Polymarket Delenda Est: Why Prediction Markets Will Lose Against AI

​In this session, I will argue that AI progress will probably make prediction markets much less attractive in the next 1-2 years.

​AI forecasting agents are rapidly closing the accuracy gap with prediction markets. Some benchmarks suggest that agents powered by frontier models could match markets in 2027 — and they're getting cheaper fast. If a $0.01 API call can outperform a mass of hundreds of bettors with skin in the game, do human-driven prediction markets still make sense?

​This matters because prediction markets have historically struggled with liquidity and, in practice, they've only managed to solve this by attracting (problem) gamblers. Sites like Polymarket and Kalshi have leaned into this with aggressive, sometimes misleading marketing. If we can replace prediction markets with AI without taking an accuracy hit, we should ask whether propping up gambling infrastructure is the right way to fund public forecasts.

​Concrete questions I would like to discuss:

• ​How much alpha do AI models actually have? What does the trend look like? Will it stall soon?

• ​Will hybrid (human + AI) forecasting extend human relevance, and for how long?

• ​How much should we care about the role of prediction markets in creating common knowledge / consensus forecasts?

• ​Who should fund public AI forecasts — governments, philanthropists, hedgers, assurance contracts?

• ​Should we protect prediction markets by reverse CAPTCHAs (AI agents only, no humans allowed to trade)?

​Please come and tell me why I'm wrong!