The argument that AI will save prediction markets has the appeal of every technology-fixes-coordination-failures story: a clean technical answer to what looks like a clean technical problem. The argument is wrong. Prediction markets have not failed for the reasons AI can fix.

A widely shared X thread on May 26 made the AI case directly. Polymarket, Manifold, and Kalshi each operate at a fraction of the scale Robin Hanson’s 1990 Idea Futures proposal envisioned. The diagnosis offered: liquidity is thin, market resolution is fuzzy, and forecasting baselines are weak. AI fixes all three. Market makers can be automated. Resolution criteria can be parsed and adjudicated by a language model. Pricing baselines can be supplied by AI forecasts that move the market efficiently.

Each of those technical claims is plausible. None of them addresses the underlying problem.

The structural reason prediction markets have not delivered on Hanson’s vision is that most people do not want to bet on truth. They want to bet on outcomes that already feel resolved (sports, elections in the final week) or on outcomes that feel like fun gambles (celebrity divorces, weather extremes). The truth-finding use case (whether a policy will work, whether a forecast is accurate, whether a claim is empirically true) is a niche even among the small population that participates in prediction markets at all. Liquidity follows interest, and interest in truth-finding markets is structurally low.

The historical record confirms this. The Iowa Electronic Markets, the academic prediction market that has been running for elections since 1988, has had liquidity an order of magnitude below what general election betting markets attract. The same questions are easier to answer with prediction markets, but the participation gap remains. PredictIt, before its regulatory shutdown, demonstrated the same pattern: election volume dwarfed any other category.

The regulatory and counterparty issues compound this. The Commodity Futures Trading Commission has consistently treated truth-finding prediction markets as unauthorized derivatives. The legal questions create an overhead that limits market design to either offshore venues (Polymarket pre-Kalshi acquisition), academic exemptions (Iowa Electronic Markets), or sports-and-elections-only platforms (Kalshi). Even with perfect liquidity and AI-driven market making, the legal frame would still gate the truth-finding categories.

AI does meaningfully change one piece of the picture. Resolution-criteria interpretation is a real bottleneck on truth-finding markets, since vague question definitions create disputes that erode trust. A language model that can parse contract terms, identify ambiguity in advance, and propose tighter wordings could reduce that friction. That is useful. It is not a market-creation force.

The broader pattern worth naming is that prediction markets are a coordination-mechanism solution looking for a coordination problem people actually feel. The problem people feel is not the truth-discovery gap that Hanson identified. The problem people feel is information overload, attention scarcity, and the difficulty of forming any opinion at all. AI is being deployed against the second problem (summarization, synthesis, recommendation) and is being credited with fixing the first problem (truth discovery via markets). The two problems are not the same, and the AI solutions to one do not transfer to the other.

For founders pitching AI-augmented prediction market products to investors in 2026, the practical advice is to lead with the resolution-clarity and liquidity-bootstrapping wins as concrete improvements over the existing market design, not with the Hanson Idea Futures vision as the addressable market. The Idea Futures vision has not been blocked by missing technology. It has been blocked by missing demand, and AI does not fix demand.

Reported via an X thread on 2026-05-26.