Tyler Cowen, a well-known economist, co-authors a blog called “Marginal Revolution” that covers a wide variety of topics related to economics, including politics, history, and sociology.

Like many others, Cowen has been seduced by the superficial brilliance of ChatGPT and other large language models (LLMs). In February 2023, he wrote an 811-word piece titled, “Who was the most important critic of the printing press in the 17th century?”  His surprising answer was the great English philosopher and statesman, Francis Bacon. Cowen’s conclusion was buttressed by fake quotations and made-up references, such as “the multiplication of books is a burden of the world” (Bacon, 1605, Book I, Chapter VIII, section 2). This malarkey was, of course, a “hallucination” generated by ChatGPT.

Undeterred by this deception, in August 2024, Cowen lauded the ability of an LLM-based program called AI Scientist to generate bogus, publishable science papers at a cost of $15 per paper:

I’ve said it before, and I’ll say it again. The marginal product of LLMs is when they are interacting with well-prepared, intricately cooperating humans at their peak, not when you pose them random queries for fun.

Never mind that the ability of LLMs to generate fake science papers is a negative, not a positive, when it comes to economic progress.

More recently, in a February 12, 2025, blog, Cowen argued that the principle of compound interest ensures that LLMs will soon have a substantial effect on economic growth. We are not alone in pointing out that LLM progress has reached a point where it seems to be governed more by the law of diminishing returns than by the principle of compound interest.

Law of diminishing returns

The current evidence, acknowledged even by LLM developers, is that the accuracy is still unacceptably low and hallucination rates are still unacceptably high for real-world applications that have serious consequences. They also acknowledge that improvements have slowed down dramatically—by some metrics, the newest versions are worse than earlier versions (see here and here).

Here, hot off the presses, a March 1, 2025 story reports that,

Using SimpleQA, the company’s in-house factuality benchmarking tool, OpenAI admitted in its release announcement that its new large language model (LLM) GPT-4.5 hallucinates — which is AI parlance for confidently spewing fabrications and presenting them as fact — 37 percent of the time.

Yes, you read that right: in tests, the latest AI model from a company that’s worth hundreds of billions of dollars is telling lies for more than one out of every three answers it gives.

As if that wasn’t bad enough, OpenAI is actually trying to spin GPT-4.5’s bullshitting problem as a good thing because — get this — it doesn’t hallucinate as much as the company’s other LLMs.

Regarding the value of current LLMs, Cowen’s wide-eyed jargon is remarkable:

In my view, their output is good enough that REinforcement Learning can work, synthetic data can work, time scaling can be scaled, they can grade themselves, and they are on a steady glide toward ongoing self-improvement…. I believe, for one, that this will move the AIs into the realm of being truly innovative. Stack them, have some of them generate a few billion new ideas a week, have the others grade those ideas…etc.

The overall impression Cowen gives is that LLMs are great and getting better quickly, and that the biggest impediment to LLM progress is the reluctance of people “even at top universities” to acknowledge how useful LLMs are.

In a February 23, 2025 post, “Why I think AI take-off is relatively slow,” Cohen expands on his lament that the bottleneck to LLMs improving economic growth is that humans are reluctant to accept the productivity-enhancing power of LLMs.

The role of free markets

Ironically, Cowen is a big believer in free markets. One reason a free market works better than a centrally planned economy is that, in a free market, businesses are free to incorporate new technologies that provide them with benefits outweighing the costs. Over time, successful technologies and processes are copied and adopted by others. (This adoption and diffusion has repeatedly been described and confirmed empirically; for example, by Zvi Griliches with hybrid corn and by Edwin Mansfield with 12 different technologies).

If LLMs were as reliably useful as Cowen alleges, businesses would be using them to generate profits faster than LLMs generate text. They aren’t. The slow adoption of LLMs is because the payoffs are still dwarfed by the costs (here and here).

For society as a whole, LLMs are consuming vast amounts of resources with little to show for it. They are okay at customer service and other narrow tasks. They are excellent at generating spam and disinformation. They are great at inducing wide-eyed wonder and funding. But that’s about it. So far, AI is dragging down economic growth by diverting so much human talent and natural resources away from more productive uses.