixth point is that the effect is robust to various alternatives, including: excluding tech occupations; looking separately at jobs that are, or are not, amenable to remote work; extending the pre-period back to 2018; looking separately by gender; and using the Current Population Survey instead of the payroll data. Interestingly, looking differently at occupations depending on the education level of workers, the results show that:

Occupations with a high share of college graduates have declining employment overall, with muted differences between more-exposed and less-exposed occupations compared to our main results. In contrast, occupations with a low share of college graduates have rising overall employment, with the least AI-exposed occupations growing and the most exposed occupations declining in employment.

All of this is not great news for current university students. Think about all of the results taken together. Firms have been employing fewer entry-level workers, which are your typical graduates. Occupations that have a high share of college graduates are experiencing declining employment overall, in both occupations that are more exposed to AI and occupations that are less exposed to AI. And employment in AI-exposed occupations with a low share of college graduates has also been declining. It seems that the only groups of entry-level workers who aren't experiencing negative trends are non-college-educated workers in occupations that are not exposed to AI.

Brynjolfsson et al. try to finish on a positive note though, noting that:

The adoption of new technologies typically leads to heterogeneous effects across workers, resulting in an adjustment period as workers reallocate from displaced forms of work to new forms with growing labor demand... Past transitions such as the IT revolution ultimately led to robust growth in employment and real wages following physical and human capital adjustments, with some workers benefiting more than others...

Here's hoping that we don't have to wait too long for those adjustments.

[HT: Marginal Revolution] 

*****

[*] However, decreasing employment (and wages) in the automating industry may increase the supply of workers into other industries, increasing employment (but decreasing wages) in those other industries. The general equilibrium effect is not straightforward.

[**] Arguments that increasing productivity will reduce the demand for labour, since fewer workers are needed to complete the same amount of work, run into the 'lump of labour' fallacy. They forget that firms can expand production if they have more productive workers, and will want more workers if their workers are more productive and therefore more profitable to employ.

Read more:

ChatGPT and the labour market

More on ChatGPT and the labour market

The impact of generative AI on contact centre work

Some good news for human accountants in the face of generative AI

Good news, bad news, and students' views about the impact of ChatGPT on their labour market outcomes

Swiss workers are worried about the risk of automation