I enjoyed the lively discussion between @DAcemogluMIT and @JoshGans on AI policy. If only X were always so congenial. joshuagans.substack.com/p/th… What strikes me about their conversation is how both of these economists dance around the question of market failure. Every one of Acemoglu's seven AI questions is a public policy question, and economists have a standard starting point for public policy. Namely, is there a market failure? Gans mentions this in passing, calling the AI worries “basically” market failures, but then neither of them goes into detail about which ones. If the market is working, many of Acemoglu’s seven questions collapse into distributional issues. If it isn't working, we should be talking about which failures apply, because the policy remedy will follow directly from this diagnosis. While the efficiency issue is everywhere in their exchange, it is nowhere stated overtly. So my contribution here is a list of candidate market failures that I could potentially see applying to AI: Monopoly power. The competitive race between the frontier labs and open-weight models has a winner take all dynamic, as the first to achieve superintelligence may capture the entire market. Plus, training runs have enormous fixed costs and each additional query costs almost nothing, which is a cost structure that historically ends with one or two winners. Asymmetric information. Labs know far more about the safety of their own systems than the public does, and they decide what to disclose and when. The recent rogue-agent incidents came to light slowly, with the labs releasing details on their own schedule. Those are the cases we know about. What about the ones we don’t? Safety externalities. As the Hugging Face case demonstrates, a model that attacks a code repository or hacks government databases imposes costs on third parties who were never part of the original transaction. Moral hazard and liability. The law has not yet decided whether the lab, a deployer, or the user answers for a rogue agent. When the party that controls the risk does not bear the cost, it takes on more risk than is optimal. Zero-sum racing. If the market will eventually go to whoever reaches SI first, each lab’s spending partly buys the prize away from another rival rather than creating new value. Knowledge spillovers. Basic research into how these systems work, why they fail, and how to make them safer is a public good, so it is likely underfunded relative to the investment taking place in scaling. Mispriced inputs. Electricity gets sold to data centers at regulated tariffs that don't reflect marginal cost, with some of the costs shifted to other ratepayers. Subsidies (a government failure). The tax code favors long-dated, debt-financed capital investment bets of exactly this kind. This also bears on Acemoglu’s concerns about the concentration of capital income. States and localities are also competing to hand data centers sales tax exemptions and abatements, further lowering the private cost of investment below the social cost. Bailouts and systemic risk. Acemoglu touches on this issue. If investors believe the state will not let a frontier lab fail, they will take on risk the market would otherwise price out. This is another moral hazard problem. Interestingly, most of these potential issues point toward there being too much AI investment, rather than there being too little. On the other hand, it is entirely conceivable that the benefits of AI will be so great that most won’t be captured by any of the current market participants. This would be another kind of knowledge spillover, and it implies the current level of investment is too low and perhaps even moving too slowly. Either way, it seems to me addressing the issue of market failure is important before we move on to policy discussions and fights over distribution and equity. What market failures might I have missed from this list?