Discover more from Nominal News
ChatGPT and AI - Predictions on Labor and the Macroeconomy
One of our readers asked us to discuss how Artificial Intelligence (AI) may impact the economy. What does research suggest will happen?
This is a reader suggested topic. If you have any topics or questions that you’d be interested in seeing what economics has to say about, please leave a comment below!
Since November 2022, when OpenAI, an artificial intelligence research company, released to the public the AI the online chat [ro]bot, ChatGPT, a lot of social media has been inundated with examples of ChatGPT doing a myriad of different tasks – starting up a business, summarizing text, and writing poetry. Numerous other companies quickly released or publicized their own versions of the Large Language Model (LLM) AI. Many have argued that this new technology will result in many jobs being displaced – especially many ‘white-collar’ jobs such as basic programming. A recent report by Goldman Sachs predicts 300mln jobs could be replaced by AI. However, even though they mention that other roles will appear and overall GDP will grow, this has still resulted in many individuals being worried about the future. The fear has grown to such an extent that several prominent tech leaders and computer scientists have written a letter asking to pause AI development. But what does economics have to say about all of this?
Naturally, there hasn’t been much economic research specifically talking about the new AI technologies and how they will impact the economy. However, there have been discussions and research conducted on similar past phenomena and situations that we can use to apply to this particular scenario.
The most closely related topic to AI is the concept of automation. Automation is the use of machines to do certain jobs that have been typically done by people. The issue of automation has been most talked about in the context of manufacturing automation, and more recently discussed in the context of self-driving vehicles.
Anecdotally, we know that automation did impact economies (such as the declining number of manufacturing workers over the last several decades), but also did not result in massive unemployment and upheaval in the labor markets. Regarding self-driving cars, and more importantly for the macroeconomy, self-driving trucks, we are yet to see them used anywhere beyond experiments. However, in order to understand how automation has impacted the labor force and the macroeconomy, especially in order to predict the outcomes of AI, it is important to model it. A model will allow us to think through the implications of automation and estimate its impact.
What is Automation
The first issue to consider is what exactly is automation. One way to think about it is that it is something that can do a particular job. However, jobs are usually very vaguely defined or entail a broad set of activities. For example, the titles engineer, construction worker, or consultant do not give us any specifics of what these people actually do. Thus, economists decided to approach this issue through the lens of ‘tasks’. Tasks are more specific activities performed by individuals. Some examples of task-based roles are “typists”, “cashiers”, or “tax preparers”. These well-defined tasks can now be analyzed to determine which of them can be ‘automated’ – done by a machine, robot, or software.
This exercise itself is not without controversy, as there are multiple approaches to determining which tasks and to what extent they can be automated. Autor and Dorn (2013) developed one such approach. They looked at specific tasks and created a Routine Task Index, which established how ‘routine’ a task is by looking at how much of the task involved routine actions versus manual or abstract actions. They used US Department of Labor data that contained information on what abilities or actions a task required. A routine task would score high on “finger dexterity” (usage of routine motor tasks) and “setting limits, tolerances and standards” (usage of routine cognitive tasks), and low on “eye-hand-foot coordination” (manual abilities), “direction control and planning” (managerial ability) and “GED Math” (formal reasoning requirements). A task that was more routine and required less manual or abstract skill would score high on the Routine Task Index, implying that it will be easier to automate. Below is a table of some of their results:
The column on the left shows the occupations with the highest Routine Task Index scores. The other columns show occupations with the lowest Routine Task Index scores. One thing that quickly jumps out from the middle column is that we have both bus drivers and truck drivers as occupations with low Routine Task Index scores, suggesting that they are not easy to automate – it is worth keeping that in mind as we look at another paper that created a similar index.
Frey and Osborne (2017) tackled the question of what can be automated slightly differently from Autor and Dorn. Frey and Osborne believe that computerisation and machines can not only replace routine tasks, but also any non-routine task that is not subject to ‘engineering bottlenecks’. Using findings from computer science literature, Frey and Osborne define three engineering bottlenecks (i.e. where automation will be difficult): perception and manipulation tasks (such as moving objects around in a house or warehouse), creative intelligence tasks (such as music and art)1 and social intelligence tasks (mainly roles requiring interaction with people).
Using data from Department of Labor surveys on what sort of skills are needed for an occupation, Frey and Osborne estimated the probabilities of which jobs can be automated. Below are the 3 most likely jobs to be automated:
And the 3 least likely:
Using this analysis, in their 2013 study, Frey and Osborne concluded that around 47% of US employees are at risk of computerization in the near future. As can be seen from the US labor market – this has not happened. Additionally, the US unemployment rate currently is close to an all time low at 3.7%. So what happened with this prediction?
Diving deeper, Frey and Osborne’s model predicted, for example, that the likelihood of taxi drivers and school bus drivers being automated was 89%. Truck drivers were at 69%, while postal service mail carriers at 68%. If we go back and look at the model predictions in the table above from Autor and Dorn, we see that these occupations, according to their research method, would not be easy to automate (they specifically mention that mail carriers would not be easy to automate). This implies that, at least in the last decade, the Autor and Dorn approach to automation performed much better as a predictor than the Frey and Osborne approach.
The main takeaway from this discussion is that establishing what automation and AI can do in itself is a tough question. Predicting the future outcomes to the economy are highly dependent on the modeling assumptions we make, and it is important that we are aware of them when considering the impacts to the economy. Given that in the last decade, we haven’t witnessed the rapid computerization of many roles, it does suggest that the current AI technologies, which are an incremental advancement over previous similar technologies, will not quickly revolutionize many jobs.
The next critical step in figuring out what the effects of automation and AI can be is how we think AI will translate into the production of goods and services, the ultimate goal of any technology. In most models (and in real life), production of a good or service combines two production factors – capital and labor. Capital and labor act as both complements and substitutes. That is, the more you have of one, the more productive the other one is (the complementary channel) but you can also replace one for the other in production (the substitution channel). To illustrate this with an example, one can think of a personal computer (a form of physical capital). It makes an individual that uses the personal computer more productive, but also replaces some other forms of labor, such as typists.
Economists typically assume automation does one of two things: either it makes one of the factors more productive (that is, a worker or a particular machine is more productive because there is automation) or the overall output is higher at every amount of labor and capital because the overall technology level is higher. These models, however, usually imply that demand for labor will go up! That is because a worker is now more productive under either of those assumptions. If a worker is more productive, a firm will demand more workers. So these models cannot generate opposite dynamics such as labor demand dropping.
Another way to model automation is to use the Autor and Dorn’s idea of focusing on routine tasks, where automation can directly substitute labor that does those tasks, without acting as a complement to those specific workers. Autor and Dorn (2013) and Acemoglu and Restrepo (2018) have looked at automation through that lens. Autor and Dorn found that this approach can explain well the recent job polarization trend experienced in the US and globally. The job polarization trend refers to the fact that many middle-income jobs have been gradually disappearing over the last several decades, while employment grew for “high-skill”2 and “low-skill” jobs. With the advent of machines and software, Autor and Dorn argue that many of the middle-income jobs, such as manufacturing jobs, have disappeared to automation over time. This resulted in workers moving to service sector jobs on either end of the income distribution, where automation is tougher to implement.
Acemoglu and Restrepo further develop this model, showing that the impact of automation can be ambiguous on both labor demand and wage growth. How come? Automation replaces labor tasks, reducing labor demand and lowering wages, which is the displacement effect of automation. However, if automation significantly increases a firm's productivity (or reduces costs), it can actually lead to an increase in labor demand. One such example is the ATM (Automatic Teller Machine). Even though ATMs completely replaced banking tellers, more banking ‘tellers’ were actually hired! That is because the very low cost of ATMs made opening up banking branches cheaper. Since banking branches still needed ‘tellers’ for other purposes, such as the offering of other services like loan and financial advice, more ‘tellers’ were needed. This effect is referred to as the productivity effect. Thus, if the productivity effect is larger than the displacement effect, labor demand and wages will go up for workers directly impacted by automation. It is important to stress that for the productivity effect to be large, the new technologies have to significantly increase productivity (or reduce costs). Thus, if the AI does very complex tasks (Acemoglu and Restrepo refer to this as ‘brilliant’ AI) that are expensive today, this would be preferable, than if the AI were to do only basic tasks (referred to as ‘so-so’ AI).
Lastly, Acemoglu and Restrepo discuss the reinstatement effect. That is when automation results in the creation of entirely new tasks that have not existed before. In this situation both labor demand and wages will increase, as these new tasks cannot easily be replaced by capital.
What does it all mean
With the above discussion, I gave an overview of how economics can approach the question of what will be the impact of the new AI technologies. Most, if not all, public discussion on the labor and macroeconomic impacts of AI can be easily expressed using the above theoretical frameworks. People who argue that AI developments will lead to a new economy with many more jobs believe the reinstatement effect will be very large. Others who argue that this will cause significant unemployment and fall in wages, believe that the displacement effect will dominate both the productivity effect and the reinstatement effect (additionally, they’re subconsciously assuming that these new AI technologies are not ‘very’ productive themselves believing it is more of a ‘so-so’ technology). So where do I believe this is all going?3
As I am not a computer scientist, my thoughts around the capabilities of machine learning are based on other experts in the field. The argument that I currently subscribe to is that this approach to AI (the machine learning one of which Large Language Models are a part of) will not result in the ‘brilliant’ technologies we hope it will. As articulated by Ben Dickson, founder of Tech Talks, there are significant doubts that machine learning will, for example, give us self-driving cars. The key founders behind machine learning are also skeptical that the current AI method, of which ChatGPT is a part of, will result in true AI (or AGI – Artificial General Intelligence, as it is commonly referred to, which is the replication of human intelligence).
Furthermore, a lot of the tasks these new AI bots are doing have already been automated to some extent – telemarketing, online chat bots, and simple stock market articles are automated now.4 This makes it seem like the current crop of AI bots are simply the continued improvement of automation that has been occurring for a while now.
A recent experimental study by Noy and Zhang seems to fit the prediction quite well. Noy and Zhang conducted a laboratory experiment, where they had participants complete a variety of tasks. They found that people use the recent ChatGPT software as a replacement for effort rather than a complement to their effort, suggesting the new AI will act as task replacements rather than augmenting technologies. Most participants found the tool very useful during the experiment, but the usefulness of the tool dropped in real life uses.
All of this underlines that the new AI chatbots are more likely to continue the job polarization trend described by Autor and Dorn. In the long run, these AI tools will most likely result in new tasks and jobs being created via the reinstatement effect. These impacts will be gradual, and future generations will adjust to it effectively.
To address a concern from the reader who asked about this topic – if this AI is actually far more capable than I have described, especially in writing computer code, then demand for computer scientists might actually increase! Following the ATM logic described above, if a lot of costly programming is done by AI, many firms will find it cheap to enhance their companies with self-designed software. But this will certainly significantly increase the demand for computer programmers to ensure that the AI designed software is working.
Personally, we at Nominal News are excited for this new technology, especially as it will enable us to conduct online searches more effectively and efficiently (the current search algorithms for academic work are extremely poor and it takes a lot of time to go through the economic literature to understand the whole picture on an issue). Moreover, AI has already improved the aesthetic of this post – the cover image of this post was done by an AI – Dall-E 2!
If you have any thoughts or opinions on this topic, please comment below! We’re happy to hear perspective on this issue. And if you enjoyed it this post, please spread it around!
The issue here is not the act of creating art and music, which many programs now can do, but defining what is ‘creative value’ for the software to understand.
The terms high-skill and low-skill are commonly used in economics referring to the notion that certain skills require more investment of both time and money (such as doctors or engineers), while others do not require as high of an investment.
At Nominal News, we have generally attempted to not convey opinion in these posts, but given that this refers to the future, any prediction or forecast is an opinion.