Can AI Even Be Regulated?, with Sendhil Mullainathan

Episode Summary

Executive Summary: The discussion revisits AI regulation after OpenAI turmoil and DeepSeek’s rise, arguing AI is neither a monopoly nor fully competitive but a fast-moving, widely diffusing market. The speakers debate liability, governance, safety, and displacement, concluding that regulation should focus on users and testing standards, while governance should be pluralistic, experimental, and capable of adapting to a technology whose risks and applications are unusually broad.

Main Topics: AI market structure: monopoly, oligopoly, or free-for-all (Priority: 5/5): The speakers revisit whether AI is dominated by a few firms or remains open and competitive. One side stresses a free-for-all with low barriers to entry, while the other argues a small set of firms still wields outsized influence and financing power. Regulating liability vs regulating the technology itself (Priority: 5/5): A central proposal is to place legal liability on AI users or deployers rather than on model builders, with limited safe-harbor exceptions to encourage beneficial uses such as medical support. Governance, public interest, and citizen participation (Priority: 4/5): They argue AI governance should not be left only to corporate boards or experts. Ideas include public-interest board seats, citizen assemblies, and broader governance experiments to reflect pluralistic social values. Displacement and the speed of AI adoption (Priority: 4/5): The conversation emphasizes that AI may disrupt labor and communities rapidly, so policy may need to slow deployment or create adjustment mechanisms rather than simply ‘let it rip.’ Safety and alignment as unresolved problems (Priority: 4/5): The speakers note that even experts disagree on what safety means, making alignment and oversight difficult. They worry current solutions are too weak and that experts may be systematically optimistic about their own field. Open source, diffusion, and the effect of DeepSeek/Mistral (Priority: 4/5): DeepSeek and Mistral are used as evidence that model-building costs are falling and that innovations can spread quickly across borders, lowering barriers to entry and complicating attempts to regulate only a few firms. Elon Musk, OpenAI, and the politics of the AI race (Priority: 3/5): The Musk/OpenAI conflict is discussed as both strategic and potentially beneficial if it limits oligarchic coordination. Musk’s bid is also framed as affecting OpenAI’s nonprofit-to-for-profit transition.

Key Arguments: AI should not be treated as a traditional, narrow regulatory target because its applications are broad and evolving, making governance more important than static rulemaking. OpenAI is not a monopoly; the market is highly diffuse, and even perfect conduct by OpenAI would not prevent others, including foreign actors, from developing AI. At the same time, the market is not perfectly competitive: a few firms control capital, compute, and training data, giving them disproportionate influence. A strong default is to assign liability to the user/deployer of AI, not the model creator, while creating narrow safe harbors when society wants to encourage adoption. Safe harbors should be used carefully because Section 230-style exemptions can persist far beyond their original intent and create long-term distortions. AI governance should be pluralistic rather than based on a single consensus definition of ‘alignment,’ since social views diverge on acceptable outputs and constraints. Citizen assemblies and public participation could improve governance by making it more democratic and less dependent on elite board members or captured regulators. AI may create major labor displacement and social disruption, so policymakers may need to slow rollout to allow adaptation, even if they cannot stop progress entirely. Expert-led regulation is hard because AI specialists are naturally biased toward optimism, and governments risk capture by industry voices and lobbying. Open source and falling compute costs mean that attempts to regulate only a few flagship companies will not contain the technology; diffusion across countries and labs is likely. Musk’s bid for OpenAI is interpreted as both a strategic move and a potential public benefit because it raises the floor on the nonprofit’s value and may force more value toward charity. Crash-test-style independent testing agencies could help evaluate AI systems more objectively than self-reporting by firms. The speakers worry that technology policy is being shaped by broligarchic incentives and by political environments that may further weaken oversight.

Data Points: OpenAI bid value: $97.4 billion - Musk-led unsolicited bid to take over OpenAI OpenAI nonprofit financing from Microsoft: $13 billion - Used as evidence that OpenAI had substantial capital requirements and is not fully competitive Nuclear reaction history: 81 years ago - Reference to the first controlled nuclear reaction at the University of Chicago Experiment location: squash court beneath the university football stadium - Describes where the first controlled nuclear reaction took place AI training cost range: $100,000 to $10 million+ - Used to argue that barriers to training models are falling and not prohibitively high AI training cost upper range mentioned: $50 million - Speaker notes that even this is not a very large barrier OpenAI employee count referenced: 600 people - Used in the discussion of whether staff could simply be replaced if they left Non-compete jurisdiction comparison: California vs Washington State - Used to illustrate how non-enforcement of non-competes enabled labor mobility in AI Section 230 reference year: 1996 - Mentioned as a historical safe-harbor example with long-lasting consequences

Pivotal Quotes: "Even if OpenAI behaved perfectly, that's not going to stop anybody else from developing." — Speaker: Used to argue that regulating only OpenAI would not contain AI diffusion "It's the free for all." — Speaker: Characterizes AI as a highly competitive and widely accessible market rather than a monopoly "We have socialism for the very rich, rugged individualism for the poor." — Bethany McLean: Podcast framing line highlighting inequality and capitalism debates

Implications: AI policy will likely need a mix of liability rules, testing infrastructure, and democratic governance experiments. Because the market is diffuse and fast-moving, focusing only on top firms will be insufficient; the bigger challenge is building institutions that can adapt in real time.