After work today, I came to Dupont Circle to get some fresh ideas. If you don’t know me, I like to read books, so I plan to visit Kramers to get some fresh pages.

This is objectively a questionable decision, and I often ask the question with baseless answers, which doesn’t make sense to me. There are probably more than sixty books sitting untouched at my house. Some are neatly arranged on my shelves. Others still sit in Barnes & Noble bags, looking almost exactly as they did the day I bought them. I keep telling myself I will get to all of them eventually, which may be one of the more optimistic forecasts I make as an economics student.

Still, I came to Kramers. (Honestly, it's not paid promotion, but if you are around DC and want to read some new, fresh pages and want to fuel your brain with some ideas, then I definitely suggest you ) .

There is also something about the atmosphere of a bookstore that the internet has never been able to replace for me. And yes, I understand the economics of what I am doing.

I could probably search for the same books online tonight and find them cheaper. Maybe significantly cheaper. My Amazon Prime might be telling me to order online where I can get up to 50% off these books, but my Apple Pay is ready to scan. They could arrive at my door tomorrow or the next day. No walking, no carrying books around Dupont, no standing between shelves debating whether I really need another one. If price were the only variable in my utility function, I would probably make a different choice.

But that is precisely the point. I don’t need my own economics for now.

Economics is sometimes misunderstood as the study of how people save money. It isn’t. At its core, economics is about choices: how people allocate scarce resources- money, time, attention, energy among competing possibilities.

The lowest-priced option does not automatically maximize utility. Just like my situation.

When I buy a book from a bookstore, I am not purchasing only the physical book. I am purchasing the experience surrounding it: walking through Dupont after work, browsing shelves without knowing exactly what I will find, reading the back covers, opening random pages, discovering another author beside the one I was looking for, and eventually carrying the books with me to a coffee shop.

Amazon may win the price competition. A physical bookstore can still win the utility calculation.

There is something wonderfully economic about that distinction.

People pay for experiences, convenience, atmosphere, identity, discovery, and even rituals. We reveal what we value through the choices we make, including choices that might look irrational if we examine only the price tag.

My friends know this about me. Whenever I visit a new city, I somehow end up in a bookstore. By now, I can probably recommend bookstores in many of the major American cities I have visited. For me, visiting a city without entering one of its bookstores feels incomplete.

Bookstores tell you something about a place. What is displayed near the entrance? What are people reading? Which subjects receive an entire shelf? What local writers appear everywhere? Who is sitting in the corner reading? A good bookstore is partly a business, partly a cultural institution, and partly a map of a city’s intellectual interests.

Maybe that is why I keep returning to them even when an online retailer can beat them on price.

Today I left with three.

One was What Happened to Liberal Democracy?, by Daron Acemoglu. Another was Power and Progress, by Acemoglu and Simon Johnson. And then Breakneck, by Dan Wang. If you know what I normally read, none of these choices will surprise you. Most of my reading is nonfiction. Economics is obviously a large part of it, but I am equally interested in foreign affairs, institutions, technology, artificial intelligence, economic history, and politics. Occasionally, I wander into self-help. Fiction is much rarer for me, although every once in a while I convince myself that I should become a more serious fiction reader.

The books I actually buy tend to circle around one broad question:

Why do some societies, technologies, and institutions produce progress while others fail to translate opportunity into prosperity? That question appears everywhere in economics. It is there when we talk about institutions. It is there when we study technological change. It is there when we ask whether AI will raise productivity or simply redistribute income and power. It is there when we compare countries that build infrastructure quickly with countries that struggle to execute projects. It is there when we study democracy, state capacity, inequality, growth, or industrial policy.

That is why my reading rarely feels completely separate from my economics work.

One book leads to a paper. A paper leads to a dataset. A dataset produces a question. A question sends me back to another book. I often try to visit as much evidence as possible from the books .

Before coming here, I was thinking about Difference-in-Differences. Not exactly the most romantic subject for a Tuesday evening in Dupont Circle. You might get bored reading these couple of sentences, but I need to do this anyway to survive in this world because I chose a profession and a subject that requires it. But recently I have been trying to go beyond simply knowing how to run an econometric model. I want to understand why it works, when it fails, and what assumptions are actually doing the heavy lifting.

There is a difference between being able to type a regression command and being able to defend the research design behind it.

So I have been reading newer work on Difference-in-Differences and trying to build my own understanding from the foundations.

This is also why, despite spending so much of my life around computers, I still like reading difficult material with a notebook and pen beside me.

When I encounter an equation, I stop. I rewrite it. I define the variables. I ask what changes if an assumption fails.

I draw arrows. I cross things out. Sometimes I stare at the same paragraph for fifteen minutes and realize I understood far less than I thought.

That process is slow. But I like slow learning.

In an age when almost any concept can be summarized in seconds, there is something valuable about struggling with an idea long enough that it becomes yours. AI can explain parallel trends to me. It can summarize a paper. It can generate code for an event-study specification. Those tools are enormously useful, and I use them. I often use them.

But there is still a difference between receiving an explanation and developing understanding.

For me, that difference often happens in the notebook.

After buying the books, I walked just a couple of storefronts away to Starbucks in Dupont, where I often sit for nearly two hours. That is where I am writing this.

So if you are reading these words, imagine them being typed with three fresh books beside me and a cup of coffee slowly disappearing while people move in and out around me.

Later, I am meeting a friend for dinner. Nothing particularly extraordinary happened today.

I worked. I went to a bookstore. I bought books I probably didn’t need. But I promised myself that I would definitely read this.

I sat down with coffee. I thought about econometrics.

I wrote. And I will have dinner with a friend. But perhaps that is exactly why I wanted to write this. We sometimes treat economics as something that exists inside textbooks, regression tables, Federal Reserve meetings, World Bank reports, journal articles, and classrooms.

But economics is happening constantly around us. My decision to buy a book locally instead of ordering it for less online is a question of preferences and utility.

The sixty unread books at home might reflect optimism bias or a poorly managed inventory problem. The bookstore competes not only on price but through atmosphere, discovery, location, and experience.

My decision to spend two hours reading an econometrics paper instead of doing something else has an opportunity cost. Even the coffee beside me represents a small consumption decision shaped by habit, convenience, location, and willingness to pay. Economics becomes much more interesting when you stop seeing it only as a subject and start noticing it as a language for understanding ordinary life. If you are my economics professor and reading this, smile now because your student adopted this subject as a daily way of living.

Another explanation for why I keep buying books faster than I read them.

Perhaps buying a book is partly a bet on my future self. When I purchase one, I am implicitly saying: At some point, I will have the time and curiosity to sit down and understand this.

That future version of me is apparently extremely productive. He has enough time to finish the sixty books already waiting at home, read the three I bought today, keep up with new econometrics papers, conduct research, write, work, meet friends, and somehow still sleep.

I have considerable confidence in this person.

Maybe too much. But I also don’t think unread books are entirely wasted.

A shelf of unread books reminds you how much you don’t know.

That is useful. The more economics I study, the more I realize how enormous the discipline is. Every answer seems to reveal three better questions. Every model depends on assumptions worth examining. Every empirical result rests on an identification strategy. Every policy conclusion has tradeoffs hiding somewhere.

So perhaps those unread books are not a failure to finish something. They are an inventory of questions waiting for me.

Tonight, three more joined them. Tomorrow, I could have gotten them cheaper online.

But then I would have missed the walk through Dupont, the shelves at Kramers, the unexpected books I picked up and returned, the coffee afterward, and this essay.

And economics, after all, has never been only about paying the lowest price.

It is about understanding what we value and why we choose to spend our scarce time, money, and attention on the things that make a life feel worthwhile.

Tonight, for me, that happened to be three books, an econometrics paper, a notebook, a pen, and a cup of coffee in Dupont Circle.

Note: I don’t want you to buy three extra books when you already have more to read. My case is different, so enjoy my writing, not my buying behavior, because it might hurt your bank balance. But if you want to follow and subscribe, then we can buy more books together.

I can say this is a huge change. The more interesting story, however, is not about tokens. It is about what economists have studied for decades: why some firms are much better than others at turning a new technology into productivity. This is one of the most profound questions economists have been asking

OpenAI’s new research suggests that enterprise AI is moving from assistance toward execution. We have to understand one important thing that has been changing. Workers are no longer using AI only to summarize documents, answer questions, or brainstorm ideas. Increasingly, agents can interact with tools, company information, and established workflows to actually complete parts of the work. This change has been happening very fast .

The numbers illustrate how quickly this transition is happening. As of June, Codex accounted for 64% of combined Codex and ChatGPT output tokens among enterprise customers. Since February, weekly active enterprise Codex users reportedly increased 108× in legal, 41× in sales, 41× in recruiting, 26× in marketing, and 20× in finance and accounting. Engineering, interestingly, grew only 5× partly because it was already an early adopter.

From an economics perspective, this reminds me of the productivity paradox associated with earlier waves of information technology.

When computers first entered offices, simply purchasing computers did not automatically make firms dramatically more productive. Firms had to reorganize workflows, train workers, digitize information, change management practices, and invest in complementary assets before the technology produced its full economic value.

AI may be following a similar pattern but much faster.

Think about a simple production function, which we do in our ECON 101

Output = f(Capital, Labor, Technology, Organizational Capability).

Giving every company access to the same AI model improves the technology available to them, but it does not guarantee equal productivity gains. The return depends on complementary investments: data infrastructure, employee skills, management practices, permissions, governance, and the ability to redesign workflows.

That may help explain the widening gap between frontier and typical firms.

Frontier firms are not simply using more AI. They appear to be integrating AI more deeply into production. OpenAI reports that 21% of weekly active users at frontier firms use Plugins, compared with 9% at typical firms. Skills show an even larger difference: 19% versus 3%.

Economically, we can think of this as a difference in the marginal productivity of AI capital. Two companies may have access to essentially the same model, but the company that connects it to internal data, tools, institutional knowledge, and repeatable workflows may generate significantly greater output from the same underlying technology.

There is another finding I found especially interesting.

Early-career workers appear to use AI more intensively than senior employees. Six months after enterprise adoption, early-career workers were sending substantially more ChatGPT messages than executives.

This raises an important labor-economics question: Could AI temporarily change the traditional relationship between experience and technological comparative advantage?

Historically, senior workers possess more firm-specific knowledge and human capital. But younger workers may be more willing or able to experiment with AI tools. If they learn how to combine AI with organizational knowledge, some of the productivity advantage traditionally associated with experience could shift.

This does not necessarily mean AI replaces senior expertise. More likely, it changes the optimal combination of human capital. Experienced employees provide judgment, institutional knowledge, and domain expertise, while AI-proficient workers may provide new ways of executing and scaling that knowledge.

For me, the biggest takeaway from OpenAI’s findings is therefore not that firms are simply “adopting AI.”

We may be entering a period of AI-driven organizational divergence.

The important economic divide may increasingly be between firms that merely purchase access to AI and firms that reorganize themselves around it.

If that is true, the next productivity gap will not be determined solely by who has the most powerful model. It will depend on who learns fastest how to turn that model into productive capital.

Note: I want to thank the OpenAI Economics team for this research. Figures have been imported from the OpenAI working papers; all credit goes to the authors and the OpenAI Economics Research team.

References:

https://openai.com/index/how-enterprises-put-ai-to-work/

https://cdn.openai.com/pdf/how-organizations-use-chatgpt.pdf

Image source: Will AI Help or Hurt Workers? One 26-Year-Old Found an Unexpected Answer.(Published in The Wall Street Journal before the Toner-Rodgers paper got questioned and caught ). Aidan Toner-Rodgers is in the middle of this image.

Artificial Intelligence, Scientific Discovery, and Product Innovation. This is the paper that we are going to talk about in this Substack. As someone who spends a significant part of my day reading economics papers, I rarely expect a research story to genuinely surprise me. I enjoy reading working papers, journal articles, and preprints because they teach me not only about economics but also about how researchers think. Every dataset, every regression, and every robustness check reflects months or even years of someone’s work. That is why the case involving former MIT economics PhD student Aidan Toner-Rodgers stopped me in my tracks. It was not simply another case of academic misconduct. To me, it exposed a much deeper question about the future of research in the age of artificial intelligence.

If you have not followed the story, here is what happened. Toner-Rodgers authored a paper titled Artificial Intelligence, Scientific Discovery, and Product Innovation ( which you can get at the beginning of this Substack). The paper explored how AI was changing scientific innovation and quickly became one of the most talked-about economics papers on artificial intelligence. It was presented at world-leading platforms and conferences, uploaded to arXiv, and discussed by some of the most respected economists in the profession. Nobel Prize-winning economist Daron Acemoglu promoted the paper publicly. Major publications including The Economist ( What the failure of a superstar student reveals about economics and The Wall Street Journal(Will AI Help or Hurt Workers? One 26-Year-Old Found an Unexpected Answer. covered its findings. The research was even referenced by institutions such as the European Central Bank and the Congressional Research Service of the United States Congress. I will post all the relevant sources at the bottom of this Substack, which I have found on the Stanford University website. For many people, the paper appeared to be an important contribution to understanding AI’s role in scientific discovery.

Then everything changed.

Questions began to emerge about the underlying data. MIT launched an internal investigation, and the university eventually announced that it had “no confidence in the veracity of the research.”(Assuring an accurate research record ) The paper was withdrawn, and Toner-Rodgers was removed from MIT’s PhD program. One of the most widely discussed economics papers on AI had fallen apart because the evidence supporting it could no longer be trusted. ( I mean the evidence was built up with a fabricated dataset )

I read this news last year; my reaction was not simply disappointment. My first thought was much larger than one individual researcher. I started asking myself a question that I believe every researcher should now be asking: How many datasets in the age of AI are actually real? I tried to spend a lot of time playing with AI, and I built multiple datasets that did not exist in this world, and I was surprised to find they were the same as the real dataset. (Hey, I have never used such a dataset in my writing and research!!!! )

Artificial intelligence has transformed research in remarkable ways. I use AI almost every day to summarize papers, understand difficult concepts, debug Python code, improve my writing, and brainstorm research ideas. Used responsibly, these tools make researchers more productive and allow us to spend more time thinking instead of repeating mechanical tasks. But AI has also introduced a risk that did not exist at this scale just a few years ago.

Today, a single prompt can generate an entire dataset that looks surprisingly realistic. Thousands of observations can appear within seconds. The descriptive statistics look reasonable. The regression tables make sense. The correlations appear believable. Even missing values can be simulated naturally. None of this means the evidence is real. It simply means fabricated evidence has become much easier to create than ever before. In our economics language, the marginal cost of producing such a dataset is very low.

That possibility should concern everyone in academia. Including you if you are reading right now.

What unsettles me most about this case is not that one researcher allegedly fabricated evidence. Misconduct has unfortunately existed throughout the history of science. What is different today is how convincing fabricated work can appear. If a paper from one of the world’s leading universities can be accepted, promoted, discussed at prestigious conferences, and cited by policymakers before serious concerns emerge, then perhaps our current system of evaluating empirical research is facing a challenge it was never designed to handle.

This is also a reminder that peer review has limits. Reviewers evaluate theory, methodology, identification strategies, and statistical analysis. (Fact: The paper was not published in the journal) They rarely have access to confidential datasets, and they often must trust that the underlying evidence has been collected honestly. That trust has always been one of the invisible foundations of empirical economics. Artificial intelligence makes that foundation much more fragile.

As economics students, we spend years learning increasingly sophisticated econometric methods. We debate difference-in-differences, instrumental variables, regression discontinuity, synthetic control methods, machine learning, and causal inference. We worry about omitted variable bias, endogeneity, external validity, and robustness. Yet every one of these techniques quietly relies on one assumption that never appears in the regression equation: the data are genuine. If that assumption fails, every beautiful econometric model becomes meaningless.

It would be easy to blame artificial intelligence for this problem, but I do not think AI is the villain here. AI did not fabricate research. A human being did. The technology simply made it easier to create something that looked convincing. Like any powerful tool, AI can be used responsibly or irresponsibly. The responsibility ultimately belongs to researchers, institutions, and the incentives that shape academic publishing.

As someone still early in my journey as an economist, this story did not make me lose confidence in research. If anything, it reminded me why research integrity matters more today than ever before. The value of economics has never come from elegant equations alone. It comes from the honesty behind those equations. A perfectly estimated regression means nothing if the underlying observations never existed.

Artificial intelligence is going to transform economics, and I believe that is something to celebrate. It will help us analyze larger datasets, discover new patterns, and answer questions that once seemed impossible. But the same technology also forces us to become more careful. Perhaps the next generation of researchers, including me, will spend just as much time verifying data authenticity as they do improving econometric techniques.

The biggest lesson from the MIT case is not that one PhD student failed. It is that the entire research community should take this moment to reflect. If even some of the brightest minds in economics, leading institutions, respected conferences, and major policymakers accepted a paper before fundamental concerns about its evidence were uncovered, then perhaps we need stronger systems for transparency, replication, and data verification.

The age of AI has changed the questions we ask when reading research. It is no longer enough to ask whether the identification strategy is credible or whether the standard errors are correctly estimated. Increasingly, we may also have to ask a much simpler and much more important question:

Can we trust that the data are real?

Because if we cannot answer that question with confidence, then every conclusion built on those data becomes uncertain, no matter how sophisticated the econometrics may be.

Note: Always keep your replication dataset with you and remember your dataset should be valid and exist rather than being built up and fabricated.

Important sources :

AI, Materials, and Fraud, Oh My! (BS Detector, May 16, 2025)

Artificial Intelligence, Scientific Discovery, and Product Innovation (Toner-Rodgers, A., November 26, 2024)

Assuring an accurate research record (MIT Economics, May 16, 2025)

Case Studies of How Generative Artificial Intelligence is Changing Work (Congressional Research Service, March 27, 2025)

If only Arxiv required researchers to sign at the top rather than the bottom of the page, none of this would’ve happened. (Columbia University, May 19, 2025)

Integrity concerns prompt MIT to reject AI innovation study (digwatch, May 20, 2025)

MIT Disavowed a Viral Paper Claiming That AI Leads to More Scientific Discoveries (Futurism, May 20, 2025)

MIT disavows doctoral student paper on AI’s productivity benefits (TechCrunch, May 7, 2025)

MIT Retracts Support for AI Productivity Research Paper (autogpt, May 19, 2025)

The transformative power of AI (European Central Bank, April 1, 2025)

What the failure of a superstar student reveals about economics (The Economist, May 22, 2025)

Photo: I clicked this picture in front of BFI

From the last post, you might guess that for the past three days, I have been at the University of Chicago for a research conference. Like many visitors, I expected to see beautiful Gothic buildings, famous professors, and brilliant students. Those were all there. But what surprised me was something much harder to describe. It was the feeling that economics here is not simply a subject. It is a tradition.

You can feel it while walking between the Department of Economics, the Becker Friedman Institute (BFI), and Chicago Booth. You hear conversations about incentives, institutions, markets, and policy. The questions change with time, but the way of thinking remains remarkably consistent. It is not about memorizing theories. It is about challenging assumptions and asking whether evidence actually supports our beliefs.

That is perhaps the greatest legacy of the Chicago School.

Photo: I took this photo inside the BFI building

Many people associate the Chicago School with capitalism and free markets. While that is true, reducing it to those labels misses the point. At its core, the Chicago School is an intellectual tradition built on rigorous reasoning, skepticism toward conventional wisdom, and an insistence that markets often solve problems better than governments, although they are never assumed to be perfect.

The intellectual roots go back nearly a century.

Frank Knight argued that uncertainty is fundamental to economic life and that entrepreneurship exists because the future cannot be predicted with certainty. His book Risk, Uncertainty, and Profit remains one of the foundations of modern economics.

Henry Simons defended competitive markets while emphasizing that governments still have an important role in maintaining competition and preventing concentrations of economic power. His work reminds us that free markets require strong institutions rather than the absence of government.

Then came Milton Friedman, perhaps the most influential economist associated with the University of Chicago. Friedman transformed monetary economics and became one of the strongest defenders of economic freedom. In Capitalism and Freedom (1962), he argued that economic freedom is a necessary condition for political freedom. Later, in Free to Choose(1980), he and Rose Friedman made the case that individuals, when allowed to make their own economic decisions, generally allocate resources more effectively than centralized planners.

Photo: Inside of BFI, these are the highest prizes, including Nobel and the John Bates Clark medal

George Stigler fundamentally changed how economists think about regulation. Instead of assuming that regulation automatically protects the public, he asked who benefits from regulation and why. His theory of regulatory capture continues to influence public policy debates around the world.

Gary Becker took economic reasoning into places where economists had rarely gone before. Crime, discrimination, education, family decisions, and human capital all became subjects of rigorous economic analysis. Becker demonstrated that economics was not simply about markets. It was a framework for understanding human behavior.

Interesting Fact: BFI was named by mixing Milton Friedman and Gary Becker (Baker Friedman Institute )

Robert Lucas Jr. transformed macroeconomics through rational expectations, forcing economists to rethink how policy affects behavior. The Lucas Critique fundamentally changed empirical macroeconomic research and remains essential for anyone studying monetary or fiscal policy today.

Even Ronald Coase, although not originally trained at Chicago, found an intellectual home here. His work on transaction costs and property rights reshaped law and economics and continues to influence environmental policy, corporate governance, and institutional economics.

Walking around campus, I realized something interesting. None of these economists are physically present anymore. Yet their influence is everywhere.

Students still debate Friedman’s ideas.

Researchers still test Becker’s theories.

Seminars still ask Lucas-style questions.

The Becker Friedman Institute continues to host discussions involving economists from very different schools of thought. Chicago Booth produces research spanning finance, behavioral economics, development, AI, industrial organization, and macroeconomics. The Department of Economics remains one of the most influential research departments in the world.

What impressed me most was not ideological uniformity. It was intellectual confidence.

People disagree here.

Sometimes strongly.

But disagreements begin with theory, continue with evidence, and end with better questions.

That culture matters.

Today, economics is changing rapidly. Artificial intelligence is transforming labor markets. Climate change challenges traditional policy frameworks. Industrial policy has returned. Governments around the world are intervening in ways that would have seemed unusual thirty years ago.

Some people therefore argue that the Chicago School has become outdated.

After spending three days (I have a strong connection here; I have been here every year; it’s been three consecutive 3 years. Last year, I was here for 3 months while I was working here as a Research Assistant), I think that conclusion misses the point.

The value of the Chicago School has never been that every conclusion reached in the 1960s or 1980s should be accepted forever. The value lies in its method. Be skeptical. Demand evidence. Think carefully about incentives. Question popular narratives. Be willing to change your mind when the data contradict your beliefs.

Those principles are as relevant today as they were when Friedman or Becker walked these halls.

As a young economist, the University of Chicago reminded me that economics is ultimately a conversation across generations. Every new paper builds on previous debates. Every model is a response to someone else’s question. Every empirical result either strengthens or challenges an existing idea.

History matters.

Institutions matter.

Intellectual traditions matter.

Perhaps that is why universities like Chicago remain special. They preserve a way of thinking while allowing it to evolve.

As I leave campus, I am taking home more than conference notes or presentation slides. I am leaving with a renewed appreciation that the greatest universities are not defined by their rankings or their buildings. They are defined by the ideas they refuse to let disappear.

The Chicago School is no longer just a chapter in the history of economic thought.

It is still a living conversation.

As I write this, I am sitting at the airport with about an hour and a half before my flight boards. Airports are strange places. Everyone seems to be moving somewhere important while also spending a surprising amount of time waiting. Since I have nothing better to do than stare at the departure screen for another ninety minutes, I thought today deserved another Substack.

No, today’s topic is not AI.

Well... not directly.

Instead, I want to talk about an economics idea that quietly shapes how we think about innovation, technology, and public policy: the Weak Induced Bias Hypothesis.

Tomorrow, I will be in Chicago presenting my research at the University of Chicago Booth School of Business on data centers and electricity demand. It feels fitting that while traveling to one of the birthplaces of modern economics, I am spending my waiting time wrestling with one of the most influential papers written by one of today’s greatest economists, Professor Daron Acemoglu.

The paper is Directed Technical Change.

If you have ever tried reading Acemoglu’s theoretical papers, you know exactly what I mean when I say this: every sentence demands respect. You cannot skim it. Although there might be boring stuff that you want to skip. You cannot pretend to understand it after one reading. Sometimes a single paragraph takes fifteen minutes because every equation is carrying an economic intuition that deserves to be unpacked.

While I was struggling through one particular section, one idea kept coming back into my head.

The Weak Induced Bias Hypothesis.

For anyone unfamiliar with it, the intuition is surprisingly simple.

Economies do not invent technologies randomly.

Innovation responds to incentives.

If one factor of production becomes relatively more abundant or relatively more expensive, firms have stronger incentives to invent technologies that use that factor more efficiently.

In other words, markets do not simply adopt technology.

Markets shape the direction of technology itself.

That sounds obvious today, but it is actually a profound insight.

For decades, economists mostly treated technological progress as something that simply happened. It arrived like the weather. We measured it, but we rarely asked why innovation moved in one direction instead of another.

Directed technical change asks a different question.

Why do some innovations save labor while others save energy?

Why do some complement skilled workers while others replace routine workers?

Why does one country become a leader in semiconductors while another specializes in agricultural machinery?

Technology is not just advancing.

Technology is choosing a direction.

Or more accurately, incentives are choosing the direction for it.

This is where public policy enters the conversation.

Governments often say they support innovation.

That sounds wonderful.

But innovation is not a single destination.

Innovation is a compass.

Where you point the incentives determines where the technology goes.

Subsidize renewable energy, and engineers spend more time improving batteries.

Raise the cost of carbon, and firms suddenly discover cleaner production methods.

Increase the supply of highly educated workers, and firms have stronger incentives to develop skill-complementary technologies.

Economics tells us that prices are signals.

Directed technical change reminds us that those signals influence not only production today but also innovation tomorrow.

Now let me connect this to something I have been thinking about during this trip.

Tomorrow I will present my ongoing research on data centers in Virginia.

At first glance, data centers and directed technical change seem like natural partners.

After all, AI requires enormous computing infrastructure.

The more society values AI, the greater the incentive to invest in servers, electricity, cooling systems, transmission lines, and increasingly efficient chips.

That certainly looks like induced technological change.

But I do not think the connection is perfectly clean.

My project studies how expanding data centers affect electricity demand and residential electricity prices.

It is about infrastructure.

It is about energy markets.

It is about regulation.

Directed technical change, however, is asking a much deeper question.

Why did firms choose to innovate in AI infrastructure in the first place?

Why are billions of dollars flowing into computing rather than some completely different technology?

Those are related questions.

They are not identical questions.

My research examines the consequences of technological investment.

Acemoglu’s framework helps explain why those investments emerge in the first place.

That distinction matters.

Too often we try to force every new research project into a famous theory simply because the keywords overlap.

Sometimes the connection is genuine.

Sometimes it is only superficial.

As researchers, we should be honest about that.

Good economics is not about stretching a model until it covers everything.

Good economics is about knowing where the model stops.

Ironically, that realization may be one of the most valuable lessons from reading difficult theory papers.

You do not finish them believing they explain the entire world.

You finish them understanding one corner of the world much more deeply.

Perhaps that is why theoretical economics remains so important.

Empirical papers tell us what happened.

Theory helps explain why it happened.

Neither is sufficient alone.

As I continue reading Directed Technical Change, I realize I probably understand only a fraction of what Acemoglu intended. But that is part of the process. The best papers rarely give you immediate answers. Instead, they leave you with better questions.

And maybe that is exactly what airports are good for.

Not just waiting for flights.

Waiting for ideas to connect.

Tomorrow I will be presenting research about data centers, electricity, and the infrastructure behind the AI economy. Today I am sitting in an airport reading a paper that reminds me that technology is never simply invented. It is guided by incentives, shaped by institutions, and ultimately directed by the economic environment we create.

Sometimes the most productive part of research is not running another regression.

Sometimes it is sitting in an airport, struggling through one paragraph of a theory paper, and realizing that one small idea can completely change how you think about technology and public policy.

That is a pretty good use of ninety minutes.

Image: This is how I read papers, and I love reading like this because it helps me to construct my own ideas.

As a young economics student, I have a confession.

I read papers for fun.

Yes, I know. There are healthier hobbies.

But after spending countless nights reading journal articles, something started bothering me. Not a coefficient. Not a weak identification strategy. Not even another paper that somehow “extends the literature” by changing one control variable.

The papers themselves feel... different. Not necessarily wrong.

Just suspiciously polished. Suspiciously long.

Suspiciously abundant.

Then I stumbled upon one economist’s publication record.

From 2023 to mid-2026, this professor published roughly 45 papers, with another six months still left in the year. The topics stretch across multiple areas of economics. Even more interesting, many of the coauthors remain remarkably consistent. Every academic has collaborators, of course. But when the same names repeatedly appear together, it starts looking less like collaboration and more like an industrial production line.

Naturally, my inner economist became curious.

So I compared some of the papers.

Before 2023, the average paper length was around 25 to 32 pages.

After 2023?

Around 48 to 49 pages. That is roughly a 75 percent increase.

Congratulations. I don’t know whether you deserved these congratulations or not

But

The productivity frontier has officially shifted outward.

Somewhere, Solow is smiling.

Or panicking.

Now, before anyone reaches for their pitchforks, let me be clear.

This is not an accusation that anyone is cheating.

AI is an incredible tool.

I use AI. Many economists use AI.

The question is not whether AI should be used. The question is where the human contribution actually begins. There is a massive difference between asking AI to improve grammar and asking AI to manufacture intellectual effort.

Economics has always rewarded ideas.

Today, I worry that we are beginning to reward output.

Economists constantly teach students about supply and demand.

Ironically, we seem to have forgotten that the same model applies to ourselves.

The supply of papers has exploded. I am not saying it’s bad. Listen to me again: it’s not bad. But listen again …….

When supply increases dramatically while attention remains fixed, what happens?

The value of each additional unit falls.

Congratulations.

We have created the world’s first academic inflation. Every week, another hundred working papers arrive in inboxes.

Every month, another special issue appears. Every conference has enough papers to last until retirement. Meanwhile, nobody has enough time to actually read them.

Perhaps economists have accidentally solved the scarcity problem. Unfortunately, the scarce resource is now attention.

Here is my favorite economic analogy.

Imagine Ferrari suddenly decides to produce twenty million cars every year.

Would Ferraris still feel special?

Probably not.

Scarcity creates value.

Exclusivity creates prestige. The same principle applies to research.

Economics does not become stronger because we publish twenty papers a year.

It becomes stronger when one paper changes how we think.

Keynes did not publish twenty influential papers every year. Lucas did not.

Kahneman did not. Acemoglu did not build his reputation by uploading a manuscript every other Tuesday.

Their influence came from ideas.

Not volume.

Another pattern fascinates me.

I call it the Citation Circle.

Imagine six researchers.

Researcher A cites B.

B cites C.

C cites D.

D cites E.

E cites F.

F cites A.

Repeat.

Eventually everyone looks influential.

It is the academic version of friends liking each other’s Instagram posts.

Of course, collaboration is healthy.

Citation networks naturally exist. The problem begins when citations become the objective instead of the consequence of good research.

A citation should be earned because your idea mattered. Not because your WhatsApp group remembered your latest upload. If you hate this, be my paid subscriber so we can hate this together again.

Some people argue,

“Why do you care? Their publications help their careers.”

They’re right.

Publication records matter.

Promotion committees count papers. Grant committees count papers.

Hiring committees count papers. Individual incentives are real.

But economics teaches us something important.

When incentives are poorly designed, people optimize the metric instead of the objective.

If universities reward quantity, they will receive exactly that.

Goodhart’s Law strikes again:

When a measure becomes a target, it ceases to be a good measure.

The irony is almost poetic.

Economists spend their careers studying productivity.

Now AI has made economists dramatically more productive. But productivity and value are not the same thing.

A factory producing thousands of plastic watches every day is highly productive.

Rolex still sells for more.

I am currently collecting more evidence because I want to study this systematically rather than emotionally. This professor is only one example.

I have noticed similar patterns elsewhere, including among researchers at highly respected institutions.

That observation alone proves nothing. It simply raises questions worth investigating.

If AI has fundamentally changed publication behavior, we should measure it.

How much has publication volume increased?

How has paper length changed?

Have coauthorship networks become denser?

Has citation concentration shifted?

Most importantly, has research quality improved at the same pace as research quantity?

These are empirical questions.

Economists should answer them with data. I also want to be honest about myself.

I use AI.

Almost everyone I know uses AI.

The difference I hope to preserve is this:

I want AI to help me communicate ideas I genuinely understand.

I do not want AI to replace the struggle that creates those ideas in the first place.

Thinking is slow. Research is frustrating. Writing is painful.

Those are not bugs. They are the production process. Please believe me.

Maybe I am wrong.

Maybe this is simply what technological progress looks like.

Maybe AI will produce an explosion of genuinely transformative research.

I hope it does. But I also hope we do not confuse more PDFs with more knowledge.

Economics does not need twenty forgettable papers every year.

It needs a handful of papers that people still debate twenty years later.

The marginal paper is becoming cheaper. The marginal idea remains priceless.

Thank you so much for reading. I can make sure you are reading an economics student who is human and typing on the keyboard, so you are reading organic writing with your own flavor.

Figure Credit: David Thomson

Yesterday, while exploring recent developments in artificial intelligence, I came across an excellent article by David Thomson, CTO of TechnologyChecker, titled “Open-Source AI Adoption 2026: 5.6M Projects vs Real Deployment.” His analysis immediately caught my attention because it challenged one of the most common narratives surrounding AI today. Rather than simply celebrating the explosive growth of open-source AI, Thomson asked a more interesting question: How much of it is actually being used?

Reading his article reminded me of an important lesson in economics: production and adoption are rarely the same thing. We often confuse supply with demand, invention with innovation, and creation with economic impact.

That realization inspired me to start a new series.

Over the coming months, I want to take a deep dive into the world of open-source artificial intelligence. This will not be a series about prompting ChatGPT or comparing the latest models. Instead, I want to explore open-source AI through the lenses of economics, public policy, innovation, industrial organization, and technological change.

Because I believe open-source AI may become one of the defining economic institutions of the twenty-first century.

The numbers alone tell a fascinating story.

According to the Stanford AI Index 2026, there are approximately 5.6 million AI-related open-source projects across GitHub and Hugging Face. Yet only about 206,880 projects (3.7%) have received at least ten GitHub stars, suggesting that most repositories remain experimental or attract little community engagement. At the same time, Hugging Face hosts more than 2 million public AI models, and model uploads have more than tripled since 2023.( Thank you again, David Thomson, because all this information I have grabbed from his article Open-Source AI Adoption 2026: 5.6M Projects vs Real Deployment

If we stopped there, we might conclude that open-source AI has already won.

But deployment data paints a different picture.

Technology Checker scanned more than 50 million live domains to identify AI technologies actually running in production. Their results show that the most visible open-source conversational AI framework, Botpress, appeared on only 1,558 active domains, while the OpenAI API was detected on 52,682 domains—roughly 34 times more. The implication is striking: the supply of open-source AI is exploding, but production deployment remains concentrated in a relatively small number of applications.

This distinction matters enormously.

Economists have long distinguished between invention and innovation. An invention becomes economically meaningful only when it is adopted, commercialized, and integrated into production (Thank you, Professor Michael Andrews: I learned from his economics of Innovation class). History provides countless examples. Thousands of patents never become products. Thousands of scientific discoveries never reach consumers. Likewise, millions of GitHub repositories do not necessarily translate into millions of deployed AI systems.

The gap between creation and adoption is not a failure of open source. It is simply how technological diffusion works.

Yet beneath these deployment numbers lies another remarkable trend.

Survey evidence shows that 63% of organizations already use open-source AI in some form, while 89% incorporate open-source software somewhere in their AI stack. Open models now achieve roughly 90% of the performance of leading closed models, although closed systems still account for around 80% of actual model usage because they remain easier to deploy and manage. Meanwhile, the global market for open-source AI models is projected to grow from $19.05 billion in 2025 to $23.08 billion in 2026, representing annual growth of about 21%.

These numbers raise questions that extend far beyond computer science.

Will open-source AI democratize innovation in developing countries?

Can smaller firms compete with technology giants if frontier AI models become freely available?

Will open models reduce vendor lock-in and increase competition?

How will governments regulate technologies that anyone can download, modify, and deploy?

And perhaps the most important question of all:

Who will capture the economic value created by artificial intelligence?

These are fundamentally economic questions.

As a graduate student in economics, I find this intersection between technology and economic development impossible to ignore. Throughout history, every general-purpose technology, from the steam engine to electricity, from the internet to cloud computing, has reshaped productivity, labor markets, industrial organization, and global competitiveness. Artificial intelligence appears poised to become the next chapter in that story.

This Substack series is my attempt to understand that transformation.

Over the coming weeks, I & We will examine the economics of open-source models, AI agents, digital public goods, innovation ecosystems, startup competition, industrial policy, labor markets, and global technological development. We’ll look beyond the headlines and ask not only what AI can do, but also how it changes incentives, institutions, and economic growth.

The open-source AI revolution is no longer just about writing code.

It is about rewriting the economics of innovation itself.

And this is where our journey begins.

If you enjoy my writing and want to support independent research on economics, AI, technology, and public policy, please consider becoming a paid subscriber.

In 1811, English textile workers smashed machines. Two centuries later, Generation Z isn’t destroying artificial intelligence. They’re doing something far more interesting.

They’re afraid of becoming dependent on it.

A recent Gallup survey reveals an extraordinary paradox. Nearly 79% of Gen Z believe AI makes people lazier, while 62% believe it makes people less smart. Their concern isn’t primarily unemployment. It is something much deeper: AI will replace learning by doing, critical thinking, and learning from others.

For the first time in modern history, the biggest fear surrounding a technological revolution isn’t that machines will replace our muscles.

It’s that they might replace our minds. And history suggests we should take that concern seriously.

Every industrial revolution has outsourced something distinctly human.

The steam engine outsourced muscle.

Electricity outsourced endurance.

The computer outsourced the calculation.

Artificial intelligence is attempting to outsource cognition itself. That is an entirely different category of technological change.

The Industrial Revolution multiplied physical productivity without fundamentally changing how humans learned. Apprentices still became masters through years of practice. Engineers still solved problems. Craftsmen still developed intuition through repetition.

The machine changed production.

It didn’t replace the learning process.

AI might.

Adam Smith famously argued that specialization increases productivity through the division of labor. The pin factory became wealthier because workers focused on increasingly narrow tasks.

But Smith also warned about something economists rarely quote.

Extreme specialization, he argued, risks making workers “as stupid and ignorant as it is possible for a human creature to become.”

Economic efficiency carried cognitive costs.

Governments, Smith believed, should invest in education precisely because markets alone would not preserve human capabilities.

Two hundred and fifty years later, his warning feels strangely modern.

Only today’s division of labor is no longer between workers.

It is between humans and algorithms.

Karl Marx saw machines differently.

Technology, he argued, could alienate workers from the production process, reducing skilled craftsmanship to repetitive supervision. Whether or not one agrees with Marx’s politics, his insight remains relevant.

When machines perform the meaningful parts of work, humans risk becoming operators rather than creators.

AI raises a similar question.

If ChatGPT writes the essay...

GitHub Copilot writes the code...

Claude drafts the report...

Who is actually learning?

John Stuart Mill believed civilization depended not merely on economic growth but on the cultivation of higher human faculties.

Honestly, I am trying to be equally biased towards all schools of thought because I often get criticized.

Progress was not measured only by output.

It was measured by the development of judgment, character, and intellect.

That distinction feels increasingly urgent.

AI can certainly produce more words, more presentations, more analyses, and more software. But can it produce wiser thinkers?

That remains an open question.

Ironically, economists have spent decades celebrating productivity growth while paying remarkably little attention to how productivity is acquired.

Most human capital models assume people simply accumulate skills through education and experience.

But experience itself depends on doing difficult things.

You learn to negotiate by negotiating.

You learn economics by struggling through models.

You learn to write by producing terrible drafts before producing great ones.

Remove the struggle, and you may also remove the learning.

The danger isn’t that AI answers questions.

The danger is that it answers them before we have wrestled with them ourselves.

This is where Generation Z’s concerns deserve more attention than they have received.

Their fear isn’t irrational technophobia.

It reflects an intuition that philosophers have debated for centuries.

Aristotle believed excellence comes from habit.

John Dewey argued that education is learning through experience.

Michael Polanyi described “tacit knowledge” the kind of understanding that cannot simply be written down or copied.

Much of human intelligence is procedural rather than informational.

You cannot download wisdom.

You acquire it.

None of this means AI is a mistake.

Every industrial revolution initially looked threatening before ultimately raising living standards. The steam engine created modern industry.

Electricity transformed cities.

Computers revolutionized science.

AI will almost certainly generate enormous economic prosperity.

But history also teaches another lesson.

Technological progress creates new forms of scarcity.

During the Industrial Revolution, physical labor became less scarce.

In the AI Revolution, independent thinking may become a scarce resource.

Those who can reason without relying on AI, question AI’s answers, and know when AI is wrong may command the highest premium in tomorrow’s economy.

Perhaps the debate is asking the wrong question.

Instead of asking whether AI will make us smarter or lazier, we should ask something far more fundamental:

Can a civilization remain intellectually ambitious when thinking itself becomes optional?

Every industrial revolution has changed what humans do.

Artificial intelligence may be the first to change how humans become who they are.

That is not merely an economic question.

It is a philosophical one……………………………………

If you think I am writing what you want to read, please support me by subscribing or being my paid reader.

We are living through the first era in history in which your personal information has become an economic asset not just to you but to everyone else.

Every day, millions of people voluntarily upload their thoughts, relationships, political beliefs, career frustrations, and emotional struggles onto the internet. They call it authenticity. Platforms call it engagement. Advertisers call it data.

Economists call it information. ( I often called it new gold)

And information has value.

The modern economy isn’t simply built on oil, labor, or capital anymore. It runs on something far more intimate: human disclosure.

The question is no longer “Should I post this?”

It’s becoming:

“Who profits when I do?” (I always think: someone is making money from this )

A fascinating framework from Harvard Business Review asks us to evaluate whether we should reveal personal information by weighing both the benefits and the costs.

At first glance, the framework feels psychological.

Look closer, and it’s deeply economic.

Every disclosure is an investment decision.

You pay a cost today, the loss of privacy, in exchange for uncertain future returns such as trust, opportunity, connection, or social approval.

Sometimes the investment pays off.

Sometimes it becomes a permanent liability.

The digital world has quietly changed the economics of disclosure. Twenty years ago, telling someone a personal story reached one person. Today, the same story can reach five million people before breakfast.

The cost structure has changed. Believe me, it has changed, and it changes astoundingly.

Information is almost impossible to retrieve once released. Unlike money, privacy rarely compounds after you spend it.

It depreciates.

This creates an interesting paradox. Social media rewards the people willing to reveal the most. Employers increasingly reward people who build public reputations. Entrepreneurs build billion-dollar brands by documenting their lives. Researchers grow influence by sharing their work publicly. Creators build communities through vulnerability. Disclosure has become a competitive advantage.

But so has restraint.

Because while everyone is competing for attention, very few are protecting optionality.

Every controversial tweet, emotional LinkedIn post, or impulsive Instagram story narrows the set of futures available to you. If you are reading this Substack, I’m sharing my thoughts.

Future employers can read it.

Future investors can search for it.

Future political opponents can archive it.

Even future versions of yourself may disagree with it.

The internet has an unusually long memory.

Economists often talk about opportunity cost, the value of what you give up when making a decision. Privacy has one of the highest opportunity costs in the digital economy because you rarely recognize what you’ve surrendered until years later.

Once information enters the public domain, you cannot sell it again.

You’ve already given it away.

This doesn’t mean becoming secretive.

It means becoming strategic.

The HBR framework asks one particularly powerful question:

“Am I withholding information because it’s the best decision or because it’s simply the easiest?”

I’d add another:

“Am I sharing because it creates genuine value or because the algorithm rewards it?”

Those are very different motivations.

One builds trust.

The other feeds engagement.

Ironically, the people with the greatest influence often reveal the least. The best CEOs rarely livestream every difficult meeting. Great investors don’t disclose every trade before making it. Successful negotiators understand the value of informational asymmetry.

In economics, asymmetric information is usually discussed as a market failure.

In life, it’s often a source of strategic advantage.

Not every thought deserves publication.

Not every success requires announcement.

Not every failure belongs on the timeline.

Sometimes silence isn’t insecurity.

It’s asset management.

As artificial intelligence becomes better at reconstructing personalities from tiny digital footprints, the value of privacy is likely to rise, not fall. Data that seems meaningless today may become extraordinarily predictive tomorrow.

That changes the calculation.

The real scarcity of the next decade may not be attention.

It may be the ability to control your own information.

So before you click “Post,” don’t ask whether people will like it.

Ask something much more valuable:

Ten years from now, will you be glad this became part of your permanent digital identity? Because in the information economy, privacy isn’t merely a personal preference. It may be one of the highest-return investments you’ll ever choose not to spend.

Thank you, HRB, for the content. I haven’t copied your thoughts, but I borrowed your ideas.

Artificial intelligence has become the defining technological race of the twenty-first century. Every week, we hear about larger language models, smarter AI agents, and trillion-dollar investments from Microsoft, Amazon, Google, Meta, and OpenAI. Yet behind every AI breakthrough lies something surprisingly physical: data centers.

And nowhere tells this story better than Virginia.

Northern Virginia already processes an extraordinary share of the world’s internet traffic. It has become the digital capital of the global economy. But success has created a new economic question that extends far beyond technology:

How much electricity should society devote to intelligence?

This is the question that motivated my latest research project.

As an economics student, I am less interested in whether AI is “good” or “bad.” Economists ask a different question: What are the trade-offs? Every additional data center creates benefits. Construction generates investment. Local governments collect tax revenue. Businesses gain access to cheaper computing power. AI researchers train larger models. Productivity may increase across sectors from healthcare to finance.

But every benefit has a cost.

Electricity demand rises.

Transmission infrastructure expands.

Communities worry about noise, land use, and environmental impacts.

Power grids become more stressed.

The fascinating part is that these costs are often invisible in discussions about AI. When we debate artificial intelligence, we usually focus on algorithms. Economists should instead examine the infrastructure behind those algorithms. Virginia provides a remarkable natural laboratory. It is home to the world’s largest concentration of hyperscale data centers, making it an ideal place to study how AI infrastructure affects electricity demand, local energy markets, and public policy. Rather than asking whether data centers consume “too much” electricity, my research asks a more measurable question:

How does data center expansion change local electricity demand over time?

That distinction matters.

Public debates often become emotional. Economics requires evidence.

Suppose electricity demand increases substantially after major data center investments. That finding alone does not imply data centers are undesirable. It simply tells policymakers that the electricity system must adapt. Likewise, if utilities successfully expand generation and transmission while maintaining affordable prices, the narrative changes entirely.

The objective is not to stop technological progress.

The objective is to understand its economic consequences.

This reminds me of an old lesson in economics.

Every industrial revolution has required a hidden infrastructure.

Factories needed railroads.

The automobile needed highways.

The internet needed fiber-optic cables.

Artificial intelligence needs electricity.

Perhaps AI is not merely a software revolution, but it is an energy revolution disguised as software.

That observation raises several questions worth debating. Should electricity pricing better reflect periods of peak demand? Should hyperscale data centers finance a larger share of transmission upgrades? Can renewable energy alone satisfy future AI demand, or will nuclear and natural gas become increasingly important? How should regulators balance national competitiveness with local community concerns?

These are not engineering questions alone.

They are economic questions.

One of the most interesting developments is that resistance to new data centers is growing even in places that previously welcomed them. Communities increasingly ask whether the local costs are matched by local benefits. Meanwhile, technology firms argue that expanding computing capacity is essential for innovation, productivity growth, and maintaining America’s leadership in AI. Both perspectives contain legitimate economic arguments, which is exactly why the debate deserves careful empirical research rather than slogans.

My hope is that this research contributes, in a small way, to that conversation.

Economics is ultimately about scarcity and choice.

Electricity is scarce.

Land is scarce.

Capital is scarce. I am reminding you of the basic ECON 101 here.

Yet our demand for computation appears almost unlimited.

The challenge for policymakers is not deciding whether AI should exist.

It is deciding how to allocate scarce resources efficiently while ensuring that technological progress remains economically sustainable.

Virginia may be the first place to confront this challenge.

It almost certainly will not be the last.

I’m currently building an empirical panel dataset to study the relationship between data center expansion and local electricity demand in Virginia. I look forward to sharing the results, and I would love to hear your thoughts. Should AI infrastructure be treated like any other industrial investment, or does it require an entirely new framework for economic policy?

This morning started with a cup of black coffee, a bottle of water sitting beside my laptop, and another draft of my Statement of Purpose for PhD applications. As I was taking a short break and scrolling through LinkedIn, I came across a fascinating new paper that immediately caught my attention. I don’t know the authors personally, but their research made me stop reading everything else.

The paper, “Expanding Opportunities: New Facts about the Placement of PhD Economists” (2026) by Eberhardt, Facchini, and Rueda, asks a simple but important question: Where do newly trained economists actually end up working? Expanding Opportunities: New Facts about the Placement of PhD Economists

As someone preparing for a PhD in economics, I found the results both surprising and reassuring.

For decades, many of us have viewed a PhD as a pathway that naturally leads to becoming a professor. Graduate school trains researchers, universities hire researchers, and the cycle continues. That has been the traditional story.

But the data suggest that the story is changing.

The authors assembled an impressive dataset of more than 11,000 PhD graduates from the top 30 economics departments in the United States between 1998 and 2022. The numbers reveal a clear structural shift.

Around 2009, roughly two-thirds of graduates obtained academic positions. By 2021, that figure had fallen to just above one-half. Nearly half of new economics PhDs now begin their careers outside universities.

This should not be interpreted as a crisis.

It is simply economics.

Universities are producing more highly trained economists than academia can absorb. Simple Econ 101:- Supply has expanded faster than academic demand. Markets adjust, and talented people move toward sectors where their skills create the greatest value.

That destination increasingly includes central banks, consulting firms, finance, international organizations, and, perhaps most notably, technology companies.

The rise of technology is the most striking trend.

In 2010, only about 1% of graduates entered technology firms. By 2021, that share had increased almost tenfold to around 12%. Thanks to Microsoft, Google, Amazon, and now OpenAI, Anthropic.

Think about that for a moment.

The same economists who once primarily estimated labor supply elasticities, evaluated public policies, or developed econometric methods are now designing experiments for online platforms, improving recommendation systems, studying labor markets through digital data, optimizing pricing algorithms, measuring marketplace efficiency, and helping firms make better decisions using causal inference.

Economics has not changed.

Its workplace has.

One paragraph in the paper particularly caught my attention.

The authors find that technology is the only sector where gender and ethnicity do not significantly predict placement outcomes. In contrast, academic placements continue to exhibit measurable disparities.

That is an encouraging finding.

It suggests that when firms compete intensely for analytical talent, demonstrated technical ability may matter more than traditional academic networks or historical patterns.

Another result challenges a common misconception.

Technology firms are not simply hiring economists who failed to obtain academic positions.

The opposite appears closer to reality.

Graduates advised by Nobel Prize-winning economists became increasingly likely to enter technology over time. In other words, some of the strongest students are deliberately choosing careers outside universities. It means now the AI companies (Anthropic and OpenAI are watching the candidates, ohhh not only watching, but they are also hiring all the influential economists and making them theirs )

That says something important about where frontier economic research is increasingly being applied.

Today’s largest technology companies run thousands of randomized experiments every year, analyze massive administrative datasets, build structural models, forecast demand, study labor markets, and evaluate policy changes at an extraordinary scale.

For an economist interested in causal inference, machine learning, industrial organization, labor economics, or market design, these firms have effectively become enormous research laboratories.

Of course, the paper also provides an important caution.

The growth in technology hiring is heavily concentrated. More than 40% of technology placements are at Amazon alone. That concentration creates opportunity during expansion but also vulnerability when hiring slows, as the technology sector experienced after 2022.

No market grows forever.

Even so, I think the broader lesson remains unchanged.

A PhD in economics is becoming less about choosing between academia and “leaving economics.” Instead, it is about deciding where economic thinking can generate the greatest impact.

Personally, this paper resonated with me because it reflects many of the reasons I decided to study economics in the first place.

I have always been interested in understanding how incentives shape decisions, how institutions influence growth, and how evidence can improve policy. Today those questions matter just as much inside technology companies, international organizations, central banks, and consulting firms as they do inside university departments.

Perhaps that is the biggest takeaway.

The value of an economics PhD has never depended solely on becoming a professor.

Its value lies in learning to ask difficult questions, construct credible evidence, think rigorously under uncertainty, and communicate complex ideas clearly enough for policymakers, executives, entrepreneurs, and citizens to make better decisions.

As I closed the paper and returned to editing my Statement of Purpose, I felt unexpectedly optimistic.

The destination may no longer look exactly like it did twenty years ago.

But the training itself may be more valuable than ever.

And perhaps that is exactly where the future of economics begins.

For me, honestly, I always wanted to do the research on economics. I do not have a particular point to end up, but any place that can help me grow personally and professionally, I make myself belong there.

Thanks to all……… Source: Eberhardt, Facchini, and Rueda, "Expanding Opportunities: New Facts about the Placement of PhD Economists" (2026)

While doom-scrolling on LinkedIn yesterday, I stumbled upon a fascinating new paper that completely challenged the way I think about AI and the labor market. As economists, we love when data humbles our intuition—and this paper did exactly that.

For the last two years, the dominant narrative has been simple:

AI is coming. Jobs are disappearing. Companies are replacing people with algorithms.

It almost feels like common sense now. My default mindset is always like this, without asking the counterquestion.

And honestly, I believed it too.

A lot of my friends work in tech. Many have computer science or engineering degrees. Several have been laid off recently. Every week, there’s another headline announcing thousands of job cuts from a major technology company, followed by a CEO proudly discussing their AI strategy. Naturally, many of us connected the dots: more AI must mean fewer workers.

Then I saw the figure from this paper.

I had to stop scrolling.

The graph shows something that almost feels backward. Firms making the largest investments in AI actually increased their workforce by roughly 10% over the following two years (I am not saying this; the paper I am writing this article for says this ). Even more surprisingly, entry-level employment increased by around 12% (What? Is this even true?). Meanwhile, firms making only small AI investments showed essentially no statistically significant change in employment. (No way !!!!!)

That result genuinely made me rethink today’s hiring cycle.

Now, before everyone starts celebrating and declaring “AI creates jobs,” let’s slow down for a second.

One thing I really appreciate about the authors is that they don’t oversell their findings. They explicitly point out that firms adopting AI aren’t random. These companies were already larger, more technical, faster-growing, and more likely to be in industries like information technology, finance, and professional services before adopting AI. To deal with this, the paper compares earlier AI adopters with similar firms that adopt AI later rather than simply comparing AI firms with everyone else. That’s a much stronger research design than most discussions happening on social media.

And that’s why this paper deserves attention. ( I don’t say you have to read this, but I can say it’s good to spend some time and understand this phenomenon )

For years, much of the AI debate relied on occupational exposure scores or surveys asking firms whether they “use AI.” This paper takes a different approach. Instead of asking executives, the researchers observe actual AI spending from Ramp’s payment data and link it with workforce records from Revelio Labs across more than 21,000 U.S. firms. In other words, they’re watching what companies actually spend—not what they say in interviews.

As an economist, I find this fascinating.

Economic history teaches us that every major technological revolution initially looks like labor substitution. Steam engines replaced manual labor. Computers replaced clerical tasks. Automation replaced factory work. Yet over time, productivity gains often created entirely new industries and occupations that nobody could have predicted beforehand.

Perhaps generative AI is entering that same phase.

Maybe firms making serious AI investments aren’t simply firing workers. Maybe they’re reorganizing, becoming more productive, expanding output, and hiring complementary talent. The paper even finds employment gains across engineering, sales, administration, customer service, and finance—not just software engineering.

That doesn’t mean nobody loses. ( understand this line )

Some routine jobs will almost certainly disappear. Some entry points into careers may fundamentally change. We’ve already seen painful layoffs, especially in parts of the tech sector. Those stories are real.

But maybe layoffs at a few headline companies aren’t the whole story.

Maybe we’re confusing a reallocation of labor with the destruction of labor.

That’s a very different economic question.

What excites me most isn’t whether this paper proves AI creates jobs. It doesn’t—and the authors are careful not to make that claim. What it does show is that the relationship between AI and employment is far more nuanced than the popular narrative suggests. AI adoption appears to be associated with workforce expansion among firms making substantial investments, but those firms were already different in important ways before adoption.

This is exactly why economics matters.

The economy is rarely as simple as the headlines.

A chart can overturn a thousand opinions.

And sometimes the best research doesn’t give us answers; it forces us to ask better questions.

Huge shout-out to Ara Kharazian, Lisa Simon, and Ryan Stevens for producing such thought-provoking work. Whether future evidence reinforces or challenges these findings, this paper raises one of the most important questions in modern labor economics:

Is AI replacing workers—or is it changing the kind of firms that grow and the kinds of workers they need?

That’s a question worth debating. And more importantly, it’s a question worth measuring with data rather than fear. In the words of Professor Alex Edmans, “Fact is not data”

Today, anyone can generate a 2,000-word essay, a policy brief, a research summary, or even an economic argument in seconds. The barrier to producing words has collapsed. Length is no longer a signal of knowledge. In fact, it may not even be a signal of effort.

What matters now is not how much you write. What matters is whether your ideas can justify the space they occupy.

Economics is a perfect example. Many people can now explain concepts like inflation, comparative advantage, game theory, or monetary policy with impressive vocabulary. AI can provide the definitions, the equations, and even the citations. But knowing the language of economics is not the same as thinking like an economist.

The real question is not whether AI can help us write better. The real question is whether it helps us think better.

Because in a world where everyone has access to the same tools, the scarce resource is no longer information. It is insight. And this is what I didn’t get from the 20-page policy brief and 45-page cited papers generated by the commonly used tool “AI”.

Anyone can ask AI for an answer. Fewer people can ask an original question.

Anyone can summarize a theory. Fewer people can create one.

Anyone can explain what happened. Fewer people can explain why it happened.

The economists, researchers, writers, and leaders who will matter most in the AI era are not those who produce the most content. They are the people who develop frameworks that help others see the world differently.

The future belongs less to the people with the longest reports and more to the people with the strongest ideas.

Because AI can generate words.

But it still cannot replace a genuinely original way of thinking.

Source: NBER

Every generation believes its technology is different.

The steam engine was different after the Industrial Revolution. Electricity was different. Computers were different. Now, AI is different.

Maybe they’re all right. I mean We all are probably right .

But economists have spent decades asking the wrong question.

Instead of asking whether technology destroys jobs, perhaps we should ask:

What new jobs does it create, and who gets there first?

What Makes New Work Different from More Work? Two fascinating new NBER papers by David Autor, Caroline Chin, Anna Salomons, Bryan Seegmiller, Tarek Hassan, Aakash Kalyani, and Pascual Restrepo offer one of the most compelling explanations I’ve seen for why some workers earn more than others during technological revolutions.

Their conclusion is simple:

The biggest rewards do not go to people who know the newest technology. They go to people who learn it before everyone else. And that changes how we should think about inequality, education, and even the future of AI.

The World’s Most Expensive Skill

Imagine it is 1995.

Nobody is hiring “social media managers.”

Nobody is hiring “prompt engineers.”

Nobody is hiring “cloud architects.”

Those jobs literally do not exist.

Then a new technology arrives.

Someone has to figure out how it works.

Someone has to build systems around it.

Someone has to teach everyone else.

At first, only a handful of workers possess that knowledge.

Their wages rise.

Companies compete for them.

Universities cannot produce them fast enough.

LinkedIn becomes a battlefield. ( I need a job, you need a job, she needs a job , everyone needs a job )

Then something interesting happens.

The knowledge spreads.

Courses appear.

Books are written.

YouTube tutorials emerge.

New graduates learn the skill.

The expertise becomes common.

The wage premium shrinks.

The once-exotic skill becomes another line on a résumé.

Economists call this a skill premium.

The new papers suggest that the premium follows a life cycle remarkably similar to that of a product.

Birth.

Growth.

Maturity.

Decline.

David Autor’s Evidence: New Jobs Pay More

David Autor and his coauthors study something surprisingly difficult to measure:

new work.

Instead of guessing, they identify occupations that first appeared in Census records. Think about jobs that were unimaginable decades earlier.

The results are striking.

Workers in these newly emerging occupations tend to be younger and more educated. Even after controlling for industry, occupation, demographics, and geography, new work earns higher pay.

Technology-related new work earns an especially large premium.

But the most important finding is not that new jobs pay more.

It is that the premium fades.

A recently introduced job offers a substantial wage advantage.

A job introduced decades ago often carries little advantage at all.

In some cases, the premium disappears entirely.

In other words:

The value comes from scarcity, not permanence.

Being among the first people who understand a new technology is valuable.

Being the thousandth person is much less so.

Why the 1980s Created Today’s Inequality

The second paper, by Hassan, Kalyani, and Restrepo, zooms out to the entire economy.

Using patent records, Wikipedia entries, and 300 million job postings, they track more than 6,000 technologies over time. Their findings tell an incredible story.

Before 1970, America created roughly 25 to 30 new technologies annually.

Then innovation exploded.

By the late 1980s, the number approached 250.

Think about that.

The personal computer.

Enterprise software.

Semiconductors.

Networking.

The internet.

Mobile communications.

A technological avalanche.

The authors argue that this wave dramatically increased demand for highly educated workers because these workers were better positioned to learn the new technologies first. As a result, the college wage premium surged between 1980 and 2010.

Then something unexpected happened.

The pace of new technology creation slowed.

And the growth of the college premium slowed too.

The model predicts exactly what we have observed in recent years:

A flattening of the education wage gap.

Not because technology stopped mattering.

But because the giant wave of brand-new technologies became less giant.

Why Cities Keep Winning

The paper also helps explain a puzzle that frustrates many policymakers.

Why do large cities keep attracting opportunity?

The answer is not simply that cities are bigger.

It is that new technologies arrive there first.

The researchers find that the average technology used in America’s densest cities is substantially younger than the technology used in less-dense regions.

Cities operate closer to the frontier.

Workers there learn new systems earlier.

Firms adopt innovations faster.

Knowledge spills over between people.

A software engineer in San Francisco might be discussing tomorrow’s technology while someone elsewhere is still mastering yesterday’s.

The result is higher wages and faster growth.

Not because urban workers are inherently smarter.

Because they encounter new work sooner.

What This Means for AI

This is where things become especially relevant.

Everyone is debating whether AI will eliminate jobs.

These papers suggest a different perspective.

The more important question may be:

How much new work will AI create?

History suggests that every major technology destroys some tasks while creating entirely new categories of work.

The workers who figure out these categories first tend to earn extraordinary rewards.

But history also suggests those rewards rarely last forever.

Knowledge spreads.

Training improves.

Universities adapt.

Competition increases.

The premium shrinks.

Today’s AI experts may eventually face the same fate as yesterday’s spreadsheet experts.

The winners are not those who learn a technology once.

The winners are those who repeatedly learn technologies before everyone else.

The real lesson is important to understand

For decades we have argued about whether technology helps workers or hurts workers.

These papers suggest a more nuanced answer.

Technology rewards workers who can navigate uncertainty.

It rewards adaptation.

It rewards learning.

And most importantly, it rewards people willing to step into jobs that do not yet have a clear title.

The future belongs neither to machines nor to workers.

It belongs to workers who can learn alongside machines.

That may be the most valuable skill of all.

Papers that I have used to write this article :

David Autor, Caroline Chin, Anna M. Salomons, and Bryan Seegmiller — What Makes New Work Different from More Work? (NBER Working Paper 34986)

Tarek Alexander Hassan, Aakash Kalyani, and Pascual Restrepo — The Skill Premium in Times of Rapid Technological Change (NBER Working Paper 34939)

Full credit to the authors for the original research. Any interpretation, simplification, or mistakes in this article are my own.

After the game between Spain VS Cape Verde, everyone is talking about the football ( Soccer ) World Cup match .

People are debating tactics, missed chances, controversial refereeing decisions, and dramatic results. But when I watch a football match, I often find myself seeing something else.

I see economics.

Millions of people around the world were talking about Cape Verde. Not because it is a global superpower. Not because it discovered oil. Not because it launched a revolutionary technology.

People were talking about Cape Verde because of the football match against the world number two, Spain.

An island nation of fewer than 500,000 people suddenly occupied the world’s attention. Its goalkeeper, Vozinha, produced heroic performances that transformed him from a local athlete into an international star.

Most people saw a football miracle.

I saw a lesson in economics.

Cape Verde relies heavily on tourism. Beaches, culture, music, and hospitality form the backbone of its economy. Tourism is not simply an industry there; it is one of the engines of national prosperity.

Traditional tourism campaigns cost millions of dollars. Governments purchase advertisements, hire marketing agencies, and compete for visibility in an increasingly crowded world.

But for ninety minutes, Cape Verde received something even more valuable.

Attention.

Millions of viewers who had never heard of the country suddenly searched its name online. They asked questions.

“Where is Cape Verde?”

“Can you visit it?”

“What are its beaches like?”

“How expensive is it?”

“What else is there besides football?”

Economists understand that attention is scarce. In the modern economy, attention itself has become a form of capital. Countries compete for it just as firms compete for customers.

A football match generated exactly that.

One extraordinary performance by a goalkeeper became a national marketing campaign that no tourism board could have designed any better. Of course, not everyone who searched for Cape Verde will buy a plane ticket. But economics rarely works through certainty. It works through probabilities.

If even a tiny fraction of those millions of viewers eventually become tourists, investors, entrepreneurs, or future visitors, the economic returns from that moment could extend far beyond the football pitch.

This is why sports matter.

They are not merely entertainment.

They shape national identity. They create narratives. They influence perceptions. They generate soft power. They attract investment and tourism. They introduce small nations to global audiences that traditional diplomacy often struggles to reach.

The goalkeeper was trying to save goals.

Instead, he may have helped sell a destination.

Perhaps this is why I love economics. ( Tha’s why you should - Arun always suggests )

It teaches us that the world is connected in ways that are easy to miss. A save in a football match can influence Google searches. Google searches can shape travel decisions. Travel decisions can support local businesses, create jobs, and contribute to national income.

The next time everyone is arguing about football, I will probably still be thinking about economics.

Because sometimes, behind a spectacular save, there is an entire economy quietly benefiting from the world’s attention. ( Not every time, but in general, I can say yes)

And perhaps the most valuable export of a small nation is not always what leaves its ports.

Sometimes, it is the story it tells the world. Some stories travel without scripts .

That separation is one of the biggest intellectual mistakes we make.

Economics tells us what incentives people face. Politics decides who gets to design those incentives. ( We need to understand this analogy to understand the depth of the world ). You cannot understand why housing is unaffordable without understanding zoning laws. You cannot understand inflation without understanding central bank independence. You cannot understand poverty without examining institutions, governance, and the political choices that determine who receives opportunities and who bears costs.

Markets do not exist in a vacuum. Every market operates inside a set of rules, and politics writes those rules.

The price of electricity, the quality of your education, the availability of healthcare, immigration policy, trade agreements, environmental regulations, and even the opportunities available to future generations are all shaped by the interaction between economics and politics.

This is why understanding political economy matters.

It protects us from becoming ideological cheerleaders. It forces us to ask difficult questions. Who benefits from a policy? Who pays for it? What incentives does it create? What unintended consequences might follow?

Without this perspective, we become easy targets for slogans disguised as solutions. We confuse good intentions with good outcomes.

Studying the intersection of politics and economics does not tell us what to think. It teaches us how to think and helps us build a logical foundation for ideas and knowledge.

It builds intellectual humility. It sharpens judgment. It helps you distinguish evidence from emotion and incentives from rhetoric.

In a world overflowing with opinions, understanding political economy gives us something far more valuable:

A framework.

And frameworks shape perception.

Perception shapes decisions.

And decisions shape the world we leave behind.

Reading a paper by a Nobel laureate feels like a privilege, but I want to make this one readable for everyone, not just economists (often they write papers for other economists). Acemoglu’s May 2024 NBER working paper asks one deceptively simple question: How much will AI actually grow the economy over the next 10 years? Goldman Sachs says 7% of global GDP. McKinsey says up to $25 trillion. Acemoglu says: probably less than 1%.

There is something refreshing about reading a Nobel laureate tell the world to calm down.

Over the last two years, the conversation around artificial intelligence has been dominated by trillion-dollar forecasts and civilization-changing predictions. Investment banks project an economic boom. Tech CEOs speak of abundance. Venture capitalists compare AI to electricity, the steam engine, and the internet all at once.

Then along comes Daron Acemoglu.

His paper, THE SIMPLE MACROECONOMICS OF AI , does something economists often do when the world becomes intoxicated with a story. It asks a boring question.

What are the numbers?

Not the press releases.

Not the keynote speeches.

Not the demonstrations that go viral on social media.

How much will AI actually increase economic output over the next decade?

His answer is startling.

Probably less than one percent.

Not one percent every year.

Less than one percent in total productivity gains over ten years.

If Goldman Sachs offers champagne, Acemoglu offers a glass of cold water.

And perhaps that is exactly what economics is supposed to do. If I were him, I probably could do the same thing. I was not him, so I am writing this.

The Economics of Humility

Economics has always been divided between two instincts.

The first instinct is optimism.

Adam Smith believed specialization and the division of labor could unleash prosperity on a scale unimaginable in earlier centuries. Joseph Schumpeter celebrated the entrepreneur who destroys the old world to build a better one through innovation. Robert Solow showed that technological progress accounts for much of long-run growth.

This tradition sees technology as liberation.

The second instinct is caution.

David Ricardo worried that machinery could temporarily displace workers. Karl Marx argued that technology under capitalism often strengthens the bargaining power of capital over labor. John Maynard Keynes warned about “technological unemployment.” More recently, economists studying automation have documented how certain technologies increase inequality even as they raise output.

This tradition asks who gains, who loses, and whether growth alone improves human welfare.

Acemoglu’s paper sits at the intersection of these traditions.

He is not anti-technology.

He is anti-fantasy. Which I also don’t like ( You can say F**K, you Arun, I don’t mind )

The Beautiful Simplicity of the Math

The paper relies on one of the most elegant ideas in macroeconomics: Hulten’s Theorem.

The intuition is straightforward.

A technology changes the economy according to two factors:

First, how much of the economy does it actually affect? Second, how much productivity improves in those affected areas.

If AI improves productivity in a small slice of economic activity, even dramatic improvements within that slice produce modest aggregate gains. Acemoglu estimates that approximately twenty percent of the U.S. wage bill involves tasks exposed to AI.

But exposure is not automation.

After accounting for profitability and implementation constraints, only about 4.6 percent of tasks are expected to be economically automated within the next decade.

Average labor cost savings in those tasks amount to roughly 27 percent. The resulting productivity gain is approximately 0.66 percent over ten years.

Suddenly, the miracle becomes arithmetic.

The economy is simply too large and too complicated to be transformed overnight.

Why Economists Keep Making This Mistake

History offers a fascinating lesson.

Economists underestimated electricity.

They underestimated computers.

They underestimated the internet.

But they also overestimated countless innovations that failed to transform productivity.

The lesson is not that economists are bad at forecasting.

It is that technological revolutions diffuse slowly. (I am so bad at predicting; I think it’s a good sign for the economist that is loading ………..)

Paul David famously showed that electricity did not immediately transform productivity because factories had to be redesigned to accommodate the new technology. Simply replacing steam engines with electric motors was not enough.

Brynjolfsson called this the productivity paradox.

Technology often arrives before institutions, business models, regulations, and complementary investments catch up.

AI may follow the same pattern.

Or it may not.

The honest answer is that we do not know.

Where Acemoglu Is Brilliant

The paper’s greatest strength is discipline.

Every claim is traceable.

Every assumption is visible.

Unlike consulting reports that leap directly from capability demonstrations to trillions of dollars in gains, Acemoglu forces us to ask difficult questions.

Can the task actually be automated?

Will firms find it profitable?

How quickly will adoption occur?

Will measured productivity translate into social welfare?

These questions matter.

The distinction between easy and hard tasks is particularly insightful.

Writing marketing copy is relatively easy to evaluate.

Diagnosing a rare illness, mentoring a struggling student, or negotiating a peace agreement is not.

AI excels where success metrics are clear.

Much of human value creation exists precisely where they are not.

But I Think He Underestimates One Thing

New tasks.

This may be the biggest weakness in the paper. (Sorry Professor Acemoglue but I have to point it out )

Economic history is not merely the story of replacing labor. It is the story of inventing entirely new forms of work.

Nobody in 1900 was a software engineer.

Nobody in 1950 managed cloud infrastructure.

Nobody in 1990 optimized search engine advertising.

Nobody in 2005 became a creator managing brand partnerships through social media.

Technological revolutions generate occupations that previous generations could not imagine. Acemoglu acknowledges this possibility.

Then largely sets it aside.

Ironically, his own earlier work shows that new tasks are one of the strongest forces supporting wages and reducing inequality. If AI creates complementary occupations involving oversight, customization, governance, human interaction, creativity, and coordination, then the forecast could prove too pessimistic.

This is where Schumpeter deserves a seat at the table.

Creative destruction is not merely destruction.

The Austrian Challenge :)

Economists from the Austrian tradition might raise another criticism.

Friedrich Hayek argued that knowledge is dispersed throughout society. Entrepreneurs discover opportunities that planners cannot anticipate. Forecasting the uses of AI from today’s task distribution may therefore underestimate the decentralized experimentation that markets generate. Nobody predicted ride-sharing from the invention of smartphones. Nobody predicted influencer marketing from broadband internet. Nobody predicted that entire industries would emerge around app ecosystems. If Hayek is right, the most important applications of AI may be the ones nobody is currently measuring. ( Ryan, if you are reading this, you should be proud because I learned something from you. )

The Keynesian Concern (:

Yet Keynes would ask a different question.

Even if productivity gains remain modest, who captures them?

Acemoglu predicts that capital income rises faster than labor income. The gains from AI may accrue disproportionately to those who own the models, the infrastructure, and the platforms. Low-education clerical and service workers could experience negative wage effects.

Aggregate prosperity does not guarantee broad prosperity.

A society can become richer while many people feel economically insecure.

The distributional question remains central.

GDP Is Not Welfare

Perhaps the most provocative contribution of the paper concerns what Acemoglu calls “bad tasks.”

Some AI applications may increase GDP while reducing human well-being.

Recommendation algorithms that maximize screen time.

Advertising systems that exploit cognitive biases.

Digital arms races between attackers and defenders.

Measured output rises.

Human flourishing may not.

This echoes critiques dating back decades.

I am telling you. Simon Kuznets warned that national income statistics should not be confused with welfare. ( Believe me or not )

Amartya Sen emphasized capabilities rather than consumption alone.

Even Robert Kennedy famously observed that GDP measures everything except that which makes life worthwhile. If AI helps us spend more time addicted, anxious, manipulated, or polarized, growth statistics alone cannot capture the cost.

Economics, at its best, reminds us of this distinction.

So, Who Is Right?

Goldman Sachs projects large gains.

McKinsey envisions trillions in value.

Acemoglu sees a much smaller impact.

My view is that all three may be partly correct.

Acemoglu is probably right about the next decade.

Large economies move slowly.

Organizational change takes time.

Adoption faces friction.

However, the longer horizon is more uncertain.

If agentic AI systems become reliable collaborators rather than isolated tools, if complementary innovations emerge, and if entirely new tasks are created, then today’s cautious estimates may eventually look conservative.

The question is not whether AI matters.

It already does.

The question is whether this moment resembles the spreadsheet, the internet, or electricity. Or whether it resembles the many technologies that inspired excitement without transforming the economic foundations of society.

The Real Lesson

The most valuable contribution of this paper is not the number 0.66 percent.

It is the intellectual posture behind it. Economics should neither worship innovation nor fear it.

It should investigate it :)

It should ask who benefits, who bears the costs, and what assumptions connect today’s headlines to tomorrow’s outcomes.

In an age of hype, rigor becomes radical.

So, Acemoglu’s message is ultimately simple. The future of AI is not predetermined by technological capability. It will be shaped by institutions, incentives, policies, firms, and human choices.

Technology does not write economic destiny.

People do.

And that may be the most important economic lesson of all.

My Note: The debate over AI is increasingly becoming a debate between different schools of economics. The neoclassical economist asks, "What are the productivity gains?" The Schumpeterian asks, "What new industries emerge?" The Austrian asks, "What entrepreneurial discoveries are impossible to foresee?" The Keynesian asks, "Who receives the gains?" The institutional economist asks, "Which rules shape adoption?" The welfare economist asks, "Does GDP actually improve human lives?" The answer to AI's future likely lies not in any single framework but in the conversation among them.

Good night if you are reading at 10 PM ………………………………………………………………..

After reading the Harvard Business Review article “Why People Create AI ‘Workslop’ and How to Stop It” by Kate Niederhoffer, Alexi Robichaux, and Jeffrey T. Hancock (January 16, 2026), I started thinking about a quiet shift taking place across modern workplaces. The conversation around artificial intelligence often celebrates speed, automation, and productivity. Still beneath this excitement lies a growing concern: the rise of what the authors call work-slop, AI-generated work that appears finished yet is devoid of depth, effort, or real understanding.

This phenomenon goes beyond low-quality output. It indicates a deeper transformation in how people engage with thinking itself. Generative AI tools are incredibly powerful. They can produce structured reports, summaries, code, and emails within seconds. However, the ease of generating content is beginning to blur the line between producing information and understanding it. In many cases, AI is not being used to sharpen ideas but to bypass the intellectual process required to develop them. After reading this article, I was thinking about the Arkalof paper “The Market of Lemons.” If nobody knows the quality, only lemons are in the market because good-quality goods don’t want to be undervalued.

From history writing, analysis, and research, these have played an important role in shaping thinking. When individuals write a report or analyze data, they often discuss the logic of their argument; eventually, they try to refine their assumptions and clarify their conclusions. These cognitive struggles are not inefficiencies; they are the very mechanisms by which knowledge is formed. Often, my professors tell me the same thing: AI applications are really important, but it is also important to understand the logic behind the work . I often use ChatGPT & Claude, but when I analyze their patterns, I feel it’s important to understand their capabilities and know when to use these tools. AI work doesn’t have intrinsic value because I always think of this as an intermediate good, not the final product of my work.

AI-generated outputs often sound sophisticated, but they can contain vague reasoning, recycled phrases, or surface-level insights. When these outputs circulate in teams, the burden of interpretation and correction shifts to the reader. What appears efficient on the sender’s side often creates hidden costs for everyone else. More importantly, work slop can quietly weaken trust within organizations. Colleagues can often sense when work lacks genuine effort or intellectual ownership. A performance review that feels algorithmic or a report filled with generic language signals something subtle but powerful: the human behind the document disengaged from the thinking process. Over time, this weakens the spirit of accountability and skill that strong institutions rely on.

The solution is not to reject AI. Instead, organizations must rethink how these tools are embedded into daily work. AI should function as a partner that expands human thinking, not a shortcut that replaces it. If applied thoughtfully, AI can help generate ideas, test assumptions, and organize complex information. But the responsibility for interpretation, judgment, and reasoning must remain human.

The future of work will not simply be defined by how fast we can generate content. It will be defined by whether we preserve the discipline of thinking deeply in an age when machines can generate answers instantly.

Yesterday, while reading the Wall Street Journal, I found an interesting thing that made me pause on a line from Greg Ip that felt deceptively simple but deeply important. Also, I have seen these in some of my LinkedIn connection post so I feel to write something on this.

He was pushing back against the familiar doomsday narrative around AI and jobs, which is America’s bestselling and over-consuming news nowadays. The argument, in short, was this: for the most extreme AI fears to materialize, the market economy itself would have to stop functioning the way it always has. And no evidence is happening now or at any point in U.S. history.

That framing stuck with me. Wait a minute …….

We’ve heard versions of this story before. Math teachers were supposed to be replaced first by Calculators, then by AI. It didn’t happen. I still prefer to sit in the class and take the notes written by my prof who writes on the whiteboard. The common thread is not technological failure; it’s human behavior. Businesses are risk-averse. Consumers are creatures of habit. Efficiency gains don’t automatically translate into mass displacement; they usually get absorbed slowly, unevenly, and in ways that preserve trust and continuity.

Even if we assume a more aggressive situation where AI destroys more jobs than it creates, the macroeconomic conclusion still isn’t a collapse. The money saved doesn’t vanish. It gets spent elsewhere. That’s why individual sectors can experience pain while the broader economy continues to grow. Structural adjustment, not systemic breakdown.

The discussion of AI and its Impact on the labor market is a new story which has different parts and is highly anticipated, where we won’t get any conclusion in the near future, but still we like to predict until we get proven wrong.

I am a graduating senior, and I am worried about my acceptance in the T20 PhD in Economics. When people talk about Economics PhD admissions, numbers get thrown around casually: “Top programs admit 3–5%,” “It’s harder than med school,” or “Only geniuses get in.” I wanted to go far and see and analyze acceptance rates for the Top-20 Economics PhD programs. What I found was revealing not just about selectivity, but about the limits of publicly available data. First, transparency matters. Universities do not publicly publish year-by-year PhD applications and acceptances. There is no centralized, scrapable database showing how many students applied to Harvard Economics in 2011 or how many offers Chicago made in 2020. Admissions offices treat these numbers as internal data.

I explored every realistic alternative.

The American Economic Association (AEA) publishes excellent annual surveys on economics Ph.D. programs. These reports provide aggregate statistics on total applicants, total admits, funding rates, and demographics across all participating programs. What they don’t provide is school-specific or year-by-year acceptance rates. You can learn about trends in the discipline, but not how selective a particular Top-20 program is in a given year. I also examined IPEDS (NCES) data. IPEDS is reliable and public, but it tracks enrollment, not applications or offers. You can see how many Economics PhD students are enrolled in a department, but not how many were rejected to get there. Enrollment is an outcome, not an admissions probability. But I went beyond that and tried to find the PhD student numbers on their websites and on some universities’ pages, where they list the number of applications they received and admitted. I collected as much data as I could and analyzed these trends.

Some individual universities release partial statistics in annual reports, but these are inconsistent, non-standardized, and often discontinued after a few years. Community-compiled sources like the Economics Job Market Rumors (EJMR) wiki exist, but they are unofficial and subject to reporting bias.

The conclusion is unavoidable: precise acceptance rates for Top-20 Economics PhD programs cannot be directly observed, but the real conclusion is that acceptance rates are decreasing. Any serious analysis must rely on carefully constructed estimates, grounded in AEA aggregates, cohort sizes, funding constraints, and historical placement patterns, not scraped “facts.”

Note: The data that I have collected is not the evidence. Please do not make a conclusion by accepting this information. This is just for informational purposes.