Economist Tyler Cowen Thinks ChatGPT Will Change Your Job

I’m a writer, and I think writing is hard. The first time I (glorified wordsmith) saw a bunch of algorithms (AI) output writing (aka do a hard thing), I took a beat. A mixture of surprise, fear, and curiosity washed over me. This was the end of 2022 when OpenAI released ChatGPT. 

Tyler Cowen felt some variation of this decades ago, when he saw a bunch of algorithms (AI, again) directing chess pieces across a board. 

Cowen is not a professional chess player. He’s an economist who just happens to be a chess enthusiast. But when news broke that AI had mastered chess, a game often hailed as the pinnacle of human intelligence, Cowen sensed that something important had occurred. He concluded that the next technological leap would be in AI. Armed with a deep knowledge of economics, he explored his insights further. In 2013, when the rest of the world was Instagramming their lunch, Cowen published Average Is Over, a book about how AI would change the future of work. In it, he argued that the economy will shift to reward those who can enhance the capabilities of technology.

Cowen had his AI moment a decade ago. Today, his incredible foresight is more than words and theories—it’s our reality. 

During the day, Cowen teaches economics at George Mason University. He moonlights as a prolific writer, co-writing leading economics blog Marginal Revolution, where he has published daily for over 20 years. He is also the author of 17 books, the latest of which is an AI-fueled interactive experience that analyzes the lives of influential economists and crowns one of them as the greatest of all time.

Dan interviewed Tyler Cowen to understand how Cowen uses AI in his own life and work—and what that might mean for the future of jobs and the economy. These are edited excerpts of the interview. Let’s dive in.

Tyler Cowen introduces himself

I am Tyler Cowen, and I’m an economist, professor, and writer. 

I am fascinated by the future. I’m driven by a hunger for discovering what lies beyond the curve. I articulate my findings in books, articles, and on my blog, Marginal Revolution. 

Lately, I’ve also been sharing my thoughts more through public appearances. This is a conscious decision, fueled in part by the burgeoning AI revolution. As AI grows more powerful, I believe there will be a restored emphasis on intrinsically human qualities like charisma.   

Am I charismatic enough? That’s for the world to decide. Or perhaps, if I look far out into the future, it will be for AI to decide. 

I think we will eventually be able to use AI to evaluate people’s talents. It will be highly meritocratic in a narrow way. I'm not sure we'll all be happy with it. It can be unsettling to be told just how good you are at something. On net, it can diminish human happiness. But it will also enable us to find potentially successful people easily.

My deep-rooted optimism about AI and its capabilities is rooted in chess. I’ve always liked a good game of chess. For many years, the popular conclusion was that AI could not play chess. It was too conceptual, too complex, too non-legible. That didn’t last for too long. Now, AI is almost godlike when it plays chess.

When I realized that AI could play chess—a really hard thing, in my book—I was convinced that it could learn how to do a lot of other things in time. 

My obsession with the future kicked in. I wanted to be ahead of the trend and write about where I thought it was going. So I wrote a book about how technology would change the future of work. 

With the advent of consumer AI, I have the privilege of seeing my careful predictions be tested against reality.

Predicting the short- and long-term effects of AI 

I don’t think we know yet which groups of people AI will help and harm the most. We’re still waiting. I expect many surprises.

However, I’m still willing to hazard a guess about the effects of AI. In the short term, my best guess is that AI will be an equalizer of sorts. People who can’t do things at all will soon be able to do them pretty capably, like writing an online essay to get into college. The smartest kids could already do that. ChatGPT won’t help them much. But it sure will work wonders for others who struggle with essays.  

However, I suspect that the long-term effects of AI will be far less egalitarian. Talented people brimming with ideas will derive higher benefits than the rest of society. As we learn how to use AI better, these people will leverage it to build out projects. They'll have better record keepers, translators, mathematicians, coaches, colleagues, advice givers…you get my drift. Of course, you can use large language models to help you find ideas, but they can’t go out and do it themselves. So I think the economy will adapt to rewarding some kind of hyped-up executive function.

I believe AI is excellent at managing people. Those who learn how to leverage it will be hugely productive. Companies might shrink in size, while still remaining rather potent. We’re already seeing this play out—take Midjourney. When it had its breakthrough, it had seven or eight people working there. However, an underappreciated risk of this is that AI might help poorly run terrorist organizations manage themselves better, increasing their chances of causing greater harm.

One thing that has surprised me about the current generation of AI models is their facility with words and emotions. I’ve always assumed that intelligent machines would take the form of some sort of autonomous reasoning tool. But ChatGPT is incredibly facile at taking an idea and writing a rap song about it, or a poem, or a shanty song, or whatever you might want. Large language models are also wonderful therapists that can be incredibly objective when that’s what you ask them for. 

AI can help humans understand themselves more deeply. I’m about to turn 62, so that’s of relatively less value to me, but I have been using ChatGPT for a variety of tasks, both personal and professional. 

Using ChatGPT while traveling

I love traveling and I’ve found that ChatGPT is the perfect companion, especially the ChatGPT app on my iPhone. When I’m in a foreign country, it forms a bridge of sorts, allowing me to meaningfully connect with my environment.

Let me give you an example. I was in Tokyo recently, I don’t speak Japanese, and not many people there speak English. But I used ChatGPT as my translator to effectively communicate with the world around me.

I also like sampling different kinds of food. I remember I was at a Paraguayan restaurant in Buenos Aires and wanted to eat something authentic. I’d never been to Paraguay and had no idea what that would look like. I just took a photo of the menu and asked ChatGPT: “What should I order here? Which are the classic dishes?” It just told me and I think that’s amazing.

ChatGPT is also a vehicle to fulfill my curiosities. I was in Honduras some time ago and when I saw something I wanted to know more about, I just sent a photo to ChatGPT. If you see a plant or a bird you don’t know, you can take a photo and send it to ChatGPT and ask it: “What’s this?” And it will tell you. This is objectively good because you learn new things. It’s also easier and quicker to use than Google and Wikipedia. 

Using specific prompts to do deep research

I use ChatGPT to learn obscure pockets of history. I’ve found that it happens to be far more intelligent there than in many other domains. 

I used ChatGPT to interview the 18th-century Irish writer Jonathan Swift on my podcast. I asked GPT Swift a series of detailed, probing questions for an hour. It didn’t make any mistakes. Not one. What you see online is the very first try. There were no redos or no trial runs. Presumably, ChatGPT has read a lot of Swift and read a fair bit about Swift. The interview I conducted was weird and detailed enough to point ChatGPT to the smart corners of the internet. That’s why I worry about hallucinations a lot less than most LLM users. ChatGPT just has a low tendency of hallucinating in the areas where I typically use AI.

Now, I know the kind of question that would probably make ChatGPT hallucinate. Ask it something like: “What are the three best books to read on Jonathan Swift?” My best guess is that one of three answers it comes up with won’t exist. This is likely because the prompt is too vague. But I don’t run into it so often because I’m not asking ChatGPT these questions.

Let’s say I do want to ask ChatGPT a general question, maybe about the meaning of inflation. I wouldn’t just ask: “What is inflation?” The answer it would give me probably wouldn’t be much better than Wikipedia. I might ask: “What is inflation? Answer as would Milton Friedman.” That will pretty much always get you a better answer. This helps point ChatGPT towards smarter bits in the matrices. The stuff it knows connected with Milton Friedman is smarter than the stuff not connected to Friedman.

For obvious reasons, I don’t find myself asking ChatGPT basic questions about economics. I’ve been using it to find out more about the insurance industry, which I don’t know nearly as much about.

Finding the best data by asking for comparisons

I’m planning to record a podcast about the history of insurance and insurance economics with Alex Tabarrok [the other half of Marginal Revolution]. I turned to ChatGPT to learn more about insurance. The answers weren’t perfect, but at this stage, they don’t have to be. What I’m after is context. I don’t need the exact facts.

Recently, I had a question about a common piece of investment advice: Americans shouldn’t invest through their life insurance policy because they can probably earn a better return with a direct purchase of equities through a low-cost diversified mutual fund. It made me wonder about the rationale behind whole-term insurance products, which are basically savings instruments bundled with regular life insurance. 

So I asked ChatGPT: “Why would anyone want life insurance bundled with the savings return? Doesn't it mean inferior returns?”

All screenshots courtesy of How Do You Use ChatGPT?

It came up with a fairly average answer. I would’ve probably been able to think up at least some of those points myself. I still think it’s valuable because ChatGPT’s response spurs my thoughts. 

And I can push ChatGPT to give me an even better answer by asking: “If you were to write a critique of why individuals should not buy whole-term life insurance, which points would you stress? Please answer like a sophisticated investment advisor.” 

It came up with some good points, including ones I hadn’t explicitly thought of. I’d give the answer an A. Maybe not an A+, but a solid A. 

ChatGPT does really well when you ask it to compare and contrast. It’s like crossing two swords. Even if you don’t want a comparison, you often get the best data when you ask it for one. I think there might be something about the autoregressions that explains this.

Using AI as a second screen while reading

I think AI search engines are a helpful way to satisfy my curiosity when I’m reading a book about something and want more details or insight on a particular point.

I was reading a book about the Byzantine Empire. It mentioned high inflation under one of its early emperors. I was curious about the rate of inflation, so I asked ChatGPT. It understood the query well after I referenced a familiar historian’s name, but it came back afterwards saying that there was insufficient data on this point. I wanted to push ChatGPT a little more. I asked it what the inflation rate was—if it had to guess at a specific number. 

It launched into an evasive dialogue about how its knowledge is limited by virtue of being an AI model trained up to 2023 without access to newer findings. After a few more paragraphs of that, GPT finally said that some scholars estimate that the rate might have been over a 1000 percent over the course of half a century. I think there’s a decent chance that’s a hallucination, but it was pretty clear that ChatGPT wasn’t about to commit to anything more.

These are areas where I have found that other AI tools do a better job than ChatGPT.

I turned to Perplexity, another AI search tool that I believe might have an edge over ChatGPT at answering this kind of question. 

Indeed, Perplexity came through: “Some sources say 3 percent a year. Others say 5 percent a year debasement of the currency.” One of the things I really like about Perplexity is that it gives you citations for each of its results. I’ve got one from Reddit and one from JSTOR for this query that I can just go ahead and track down.

I have found that my information funnel has evolved significantly since I leaned into AI. I still use Google to find clickable links when I’m writing an article online. But if Google is for links, AI is for learning. I’m still relatively new to Perplexity, but I already think it’s better than ChatGPT for queries where I need a reference. It also sometimes has context that GPT doesn’t because it’s looking to tie its answers to references.

Using ChatGPT inside university classrooms

I’ve already folded AI search engines into my own research habits, and I encourage my university students to do the same. I believe that it’s valuable for my students to learn how to use AI search tools. Instead of banning ChatGPT from my classroom, I’m inviting it in.

I’ve just started teaching a class on the history of economic thought. ChatGPT-4 is on the reading list. Everyone in the class is required to pay the $20 monthly subscription fee. I have a good feeling about this use case because the class is mostly about obscure history, a topic in which ChatGPT performs well.

I’ve already used ChatGPT in the classroom while teaching law students a year ago. I gave them an assignment to write a paper alongside ChatGPT. The results were very promising. I had students say they liked the experience because they were exposed to the limits and scope of GPT firsthand. The papers were also generally good. I’d say it was a successful experiment. 

I’m not worried about students using ChatGPT to plagiarize. I tell my students that the purpose of incorporating AI in the classroom is to equip them to use it in a wise, smart manner. I’ve also told them that if they copy-paste their essays from ChatGPT, not only will the paper be mediocre, I’ll probably figure it out as well. It just doesn’t make sense to do it. I want to learn new things about the technology with them, and not focus on busting them for being on one side of some vague line or the other.

Going head-to-head with my ChatGPT clone 

Before the show, Dan had created a custom GPT called Tylerbot based on my personality. The idea was to ask me and Tylerbot the same questions and see how well the machine’s answers match up to mine. It’s a play on the Turing test—the Tyler test, if you will. Just a little context before we get into it: Dan didn’t train Tylerbot on any of my books and it didn’t have web browsing capabilities. This meant that the custom GPT was only operating on what it already knew about me. 

With that, let the games begin. May the best man (or machine) win. 

Question 1: What are the core lessons of economics?

My answer: The two most important principles of economic science are incentives matter and there’s always an opportunity cost. The first mark of a good economist is trying to apply these consistently.

Tylerbot: 

Score: That’s a great answer. I’d give that an A.

Question 2: What are five words your spouse would use to describe you?

My answer: Curious, works hard, loves to travel, good father. Oh and should walk the dog more.

Tylerbot:

Score: Those are great answers too. I mean, you can’t expect GPT to know about the dog. 

Question 3: In your view, what is the most underappreciated way in which AI, particularly tools like ChatGPT, will fundamentally alter our understanding of economics and human behavior in the next decade?

My answer: I don't think it will fundamentally alter our understanding of economics in the next decade. I think it will eventually be possible to simulate small economies using GPT-like methods and we might learn a lot from those simulations. However, those seem like a distant possibility because of difficulties in data collection. So I don't think it will matter for quite a while. 

Tylerbot:

Score: This answer doesn’t sound like me. It’s not a bad answer, but it seems like it couldn’t find what Tyler might think, so it came up with a pretty good generic answer. It’s fine for accuracy but on the Tyler-scale, I’d give it a C minus.

AI will shape how we write and think

When I receive questions about an article that I’ve written, I’m often surprised by what my readers are interested in. It’s sometimes different from the focus I had when I wrote the piece. I think AI is going to enable new, interactive content formats that will surface to readers what they really care about. That is the insight fueling my generative book experiment, and I’m excited to see where it leads.       

Watch Dan’s whole conversation with Tyler Cowen on X or YouTube, or listen on Spotify or Apple Podcasts.