AI and the research production function

I've been a bit quiet on the blog lately, due to travelling, and attending the North American Regional Science Conference in San Diego. Regional science is a multidisciplinary field that overlaps economics, geography, planning, and many other disciplines. There were many interesting sessions at the conference, although apparently not the session that I was presenting in, which had an audience of two (one of which was another speaker in the session, with the other three speakers in my session no-showing).

Anyway, one of the most interesting sessions was the day before the conference proper started, on "The Potential of AI for Regional Science", hosted by The Regional Science Academy. Speakers included many of the contemporary great minds in regional science such as Rick Church (University of California, Santa Barbara), Tomaz Dentinho (University of Azores), Peter Nijkamp (Open University), and John Östh (Oslo Metropolitan University), among others. They noted (with examples) the great opportunities that artificial intelligence offers, especially as an aid for research. For example, Östh demonstrated a really interesting application of AI to generating data on neighbourhoods, while Patricio Aroca (Universidad Andres Bello) showed how AI could help researchers whose first language is not English to improve their chances of publication in top-ranked English language journals.

I came away from this session in equal parts excited and concerned (which might be an appropriate mix of reactions to any sufficiently disruptive technology). There is great potential in using AI as a research tool, and I've already seen many examples (and blogged about some ideas here and here). However, my concern comes from how AI will change the research production function.

Before we get that far, let's go back to an earlier technological revolution that greatly changed how research in economics was conducted. This story is not mine, but I forget where I got it from. Consider a simple research production function in economics, with two inputs: (1) econometric modelling; and (2) economic explanations. As computers increased in computational power, the relative price of econometric modelling decreased. So, researchers reallocated their scarce research resources from economic explanations (which had become relatively more expensive as an input) towards econometric modelling (which had become relatively less expensive as an input). The result was a rapid increase in econometric modelling, and a corresponding decline in the quality of economic explanations to go along with the econometric modelling. It is not clear that this was a positive change for the overall quality of economics research (but there certainly was an increase in the quantity of research).

Now consider a similar research production function in economics, where the inputs are: (1) human intuition and explanations; and (2) artificial intelligence. New AI models have made AI radically less expensive to use. We can probably expect research in economics (and in other fields) to adjust to much more use of AI, and much less use of human intuition and explanations. It remains to be seen whether that leads to an increase or a decrease in the quality of research overall. My money is on an overall increase in the quantity of research, but a reduction in the average quality and an increase in the variance (with some very high quality research resulting from AI, as well as a lot of very low quality research).

Unlike the science fiction and fantasy magazine Clarkesworld, which shut off submissions earlier this year due to a flood of AI-generated stories were submitted, I'm not aware of any academic publishers that are feeling the strain, yet. It will be coming though, and may intersect with the increasing volume of predatory publishers (see here and here). As the Managing Editor of a journal myself (the Australasian Journal of Regional Studies), I'm really not looking forward to policing the submission of AI-generated articles of low quality.

Taken altogether, it is clear that there is a trade-off inherent in the impact of AI on research (in regional science, in economics, and in other fields). We will need to accept (and act to mitigate) the negative impacts, while endeavouring to maximise the good.

A simple production model of my breakfast

This morning, I had a vegetarian quiche for breakfast. It was delicious, but also surprising. The filling of the quiche was mostly mushroom and spinach, with very little egg. I love mushrooms. However, the defining feature of a quiche is egg, and normally the egg to mushroom ratio in a quiche would be much more in favour of egg, rather than mushroom. What led to this delicious and unexpected treat? It turns out that a very simple model of production, as covered in the first week of my ECONS101 class, can help explain.

That model is shown in the diagram below. It shows different combinations of egg and mushroom that can be used to make a vegetarian quiche, with mushroom measured on the y-axis and egg measured on the x-axis. Let's say that there are only two recipe options for the bakery, A and B [*]. Recipe A uses a lot of egg (EA) and not much mushroom (MA), while Recipe B uses not much egg (EB) and a lot of mushroom (MB).

Which recipe should the bakery use? If the bakery is trying to maximise profits, then given that both recipes produce the same quantity (one quiche), the bakery should choose the recipe that has the lowest cost. [**] We can represent the bakery's costs with iso-cost lines, which are lines that represent all the combinations of mushroom and egg that have the same total cost. The iso-cost line that is closest to the origin is the iso-cost line that has the lowest total cost. The slope of the iso-cost line is the relative price between mushroom and egg - it is equal to -Pe/Pm (where Pe is the price of eggs, and Pm is the price of mushrooms).

Now think about that relative price. Eggs are currently priced very high (see here and here), so the relative price (-Pe/Pm) is a large number (in absolute terms). That makes the iso-cost lines relatively steep, as shown in the diagram below by ICA and ICB. In this case, the iso-cost line that passes through B (ICB) is the lowest iso-cost line. It is closer to the origin than the iso-cost line that passes through A (ICA). So, Recipe B is the lowest cost recipe, and the bakery should choose Recipe B.

However, eggs are not usually as expensive as they are right now. In 'normal' times, the relative price (-Pe/Pm) would be a smaller number (in absolute terms), because the price of eggs would be lower. That would mean that the iso-cost lines would be relatively flatter, as shown in the diagram above by ICA' and ICB'. In that case, the iso-cost line that passes through A (ICA') is the lowest iso-cost line. It is closer to the origin than the iso-cost line that passes through B (ICB'). So, Recipe A is normally the lowest cost recipe, and the bakery should normally choose Recipe A.

So, that explains my surprising and delicious breakfast. Normally, if I got a vegetarian quiche, it would be mostly egg with a bit of mushroom (Recipe A). But, now that eggs are relatively more expensive, I got a quiche that was mostly mushroom with a bit of egg (Recipe B). Delicious!

*****

[*] In the standard version of this model, A and B would be referred to as production technologies. In this case, it makes more sense to refer to them as recipes.

[**] We're assuming here that the bakery wouldn't change its price, depending on which recipe they used. It is costly to change prices (economists refer to these as menu costs), so the bakery would prefer to keep prices the same, even if the prices of eggs and mushrooms are fluctuating from week to week.

Relative prices, incentives, and the move to automated telephone switching

This week in my ECONS101 class, we talked about the Industrial Revolution. Specifically, we used a relatively simple production model to show how changes in the relative price of coal and labour in England created incentives for firms to switch from using labour-intensive production technologies to using coal-intensive production technologies.

That model, and the associated explanation, can be used in a range of situations. For example, I previously used it in a post about taxing robots. However, in this post I want to talk about telephone switching. The Federal Reserve Bank of Richmond's Econ Focus had an article about telephone operators late last year:

Users of the telephone in the late 19th century and early 20th century couldn't dial their calls themselves. Instead, they picked up their handset and were greeted by an operator, almost always a woman, who asked for the desired phone number and placed the call. Technology to automate the process emerged quickly, however: The first automated telephone switching system — a replacement for human operators and their switchboards — came into use with much fanfare in La Porte, Ind., on Nov. 3, 1892, 16 years after Alexander Graham Bell's patent on the telephone.

Yet telephone companies continued relying on the women long afterward. In 1910, only around 300,000 telephone subscribers had automatic service — that is, service in which they dialed calls themselves rather than interacting with an operator — out of more than 11 million subscribers total. The companies of the Bell System did not install their first fully automated office until Dec. 10, 1921, and did not install an automated system in a large city until the following year, two decades after the technology had been demonstrated. Those telephone users who did have access to automated calling were customers of independent phone companies, mostly in small towns and rural areas.

This raises a couple of questions. First, why did the telephone companies switch from human operators to automated switching systems? Second, why did this change happen in small towns and rural areas before big cities? The simple production model can help us explain.

Let's start by setting up the simple production model. It is shown in the diagram below, with capital (e.g. automated switching) on the y-axis and labour on the x-axis. Let's say that there are only two production technologies available to telephone companies, A and B, and that both production technologies would produce the same quantity (and quality) of output (connected telephone calls) for the combination of inputs (capital and labour) shown on the diagram. Production technology A is a labour-intensive technology (human operators) - it uses a lot of labour, and supplements the labour with a bit of capital. Production technology B is a capital-intensive technology (automated telephone switching) - it uses a lot of capital, and a small amount of labour (for keeping the switching machines operating).

How should a telephone company choose between the two competing production technologies A and B? If the firm is trying to maximise profits, then given that both production technologies produce the same quantity and quality of output (connected telephone calls), the firm should choose the technology that is the lowest cost. We can represent the firm's costs with iso-cost lines, which are lines that represent all the combinations of labour and capital that have the same total cost. The iso-cost line that is closest to the origin is the iso-cost line that has the lowest total cost. The slope of the iso-cost line is the relative price between labour and capital - it is equal to -w/p (where w is the wage, and p is the cost of a 'unit' of capital).

First, consider the case where labour is relatively cheap and capital is relatively expensive. The iso-cost lines will be relatively flat, since w is small relative to p (so -w/p is a small number). In this case, the iso-cost line passing through A (ICA) is closer to the origin than the iso-cost line passing through B (ICB), as shown in the diagram below. So production technology A is the least-cost production technology, and firms should use the relatively labour-intensive production technology (human operators).

Now consider what happens if wages are higher, or the cost of capital is lower. The relative price between labour and capital (-w/p) would increase, and the iso-cost lines would get steeper. This is shown in the diagram below, where the iso-cost line passing through B (ICB') is now closer to the origin than the iso-cost line passing through A (ICA'). So, now production technology B is the least-cost production technology, and firms should use the relatively capital-intensive production technology (automated telephone switching).

So, we already know that telephone companies switched from A to B, and we have an explanation of why it might have happened. Is our model's explanation consistent with the facts? In terms of wages, most telephone operators were women. Going back to the article in Econ Focus:

At first, the telephone industry hired men and boys as operators. But the practice was short-lived. The first woman operator, Emma Nutt, was hired by a telephone service in Boston in 1878, and the hiring of women spread quickly. Women operators were viewed by the companies as more polite to customers, more patient, more reliable, and faster — not to mention cheaper.

That last point is relevant to our model - wages of telephone operators were low. What about the cost of capital? From the article:

The difference with automatic telephone switching was that the cost structure, perhaps surprisingly, favored the smaller firms with their smaller customer bases. With the electromechanical systems of the day, each additional customer was more, not less, expensive. Economies of scale weren't in the picture. To oversimplify somewhat, a network with eight customers needed eight times eight, or 64, interconnections; a network with nine needed 81.

"You were actually getting increasing unit costs as the scope of the network increased," says Mueller. "You didn't get entirely out of the telephone scaling problem until digital switching in the 1960s."

So, the cost of automated telephone switching was higher in larger markets (e.g. large urban areas) than in smaller markets (small towns or rural areas). So, it seems likely then that the relative price of labour to capital (-w/p) was low (low wages, high cost of capital), and even more so in large urban areas. This makes the iso-cost lines flat (and flatter in large urban areas), which favours the labour-intensive technology.

But later:

Together with refinements in the technology, probably the foremost factor was wage inflation during and after the Great War — what is known today as World War I.

Following the war, a steep rise in the wages of the labor pool from which telephone companies drew telephone operators was enough to jolt Bell management into rethinking its attitude toward automatic switching. Thus the Bell System began planning in 1919 to adopt automation.

The increase in wages increases the relative price of labour to capital (-w/p), making the iso-cost lines steeper, and favouring the capital-intensive technology. The change to capital-intensive technology didn't happen overnight though. Sometimes incentives take a while to be acted on:

In 1965, the year after Baker's forecast, the Bell System installed its first permanent fully electronic switching system in Succasunna, N.J.

Digital switching, which was introduced in the 1960s, lowered the price of capital, which in our model makes the iso-cost lines even steeper, and increasing the incentives for telephone companies to switch to automated (digital) telephone switching. As you can see, the simple production model that my ECONS101 class uses in the very first week is quite versatile in explaining changes in production over time.

[HT: Marginal Revolution]

Pornography actress productivity

Certain research article titles can make you wonder, "how...?". For example, take this 2017 article by Shinn-Shyr Wang (National Chengchi University, in Taiwan) and Li-Chen Chou (Wenzhou Business College, in China), published in the journal Applied Economics Letters (sorry, I don't see an ungated version), and entitled "The determinants of pornography actress production". It certainly made me wonder, how did they collect the data for that?

It turn out that it wasn't as dodgy as you might first imagine. Wang and Chou used data between 2002 and 2013:

...released by the Japanese ‘Digital Media Mart’, which on a regular basis collects data regarding videos and personal information of the Japanese AV actresses.

Essentially, the study tests whether actresses' physical appearance (proxied by cup size and whether they a side job [*] as a model or entertainer outside the pornography industry), and engagement in risky sex, affect the number of movies the actresses appear in. They find that:

...the later an actress commences her career, the fewer video shots she produces. Cup sizes and experiences as models or entertainers have positive effects on the number of video shots... having acted in risky sex videos could increase the production of an AV actress by more than 60%, which implies that, if the actress is willing to perform risky sex, her production may be significantly increased.

I'm unconvinced by their proxies for physical appearance, and a more interesting study would have addressed this by rating the actresses appearance directly (and no, that wouldn't necessitate the researchers watching loads of porn). The finding in relation to risky sex might be the result of an upward-sloping supply curve (there may be greater demand for riskier sex, so riskier sex attracts higher pay, so the actress works more), but of course there would also be a compensating differential to overcome as well (since riskier sex is probably a negative characteristic of the work, the actress would want to be paid more to compensate them for engaging in it). It would be more interesting to know the results in relation to earnings, rather than production, but I suspect that data is particularly difficult to obtain.

*****

[*] It seems to me that the pornography might well be the side job, and the modelling or entertainment the main job.

Immigrant restrictions and wages for locals

A simple economic model of demand and supply tells us that, if there are two substitute goods and the supply of one of them decreases, then demand for the other substitute will increase. This leads the price to increase for both goods. If the two 'goods' here are the labour of immigrants and the labour of locals, then a decrease in the supply of immigrant workers should lead to an increase in the demand for local workers, and higher wages for both groups. However, that simple analysis ignores that there is often another substitute for labour - mechanisation (or capital). So, it is by no means certain that restricting the number of immigrant workers will raise the wages of local workers, because employers might substitute from immigrant workers to technology, rather than from immigrant workers to local workers.

Which brings me to this new paper (ungated earlier version here) by Michael Clemens (Center for Global Development), Ethan Lewis (Dartmouth College), and Hannah Postel (Princeton), published in the journal American Economic Review. In the paper, Clemens et al. look at the effect of the 1964 exclusion of Mexican braceros from the U.S. farm labour market. At the time, the exclusion was argued for because it would lift wages for domestic farm labourers. However, looking at data from 1948 to 1971, Clemens et al. find that it had no effect. That's no effect on wages and no effect on employment of domestic farm workers.

They argue that the reason for the null effects is that farmers shifted to greater use of mechanisation (which they had not adopted in great numbers up to that point). They provide some empirical support for this. Crops where there was an existing technology that was not in wide use (e.g. tomatoes, where expensive harvesting machines were available that could double worker productivity) didn't suffer a drop in production after the bracero exclusion, because farmers simply adopted the available technology. In contrast, crops where there was no new technology available (e.g. asparagus) suffered a large drop in production (because farmers couldn't substitute to new technology, and fewer workers were employed).

The lesson here is that when prompted to change, producers will usually adopt the cheapest available production technology (as I have noted before). But that isn't necessarily the production technology that policy makers want them to adopt. In this case, instead of a production technology that made use of more local workers, the farmers opted for a production technology that made greater use of mechanisation. So, even if as a policy maker you believed that reducing immigration would improve wages for local workers, it isn't certain that would be the result of polices that reduce immigration (more on the effect of immigrants on local wages in a future post).

[HT: Eric Crampton at Offsetting Behaviour]

Taxing robots may be a good idea, but it won't keep people employed

Right now, we barely go a week without another article about robots (or algorithms) taking our jobs. One of the latest is this article from the South China Morning Post, which quotes Cai Fang (vice-president of the Chinese Academy of Social Sciences) as advocating for taxing robots:

For years, we have been warned that the day will come when machines will be able to do our jobs better than we can. Now a leading Chinese economist is offering a time frame.

Cai Fang, vice-president of the Chinese Academy of Social Sciences, the country’s top think tank, and former head of its Population and Labour Economic Research Institute, robots will “definitely” surpass humans in many job skills in 10 to 20 years.

Like Microsoft founder Bill Gates and other technology titans, Cai is an advocate of tax policies and other measures to keep robots from putting human workers out of jobs.

In February, Gates said governments should levy a tax on the use of robots to fund retraining of those who lose their jobs and to slow down automation.

“For a human worker who does US$50,000 worth of work in a factory, the income is taxed,” Gates said. “If a robot comes in to do the same thing, you’d think that we’d tax the robot at a similar level.”

Cai, a delegate to the National People’s Congress in Beijing, said the idea made sense.

In the first week of ECONS101 this semester, we used a very simple production model to explain how rising wages in England provided incentives for automation and primed the Industrial Revolution. Now we're seeing something very similar, except it isn't rising wages but the falling cost of robots (and algorithms). However, the effect in terms of the cost of wages relative to capital are the same.

Consider a simple production model, as in the diagram below, with capital (robots) on the y-axis and labour on the x-axis. Let's say that there are only two production technologies available to firms, A and B, and that both production technologies would produce the same quantity (and quality) of output for the combination of inputs (robots and labour) shown on the diagram. Production technology A is a labour-intensive technology - it uses a lot of labour, and supplements the labour with a few robots. Production technology B is a robot-intensive technology - it uses a lot of robots, and a small amount of labour (perhaps for servicing the robots).

How should a firm choose between the two competing production technologies A and B? If the firm is trying to maximise profits, then given that both production technologies produce the same quantity and quality of output, the firm should choose the technology that is the lowest cost. We can represent the firm's costs with iso-cost lines, which are lines that represent all the combinations of labour and capital that have the same total cost. The iso-cost line that is closest to the origin is the iso-cost line that has the lowest total cost. The slope of the iso-cost line is the relative price between labour and capital - it is equal to -w/p (where w is the hourly wage, and p is the hourly cost of robots).

First, consider the case where labour is relatively cheap and robots are relatively expensive. The iso-cost lines will be relatively flat, since w is small relative to p (so -w/p is a small number). In this case, the iso-cost line passing through A (ICA) is closer to the origin than the iso-cost line passing through B (ICB). So production technology A is the lowest-cost production technology, and firms should use relatively labour-intensive production methods.

However, as the hourly cost of robots (p) falls, the relative price between labour and capital (-w/p) increases, and the iso-cost lines get steeper. Eventually, we end up in the situation in the diagram below, where production technologies A and B have the same total cost. At this point, firms are indifferent between choosing the labour-intensive production technology or the robot-intensive production technology. From this point on, if robots become any cheaper, firms would be better off to choose the robot-intensive production technology, as it will be the lowest-cost production technology.

What happens if we tax robots? Taxing robots has the effect of increasing the hourly cost of robots (p), and flattens the iso-cost lines. That will reduce the incentive for firms to shift to the robot-intensive production technology, but only temporarily. If robots keep getting cheaper relative to wages (which seems likely for now), then the size of the tax necessary to keep labour-intensive production technologies competitive will continue to increase over time. That seems unsustainable in the long term.

Fortunately, the most thoughtful advocates of taxing robots are not arguing for the tax in order to keep humans employed and competitive with robots in terms of cost. Instead, they are arguing for the tax to raise revenue to fund a universal basic income to ensure a minimum standard of living for all, or they are arguing that the tax could pay for re-training or up-skilling workers who lose their jobs to the robots.

At the time, many people worried that the first Industrial Revolution would lead to widespread job dislocation and poverty, and during that time there were certainly many workers who were negatively affected. This time will not be different. However, once we came out the other side of the first Industrial Revolution, there turned out to be just as many (if not more) jobs available, as new opportunities for workers opened up. My crystal ball is offline at the moment, so I'm unsure if that will be the case this time. But why take the risk? If we can think through the (probably many) practical issues, taxing robots might be one way for the winners from this transition to robot-intensive production technologies to fairly compensate the losers.