Are We Missing the Real Danger from AI?

What happens when no one can afford to buy what robots produce?

Photo by Steve Knutson on Unsplash

Economics Matters — Blog/Podcast/Financial Riddler/MaxiFi Puzzler

Please sign up for a free subscription to all four products at larrykotlikoff.substack.com. I’ll immediately make it a lifetime subscription. But please support Economics Matters. Just unsubscribe and resubscribe here on a paid basis. Paid subscriptions come with a one-hour, free financial consultation. Email me to find a time at kotlikoff@gmail.com. You can unsubscribe and resubscribe for free. Also, please sign up kids, friends, colleagues, students, … !

If We All Lose Our Jobs, Will Demand Destroy Supply?

Jean-Baptiste Say is the famous French economist who, in his 1803 Treatise on Political Economy, proclaimed what came to be called Say’s Law.

Supply Creates Its Own Demand

The phrase captures the point that if you produce something you have the ability to sell/swap it for something that someone else produces. Thus, in producing shoes for Sally, Joe has the wherewithal to demand something, say, pizza, from Sally in exchange. If the swap occurs, the value of what’s supplied obviously equals the value of what’s demanded. If the swap doesn’t occur, Joe ends up stuck with and, thus, by default, demanding the shoes he made. And Sally ends up stuck with (demanding) her pizza. Again, supply equals demand.

Why is this tautology considered economics, let alone an economics “law”?

It tells us something about the operation of markets. If you take the initiative to produce something of value to others, others will have an incentive to produce things of value to you. But this raises the coordination problem. If Joe and Sally aren’t sure the other will demand what they supply, both may simply choose to pick wild berries for a living.

Concern with insufficient and uncoordinated demand is at the heart of Keynesian economics. Interestingly, deficient demand can arise in a very different context, namely the one we’re experience now, in which code is taking our jobs or at least our good jobs. In the process, it’s limiting our ability to demand what code can supply.

Will AI Leave Us Picking Berries?

I asked AI (Chatgpt, in this case) how many US jobs it intends to destroy over the next two years. It thought a nanosecond and told me 2 million. That’s more than 1 percent of our nation’s workforce. I then asked AI “If AI destroys too many jobs, will it destroy the economy?” Its answer was “Yes.” followed by “It has the potential to reduce consumer spending.”

AI wasn’t hallucinating. It understands the root problem that it presents. That problem is not rogue robots killing the human race. It’s code, embedded in machines, that does to humans what the combustion engine did to horses — makes us redundant. A century ago, 25 million horses were earning a daily vegan paycheck. Today, there are fewer than 7 million horses, almost all of whom are out to pasture and on the dole.

Code is my codeword for AI (e.g., the neural networks underlying LLMs), instructions for robotic assembly arms, rules for self-driving vehicles, software controlling robo lawnmowers, the QR sensors at self-checkout counters, the GPS signals directing our commutes, our devices and internet connections that let us face-time with anyone, anywhere, at any time, self-correcting drone-navigation systems, and, well, add to the list.

Most academic economists appear rather sanguine about codemization — code-based automization. This equanimity reflects a seminal paper — “The Race between Man and Machine” — written by my brilliant former colleague, Pasqual Restrepo (now at Yale), and beyond brilliant economics Nobel Laureate, Daron Acemoglu (at MIT). The paper presents a task model of automation. The model assumes that humans alone can develop new tasks, be they producing new products or improving existing products. Yes, with time, new tasks are automated and performed at scale. But continual product development and improvement ensures ongoing demand for the unique thing humans supply — creativity.

Say it is so.

It ain’t.

Pasqual and Daron’s paper was written eons ago — in 2018. Today, AI is at the forefront of product and task innovation. I asked Chatgpt, in this regard, whether AI can produce new drugs. Its answer:

Yes, artificial intelligence (AI) has made significant inroads into the field of drug discovery and development, including the potential to invent new drugs. … AI models can predict key properties of potential drug candidates, … helping prioritize the most promising molecules and reducing the need for extensive experimental testing.

For its part, Wall Street has invested over $1.6 trillion in AI in the past decade. Goldman Sachs foresees a 7 percent higher level of AI-generated global GDP over the next decade despite the IMF’s predicting that AI will undermine 40 percent of global jobs.

“Not so fast,” says Acemoglu, in a recent article. He estimates that AI, including generative AI, which creates its own content, will have only a modest impact on the world economy in the near term. In his words,

To put it simply, it remains an open question whether we need foundation models … that can engage in human-like conversations and write Shakespearean sonnets if what we want is reliable information useful for educators, healthcare professionals, electricians, plumbers and other craft workers.

Clearly, no one knows for sure where our increasingly coded world is headed nor when it will get there. Nor do economists know the right way to model the infusion of code into our personal lives, the goods and services we consume, or the means of producing our world’s voluminous list of products. Yet, the Acemoglu-Restrepo (AR) task model has captured the profession’s favor — for two reasons. First, economists like mathematical elegance and the task model is highly elegant. Second, economists favor models with happy endings — in this case, long run (steady-state) outcomes in which humans and code are productive complements. Permanently granting mankind exclusive power to innovate permits steady-state mutually beneficial economic co-existence.

Of course, economists don’t all agree. Many can’t even agree with themselves. Personally, I’m a task-model sceptic. Generative AI, including the ability to write its own code, can already out-innovate homo sapiens in a range of tasks, including development/discovery of certain types of new products. And the share of humans that can invent new tasks/products is distressingly small. Finally, in favoring elegance over reality, the AR framework ignores codemization’s major downside — impoverishing future generations to the benefit of current generations and, thereby, limiting their demand for what code can supply.

Robots Are Us

“Robots Are Us” (RAU) is the title of a paper I co-authored with Seth Benzell and Guillermo Lagarda (both former PhD students) and Jeff Sachs. Seth is at Chapman University, Guillermo is at the InterAmerican Development Bank, and Jeff is at Columbia University. AR is a single agent model, whereas RAU is a life-cycle model. The two have very different production and preference structures.

AR treats output as the combination of tasks and treats households as infinitely-lived agents — so called single agents — who care about their progeny. This assumed intergenerational altruism leads current households to automatically share the economic gains from codemization with economic losers — the young and future generations who this form of technological change will surely, in the main, make redundant. In spreading their wealth intergenerationally, primarily via bequests, AR’s benevolent single agents ensures that tomorrow’s job- and labor-income losers will have the means, in the form of inherited wealth, to purchase (demand) what code can supply.

Including single agents in macroeconomic models is commonplace. It simplifies things. And simplification is the sine qua non of mathematical elegance and computational ease, each of which is highly favored in the academic rat race — getting your work published. The messy/difficult alternative to effectively assuming one intergenerationally beneficent social planner is keeping mathematical track of a plethora of current and future generations, each of which is solely out for itself.

Unfortunately, such selfish, “life-cycle agents,” not pretend economic gods, are the actors that inhabit our planet. Indeed, it’s hard to find a shred of evidence in advanced Western economies supporting widespread intergenerational altruism — caring for the next generation. Instead, take as you go is the operative policy, with older generations leaving their heirs massive fiscal, environmental, and other burdens, including automation.

Robots Are Us (RAU) posits selfish generations and makes the technologically limiting assumption that output is produced purely by code and capital (physical stuff). As for humans, they work when young and play pickleball when old. Each generation of young workers can either produce new code or prayers (serve as priests). But regardless of their profession, they care nothing about their progeny. Their goal is to earn as much as possible when young and spend every penny before meeting their maker.

Goosing Up Technology

The model’s economic excitation and potential doom arises from technological change that extends the longevity, quality, and capacity of code. Real world examples include the development of transistors, the invention of the integrated circuit, the creation of the microprocessor, the expansion of semiconductor memory, AI transformers, etc. When this happens, there is a boom in the market for coders. Intuitively, if your code can suddenly last longer, do more, and even endogenously update itself , your enhanced capacity to supply it will elicit higher demand.

Higher wages paid to early coders poised to benefit from the codemization breakthrough leads their brother priests to unrobe and grab their laptops. In the short run, there are more coders producing more durable, better, and, even self-generating code — all earning far more money. Those that remain in the priesthood face less competition and benefit equally in terms of wages.

The improvement in code leads to an extended economic boom. There is more output and more saving by the young out of their higher wages. This spells more investment in capital for use with the improved code, spelling yet more output. Times are great until something unexpected happens — the demand for coders plunges as the stock of existing code accumulates and the need for new code and new coders falls.

This is the meaning of our paper’s title — Robots Are Us. The code embedded in robots and all other smart machines is the stuff of humans. It comprises code that was primarily or largely written by humans who are now dead. Consequently, new coders effectively find themselves competing with dead coders. A decline in the demand for coders appears, by the way, already underway. According to the Washington Post, employment of computer programmers is at a 45-year low. Certainly, no one is hiring young programmers to code chess programs to beat humans.

In 1997, IBM’s Deep Blue computer, built by currently dead or effectively dead (retired) IBM coders, beat Gary Kasparov, then the world’s top-ranked chess player, in just 19 moves. More important, as of 1997, we haven’t needed a better computer program to beat humans. Deep Blue, fully manned by the living dead, is ready and able to beat any warm-bodied one of us through the end of time.

Deep Blue’s code has turned us into useless horses when it comes to developing better chess-playing programs. Indeed, humans no longer compete with computers in chess. Instead, LLMs meet each other, electronically, in tournaments, with, no doubt, both winners and losers reprogramming themselves autonomously the second each match is completed.

Codemization, Deficient Demand, and Economic Immiseration

In RAU, economic boom is followed by economic bust — or, in the extreme, complete economic collapse — in a true Kondratieff business cycle. The eventual decline in wages — as current coders compete, in effect, with dead coders and, as we’re now seeing, existing code codes for itself — limits what future young generations can earn, save, and invest. This ultimately leaves the economy with less capital since in a life-cycle model each generation resupplies the economy with capital based on their saving when young. But what happens to the capital accumulated by older generations. They spend it on consumption before taking their leave. And if the government threatens to confiscate and preserve capital holdings or if older generations own too much to spend before they end? They move their capital offshore, leaving it for their issue to spend over time. This is particularly likely if code ownership is highly concentrated, which, of course, is exactly the case.

As RAU shows, if capital’s reduction — crowding out is the inside-baseball term — is sufficient, there will be less output despite the increase in quality and quantity of code. In short, the economy can experience immiseration with those alive in the long-run worse off than had the code improvements never occurred.

RAU has a specific structure, but variants, which, in the extreme, feature production based solely on code — imagine a robot able to transform air molecules into other physical entities — can also feature long-term economic doom.

The story here entails demand limiting/destroying supply. In this case, future young generations can’t afford to save enough — demand enough capital to bring into old age — to support the economy’s prior level of output. This process won’t, in general, lead the economy to totally implode. Over time it will, ignoring extreme cases, settle down to a permanently lower level of output and economic welfare. The arguable silver lining here is that a larger share of young people return to doing what humans in RAU are uniquely able to do — pray.

RAU was written a decade ago. I’m hoping it will be published soon. It’s been rejected by one journal after the next as too conjectural. Let’s hope that’s true, but crowding out and immiserizing growth are old themes in economics. Codemization appears to combine both elements, in messy and subtle ways, to the enormous peril of our supposedly most precious possessions — our children and grandchildren.

---