From the Archives-written by ChatGPT.

Both became dramatically more confident in AI’s capabilities while remaining optimistic about its economic potential. Their differences concern how benefits arrive, how seriously to weight emerging dangers, and whether AI might possess consciousness. Cowen emphasizes institutions, bottlenecks, and geopolitical competition. Tabarrok emphasizes inexpensive intelligence, scientific discovery, and—especially in his later writing—concrete safety failures.

The dated passages below trace those changes across the public archive. Quotations are the authors’ own words, including excerpts from their columns; outside material they reproduce is identified separately.

Early views on unequal gains and abundance

Cowen’s early framework grew out of chess. On March 11, 2011, he observed that stronger computers required stronger human partners:

“The required education level — for the team’s ‘wage premium’ — is ratcheted up.”

His concern was already distributional: computers could increase productivity while changing which human skills commanded a premium. He also emphasized learning how to work with computers. What I learn from chess and computers.

Tabarrok’s April 12, 2016 argument was more explicitly reassuring:

“AI will be cheap and ownership will be widespread.”

He compared AI to an increase in the world’s supply of educated people: more intelligence produces useful ideas, and competition spreads the benefits to consumers. But his forecast of comprehensive substitution was cautious—AI becoming a perfect substitute for human intelligence could be “100 years or more” away. Will the proliferation of affordable AI decimate the middle class?

Neither position was absolute. Cowen warned in February 2017 that the Industrial Revolution involved decades of painful adjustment; historical progress did not guarantee quick wage gains. Tabarrok, in October 2019, praised the argument that mediocre automation could displace workers without creating enough additional wealth to compensate. See Cowen’s historical warning and Tabarrok’s Beware the Mediocre Robots!

Rising confidence in AI capabilities from 2020 to 2025

Cowen was already optimistic about language models in July 2020. After publishing a skeptical reader’s assessment of GPT-3, he added:

“I remain bullish, but it is always worth considering other opinions.”

That attribution matters: the skeptical arguments in the post were the reader’s, not Cowen’s. The case against the import of GPT-3, July 23, 2020.

Tabarrok’s shift toward near-term transformative AI became unmistakable by March 2, 2023:

“I never expected to witness the birth of aliens.”

Later that month, he endorsed Microsoft researchers’ description of GPT-4 as exhibiting an incomplete form of general intelligence: “They are correct.” This is a substantial acceleration from his 2016 outlook, although perfect substitution for humans and early general intelligence are different thresholds. See AGI is Coming and Baby AGI is Here, March 24, 2023.

Cowen reached an especially explicit formulation on April 16, 2025, discussing o3:

“I think it is AGI, seriously.”

But he immediately separated capability from adoption:

“It still will take us a long time to use it properly.”

His growing enthusiasm about the technology did not imply an equally dramatic forecast for immediate economic growth. o3 and AGI, is April 16th AGI day?

Economic transformation and the pace of adoption

On January 6, 2025, Tabarrok wrote:

“The intelligence revolution is going to be bigger, more impactful and more wrenching than the industrial revolution.”

This is his own concluding sentence, following a quotation from Sam Altman. The Intelligence Revolution.

Cowen’s February 23, 2025 response to rapid-takeoff expectations focused on a different constraint:

“Human bottlenecks become more important, the more productive is AI.”

Energy, regulation, organizational habits, and weak human collaborators can constrain the economic value of better models. He offered an explicitly tentative estimate of an additional half percentage point of annual growth—large when compounded, but far from an immediate economic explosion. Why I think AI take-off is relatively slow.

By September 3, 2026, Tabarrok was developing the distributional implications:

“A smaller share of a much larger economy can still mean more income for labor.”

He argued that even substantial displacement could leave aggregate labor income higher, or make compensation affordable. He explicitly acknowledged painful losses for particular workers. His high-growth calculations are scenarios, not forecasts. How Much Redistribution Will AI Require?

They also converge more than their rhetoric suggests. In May 2026, Cowen rejected mass unemployment while identifying transitional hiring problems. In August 2026, Tabarrok interpreted business-survey evidence as consistent with their shared position that AI had not yet broadly reduced employment. See Cowen, May 13, and Tabarrok, August 31.

Growing attention to safety and institutional limits

Cowen stated his central position on November 19, 2023:

“I view AI as more likely to lower than to raise net existential risks.”

His reasoning was comparative: the world already faces serious dangers, and stronger science could improve our defenses. My summary views on AI existential risk.

Tabarrok’s later writing places more weight on specific vulnerabilities. On April 8, 2026, he described:

“the worrying combination that very powerful AI’s are very vulnerable.”

He was discussing cybersecurity capabilities alongside agents’ susceptibility to manipulation. AI Risks.

On September 17, 2026, he sharpened the institutional objection to assuming that beneficial science will automatically offset dangerous capabilities:

“But faster science does not automatically mean faster defense.”

This can be read as a qualification of Cowen’s earlier optimism: defensive discoveries must pass through institutions that may move slowly. It also applies Cowen’s own bottleneck logic to safety. Pandemic Preparation is AI Safety.

Cowen nevertheless supports safeguards. In August 2026, he proposed government-overseen, industry-led safety audits. In September, he opposed restrictions on recursive improvement as impractical to enforce internationally. His position is therefore selective regulation combined with continued development. See his August 27 proposal and September 22 critique.

The September 2026 exchange about AI leaders

On September 16, 2026, Tabarrok rejected the claim that executives’ safety warnings were merely marketing or regulatory capture:

“Amodei actually believes what he’s saying.”

He did not thereby establish that the warnings or proposed policies were correct. His argument concerned sincerity and incentives. What Regulatory Capture Actually Looks Like.

Cowen replied the next day:

“I agree with Alex that the standard model of regulatory capture does not apply here”

But he emphasized competition for power: each executive sincerely thinks their own company should win. For Cowen, that complicates treating their warnings as disinterested policy advice. They agree that concern is sincere; they differ over which incentives best explain behavior. The correct model of AI CEO behavior, September 17, 2026.

Consciousness and Tabarrok’s warning about infant pain

On March 17, 2025, Tabarrok’s What Did We Learn From Torturing Babies? drew a moral lesson from medicine’s historical dismissal of infant pain. He described how theories about immature brains allowed physicians to discount babies’ observable distress and perform procedures without adequate anesthesia. His warning explicitly extended to judgments about animals and AI:

“If we must err, let us err on the side of compassion.”

Tabarrok’s position is stronger than mere openness to AI consciousness: he argues for moral caution under uncertainty. The analogy does not establish that AI suffers. It challenges confidence that a theory ruling out suffering is sufficient grounds to disregard possible evidence of it.

On October 14, 2025, he went further in his own commentary following an essay by Jack Clark:

“these systems are becoming self-aware”

He distinguished self-awareness from sentience, while warning against confidently dismissing the latter. His link accompanying that warning points back to the infant-pain essay, making the connection explicit. We Turned the Light On—and the AI Looked Back.

Cowen’s position is much more categorical. On August 31, 2026, he wrote:

“I am very much opposed to the view that the AIs are sentient, or might be sentient.”

He treats AI sentience as a category error, although he accepts that human-like descriptions can sometimes help predict AI behavior. Anthropomorphizing AI?

The disagreement concerns both what AI might experience and how to act when we cannot confidently know. Tabarrok’s historical analogy places greater weight on the moral cost of mistakenly denying consciousness; Cowen expresses much greater confidence that sentience is the wrong category to apply. These are their philosophical assessments, not established findings about AI consciousness.

The post How Tyler Cowen and Alex Tabarrok Have Changed Their Views on AI and Where They Disagree appeared first on Marginal REVOLUTION.

#Tyler #Cowen #Alex #Tabarrok #Changed #Views #Disagree