Prediction Markets Meet the AI Forecaster预测市场遇上 AI 预言家

The Setup

Prediction markets have finally become visible. Polymarket and Kalshi can move billions in volume, show up in mainstream media, and give the internet a live probability for elections, lawsuits, wars, product launches, and celebrity nonsense. The old dream was bigger than gambling: markets would become public truth machines, aggregating dispersed information into prices that institutions could use.

Dan Schwarz’s Asterisk piece is useful because it checks that dream against the data. The answer is not “prediction markets are fake.” It is sharper: the useful version exists, but it is smaller, thinner, and more fragile than the hype suggests.

Key Takeaways

• Most volume is not public intelligence. Roughly 90% of Kalshi volume is sports betting, and more than 80% of Polymarket volume is sports, crypto prices, or elections. Liquidity is real, but it does not automatically spill into useful questions.
• Risk monitoring is the best current use case. Markets tracking conflict escalation, disease outbreaks, bank failures, and geopolitical shocks are where the format looks most socially valuable.
• Liquidity only helps in the right window. In Schwarz’s analysis of about 6,800 potentially useful markets, higher volume correlated with better accuracy for markets lasting 90+ days. For short-lived markets, the relationship largely disappeared.
• Prediction laundering is a problem. Some markets look like accountability but are really entertainment wearing an epistemic costume: the Epstein-files style of speculation produces volume without much decision value.
• AI changes the bottleneck. The hard part may not be aggregating wisdom through traders. It may be distributing probabilistic reasoning to people who do not naturally think in probabilities.

Why It Matters

For Rex, the interesting question is not whether prediction markets survive. They will. Gambling, hedging, and election speculation are enough to sustain the category. The bigger question is whether “info finance” becomes a serious research and decision layer, or whether it gets eaten by AI interfaces.

A market gives you a number. Claude or another strong forecaster can give you the number, the reasoning, the base rates, the objections, and the follow-up questions. For most users, that is simply the better product. The market may still be the underlying signal source, but the user-facing interface could become an AI analyst that reads markets alongside news, filings, expert writing, and historical analogs.

That creates a better investment frame: prediction markets are less likely to be the final consumer interface for foresight. They are more likely to become one data layer inside a broader forecasting stack.

What to Watch

• Whether Polymarket and Kalshi build AI-native research interfaces instead of only more markets.
• Whether “useful” markets gain durable liquidity outside sports, elections, and crypto prices.
• Whether media citations keep turning market prices into public narratives.
• Whether AI forecasters begin citing prediction markets as one input, not the final answer.