Generative AI and the future of the university

Universities and other educational institutions have had a few years now to come to grips with generative AI and its implications. For universities, what I have seen is a lot of angst about student cheating and the impacts on learning, courses set up to take advantage of the demand for students with AI skills, but mostly a lot of tinkering around the edges and hoping for the best.

A new working paper by Jason Potts (Alfaisal University) outlines the fundamental challenge to the entire business model of universities. Potts notes that universities are a very complex multi-sided platform market, matching at least fifteen disparate groups including, but not limited to, students, employers, and government.

Potts also notes that what universities offer to students is not content at all, but a 'warrant' - a signal to employers of the student's quality (see my post here for more on that point). This overstates the case of course. Universities do offer students the opportunity to develop important knowledge and skills. Even Bryan Caplan, in his book The Case Against Education (which I reviewed here), acknowledges that education develops human capital, despite arguing that most of the financial return to education comes from signalling.

Now, the challenge is that generative AI makes content delivery a commodity, so Potts argues that universities should concentrate more on providing the warrant of quality, essentially becoming a certifying organisation that provides a signal of the quality of students to future employers. He notes that:

The strategic form this implies is a university that examines and certifies as its primary business, and that treats instruction as a service which may be produced internally, purchased, or left to the student. Almost no one working inside a university right now wants this. But it is the logical direction this new wave of technologies is pushing the university.

I think Potts is right that no one wants this outcome. But he is also right that the value that universities provide to students will be increasingly embedded within the quality of the signal. And it is there that generative AI presents another challenge. Effective signals need to be able to separate high-quality and low-quality students, and one of the ways that occurs is by education being more costly (more difficult or time-intensive, rather than necessarily more financially costly) for lower quality students. Generative AI lowers the cost of producing assessment work that is of high quality, meaning that low-quality students are more difficult to distinguish from high-quality students. The signal value of education is reduced, unless universities make changes to preserve the value of the signal. On that point, Potts argues to:

...move the assessment to a modality where the production cost remains type-dependent. At this stage, the best candidate is probably a return to the viva, common in earlier times as an observed, embodied, high-bandwidth, unscripted examination conducted as an oral defense, or as live problem-solving before an examiner... This is expensive per student and is complementary with cheap AI-delivered instruction.

I'm not convinced that oral examination is the only way to achieve this aim. Properly invigilated in-person written exams are similarly resistant to unauthorised AI use, and are considerably more efficient than oral examinations. However, they are not appropriate to all courses, and may best suit courses where basic foundational knowledge needs to be assessed (although noting that it is possible to assess higher-level skills in a written exam as well). Potts notes that, in terms of university revenue, this means:

...open entry, examination-only pathways to a degree, and pricing on the assessment rather than on tuition or contact hours.

So, in Potts' view, universities wouldn't necessarily do any teaching, only assessing. Or rather, they might do teaching, but it would be complementary to their main activity of assessing students and certifying their knowledge and skills. 

There is a positive element to this for some university academic staff though. If you think through the incentives here, it becomes clear that the endpoint of this transition could easily be for universities to become diploma mills, offering certifications to paying students that bear no meaningful relation to students' actual learning, knowledge, or skills. Universities can only avoid that outcome if they can credibly demonstrate to employers that the credentialling process is not influenced by student demands. And so, academic staff retain some academic freedom:

Tenure exists so that an examiner can fail a candidate, reject a result, or refuse a certification without personal consequence. It is therefore the certifying body that must inherit the collegiate form (i.e. peer judgement, protected tenure, academic freedom) while the instructional arm can be run on ordinary managerial lines.

I'm not convinced that the market will discipline universities to the extent that Potts believes. Diploma mills already coexist alongside credible institutions, demonstrating that there is demand for credentials of questionable value, although some employers may already recognise those differences. The more serious problem is that employers may not recognise gradual changes in assessment standards and the quality of certification. Universities might slowly drift towards satisfying student demand for credentialling, rather than learning, while continuing to benefit from reputations built up over many years. By the time employers recognise that the qualifications no longer signal what they once did, their value may already be substantially less than it was.

Complementary to credentialling, Potts notes that the decline in junior positions at many employers opens an opportunity for universities. If employers are unwilling to invest in students to train them in what were previously junior roles, universities could step into that gap. In part, universities have already started moving in that direction - witness the rise of 'work-integrated learning', which is now a feature of many universities' offerings, and increasingly compulsory for students (including at Waikato). 

Potts notes that the challenge here will be to find a financing mechanism. However, to me there is an obvious solution. Traditionally, junior roles have bridged the gap between university study and the point at which employers see graduates as sufficiently experienced and skilled to be fully productive. Employers have effectively funded that transition through on-the-job training. If employers are increasingly unwilling to employ graduates in these junior roles, universities could provide that bridge instead, through postgraduate qualifications involving work-integrated learning, either with an employer or in a 'simulated' work environment. Employers may not be willing to pay for this (after all, they are increasingly unwilling to employ juniors), but students may, especially if it smooths their pathway into employment. Again, this might not be something that we like, but it may be the reality in the future.

What about research? Potts has some interesting things to say about that as well, in particular noting that the minimum efficient scale for research is declining and fast approaching one. What would previously have required a large lab team or research group, can increasingly be undertaken by a sufficiently motivated researcher working with agentic AI. Of course, Potts is overstating the case here, because there remains a lot of research that relies on expensive infrastructure, specialised equipment, labour-intensive fieldwork, or collaboration among researchers. However, Potts' view suggests a fundamentally different relationship between research-intensive academic staff and universities. Specifically, Potts proposes that universities:

...unbundle affiliation from employment, making the university more like a club... Rather than the binary of tenured staff or nothing, universities can offer graded affiliation over institutional standing, grant-holding capacity, facility access, and the right to publish under institutional imprimatur, priced as a service and subject to quality review. This converts a fixed cost into a revenue line, retains scholars the institution could not otherwise afford to employ, and extends the warrant into the market for independent research.

This may work for high-profile researchers who can maintain an ongoing flow of external research grant money. However, most researchers are not in that privileged position. Also, Potts ignores that research funding is under pressure, and his proposal would make researchers even more beholden to grant funders than they are now. And it would create an even starker division among academic staff, between those who are highly-research active, bringing in substantial external research funding, and maintaining independence from the institution they are currently affiliated with, and those who are reliant on a role that simply certifies students. This problem is going to affect disciplines to different extents. For example, in a research environment like New Zealand at the moment, where funding changes have severely undercut research funding in humanities and the social sciences, this would be particularly problematic.

Potts also notes that the type of research that universities should specialise in should change:

AI has commoditized knowledge and particularly its search and discovery, including in domains that were previously high in expertise and difficulty such as literature review, statistical analysis and theorem proving. The implication is that the comparative advantage of universities has moved further afield, into the harder to conceptualize and reach corners of new ideas and possibility.

However, that proposal really relies on the willingness of research funders to fund frontier research. That might work in the US, but New Zealand's current research funding environment is increasingly emphasising economic benefits rather than fundamental research. That may make it more difficult to obtain funding for speculative frontier research, where economic benefits are uncertain.

Potts' working paper has a lot of other proposals for universities. However, I think the most important ideas are those I have outlined above. Potts is clear that not all universities are equally affected by generative AI. In particular, he notes that the elite universities are somewhat insulated from changes by their reputations. However, the vast majority of universities are not elite institutions. Potts' vision of universities is radically different from the status quo. We might buy into his vision, or not. But regardless, the economics of higher education are changing, and universities will need to adapt. Whether that requires moving away from the traditional model that combines teaching, credentialling, and research, remains to be seen.

Read more:

• Some notes on generative AI in higher education
• Some notes on generative AI and assessment (in higher education)
• More notes on generative AI and teaching (in higher education)
• Universities' (and teachers') cheap talk on generative AI and assessment
• Simas Kucinskas on AI, university education, and the 'mushy middle'
• The wicked problem of generative AI and assessment
• Matthew Kahn on generative AI and the economics major
• David Deming on generative AI and commitment to learning, and the impact of generative AI on signalling in education

The declining returns to higher education in Australia

In yesterday's post, I discussed a recent meta-analysis on the returns to compulsory schooling. That research was necessarily focused on secondary education. What about the returns to higher education? Two recent articles tell an interesting story about the change over time in returns to higher education in Australia.

The first article is this one by Elisa Birch and Alison Preston (both University of Western Australia), published in the journal Economic Record (open access). Using data from the Household, Income and Labour Dynamics in Australia (HILDA) survey, they document a striking decrease in the returns to higher education over the period from 2001 to 2023. They apply a fairly standard Mincer wage equation, comparing hourly earnings between workers aged 28 to 38 years with no post-secondary qualification to those with one of seven categories of education: PhD; Master’s; Graduate Diploma; Graduate Certificate; Honours; Bachelor’s; and Diploma/Certificate. Strictly speaking, these estimates are wage premiums associated with different qualifications, rather than necessarily causal estimates of the returns to obtaining those qualifications. Nevertheless, they do show some interesting trends, and their headline results are summarised in Figure 2 from the paper:

Notice that, in all cases (except, arguably, Graduate Certificates), the general trend in the returns to higher education has been downward. The wage premium between higher education and no post-secondary education is lower in 2019-2023 than it was in 2001-2005. Birch and Preston also look in more detail at the difference in returns between the early period in their data (from 2001-2011) and the later period (from 2011-2023), and the differences by gender, and find that:

...over the study period, the wage premium for a Master’s degree declined by 13.2 per cent for males. For females, the decline in the Master’s wage premium was even more pronounced, falling by 16.6 per cent... The reduction in the return to a Bachelor’s degree was similar for both males and females, with a decrease of approximately 13 per cent in both groups. In Period 2, the wage premium on a Master’s degree was equal to 35 per cent among men and 26 per cent among women (a gender gap of 9 percentage points). Among Bachelor’s degree holders the wage premium in 2012–23 was equal to 27 per cent among men and 22 per cent among women (a gender gap of 5 percentage points).

So, there isn't even anything positive to say about the changes in relation to the gender wage gap. The higher wage premium for male graduates remained the same, or increased, compared to the wage premium for female graduates. Even worse, when looking across the whole wage distribution, Birch and Preston find that:

While the male wage structure appears to have shifted downward and flattened, the female wage structure shows both a downward shift and a notable decline in returns at higher wage levels.

So, the returns to higher education have been declining in Australia (a result that other studies have also found), particularly at the upper end of the wage distribution for women. Birch and Preston largely leave open the question of why. That is where this new article by Michael Coelli and Jeff Borland (both University of Melbourne), published in the journal Australian Economic Review (open access), comes in.

Coelli and Borland first note that the decline in the returns to higher education is concentrated in the period after 2001, using five-yearly Australian Census data from 1981 to 2021, and using income as a measure (rather than hourly earnings, which isn't available in the Census). Then, they explore different explanations for why the returns to higher education changed, focusing on explanations that might explain a decrease starting around 2001. They find little support for a slowdown in skill-biased technical change relative to increases in the supply of skilled workers. Instead, they find that the decrease is associated with a relative increase in the wages of workers with no post-secondary education since 2001, and link that to two main probable causes.

First, the Federal Minimum Wage began increasing in real terms from 1996. On top of that:

The method of setting wages for employees covered by awards during 1993 to 2010 may also have raised the relative wages of low‐skill employees. During that time, the annual increases to rates of pay for employees covered by awards were made in flat dollar amounts, resulting in higher percentage growth in wages for lower wage earners. This may have underpinned stronger wage growth in low‐skill occupations... Award rate increases were above inflation at the lower end of the wage distribution (especially at the minimum wage), but below inflation further up the distribution. Since 2011, a common percentage increase has been set for employees covered by all awards.

So, the process of wages being determined by awards, alongside an increasing real minimum wage, may have compressed the wage distribution, especially before 2011. This would have the effect of raising wages for workers at the bottom of the wage distribution, including many workers without post-secondary education, relative to workers at the top of the wage distribution, including many workers with higher education.

Second, Coelli and Borland note that:

Starting around 2001, Australia experienced a large and extended boom in mining, coinciding with the rise of China in global trade after its accession to the WTO. This affected employment in mining and in construction. Its impact was felt more strongly among low‐skilled workers than among the higher‐skilled.

Since workers with no postsecondary education are more likely to be working in mining and construction than workers with higher education, the former workers likely benefited disproportionately from the mining boom.

Alongside those two explanations, Coelli and Borland also note a phenomenon that they call 'occupational downgrading':

Growth in the supply of workers with a bachelor's degree generally exceeded demand growth throughout the 1981 to 2021 period, leading to them moving down the jobs ladder.

These three explanations (minimum wages and wage awards; the mining boom; and occupational downgrading) are Coelli and Borland's explanations for the decrease in the returns to higher education. They do leave some questions from Birch and Preston unanswered though - why has the decrease in the returns to higher education been concentrated among the top of the wage distribution for female workers? The top of the wage distribution is least likely to be affected by minimum wages and wage awards. However, that leaves the other explanations for the decrease. Is it because female workers are more likely to 'occupationally downgrade'? Or, did female workers miss out on the mining boom? Clearly, there is more research to be done here.

Read more:

• Internal rates of return for college majors in the US
• The returns to each additional year of compulsory schooling

The returns to each additional year of compulsory schooling

This week, my ECONS102 class is covering the economics of education. Part of that topic considers the private education decision - the decision each individual makes about the amount of education they receive. This depends on the private costs and benefits of education. Arguably the primary benefit of education is incremental income - the additional lifetime earnings that education brings.

How big are those gains? Previous studies I have seen (such as a study reported in Paul Oyer's book Everything I Ever Needed to Know about Economics I Learned from Online Dating (which I reviewed here)) have estimated effects in the order of 10 percent per additional year of education. However, that was a single study, which could easily provide a biased view if considered in isolation.

What are the effects more generally? This recent article by Gregory Clark and Christian Nielsen (both University of Southern Denmark), published in the journal Kyklos (open access), provides an answer based on a meta-analysis of 79 causal estimates of the effect of an additional year or years of compulsory schooling on earnings, drawn from 53 other papers that apply methods consistent with causal inference. Meta-analysis provides a method of combining the estimates across many studies into one overall estimate. Importantly, Clark and Nielsen apply various methods to correct for the effects of publication bias (which is the tendency for statistically significant results to be more likely to be published, while statistically insignificant results tend to be missing from the research record).

Clark and Nielsen report baseline results that are somewhat lower than I expected:

With these procedures the average percentage gain in earnings from an additional year of schooling was 8.2% for 79 independent estimates, from 53 papers... If these estimates are weighed by their precision in a random effects estimation, the gains are reduced to 6.0%, and similarly 6.2% for a fixed effects weighting.

Given that the true effect is likely to vary across populations and study settings, I would consider the random-effects estimate of 6.0 percent for each additional year of schooling as the more plausible of the three. However, Clark and Nielsen also employ meta-regression, using PET (Precision Effect Test) and PEESE (Precision Effect Estimate with Standard Error) regression models, which are designed to address imprecision and publication bias in the overall sample of results. After removing some outliers, these approaches suggest:

...a true effect of 6.1%–6.4% to an extra year of education and an insignificant effect of publication selection...

So again, it seems like the effect may be in the order of six percent per additional year of schooling. Interestingly, Clark and Nielsen themselves note that the PET-PEESE evidence for publication bias is highly model-dependent, and that these results could instead reflect genuine heterogeneity in the true effect. However, Clark and Nielsen go a little further, looking in more detail at the distribution of effects across the 79 estimates. They note that if the estimates of the earnings return were normally distributed, then there are a lot of 'missing' studies with negative returns. They also consider whether the underlying returns might be log-normally distributed. Once you consider sampling error, the log-normal distribution can still generate negative estimates, and Clark and Nielsen argue that there are fewer negative estimates in the published literature than the model would predict. Consider Figure 10 from the paper:

In the figure, the yellow bars show a histogram of the 79 estimates in the sample. The blue line shows a log-normal distribution based on the estimates. Notice that, compared with the log-normal distribution, there are more studies than expected with estimates in the 0-4 percent range, and too few in the negative range. Clark and Nielsen argue that, if the 'missing' studies were included, a better estimate of the returns to each additional year of schooling may be in the range of 0-3 percent.

However, I’m not convinced that Clark and Nielsen’s distributional assumptions are justified here. Figure 10 (like Figure 8 in the paper) shows that the observed distribution of estimates doesn’t match their chosen distribution. But they haven’t established that their chosen distribution is what we should expect in the absence of publication bias. That distribution would depend on differences in estimation methods, the precision of the estimates, and genuine differences in returns across the populations studied. A mismatch between their chosen distribution and the observed study estimates therefore does not, by itself, demonstrate that studies with negative estimates are missing. So, while publication bias remains a possible explanation, I would treat Clark and Nielsen’s proposed 0-3 percent range for the average return very cautiously.

Nevertheless, overall it does seem clear from their other results that the returns to an additional year of schooling may be somewhat lower than the 10 percent found in other studies. What would be interesting to know next is whether the estimated returns have been changing over time (which my next blog post will look at, for Australia), and what study features lead to different estimates. Clark and Nielsen do note that instrumental variable regression leads to higher estimates than difference-in-differences or regression discontinuity estimates. However, there are other contextual variables that may matter as well. Hopefully, someone else can pick up on this work and take it to the next step. That might help to distinguish publication bias from the more mundane possibility that there simply isn't a single 'true' return to an additional year of schooling.

Read more:

• Internal rates of return for college majors in the US

Can a simple email get students to study more economics?

This year, I started sending the top students in my ECONS101 class a 'student recognition letter', congratulating them on their performance and noting their overall grade and their ranking in the class. The purpose of this letter was not to try and sell these students on studying economics, but to provide them with some well-earned recognition, as well as a valuable signal for future employers that their A+ grade actually meant something (whereas in some papers, it clearly doesn't mean much).

I could have added a sentence or two encouraging those students to study more economics. They certainly have shown an aptitude for it. However, some years ago I did try to encourage the top students to study more economics, but it seemed like those efforts didn't have much impact. I stopped doing this in about 2018. And it seems I might have made a good choice, at least according to this forthcoming article in the journal Economics of Education Review (open access) by Olivia Edwards and Jonathan Meer (both Texas A&M University).

Edwards and Meer study the effect of a simple email from the professor to the top ten percent (or so) of the class in introductory economics at Texas A&M University:

...praising their performance and encouraging them to take more economics courses and considering majoring or minoring in economics if they were not already doing so.

Their analysis is based on 10,600 students who took the course between 2017 and 2023, of which 1802 received the encouragement email. Edwards and Meer apply a regression discontinuity design (RDD) approach to the analysis, essentially comparing students just above the cutoff mark for receiving the encouragement email with those just below the cutoff mark. The outcome variables Edwards and Meer look at are whether the student went on to study intermediate microeconomics, whether they majored or minored in economics, and whether they majored in an 'economics-adjacent field' (by which they mean agricultural economics or business). They also test whether the results differ by gender, under-represented minority (URM) status, and whether the student was first-in-family to study at university.

There did seem to be some effect on students going on to study intermediate microeconomics. For students around the cutoff point for receiving the encouragement email, there was:

...a relatively large and statistically significant discontinuity of 8.8 percentage points (s.e. = 3.3 percentage points), an increase of about 40 percent over the baseline level below the cutoff.

However, there was little evidence of an effect on majoring or minoring in economics, with both effects being statistically insignificant. Turning to whether there were heterogeneous effects based on demographic characteristics, the effect of the email on taking intermediate microeconomics did not differ significantly by gender, but the overall positive effect appears to have been concentrated among URM students and students who were first-in-family to attend university. The encouragement email had no effect on going onto an economics major or minor for any of the demographic groups.

Edwards and Meer describe the effect of the encouragement email as "modest". I would say it was disappointing, but not surprising. If Texas A&M is anything like Waikato, then once a student has chosen a particular major (or minor), it is actually quite difficult to induce them to switch, even if they later find that some other major or minor would be a better fit for them. In part, this is due to institutional barriers, such as the difficulty in navigating the change-of-enrolment process, or the necessity to take pre-requisites that students might have missed, or that the student may have 'spent' papers on a particular major that would be lost if they switch. It could also be that students simply don't like changing their minds (a 'status quo bias').

So, while a simple information intervention may not be enough to change students’ choice of major, it does seem capable of nudging some students into taking more economics. And on the plus side, sending an email doesn't cost the professor much, so even if it only encourages some additional enrolments in intermediate microeconomics, on a cost-benefit basis for the Economics Department, it is probably a net positive. Given these results, perhaps I was too hasty in stopping the explicit encouragement of my top students to study more economics? I may have to re-draft my student recognition letters for this trimester.

[HT: Tim Harford, last year]

Read more:

• Could closing the gender gap in economics be as simple as providing students with information?
• Disappointing results on simple information interventions to close the gender gap in economics

Taking advantage of loss aversion in education

Many years ago (I forget exactly when), I introduced extra credit into my ECON110 class (which is what is now ECONS102). The idea was to provide an incentive for students to attend class, since they could earn extra credit for completing various in-class exercises. A couple of years later, I briefly changed the way that I framed the extra credit, from being "extra marks that would be gained from attending", to "extra marks that would be lost by not attending".

If students were purely rational, the change from 'gain framing' to 'loss framing' the extra credit should have had no impact on student attendance. However, I was looking to exploit the fact that most people are quasi-rational, rather than purely rational. Quasi-rational decision-makers are loss averse, meaning that they value losses more than equivalent gains. For a loss averse person, losing $20 makes them unhappy to a greater extent than winning $20 makes them happy.

Does a change from 'gain framing' to 'loss framing' work? That is the question that this new article by Antal Ertl, Éva Holb (both Eötvös Lóránd Science University), and Barna Bakó (Corvinus University of Budapest), published in the Journal of Economic Behavior and Organization (open access), tries to answer. They use data from a field experiment at Corvinus University of Budapest, where students enrolled in a compulsory macroeconomics course for business students were randomised into one of three conditions: (1) Gain group, which earned points in each of four tests and the final examination as usual; (2) Loss group, which started each test and the final exam with full points, but had points deducted for each incorrect answer; and (3) Hybrid group, which was the same as the Gain group for the tests, but switched to the loss framing for the final examination.

Ertl et al. have a sample of 321 students who consented to be part of the research, completed an initial questionnaire at the start of the term, and earned a non-zero grade. Randomisation was conducted at the level of the tutorial group (so all students in a tutorial were in the same treatment), in such a way that each teacher had groups across more than one treatment. One wrinkle in their analysis is that the best three out of the four tests would count towards a student's grade, meaning that students may end up putting differential effort into each test, depending on how they have performed in the other tests already completed. So, in addition to looking at the effect of treatment on each test mark individually, Ertl et al. look at the effect on the 'best three' tests collectively, as well as the exam mark.

If randomisation were perfect and the treatment groups were balanced, the comparison between the Loss group and the Gain group would demonstrate the overall effect of loss framing on student performance. The comparison between the Loss group and the Hybrid group for the final exam, compared with the same comparison for the best three tests, would demonstrate whether students adjust in such a way that the loss framing has less impact over time (because the Hybrid group would be in their first loss-framed assessment, while the Loss group would be in their fifth such assessment). The treatment groups weren't perfectly balanced, with students sorting into tutorial groups in part based on whether they worked part-time. So, Ertl et al. control for working part-time, the tutorial day and time, and the tutorial group teacher, as well as other demographic and background variables.

In their main analysis, they find support for the positive effects of loss framing:

For the Loss treatment, the effect on the average of the Best 3 Tests is 3.2 percentage points, although the difference is not statistically significant. The treatment effect on the Final Test score, however, shows a large difference of 9.6 percentage points when not controlling for Best 3 Tests’ scores, i.e., how well students did throughout the semester before the Final Test.

After controlling for performance in the best three tests, the effect of the loss framing on performance in the final examination is a statistically significant 7.8 percentage points. Turning to the comparison of the Loss and Hybrid groups, Ertl et al. find that:

...the estimated effect sizes for Loss and Hybrid are essentially the same for the Final Test, once we take into account how well students did perform throughout the semester...

These results are consistent with loss framing leading to better student performance, and there being no novelty effect - the effect of loss framing doesn't appear to decline over time. Ertl et al. go on to show that the effects are similar for both male and female students, but larger for students who did not take advanced mathematics in high school than for those that did. They also show that the treatment did not seem to negatively affect students' perceptions of the course, because the teaching evaluations were similar for the different treatment groups.

Finally, Ertl et al. do provide a note of caution in their conclusion:

previous studies have highlighted possible psychological and motivational costs associated with loss framing... These findings suggest that the mechanism by which loss framing improves performance may, at least in part, operate through heightened tension and concern about avoiding mistakes rather than through enhanced intrinsic motivation. Moreover, in extreme cases, loss-framed grading may even produce adverse effects — for example, low-performing students might become discouraged early in the semester after ‘‘losing’’ too many points. Once it becomes apparent that only a passing grade is attainable at best, the loss-framed structure may make this limitation increasingly salient, potentially exacerbating anxiety and disengagement. Over time, this could have broader implications for students’ well-being and their willingness to enroll in courses or programs that employ such systems.

Ertl et al. don't directly test for these effects, but they should be a concern. We may be able to improve student performance through loss-framing assessments, but that might come at a cost to student mental health and wellbeing.

And that brings me back to the example I started with, from my ECON110 class. When I switched extra credit from gain-framed to loss-framed, student attendance in class did improve slightly. However, the bigger impact seemed to be the number of students who would contact me by email, seeking special consideration for missing the extra credit, offering to provide medical certificates or other evidence to explain their absence, and asking for extra chances to complete the in-class exercises. It turned out to be administratively much more costly for me, and so the change was short-lived (to the extent that I cannot even remember which year I tried this in). Those reactions could suggest a negative psychological effect of the switch from gain framing to loss framing.

So, not all interventions that are effective for promoting student performance should be adopted. We need to carefully consider both the benefits and the costs of the intervention first. Taking advantage of student loss aversion might be worth exploring further, but I would want to see a wider evaluation that included student wellbeing outcomes before adopting it.

The impact of using the CORE textbook in Uruguay

We introduced the CORE textbook The Economy at the University of Waikato when we recoded the compulsory economics paper in our management degree from ECON100 to ECONS101 (see here). We were early adopters, as the CORE textbook was only released in 2017. It was a big change, and largely a positive one. The CORE textbook was free, substantially lowering the cost for students to access an important learning resource. Because the textbook was online, it could be constantly updated. And I really liked the way that it turned the traditional approach to the teaching of microeconomics on its head. Instead of starting with perfect competition and the supply and demand model, and then teaching imperfect competition as an exception, the CORE textbook started with monopolistic competition (where firms sell products that are differentiated from those of their competitors) and teaches perfect competition as an exception. Since many firms operate in monopolistically competitive markets, the approach that CORE adopted seems more attuned to the real world that students see.

I've often wondered whether the CORE textbook improved students' learning though. So, I was interested to read this recent article by Federico Araya (Universidad de la República, Uruguay) and co-authors, published in the journal Economica (sorry, I don't see an ungated version online [*]). They evaluate the impact of adopting the CORE textbook for the introductory economics course at the Faculty of Economic Sciences and Administration (FCEA) at the Universidad de la República, the largest university in Uruguay.

Although the CORE textbook was introduced at FCEA in 2020, Araya et al. start their analysis from 2021, to avoid the impacts on online teaching during the pandemic. FCEA offered two introductory microeconomics courses, one of which used CORE and the other continued to use their traditional textbook. Students were randomly assigned to either course based on the last number of their identification document, with 30 percent of students assigned to the course that used CORE.

However, the textbook was not the only difference between the two courses. As Araya et al. explain:

Although attendance is optional in both courses, in 2022 and 2023, the CORE course introduced a modification to its evaluation system, assigning 10% of the total grade to group activities conducted during class sessions. This change may have created an incentive for higher attendance.

So, their evaluation is not a clean comparison of the same course taught with two different textbooks, but will compare two different pedagogies, one which uses the CORE textbook and in-class group activities that are worth grade points, and one that uses a traditional textbook without the in-class group activities.

Araya et al. then compare the two groups in terms of whether students passed the introductory microeconomics course, as well as whether students passed an introductory calculus course and whether they passed the intermediate microeconomics course that follows on from the introductory course, while controlling for a range of demographic and socioeconomic variables for each student. They find:

...no statistically significant differences in pass rates between CORE and the conventional course, with the exception of the 2022 cohort.

I was initially surprised that they decided to evaluate each cohort separately, rather than pooling them. However, the 2021 cohort is different because that year the CORE course didn't have the in-class group activities, whereas it did for the 2022 and 2023 cohorts. Combining the 2022 and 2023 results would give us a better sense of the overall effect (of the combined CORE textbook plus group activities intervention). Instead, we are shown a statistically significant positive effect in 2022, but no statistically significant effect in 2023. That doesn't tell us whether the effects in those two years were actually different from each other, or whether the combined intervention had a positive overall effect across the two cohorts. One further problem here that muddies the comparison is that the CORE and traditional courses didn't use the same assessment, and so passing one course may be different from passing the other. And that might also explain the different cohort-specific results (if the 2022 traditional course had more difficult assessments than the CORE course, for example).

That won't be a problem for comparisons in terms of student performance in introductory calculus and intermediate microeconomics. For those courses, Araya et al. also find no statistically significant effects on passing.

So, at least there is no evidence from this study that the CORE textbook (with or without in-class group activities) made students worse off. Although equally, there is no evidence that it made them better off either. That allows me to raise an issue that is general to much of the similar research on educational interventions (including my own research on the impact of AI tutors in ECONS101). We might expect to see no significant effect on student performance even from a successful intervention. That's because a successful intervention may make studying easier for students, freeing up time that they can then devote to other activities. That might be studying for their other courses (although notice that in this case any reallocation of study effort doesn't appear to have affected the probability of passing introductory calculus), or something entirely different (maybe working more, or having more leisure time). So, I'm not surprised to see no effect of CORE on student performance in this study.

We continue to use the CORE textbook in my ECONS101 class (although last year we moved to the new edition, The Economy 2.0). We don't follow the text very closely, at least not in the microeconomics section of the paper that I teach. Nevertheless, it continues to provide the base material for a lot of what we teach. And it's good to know that at least one study says that there is not evidence that continuing to use a 'non-traditional' text is doing harm to students.

*****

[*] It's kind of ironic that a paper evaluating the impact of an open-access teaching resource is not itself published open-access.

Read more:

• Goodbye to ECON100, hello to ECONS101
• Differing perspectives on introductory economics textbooks
• Does the CORE textbook not go far enough?

Can a simple email nudge students to seek study help?

Students don't always access the help that they need. Often, it's because they don't realise what support is available to them. I'm interested in any way that we can make it easier for students to access support that enhances their learning and helps them to succeed. So, I was really interested to read this 2020 article by Rita Balaban and Patrick Conway (both University of North Carolina at Chapel Hill), published in the AEA Papers and Proceedings (ungated earlier version here).

Balaban and Conway evaluate the impact of emails (passive nudges) reminding students about two economics support programmes that were available on the UNC Chapel Hill campus. The two programmes were the 'EconAid Center', which provided walk-in peer tutoring during weekdays and early evenings, and a series of five hour-long learning strategy seminars run by the co-authors. Students were randomly assigned to one of two email groups, one of which received weekly reminders about the EconAid Center, while the other received reminders before each learning strategy seminar. The evaluation focuses on the 199 students who consented to participate.

Those emails represent a very light-touch intervention. But did they work? The nudges did get students to engage with the programmes, but only for female students and first-year students, who were more likely to engage with the EconAid Center if they received the email nudges. Female students and students of colour were actually less likely to attend the learning strategy seminars if they received the email nudges. So, at best a mixed result.

Turning to whether the two programmes worked to improve grades:

The coefficients corresponding to the effects of the support programs on performance are positive but insignificantly different from zero.

Now, the small sample of 199 students (and the much smaller samples for subgroups) may simply have been too small to estimate the effects precisely. Participation in the programmes was also relatively low. A larger study might therefore do a better job of telling us whether the programmes actually had an effect or not.

So, file this study away under possibilities for the future. Light-touch nudges may help some students take the first step towards seeking support. A natural extension today would be to test whether similar nudges can steer students towards effective human or AI-assisted study support.

Is it worth starting a Division III college football programme?

College football starts towards the end of August. The big and successful college football programmes attract millions of dollars and hundreds, if not thousands, of additional student enrolments, as well as keeping alumni engaged. It's not just college students who care about the result of a Michigan vs. Michigan State matchup!

But does it pay off for colleges further down the NCAA ladder to have football programmes? In the NCAA divisional system, Division I contains the powerhouse athletic programmes, Division II consists mainly of smaller public and private schools, and Division III is reserved for schools that don't offer athletic scholarships. Is it worthwhile for those Division III schools to have a football programme?

That is the question addressed in this 2025 article by Bryan McCannon (Illinois Wesleyan University), published in the journal Economics of Education Review (ungated earlier version here). McCannon starts by noting the growth in the number of schools playing Division III football, as shown in Figure 1 from the paper:

I was surprised that so many of these schools have added football programmes over the last thirty years, but note that the increase has levelled off since the mid-2010s. McCannon looks at data from 1984 to 2021, for all schools that had a Division III programme in 2022. He is interested in whether there are changes in student enrolment, gender balance, and endowment. He employs a two-way fixed effects (TWFE) approach, which essentially compares changes at schools before and after they introduced Division III football with changes at schools that did not introduce it. Conventional TWFE estimates can be biased when schools adopt football at different times and its effects vary across schools or over time, so McCannon applies a correction for that problem and also uses a synthetic difference-in-differences approach.

It turns out that both approaches produce similar results, and those results are not favourable, and McCannon reports that:

I fail to provide evidence that the adoption of Division III football has any effect on undergraduate enrollment. The change is statistically indistinguishable from zero. In addition, I provide evidence that the proportion of the student body that is female reduces. Taken together, this suggests that any increases in the male student population attributed to the introduction of football is offset by either reduced demand from female students or changes in the institution’s admissions practices.

And in terms of endowments:

Schools which added college football were overdrawing their endowments prior to adoption... I fail to find evidence that the addition of college football reverses these downward slides. This suggests that it did not sufficiently energize alumni giving or reduce financial pressure on the institutions.

So, adding a Division III football programme appears to offer little measurable reward for a college or university, at least in terms of the outcomes McCannon looked at. Why then would these schools start football programmes? One possibility is that they are caught in a competitive arms race, which is a type of prisoners' dilemma (which I covered in my ECONS101 class this week). A Division III college may believe that introducing football will attract students away from rival institutions, or prevent it from losing students when other Division III colleges introduce football. But if every Division III college introduces a football programme, none of them gains relative to the others, but they all incur the cost of running a football programme. The result is that many Division III colleges introduce football programmes, only to leave them all worse off (or at least no better off, based on McCannon's results).

Alternatively, McCannon suggests in his conclusion that it may be attractive for these schools to add a football programme in times of financial distress, in order to attempt to reverse the decline. However, these results suggest that such efforts would be largely unsuccessful in reversing declining enrolments or financial pressure.

These Division III schools are typically small, and education-focused. Perhaps they should stick to their strengths, and leave the expensive football programmes to the larger schools?

Could student-run social media groups reduce university student dropout?

Universities are quite focused on student retention, and as I noted in this 2018 post, if we can identify at-risk students, perhaps we can help to find ways to ensure they succeed. However, around that time I had a summer research scholarship student looking into the reasons that students drop out, and it turned out that each student dropped out for quite different, and difficult to predict, reasons. I call this the Anna Karenina principle of dropout: 'all students who persist are alike; each student who drops out does so in their own way'.

What if there were a simpler way of reducing student dropout, that did not require universities to identify at-risk students in advance, but instead reduced the risk of dropout from the outset? That would seem to be an attractive proposition.

So, I was interested to read this 2021 article by Lucio Masserini (University of Pisa) and Matilde Bini (European University of Rome), published in the journal Socio-Economic Planning Sciences (ungated version here), which evaluates the impact of student-created social media groups, such as Facebook pages, on student dropout. Masserini and Bini use survey data from 1879 first-year students from a major university in Central Italy.

Why would joining social media groups reduce dropout? Masserini and Bini suggest that these groups may help students form social connections and feel greater 'belonging' within the university community, while also providing a way for students to share information about courses, assessments, and study materials.

The key challenge in the analysis is that students are not randomly assigned to join social media groups - they choose whether or not to do so. And students who join these groups may differ from those who don't in ways that would bias a simple comparison of the students who joined social media groups and those who didn't. For example, more engaged students, who are less likely to drop out, might also be more inclined to join university-related social media groups run by other students. Masserini and Bini deal with this using propensity score matching - which involves identifying 'control' students who didn't join a social media group but who are most similar to each 'treated' student who did join a social media group. Then, comparing their matched control and treated students deals with any observable differences between the students who did, and did not, join social media groups.

Masserini and Bini then report a range of results of the estimated impact of social media groups on dropout, based on different assumptions used to do the matching, and:

...with the exception of k=1 nearest-neighbour, all the estimates indicated that students joining groups or Facebook pages had, on average, a lower probability to dropout, compared with those who were not part of such groups. The results also showed that the extent of the difference between the treated and control groups was not negligible, as it varied from 0.081 to 0.113, depending on the matching algorithm.

So, the results suggest that joining student-run social media groups or Facebook pages reduces the probability of a student dropping out by between 8.1 and 11.3 percentage points. Now, I should note that I don't in general find propensity score matching to be terribly convincing as a way of dealing with selection bias. 

Now, I should note that I do not find propensity-score matching entirely convincing as a way of dealing with selection bias. Although matching can make the treatment and control groups similar on observed characteristics, there is still something that is different about the treated and control students that leads the treated students to choose to join social media groups and the control students to choose not to join. That something is an omitted variable in the propensity score matching approach, and it is unclear how big the omitted variable bias will be. If, for example, joiners are more motivated or feel more connected to university life, then some of the apparent effect of joining the group on the probability of dropping out may instead reflect those underlying differences. Masserini and Bini's results are robust across several matching methods and sensitivity checks, which is reassuring, but robustness checks cannot establish that there isn't some omitted variable bias in the matching.

Having said that, if we take these results at face value, then there may be some merit in having student-run social media groups that university students can join. We must bear in mind that these results come from a survey in 2016, and they may not have aged well. But social media groups still exist, and students still participate in them. It could be worth exploring whether these effects still hold, given that increasing student retention remains a key focus for universities.

Having said that, student-run social media groups would probably be a relatively inexpensive way for universities to reduce dropout. Now, these results come from students surveyed in 2016, and both social-media use and the university environment have changed considerably since then. Nevertheless, the basic idea remains plausible. Universities could support the creation of student-run groups and randomly encourage or 'nudge' some students to join, then compare their subsequent retention with that of students who were not encouraged. That experimental approach would provide more contemporary and causal evidence of whether the groups reduce dropout, rather than merely attracting students who were already less likely to drop out.

Read more:

• Predicting student success, and failure
• Big data, student dropouts, and failing fast

How career stereotypes shape students' choice of major

What jobs do accounting majors get? How about psychology majors? Or economics majors? If you answered, respectively, 'accountant', 'psychologist', and 'economist', you're probably far from alone. When most people think about particular fields of study, they have stereotypical jobs in mind, and they're far more likely to believe that majors get the stereotypical job than any other job. Even in the case of economics, where very few graduates will go into a job with the title 'economist' (many will go into a job with some sort of 'analyst' title, like a business analyst, market analyst, or financial analyst, etc.).

A new article by John Conlon (Ohio State University) and Dev Patel (Brown University), published in the Quarterly Journal of Economics (open access) demonstrates the extent of this stereotyping. They also show using a simple survey experiment that students' stereotypical views can be changed, affecting their choice of major.

The first part of the paper compares students’ beliefs about the careers associated with different majors with actual major-career combinations in the 2017-2019 American Community Survey. Conlon and Patel then use the CIRP Freshman Survey, covering more than nine million first-year students between 1976 and 2015, to compare students’ expected careers with the occupations subsequently observed among graduates from the same cohorts. In that, they find:

...large, systematic, and persistent differences between the careers that freshmen expect to attain and the actual occupations they go on to have... We see that twice as many students expect to become artists, counselors, and lawyers (about 5% each) than actually do (2%–3% each). Four times as many students expect to become writers and doctors (2.7% and 11.1%) than do (0.7% and 2.8%).

Interestingly, the occupations that students most overestimate themselves as having are those that are rare and representative of particular majors (like writers or artists), while those that are most underestimated are those that are common alternative occupations that many majors may later hold (like teaching or business). Conlon and Patel also show using implicit association tests that people:

...strongly and systematically associate majors with their representative careers: implicit associations are 0.30–0.36 standard deviations higher for representative major–career pairs than for nonrepresentative pairs...

This supports Conlon and Patel’s interpretation that stereotyping contributes to students’ exaggerated beliefs about the connection between majors and the corresponding representative careers. Conlon and Patel then show using a theoretical model that stereotyping can increase misallocation of labour. What that means is that, given the occupation in which a graduate eventually works, that person might have been better off studying a different field. They also present suggestive evidence that links greater stereotyping with job dissatisfaction and regrets about the chosen field of study.

The welfare loss, job dissatisfaction, and regret, then motivated a survey experiment where Conlon and Patel attempt to correct for the stereotyping. The survey experiment proceeded as follows, using students from Ohio State University:

The survey began by asking students the percent chance that they would graduate with the two majors they selected as being most likely to pursue (their “top-ranked” and “second-ranked” majors). It then asked their self and population beliefs about the likelihood of each career group conditional on these two majors. Students were randomly sorted into a control group and a treatment group. Those in the control arm answered questions about their classes so far that semester and how they had (or had not) contributed to their major and career plans...

In the treatment arm, information modules provided students with the actual distribution of careers conditional on each of their top two majors according to data from the [American Community Survey]. For each major, it told them several headline numbers about the frequency of the careers they had listed as their most likely jobs if they graduated with that major...

Conlon and Patel then looked at whether assigning a student to the treatment group, where they received accurate information about the chance their top-ranked majors would lead to particular careers, affected students' beliefs about their own chances of working in each major's 'representative career'. They found that the information partially (but not fully) corrects students' misbeliefs about the chance of attaining the representative career, and that:

...reducing students’ self beliefs about their chance of having their top-ranked major’s representative job by 10 percentage points decreases intentions toward that major by 0.11 standard deviations (about 3.5 percentage points, p < .05).

So, when accurate information lowers students' beliefs about their chances of obtaining the representative career associated with their top-ranked major, their intentions towards that major, and their enrolment in that major, decline. Interestingly though, correcting beliefs about a student’s second-ranked major can make that alternative more attractive. The response therefore depends on both the information that students receive and how much they value the different careers.

That would seem to be good news for some university majors, where the 'representative occupations' are relatively common (e.g. accounting or marketing) and bad news for other majors where the 'representative occupations' are less common (e.g. journalism or film studies).

Where does economics fit in as a major? Sadly, Conlon and Patel don't answer that question, as they combine economics with accounting, finance, marketing, and other business fields. In that broad group, 73 percent of prospective majors expected to enter a business career, compared with 47 percent of graduates who actually did so. So economics may not escape stereotyping. It may simply be stereotyped as a route into 'business', rather than more narrowly as a route to a job called 'economist'. However, we don't know for sure from these results.

Would better information help prospective economics students to make better choices? On the one hand, it might deter students who see economics as a guaranteed path to one particular career, but it might attract others by showing that economics is not tied to a single occupation. Perhaps the problem for economics is not a lack of possible career destinations, but that its many potential destinations are less vivid than the single job title of 'economist'.

[HT: Marginal Revolution, back in 2022 when it was Conlon's job market paper]

Generative AI and grade inflation

If generative AI can produce assessed work that earns higher marks than a student would earn by themselves, widespread use of generative AI should increase measured grades even in the absence of equivalent learning gains. In other words, we might expect grade inflation to occur as a result of students using generative AI, and that grade inflation to not reflect improved learning. To what extent is this generative AI-induced grade inflation occurring? That is the question addressed by this recent working paper by Igor Chirikov (UC Berkeley). 

Specifically, Chirikov looks at the change in the grade distribution across 319 courses (and over 500,000 enrolments) at "a large, selective public research university in Texas", covering the period from 2018 to 2025. He follows the labour economics literature by measuring 'task exposure' to AI, using the share of required tasks in each course's Fall 2022 syllabus that involved writing or coding - areas in which generative AI is particularly capable and could be used as a substitute for students' own efforts. Using a difference-in-differences (DID) approach, Chirikov compares the difference in the share of A grades and GPA between years before and including 2022 (when ChatGPT was released) and more recent years up to 2025, between courses more or less exposed to generative AI. He finds that:

...grades rose substantially in high-exposure courses after 2022: the share of A grades increased by 13 percentage points (about 30% relative to the 2022 baseline) and GPA by 0.12 points, accompanied by compression of the grade distribution.

Looking at the shares of other grades, Chirikov finds that:

The share of A- grades fell by 4 percentage points, the share of B+ grades by 3 percentage points, and the shares of B and below show smaller and mostly insignificant changes.

So, courses that were more exposed to generative AI have seen a greater increase in the share of A grades and a greater decrease in the share of lower grades than courses less exposed to generative AI. Chirikov then turns to exploring the mechanism that underlies the change, by extending the DID approach to also compare courses that place greater or lesser assessment weight on homework tasks (in what is called a 'triple differences' approach). In that analysis, he finds that:

...above-median homework courses show an additional 16 percentage point increase in the share of A grades relative to below-median courses with the same level of AI exposure. The effect on GPA follows a similar pattern, with an additional 0.13 point increase in high-homework courses, though this estimate is less precisely estimated...

In above-median homework courses, the share of other grades declines significantly with AI exposure relative to below-median homework courses.

So, the observed pattern of grade inflation, with grades near the top of the distribution shifting towards As to a greater extent in courses that have greater homework weight, provides strong evidence consistent with generative AI contributing to the observed grade inflation, principally by substituting for student effort on unsupervised assessment tasks.

We should be cautious about over-interpreting the results from this study. It is based on results from a single university, and it measures course-level task exposure to generative AI, rather than students' actual use of generative AI. Nevertheless, the results are consistent with what we would expect, and are likely to hold in other contexts.

Given that, how should universities adapt to solve this issue? The obvious response is to move to more invigilated, in-person assessment. However, Chirikov cautions that:

Not all skills can be meaningfully evaluated under exam conditions: the ability to produce a well-researched essay, develop a software project, or conduct an empirical analysis requires sustained engagement that timed in-person assessments cannot capture. Restricting assessment to formats that are AI-proof risks measuring a narrower set of capabilities than the ones courses are designed to develop, potentially undermining the learning goals that graded work is meant to serve. A more promising direction is to redesign assessments so that AI use is either structurally constrained by the task or purposefully incorporated into it, for example, by requiring students to document their process, justify their choices, or demonstrate understanding through follow-up interaction.

In my view, the optimal response for universities is a combination of invigilated in-person assessment and authentic assessments in which AI use is permitted (or even encouraged) and evaluated. The appropriate balance depends on the specific learning outcomes for each course. However, as I have argued before, one approach is to scaffold students through their studies, with lower-level courses that rely more on developing core knowledge, relying less on generative AI and therefore using more secure assessment, while higher-level courses increasingly incorporate generative AI use explicitly into the assessment. This also requires scaffolding students through learning how best to apply generative AI at each level.

It almost goes without saying that we cannot simply continue to assess students as we always have done. Generative AI has broken key elements of the assessment toolkit that we previously used. This is not a new observation. However, evidence that generative AI may be accelerating grade inflation makes it even more imperative that assessment practices are updated to better reflect the availability of generative AI.

[HT: The Dangerous Economist]

Read more:

• Teaching evaluations and grade inflation
• Grade inflation is harming students' learning
• A few papers on grade inflation at universities
• Grade inflation and college completion in the U.S.
• More on teaching evaluations and grade inflation
• Grade inflation at New Zealand universities, and what can be done about it
• Who is morally responsible for grade inflation?

Is it better to have a more educated mayor?

It seems somewhat self-evident that having a more educated mayor would be better than having a less educated mayor. However, whether education is a positive attribute for a mayor really depends on whether, and to what extent, more educated mayors act differently than less educated mayors. Do they spend more, or less? How do they spend the public budget?

This new article by Alessio Mitra (University of Kent), published in the European Journal of Political Economy (ungated earlier version here) directly addresses the second question - how does mayoral education affect public finance? Mitra uses data from municipal elections in Italy over the period from 2000 to 2015, focusing on municipalities with a population of less than 15,000 (because larger municipalities use different electoral rules). He defines a more educated mayoral candidate as one with a university degree, and a less educated mayoral candidate as one without a degree.

Mitra applies a regression discontinuity design (RDD), which involves comparing municipalities that narrowly elected a more educated mayoral candidate over a less educated candidate with similar municipalities where the more educated candidate lost to the less educated candidate. In very close elections, the identity of the winner is plausibly as-good-as random, provided there is no manipulation around the threshold related to the education of the candidates. In other words, since the difference between getting 50.01 percent of the vote and getting 49.99 percent of the vote is essentially random, the education of the winning mayoral candidate is basically determined randomly in these close elections between candidates with different education levels. With that assumption in mind, observed differences between the municipalities where a more educated candidate won with those where they lost can be attributed to the difference in mayoral education.

Mitra's dataset includes more than 18,000 mayoral elections, of which 1211 have a margin of victory of less than five percent (which he defines as a close election, and includes in the analysis). He looks at the differences in public expenditure, initially focusing on changes in the share of spending devoted to operational expenses (or 'current expenditure' as he terms it) or public investment. In this, Mitra finds that:

When an educated mayor is elected by chance, public investment rises by 3 percentage points of total expenditure compared to a less educated counterpart.

Digging down into the allocation of that public investment, he finds that:

...educated mayors allocate an additional 1 percentage point of total expenditure to education investment, accounting for one-third of the overall increase in public investment.

And going a bit deeper than that:

Among education investments, immovable assets dedicated to nurseries receive the largest increase in resources.

Consistent with Italy’s balanced budget requirement on municipalities, there is no significant change in fiscal deficit. That means that the additional spending devoted to public investment must mean a corresponding reduction in operational expenditure. Mitra doesn't really dig into that at all.

What we take away from this paper is that more educated mayors devote more spending to education. In the Waikato Economics Discussion Group today, we discussed what mechanisms might underlie this difference, which is something that Mitra didn't explore. Perhaps more educated mayors see more value in education. After all, they invested more in their own education than a less educated mayor did. However, that's not entirely consistent with spending more on public investment in early childhood education.

A second possibility is that more educated mayors have a lower intrinsic discount rate, increasing their willingness to make long-term investments, both in their own education and in the education of their citizens. This is more consistent with devoting spending to public investment in early childhood education.

A third possibility is that more educated mayors may be better at the administration of public investment, such as project approvals, capital budgeting, grants, or procurement. This means that they have greater capacity for public investment projects. However, that greater capacity wouldn't necessarily be more apparent for public investment in education, or early childhood investment.

However, an intriguing but speculative fourth possibility is that more educated mayors understand that public investment can be used strategically to affect demographics. Many municipalities in Italy are facing extreme population ageing and/or declining populations. Mitigating (but probably not reversing) those population changes may be possible through creative policy. If the municipality invests in early childhood education, that may make the municipality more attractive for young parents to relocate to, and may reduce cost pressures that hamper fertility. The problem with this as an explanation is that it isn't clear that these trends and policy solution would be more apparent to a more educated mayor than to a less educated one.

The second possibility seems to me like the most promising. However, exploring the reasons why more educated mayors spend more on public investment, particularly in education, is a promising exercise for future research.

One last point is that the effects are actually quite modest. The total budget for a municipality of 15,000 population would be around €15-25 million per year (based in part on this and this, both in Italian, but see also here for public finance data for all Italian municipalities). A reallocation of three percentage points to public investment represents up to an additional €750,000 per year. And if one-third of that is spent on public investment in education, that is an additional €250,000 per year. It's not nothing, but it's certainly not building multiple new schools. Maybe it's an additional small school building per year.

So, is it better to have a more educated mayor? This research suggests yes, but that relies on a normative view that more spending on public investment, particularly in education, is overall a good thing. However, the size of the effect doesn't suggest transformational change, and we don't really know what the trade-offs are in terms of what categories of operational spending were reduced. A university degree does not necessarily make someone a better mayor, and this paper cannot tell us whether more educated mayors have better preferences, longer time horizons, or simply greater administrative capacity. What it does show is that who gets elected can change not just how much is spent, but what kind of future a municipality chooses to invest in.

Does the future of higher education look more like a mentoring pyramid scheme?

In response to my recent post about the future of higher education and one-on-one mentoring, one of my students from last year, Yunze, got in touch via email to offer a potential solution:

...I wonder whether it is possible to set a clear academic threshold within each discipline. If students who reach this threshold could mentor upper‑middle‑level students, while professors spend only a small amount of time supervising the overall direction, the system might become more sustainable. However, I suspect this could harm the interests of the top students, since they might otherwise use that time to further advance their own academic achievements, and If [sic] they fail to successfully train students with real research ability, it would likely damage both the university’s reputation and the professor’s own reputation.

You know, I think Yunze is right on the money here. Consider the problems I outlined in the earlier post: (1) the signalling value of education is falling due to generative AI; (2) a one-on-one mentoring approach may be a solution; but (3) one-on-one mentoring doesn't scale due to limited faculty time. If one-on-one mentoring is not conducted between faculty and students, but works more like a pyramid mentoring model, then this might actually work, not just for students, but for faculty and for universities as well.

So, let's think it through. But first, remember that the mentoring model I introduced in the earlier post is not simply a model of small classes, where senior students perform limited teaching roles, such as tutoring. This is a model of genuine mentoring, where the mentor encourages the mentee to become a builder, in the words of Auren Hoffman. A builder creates things, and it is the act of building, and the learning alongside that, which will be a durable signal to future employers. In relation to mentoring, I said in that post that mentors should do the following for their mentees:

Teach them to be builders. Encourage them to create things. Work with them and chart a path forward for their success.

If faculty provide one-on-one mentoring to a small number of senior students, then that makes better use of faculty time than them mentoring hundreds of first-year students. The senior students can then each mentor several second-year students, who in turn can then mentor several first-year students. [*] In this model, faculty time is targeted at the senior students, where the impact of faculty on student employment outcomes may be greatest.

Students benefit from helping junior colleagues to become builders, where the signalling value may remain even in the face of generative AI. Even better, mentoring provides student mentors with an opportunity to build - they may be able to point employers to the success of their mentees as an example of their building, talking also about what went wrong in the mentoring relationship, and what they learned from the experience.

In this mentoring pyramid model, universities retain a key role, but that role becomes very different. Universities essentially become a platform, connecting students with mentors - first-year students with second-year mentors, second-year students with senior student mentors, and senior students with faculty mentors. In the terms of my earlier post, the university runs their own OnlyStudents platform.

Of course, this platform role creates a new problem for universities. If mentoring works mainly as a way of matching students with mentors, then the market may not need eight OnlyStudents platforms in New Zealand, or thousands of OnlyStudents platforms worldwide. A small number of large platforms could have a big advantage in that case - more students attract more mentors, more mentors improve the quality of matching, and better matching attracts still more students. Those network effects could create a winner-take-all dynamic, in which universities would struggle to differentiate themselves simply by running their own mentoring platforms, and where a single surviving OnlyStudents platform might be the ultimate outcome. However, that conclusion depends on the strength of the network effects. If effective mentoring also depends on institutional trust, disciplinary reputation, local employer connections, pastoral care, or an in-person community, then universities may retain some defensible advantages. Geography alone probably won’t be enough, especially if online mentoring is close to being as effective as in-person mentoring, but local connections might still matter. So the question for universities is not just whether they can build their own OnlyStudents, but whether they can attach that platform to something that a larger, more generic OnlyStudents cannot easily replicate.

Universities may also retain a role in the initial and ongoing training of mentors. Since each student, and each faculty member, will need to be a mentor to one or more others lower down in the pyramid, they will need to understand how to mentor. That means universities would not simply be matching students with mentors. They would also need to train mentors, monitor the quality of mentoring relationships, and intervene when mentor-mentee relationships are not working well. Moreover, the adoption of a mentoring pyramid model is likely going to change who the most successful students (and faculty members) are. The top students do not necessarily make the best mentors (or the best tutors, as I have learnt across years of coordinating tutors in my first-year papers). Good mentoring requires a specific skill set, but it is those skills that may also demonstrate the quality of the student as a builder - a signal of high quality for employers.

A further point about the pyramid mentoring model is that it likely requires a strong filtering effect to be financially viable. Since each faculty member can only mentor a limited number of senior students, and each of those senior students can only mentor a limited number of second-year students, who in turn can only mentor a limited number of first-year students, each level of the pyramid probably needs to be somewhat wider than the levels above it. To achieve that, student progression needs a strong filter, limiting the number of students who progress from first-year to second-year, and from second-year to senior.

Let's consider some simple numerical examples that illustrate why filtering is needed. If each faculty member mentors five senior students, and each senior student mentors five second-year students, who each mentor five first-year students, then the pyramid contains 155 students per faculty member - five senior students, 25 second-year students, and 125 first-year students. A model where each faculty member’s salary is covered by fees from 155 students, setting aside any contribution to central university costs, seems likely to be financially viable to me. However, in this model only one-fifth of students could be allowed to progress each year. That means the model would also need some form of orderly exit for students who are filtered out - perhaps an exit qualification, or a pathway into a non-mentored track. The problem is that both options may provide negative signals about the student who is filtered out.

If all students were to progress, then that would require each student to mentor at most one student at the level below. Keeping five senior students mentored by faculty, then the pyramid would contain 15 students per faculty member - five senior students, five second-year students, and five first-year students. That system would be much less likely to cover the cost of faculty time. So, it's unlikely that the pyramid mentoring model would be viable to run without some form of filtering - perhaps not as extreme as only one-fifth of students progressing each year, but clearly not all students could progress every year.

So, to return to my conclusion from the previous post, the current mass higher education model still looks increasingly fragile, but perhaps one or a few universities might be able to navigate their way through. However, the survivors are likely to be first-movers or fast followers in developing a platform market strategy that leverages a pyramid mentoring model. This model is still going to cost students a lot, and the filtering effect would make higher education more elitist as well.

And thanks to Yunze for inspiring this post with his perceptive email comments.

*****

[*] For simplicity, I'm assuming a three-year higher education degree structure, as we have in New Zealand. For a four-year degree structure, you would of course need to add an additional level.

Read more:

• Why all students need to understand adverse selection and signalling
• Signalling theory suggests that we may need two categories of university degree
• Does the future of higher education look more like one-on-one mentoring?

Does the future of higher education look more like one-on-one mentoring?

It almost seems a cliche to say that generative AI is both the greatest threat, and the greatest opportunity, for higher education. That doesn't make the statement any less true. And there are many commentators who are trying to work out what happens next for higher education. I think one of the best is Hollis Robbins, whose Substack Anecdotal Value is well worth subscribing to.

Robbins was recently interviewed by Jay Caspian Kang of The New Yorker (paywalled, but you can find an ungated version here) about the future of higher education. The whole interview is worth reading, but I want to highlight this bit in particular, where Robbins says:

I was in Austin, Texas, a couple of times in March with a bunch of twenty-five-year-old billionaires. This is what they’re looking at. Instead of having the credential from the institution, why not have the credential from the professor? If you have a Hollis Robbins education, what would that signal? What would that credential mean as opposed to a degree from a university? There was some conversation about what that would look like, and one guy at the end of the dinner said, “Instead of OnlyFans, it’s like OnlyProfessors.”

The correct analogy here would be that it would be OnlyStudents, not OnlyProfessors (since the name contains the audience, not the performer). However, Robbins makes a good point. Higher education is, in part, an exercise in signalling (as I've noted before here and here). In fact, Bryan Caplan argued in his book The Case Against Education (which I reviewed here) that one third of the benefit of higher education is signalling (the other two-thirds is made up of genuine learning, socialisation, and transferable skills). 

The diploma that a student receives at the end of their higher education journey is a signal to employers of the student's quality as a future employee. The signal is credible to the employer because it is costly for the student to obtain, and costly in such a way that low-quality students wouldn't attempt the signal. University reputation matters here, because it is an assurance of the second of those conditions - low-quality students wouldn't attempt the signal of a Harvard degree, firstly because they wouldn't gain admission to Harvard in the first place. So, part of the signal comes from getting into Harvard. Second, low-quality students wouldn't attempt the signal of a Harvard degree because passing courses at Harvard is hard (or, at least, harder than at many other universities). The quality of the education at Harvard has traditionally been higher than elsewhere.

In the interview, Robbins makes a further important point, which is that we've spent the last several decades making higher education a commodity. Students studying a degree in a particular subject learn the same things, often using the same teaching materials, the same textbook, and the same style of assessment, regardless of which university they go to. That means that the quality of signal that arises from the education itself has reduced over time, meaning that most of the value of the signal arises from the admission process. Once a student has been admitted to Harvard, their signal is in place, and the further signalling from their education is lower than for comparable students in years past.

Robbins argues that generative AI is accelerating and expanding this commodification of education. Since generative AI has access, through its training corpus, to a large store of human knowledge, to add value over and above generative AI a professor must be a true expert in a very specific subfield. For students, learning a subject from a true expert still retains value, because generative AI cannot as easily replicate the learning that would occur from the expert. At that point, the university becomes less important as a mediator of education.

In other words, Robbins is arguing that the university could be disintermediated, with individual professors issuing credentials instead of universities. Who needs a Harvard degree, when you could have a Hollis Robbins degree? And since Robbins is an expert (in African American sonnet tradition, she says), the signal retains high value. However, that is true only to the extent that employers find the Hollis Robbins degree a credible signal.

That brings me to this Substack post by Auren Hoffman. Hoffman also argues that generative AI has diminished the value of the higher education signal. However, Hoffman argues for a different solution, this time from the perspective of the graduating student:

you have to show you can add value. that is it. that is the only thing.

the test the smart hiring manager applies in 2026 is simple. can you learn something on your own? can you finish what you started? can you do what you said you would do? these were always the skills that mattered. the difference is they are now the ONLY skills that matter, because the credential stopped doing the screening.

if you have a few years of experience, your resume can show you have these skills. if you are a new grad, you have to show what you have created and built.

Hoffman is overstating things a little, as university degrees are unlikely to disappear overnight. However, his critique is still important, as is the implication he draws. Hoffman argues that graduates (or young people, generally, since the solution doesn't depend on a student attending a university or completing a degree) need to become builders:

the most valuable thing a 22 year old can do in 2026 is create something. an app. a screenplay. a side business. an internal tool you wrote for a club you were in. a dinner series. a script that automates something annoying. a website. a chrome extension. a sculpture. a dance party. a discord bot. a substack with 32 readers and a real point of view. anything that moved from idea to working.

will the thing make money? probably not. that is not the point.

the point is that you taught yourself something. you finished it. you can describe what you learned, what broke, what you fixed, why you made the calls you made. that story is the new resume.

every hiring manager would rather interview a 22 year old with a launched app and a github full of weird side projects than a 22 year old with a 3.9 GPA from a top 50 school. it is not close. when one candidate has tangible evidence of what they can ship and the other has a transcript, the transcript will lose every time.

According to Hoffman, the best signal for young people to be sending in future is that they can build. Our students will ultimately be more successful if they can show off their skills (both technical and transferable skills) by building something. That is a signal that is costly, and costly in such a way that low-quality students will not attempt it. The signal is credible, and for the most part it retains currency even in the face of generative AI. Indeed, if the student builds something while effectively leveraging generative AI, then the signalling value to employers may be even greater. The key distinction is between the student using generative AI as a tool and the student using it as a substitute for doing the work. A student who can explain what they built, what broke, what they learned, and why they made the choices they made, is still sending a costly signal. A student who simply lets generative AI build for them is not. Employers would likely see through the latter pretty quickly.

Hoffman's idea isn't exactly new. When I think about some of my best students over the past two decades, they tend to be those that built something either while they were studying, or immediately after. For some, this was their own small business or entrepreneurial activity. For others, it was writing and publishing a research paper. Those were challenging tasks that set them apart from other students - a clear signal of quality. What Hoffman is essentially saying is that, with the signal from higher education itself being removed, the only remaining signal of quality is the signal from being a builder.

Where does that leave higher education staff? I think we can combine Robbins's and Hoffman's ideas, and chart a path forward. We don't need to start an OnlyStudents, and issue our own degrees, but we do need to cultivate closer relationships with our best students. Teach them to be builders. Encourage them to create things. Work with them and chart a path forward for their success. In other words, be a mentor.

Universities are absolutely going to hate this. Mentoring is not an activity that can be offered at scale. For example, there are simply not enough hours in the day for me to individually mentor all of the 350-plus students in my first-year economics class. Nor is mentoring easy to timetable, measure, standardise, or reward under current academic workload models. Universities are built around papers, credit points, learning outcomes, assessment rubrics, and student evaluations, all of which can be offered at scale. One-on-one mentoring doesn't work so well in that system. For example, postgraduate supervision already sits awkwardly within the system, with it being unclear whether it counts as teaching, or research. Nevertheless, mentoring may be one of the few ways that higher education can continue to offer something that is both valuable and difficult for generative AI to replicate.

If generative AI significantly reduces the signalling value of university education, and if students increasingly use it to avoid genuine learning, then the current mass higher education model looks increasingly fragile. Moreover, what remains is going to cost students a lot more. If having a high-quality mentor who can encourage a student to build is the path to their future employment, then it may be worth it. After all, high-quality signals are costly. That aspect of education, at least, won't have changed.

[HT: Marginal Revolution, for both the Hollis Robbins interview and the Auren Hoffman post]

Read more:

• Why all students need to understand adverse selection and signalling
• Signalling theory suggests that we may need two categories of university degree