S7 Ep20: How quickly should we adopt AI?

Episode Summary

Executive Summary: Joshua Gans argues that AI regulation should focus less on pausing development and more on learning quickly and safely about harms. He challenges the 2023 pause letter, warning that broad delays can hurt startups and slow beneficial uses, while advocating targeted, reversible regulation and smaller-scale experimentation to understand risks and unintended consequences.

Main Topics: Why the AI pause letter was unusual (Priority: 5/5): Gans explains that calls to pause AI development were surprising because innovation economics usually suffers from too little adoption and too little information, not too much speed. Learning about AI harms through adoption (Priority: 5/5): The core research question is how regulators can learn about harms such as misinformation and labor-market effects; Gans argues that real-world deployment may be necessary to observe full effects. Regulatory uncertainty and startup costs (Priority: 5/5): He warns that broad AI rules create uncertainty that small firms cannot absorb, unlike large tech companies, and can deter funding, launches, and market entry. Reversibility and lock-in (Priority: 4/5): A major concept is whether harms can be rolled back once discovered; Gans says software-style reversibility exists now, but durable adoption could eventually create stronger lock-in. EU AI Act and safe harbors (Priority: 4/5): Gans is broadly supportive of transparency and oversight, but says regulators need clear safe harbors for startups so compliance does not chill innovation. Global competition and policy scale (Priority: 3/5): He notes that AI regulation is complicated by country competition and suggests learning on smaller jurisdictions when possible rather than treating entire blocs as test cases. Risk of overreaction after a harm event (Priority: 4/5): Gans fears a major AI-related incident could trigger sweeping backlash against all AI-related tools, even ordinary computational methods.

Key Arguments: The pause letter was notable because the usual innovation problem is insufficient adoption, not excessive speed; pausing a major technology is therefore economically unusual. AI harms are still uncertain, but absence of evidence so far is not proof of safety; regulators need a framework for learning under uncertainty. Lab-based testing can reveal some issues, but many harms—especially misinformation and labor-market effects—require real-world deployment to observe equilibrium outcomes. Faster adoption can accelerate learning, which may help society decide sooner which AI uses are safe and valuable. Regulation should be targeted: allow quicker deployment in markets where learning is most informative, rather than freezing broad adoption. Large firms can absorb compliance burdens and reputational risks better than startups, so uniform regulation disproportionately harms the startup ecosystem. Clear, reversible regulation and safe harbors could reduce uncertainty while still allowing oversight of high-risk AI applications. If AI causes a visible harm, policymakers may overreact and suppress benign uses of AI or even ordinary statistical software; gradual learning is preferable.

Data Points: Pause proposal duration: 6 months - The March 2023 open letter called for a six-month pause in giant AI experiments. Productivity increase from AI tools: 30-50% - Gans cites studies showing massive productivity gains for some tasks when people are simply given AI tools. Podcast source event: 79th Economic Policy Panel - The interview was recorded at the panel in Brussels. Publication title: How Learning About Harms Impacts the Optimal Rate of Artificial Intelligence Adoption - This is the title of Joshua Gans's paper discussed in the episode.

Pivotal Quotes: "we should stop. we should pause for a while and take a breath" — Joshua Gans: His reaction to the 2023 open letter calling for a pause in large AI experiments. "not only is there not a case for a pause of AI, there is now a case for an acceleration of it" — Joshua Gans: He summarizes his paper's implication that faster adoption can improve learning about harms. "the concerns I have is for that startup ecosystem" — Joshua Gans: He explains why broad AI regulation may disproportionately burden startups compared with big tech firms.

Implications: Policy should favor staged, reversible, and geographically limited AI rollouts that maximize learning while minimizing startup harm. Overly broad pauses or heavy compliance can slow beneficial innovation and still fail to prevent future overreactions.