MarginalRevolution, the long-running blog of Tyler Cowen and Alex Tabarrok, recently published their archive so that AI agents can more easily access their content. As an example of an AI analysis, Alex published a post on how they Changed Their Views on AI and Where They Disagree, but it’s very brief.
As is becoming my habit when George Mason economics professors publish their AI analyses, I decided to investigate further. I asked ChatGPT for Work to investigate Tyler’s views over time (and how they held up) for 6 questions:
How rapid will AI progress and how capable will it become?
How dangerous could AI be, should we concerned about catastrophic harm from AI?
How harmful could AI be, should we be concerned about its effect on youths, or on education, etc.
Would it be a good idea to do strong regulation of AI or slow down its development?
Would people and politicians support regulating AI? Is it practical to push for that?
What to think of AI and China? Is there any way the US could coordinate regulations or slowdowns with China?
The main analyses ran in approx. 20 minutes on GPT 6.1 Sol Extra high. Below is its summary. See also the full report, with citations.
Five surprises from a decade of posts, columns, and interviews
Tyler Cowen’s AI commentary is prolific, provocative, and sometimes cryptic. Looking back over his writing from 2015 through October 2026, the biggest change is in how much intelligence he expects from the technology. His instincts about markets, institutions, and government power have been much steadier. Five findings stood out.
1. His change on AGI was substantial
In December 2022, Cowen wrote: “No, it is not converging upon human-like intelligence or for that matter AGI.” In April 2025, discussing o3: “I think it is AGI, seriously.”
Definitions of AGI complicate the comparison. Still, he explicitly acknowledged that progress had exceeded his expectations. He had been bullish on useful language models for years, but his assessment of their intellectual reach changed considerably. Throughout, he distinguished smarter models from faster economic growth: institutions and human bottlenecks could slow the latter.
2. One of his best warnings came in 2016
Long before ChatGPT, he imagined an AI mimicking your voice and calling your loved ones. In 2023, the FTC warned about that precise family-emergency scam mechanism.
That is a useful corrective to the idea that skepticism about AI extinction means dismissing AI harm. Cowen has worried about fraud, manipulation, and military misuse. His distinctive bet is that AI’s benefits will outweigh its dangers at the level of civilization.
3. His education optimism comes with a firm limit
In 2023, he predicted personalized child assistants and faster learning. By January 2026, he was emphatic about requiring students to “write (and thus think) without AI.”
That tension matters. A randomized high-school mathematics study found that unrestricted GPT assistance improved practice performance while hurting later unaided exam performance; a tutor designed with safeguards largely avoided that loss. Better answers during homework and better learning are different outcomes. His broad educational optimism needs those design details.
4. He has supported targeted regulation for years
Cowen endorsed privacy standards for children’s interactions with robots in 2017 and considered targeted liability in 2023. In August 2026, he proposed an industry body, authorized and overseen by government, to audit models and offer qualifying companies liability protection.
His consistent opposition has been to broad slowdowns and intrusive controls. The interesting debate concerns whether his preferred safeguards would work, including whether industry auditors would be sufficiently independent.
5. Some bold China claims have aged poorly
His March 2023 headline, “Yes, the Chinese Great Firewall will be collapsing,” remains unfulfilled in the evidence reviewed through October 2026. He supplied no deadline, so this is an outstanding prediction rather than a cleanly resolved wager.
His claim that no international cooperation options were on the table was also too broad. China and the US subsequently joined the Bletchley Declaration. That demonstrates limited cooperation; it does little to establish the feasibility of a verified development slowdown.
Cowen’s strongest calls concern useful capabilities and concrete mechanisms of misuse. His social and geopolitical extrapolations are less secure. And his biggest safety claim—that AI is more likely to lower than raise net existential risk—remains unresolved. Progress in AI capability cannot, by itself, vindicate that judgment.
The underlying research report contains 174 annotated entries, a timeline, and source links. Assessment cutoff: October 29, 2026.