Introduction

The age of artificial intelligence is dawning. In contrast to AI’s many benefits is the fact that it will displace millions of people around the world from their current workplace roles, especially those in white-collar jobs such as customer service, copywriting, and computer programming. It has already started to do so. Yet AI also presents a wonderful opportunity to rethink how we develop new skills. Those that seize this opportunity will move forward into exciting roles of their choosing, equipped with new skills they learned with the support of AI. To mitigate the negative impacts of AI on our careers, we must evolve our methods of acquiring new skills.

In this post, I share my recent experience of interacting with ChatGPT while exploring the impact of permanently closing the United States stock exchanges. Although I have no formal economics background, I was able to go from this question through the process of learning how to ask it in the context of the macroeconomics field. This included exploring a hypothetical world where the stock market was closed, uncovering the related assumptions and potential impacts, and generating the program code required to do some of the math. This exercise astounded me and left me very excited about the potential for AI’s use in education, if we use it properly.

One of the dangers of ChatGPT and similar AIs is that, for now, they are wildly inaccurate when it comes to specific details. They are constantly “hallucinating” alternative versions of reality, and then passing that on in a very convincing manner for the user to absorb. It is an unfortunate limitation, but one that can be mitigated through careful use of the AIs.

Using these AIs effectively in educational settings depends largely on the ability of the learner to know what statements to trust and which to validate. In my experience, all quantifiable data should be regarded as inaccurate by default. Names should be verified when they are a crucial part of the narrative being explored. Concepts are the most trustworthy because they stem directly from what the AIs do best: finding the connections between things.

As odd as it may sound, in the experience I am about to describe, the inaccurate details provided by ChatGPT do not matter. I intended to use the AI to learn how to approach answering my question by observing how it did, not to get concrete answers. I disregarded the bulk of the details because they are simply stand-ins that will be verified and updated later when I go to apply what I have learned.

Here is how it went.

Format of Post

In this post, I will include my prompts to ChatGPT in the following format so others can easily reuse them if they would like to:

This is a prompt I provided to ChatGPT.

ChatGPT’s responses I will share as screenshots because they contain formatting (especially tables) that this publishing platform does not support. I will also paraphrase some of the answers to remove the AI’s superfluous rhetoric. For longer replies, I will summarize the important elements below the screenshot.

Using ChatGPT

Initial Question

Lately, I have been reading about the development of the U.S. financial system. One thing that struck me was the recency with which the public stock market grew so prolific and accessible to everyone. Many view the stock market as a long-standing behemoth when in fact it is a relatively new phenomenon. Rewind a mere century, and the stock market as we know it today barely existed.

This brought a question to mind: what if we closed the U.S. stock market for good? It is a big question, and as I have a limited understanding of macroeconomics, I was unsure how to approach it. I thought it would be fun to see how ChatGPT might do this.

I started a new ChatGPT session and I told it what behavior I wanted from it:

You are an extremely advanced economic projection chatbot. When asked questions about future states of the economy, you provide detailed responses with recommendations for implementation.

Whether or not this did anything is debatable. I think an initial prompt like this would be more effective at specifying the user’s academic level. For example, I probably could have instructed it to “respond as if I am in middle school”, which would have toned down some of the more advanced language encountered later on.

However, ChatGPT was happy to oblige:

So I asked it:

What would the impact of removing the stock market from the United States financial system be?

As expected, it demurred, stating that there would be a large impact and that there are many variables and therefore impossible to predict.

I asked it to provide me with additional data, trying to solicit a more detailed response. It provided some high-level metrics about how much money is raised by initial public offerings. As I mentioned, whether or not this data was correct is irrelevant. The goal of this session was to see how the AI approached the problem, not to get accurate information from it.

Using an Example

I then asked ChatGPT to create a hypothetical case for if the stock market was removed:

Create a hypothetical case for the impact of removing the stock market on the U.S. GDP, complete with exact numbers. Be precise in your creation of this example scenario.

This “create an example” scenario is one way to get ChatGPT to stop dodging questions with “it depends”-type answers.

It started to provide some interesting ideas:

The key takeaways from this response are:

• If the public market is removed, companies will need to rely on private financing, which ChatGPT scopes down to debt financing.
• Business and consumer spending could decrease.
• Some important metrics to consider are declines in business investment, consumer spending, and GDP (“Gross Domestic Product”).

Again, the accuracy of the numbers is not important.

The mention of these various components started the formation of a general picture of the approach in my mind. I now knew some of the high-level impacts and factors to consider. Already, this helped me identify some gaps that existed in my understanding of this space, solving the classic “you don’t know what you don’t know” conundrum that is a major barrier to learning a new skill.

Digging in Deeper

Above, ChatGPT stated that its estimate was based on a macroeconomic model, so I asked it which one:

What macroeconomic model did you use?

It replied:

The “simple multiplier model” it mentioned sounds important. At this point, I broke out to good-ol’ Google to see if ChatGPT was already spouting nonsense. But this macroeconomic model does indeed exist. Study.com states:

The simple spending multiplier shows us how much economic output increases with an increase in spending. Economists ask the question this way: how much did real GDP change when a component of aggregate demand changed?

I then asked ChatGPT for another example, and for more details:

Provide another example with more specific details. Show your math.

It did so:

In this response, ChatGPT revealed some of the assumptions it made to create this example, but they were mixed in with a lot of other text. I’m a fan of structured data, so I asked for the same information in a table format, which it provided:

To be able to manipulate data on the fly like this is very useful. Hopefully, OpenAI will eventually add the ability to graph data to ChatGPT. It would be incredible if you could ask to visualize time series data, for example.

Keep in mind that what ChatGPT has provided are assumptions it made that fit into the example case, not actual data. This is analogous to a case study in a traditional textbook, although much more dynamic, as shown below.

I wanted to understand what else might be worth consideration, so I asked it several times to provide additional assumptions that fit into this example, which resulted in the following:

Again, the goal is to see how the AI is approaching the problem. What does it deem important that perhaps merits further research and/or inclusion when I do the math for real?

It certainly made some interesting assumptions. A quick scan showed that the largest assumed declines were in areas like taxes, the value of financial firms, and corporate profits, with metrics like productivity and unemployment barely changing at all.

I wanted to know more. I asked it:

Provide some specific numerical outcomes of the social impacts that would occur in this example. Even if you do not know for sure, provide a specific number and the assumption you used to create it.

This last sentence helps avoid the “it’s hard to say” rhetoric because it gives ChatGPT even more flexibility to invent numbers, provided it shows how it got there.

It replied with a few examples, then I asked for more, and eventually requested it in a table format:

Each row in this table represents an impact worth researching further, the AI’s projected quantifiable impact, and the assumption it made to get there. Even though the numbers are likely somewhat inaccurate, it still highlights possible and previously unknown associations between the assumptions and impacts.

Being a nature lover and hobbyist gardener, I was also curious about the agricultural impacts. I ran the same exercise for this sector, which resulted in the following assumed impacts:

Again, I am awed by the specificity of the AI. It is referring to things like research and development, specific crops, and farmer extension (educational) services.

I then asked it to add the macroeconomic model used to these tables. Here is the one for social impacts:

In some places, the response was an unhelpful “Macroeconomic model”, so I asked what model was used exactly for the first row. It said:

This reply specifies that a “linear relationship” was assumed to exist between a decline in personal income and an increase in unemployment. That makes sense: fewer people working would result in less income generated. It also describes other economic models that may be worth considering in this scenario.

Alright, now I know.

Learning How to Validate ChatGPT’s Assumptions

Now, I wanted to learn how to check these assumptions for myself:

I would like to learn how to use the “Endogenous growth model” to validate the “Reduction in funding for agricultural research and development” assumption. What is the first thing I do?

The AI gave me a high-level response referring to the need to incorporate data into the model and test its predictions, as well as a concise description of the model:

I then asked for definitions of “human capital” and “physical capital”, which it accurately provided. Human capital is a fascinating subject tied closely to skill development, and will perhaps be the subject of a future post.

For now, I wanted to start exploring how to apply this skill myself. I asked:

I would like to do the math to test this hypothesis. What is the first thing I do?

It dove right in:

This response read much more like something out of a textbook, complete with mathematical formulas and how to use them. I could have asked for it to restate the answer in more approachable language, but I stuck it out and worked my way through what I had been given.

I did not check if the formulas were accurate, because like the example data, at this point they do not need to be. If I go to use them, I will validate their accuracy. For now, all I need to know is that there are established formulas you use to apply these macroeconomic models.

I then asked a question about adding another term to the model, then about the meaning of the “elasticity output” variable mentioned in the initial response. It provided some helpful context not applicable to this post. However, it applied to me and my interests at the time. The ability to drill down into specifics like this on-demand is yet another powerful feature of ChatGPT.

Time for the next steps:

What is the next step in doing the math to test the agricultural impact example?

It replied:

Here, ChatGPT started to substitute variables for actual values, which is another helpful way of illustrating the actual application of the formula.

However, because I did not want to do the math myself at that point, my key takeaway was the need for data. Given we have never (permanently) closed the U.S. stock market before, there is unlikely to be enough meaningful data for this analysis. I asked what I should do in this situation.

It said:

It suggests using “sensitivity analysis”, which sounded promising. What next, I asked, and it replied:

To summarize, ChatGPT said I should:

1. Collect data for the relevant variables.
2. Use software like R or Stata to perform a regression analysis.
3. Vary the numbers to see what data is useful in the analysis.

As a former computer programmer, the mention of the R programming language got me excited.

The response referred to several different types of data. I asked for some more details:

Provide an example table representing recommended data to collect, including the name, type, and relevancy of each attribute.

It obliged:

I asked it to expand to “many more attributes” to get a better sense of the types of data that would be relevant to these calculations, and it did exactly that:

Again, I am using ChatGPT to paint a picture of the problem space and by doing so reveal the elements that may be relevant to my task. Going into this session, I knew none of this.

I then asked it to provide some example data for R&D investment, which it did:

Although far from a definitive comparison, this report states that the total average annual R&D funding from 2017 to 2021 was about $240 million, so ChatGPT is at least in the right range. This seems to often be the case with quantifiable answers, which makes me even more confident in its usefulness in education.

I then asked it to provide tables with example data for other attributes, which it did.

Now, time to bring it all together using some code:

Provide some example R code for testing the models and parameters you have described thus far for this example.

The following streamed onto the screen:

I love the succinct comments and descriptions included in the code. Here are some of the key takeaways:

• The R&D investment data imported matches the example data provided earlier.
• ChatGPT is using the aforementioned linear regression model.
• It is computing the elasticity of output, the thing I asked to learn how to do in the first place.

Just for fun, I said:

You are now a computer. Run the R code provided and respond with the result.

And it (sort of) did:

Unfortunately, while this is fascinating, the output was not correct. When I ran it myself, I received an elasticity of output value of 1.3017.

That is okay because, like the earlier data, the “computed” value does not matter at this point. Running this R code for myself is a completely reasonable operation for me to perform outside of this chat. In fact, that is the jumping-off point for me as a learner to begin to apply my new skills.

I threw ChatGPT another curveball:

Check all additional assumptions for the “R&D investment and productivity” attributes by producing the R code required to do so, then becoming a computer and running that R code, then printing only the attribute and result in a table. Do not print the code or any response other than the table.

Here it paused for about twenty seconds. All other responses had started generating text in only a second or two. That is very interesting, given that I am requesting it to perform “computations”, which conceivably would take some time to “perform” in a real-world context.

It printed a table of results, but for each data type’s p-value, not the elasticity of output:

I was not sure what a p-value was, so I asked and was told that it is the measurement of the chance of obtaining the same results by chance. In research, data with a p-value of less than 5% is considered to not be the result of chance, and therefore safe to use in the modeling.

My final request was for the ideal search queries to find the real data for use in my analysis. It provided me with the following table:

These searches were quite accurate and will provide a great starting point for my own data collection.

This is where I left off with ChatGPT. My overall question about closing the U.S. stock market remained open, but I did not expect to come away from this session with an actual answer. Now, I have the skills required to take the first steps in performing this analysis on my own. I can use what I have learned and begin to apply it myself. I will then come across new gaps in my knowledge, which I will use ChatGPT to address in the careful, deliberate fashion described in this post.

Summary

This post covered a lot of ground. Here are the highlights:

1. I wanted to know what would happen if we permanently closed the U.S. stock market.
2. I asked ChatGPT to explain what would happen, and it gave me some high-level things to consider.
3. I asked for an example case to free it from most of the “it depends” answers, which was used for the rest of the session.
4. I asked how the example case was calculated to see how ChatGPT approached the problem.
5. I learned about the potential financial, social, and agricultural impacts in this example.
6. I learned the basics of how to use the Endogenous growth model to test these impacts for myself.
7. I learned how to handle situations where there was no data present using sensitivity analysis.
8. I learned about the data required to perform this analysis.
9. ChatGPT wrote the R code required to start my analysis.
10. ChatGPT performed the calculations with an impressive level of accuracy, given that it is not actually running the code (we think).
11. I learned where to go find the data myself.

All of this took place in about an hour.

This illustrates how AI can accelerate the skill development process. AI can paint a picture of the surrounding information, show us our blind spots, and open up new lines of inquiry. Although some details will miss the mark, that is a mere shortcoming in a brand new technology. When utilized properly, AI will hasten and guide our steps towards learning, applying, and monetizing skills that interest us.

Closing Thoughts

After I was done with this ChatGPT session, I sat for a while contemplating the gravity of the interaction. I realized I had found a wonderful tool that can support me in developing new, exciting skills I never dreamed I could have. Not because I thought them unattainable, but because the act of learning a new skill often takes so long that we have to pick and choose what skills we invest our time in. Previously, becoming productive with a new skill required a broad understanding of the entire field, from which you eventually carve out a sliver of applicable information.

That is no longer the case.

I am excited for a world where our pedagogy shifts from one of rote to that of hyper-individualized learning adventures. We will journey through the subject matter in whatever way works best for each of us, whether that be in exhaustive detail or using playful dinosaur-themed examples. Chatbots will soon use the AIs already capable of generating images and videos to provide on-demand material for visual learners. As accuracy improves, engaging knowledge checks will be generated to improve knowledge retention.

It is truly an exciting time to be alive.

I believe we are off to a great start with AI. Unlike the Internet, the release of AI is being done with caution, with true consideration given to its value and its dangers. Where we once had social media executives dodging questions, we now have AI companies at least open to working with regulators. This is a big moment for the human race, and I think the right people are keeping that in mind.

Thank you for reading. Be well.