We need well-capitalized prediction markets for AI impacts

And how to create them

As our readers have surely noticed, there’s enormous disagreement about the likely economic effects of AI. Some AI forecasters, especially those emphasizing recursive self-improvement, expect AI capabilities to soon surpass that of most human workers. As a result, they expect large-scale labor market displacement, beginning with white-collar work and spreading to all of labor in the long run. The AI 2027 scenario lays out a year-by-year path to transformative AI and implies large-scale labor displacement. Other economists, in contrast, are more skeptical of such extreme rapid take-offs, emphasizing adoption frictions, organizational bottlenecks, complementary investments, and the slow diffusion of new production processes.

The Forecasting Research Institute’s recent survey of leading economists, AI experts, and superforecasters quantifies several dimensions of disagreement. The figures below show that there is enormous disagreement among experts and that AI experts tend to forecast higher economic growth rates than economists. Tom Cunningham nicely describes a variety of forecasts in his blog post.

Some disagreement reflects genuinely different beliefs about AI progress and the transition path of economic adjustment. Disagreement is healthy, and reflects diversity in economists’ world models. But a deeper and more fundamental problem is that few people have strong incentives to produce granular, decision-relevant forecasts. Few have done the detailed work of trying to predict what will happen to the macroeconomy. Even fewer have worked at the granular level required to predict which industries will be affected and when, which jobs will be diminished and which will grow, and what will happen to wages in those jobs. The financial incentives are too weak for these experts to invest the effort.

Even structured modeling efforts illustrate how wide the uncertainty is. Epoch’s GATE model is a macro model of AI’s economic impact, built around an automation feedback loop between compute investment, AI capabilities, and task automation. But its outputs are highly sensitive to a handful of parameters whose true values are deeply uncertain. The model can produce qualitatively different trajectories depending on plausible parameter choices.

So we’re operating in a world of great uncertainty. This uncertainty is high-dimensional and will never be fully resolved. But we can, in principle, make good predictions about slices of the state space -- specific events that are both learnable: we can go out into the world and collect information about what will happen; and decision-relevant: the information is valuable for making better decisions. We think our best bet for doing so will be through prediction markets that, despite their shortcomings, are powerful technologies for incentivizing information acquisition and aggregating dispersed information.¹ This is a modern rendition of a classic Hayekian case for prices.²

The Proposal

We propose that AI labs, Big Tech companies, and philanthropic donors contribute funding to create well-capitalized prediction markets to understand the likely impacts of AI. We need this information because individuals, firms, and policymakers are already making decisions right now – with or without this information.

Consider a college graduate deciding which industry to enter, or even a highschooler entering college deciding what to major in. They probably want some information about which jobs are likely to exist in the future and which jobs are likely to be less common. It would be very useful to have, based on the most recent up-to-date information, some signal for how to make this decision. Firms are in a similar position, when deciding which occupations to hire and at what level of seniority.

Policymakers face an even harder version of this problem. Decisions about retraining programs, extended unemployment insurance, and sector-specific support require knowing not just whether AI will displace workers, but which workers, in which industries, and how quickly. Retraining programs in particular are highly sensitive to these forecasts: training a displaced bookkeeper takes years, and the value of that investment depends on which destination occupations will still exist when the training is complete.

The FRI survey found that economists ranked job retraining as their most-supported policy response, with a forecasted boost to labor force participation in a rapid AI progress scenario, but designing such a program requires answering granular questions that current forecasts simply don’t address. Extended UI faces a similar problem: the case for sector-specific extensions depends on whether displacement is temporary (workers transitioning between jobs) or structural (occupations disappearing), and the appropriate duration depends on expected re-employment rates that vary enormously by industry.

How Prediction Markets Would Help and Why They Don’t Yet Exist

For prediction markets, we envision a variety of contracts. Start with the labor market: aggregate labor-force participation, and occupation-level employment and wages, are measured in official BLS data. For more detail on how these contracts could resolve, see the appendix on contract resolution using labor-market data. We would like to have a forecast of the future labor force participation rate and of the employment rate and wages across different occupations.

Suppose that we had prediction markets at different time horizons (1, 2, 5 years) for the total labor force size of different occupations in the United States, and analogous contracts on wages and employment rates. With an appropriate and liquid contract, we should have a good signal about the future labor market. If this exists, when a new graduate is choosing a career, they could use this information to think about which occupations have continued demand. Note that these contracts don’t forecast the impact of AI directly, but to the extent that AI is an important force in determining future employment rates, its impact will be a part of the forecast.

But labor outcomes are downstream of AI capabilities, and we’d also want contracts directly on capabilities themselves. The questions currently trading on Polymarket are coarse, such as “Which company has the best AI model by the end of June?”, and don’t track the underlying technical progress in a way that’s useful for forecasting economic impact. We’d want contracts on benchmarks (e.g., ARC-AGI-3) or capability thresholds (e.g., passing Steve Wozniak’s coffee test) at different horizons. These capability forecasts are valuable on their own and would also help calibrate the labor market contracts: if traders price in rapid capability gains but slow labor market changes, that discrepancy is, itself, informative about the market’s belief about passthrough from capabilities to labor market impacts.

The next question is: what will it take for these prediction markets to have meaningful information aggregation, and why don’t they exist yet?³

There are two reasons people might participate in prediction markets: speculation and hedging. So far, firms haven’t created enough demand to hedge for AI’s labor market impacts. Furthermore, for contracts that are longer-term, traders need to be compensated at least at the risk-free interest rate, meaning that incentives to trade are diminished for contracts further in the future. As a result, the chicken-and-egg problem hasn’t been broken, and we lack prediction markets on these questions. So the question becomes how to bootstrap them (see Andy Hall’s essay on some ideas).

What It Would Take

Prediction markets need to be seeded. The reason is that prediction markets have increasing returns to participation. With a thin market, prices are noisy, spreads are wide, and traders cannot enter and exit positions without moving the market. But without traders, the market never becomes informative. Seeding solves this coordination problem by paying for early liquidity, narrowing spreads, and making it worthwhile for those best-positioned to acquire information to do so.

There are two ways to seed prediction markets. The direct way is to provide liquidity by staking a market with capital on both sides of contracts, so that other participants have something to trade against. This is how most prediction markets historically got started, but it requires the seeder to take on inventory risk.

A more natural way is sponsorship. The way this works is that a principal (the entity that wants the information) deposits money into a contract. Sponsorship adds a way to bootstrap the market: the sponsor pays participants who improve liquidity, for example by posting orders that narrow spreads or add depth. If the sponsorship is large enough, then traders have incentives to acquire information and to participate in the market. Once there is already active trading volume, the market may sustain itself even without additional sponsorship funding. Sponsorship is already a feature on Polymarket and may be available soon on Kalshi. Technically, sponsorship is one way to subsidize liquidity provision. This connects to the market-design literature on automated market makers and scoring rules, especially Hanson’s logarithmic market scoring rule and Chen and Pennock’s bounded-loss market-maker framework, which formalizes the tradeoff between liquidity provision and worst-case loss (Hanson 2003, 2007; Chen and Pennock 2007).

It’s worth being clear about what sponsorship does and doesn’t do, because the FRI survey is instructive here. FRI paid economists to participate, and got back the most comprehensive elicitation of expert beliefs about AI’s economic impact that we have. But the economists essentially had no skin in the game. Sponsorship has a different structure. The sponsor pays for the market to exist, but the traders who actually set prices are putting their own capital at risk.

The natural candidates to sponsor these markets are large players, either the AI labs themselves, or firms with significant exposure to AI that are interested in getting better information. They have the budget, and they have the strongest reason to want this information to exist. We don’t have a good sense of how much sponsorship volume is necessary for good information aggregation. That’s an open question, and one we’re interested in experimenting with. That said, we can’t do this alone. A big player needs to participate, or at least fund research on how to create these contracts.

Thanks to Tom Cunningham, Andy Hall, John Horton, and Scott Kominers for comments.

Appendix: What Makes a Good Contract

Not every interesting question can be turned into a useful prediction market. A workable contract needs four things:

Verifiability. The resolution criterion has to be objectively checkable. “By 2030, will AI have caused Cheerios sales to double?” fails immediately because “caused” isn’t observable. You can’t separate the AI-driven counterfactual from everything else moving in the world. “By 2030, will the BLS-reported employment in occupation X be below Y?” is verifiable: the number gets published, and you read it.

Stability. The underlying data series has to keep being published and measured the same way over the life of the contract. Niche or proprietary statistics fail this. BLS occupation employment, unemployment rates, and labor force participation are highly stable. Anything that depends on a single company’s voluntary disclosure is not.

Robustness to gaming. The bigger the underlying quantity, the harder it is for any trader to move it. Someone with a few million dollars and an axe to grind could plausibly distort sales of a single product or the headcount of a small firm. They cannot plausibly distort BLS employment statistics for a major occupational category. Sector-level labor market contracts are robust for the same reason: the sector is too large to move by hiring or firing at the margin.

Attention. Even a verifiable, stable, robust contract is useless if no one cares enough to trade it. Cheerios sales fail this test too. Unemployment contracts have so far failed, but we think that this is for reasons that are fixable.

Appendix: Contract Resolution Using Labor-Market Data

For labor-market contracts, the natural resolution sources are large, official, regularly published data series, and not tailor-made AI-impact metrics. The Occupational Employment and Wage Statistics (OEWS) program under the Bureau of Labor Statistics (BLS) provides annual estimates on employment and wages for approximately 830 occupations at the national level, state, metropolitan, and nonmetropolitan areas, as well as for specified industries. This makes OEWS a natural source for occupation-level employment and wage contracts. Such contracts would not identify the causal effect of AI. Instead, they would rely on a set of observable labor-market outcomes — employment, wages, or labor-force quantities, which directly influence students, firms, and policymakers. For example, a contract could be settled on whether employment reported by BLS in a specified SOC occupation falls below a pre-specified threshold in a given release year.

References

Hayek, F. A. 1945. “The Use of Knowledge in Society.” American Economic Review.

Wolfers, Justin, and Eric Zitzewitz. 2004. “Prediction Markets.” Journal of Economic Perspectives.

Arrow, Kenneth J., et al. 2008. “The Promise of Prediction Markets.” Science.

Hanson, Robin. 2003. “Combinatorial Information Market Design.” Information Systems Frontiers 5 (1): 107–119.

Hanson, Robin. 2007. “Logarithmic Market Scoring Rules for Modular Combinatorial Information Aggregation.” Journal of Prediction Markets 1 (1): 3–15.

Chen, Yiling, and David M. Pennock. 2007. “A Utility Framework for Bounded-Loss Market Makers.” In Proceedings of the 23rd Conference on Uncertainty in Artificial Intelligence (UAI 2007), 49–56

1. For example, if traders are risk averse and use prediction markets for hedging then the price may be a biased estimate of the expected outcome. Alternatively, if retail meme traders herd on the contract, then its value could become disjointed from the underlying value.

2. See Hayek (1945); Wolfers and Zitzewitz (2004); Arrow et al. (2008); Hanson (2003, 2007).

3. There are prediction contests on Metaculus relating to the labor force participation rate and other outcomes, but these have few participants and have small financial incentives to be correct.