What is a prediction market, and can you trust it?
Every number on SafeASI comes from a prediction market. Here is how these markets work, why they are often surprisingly accurate, and why you should still be careful.
Published 6 October 2026 by Omar Nouha Sane
The basic idea
A prediction market lets people trade on the outcome of a future event. Each question has shares that pay out if the answer turns out to be yes. If a “yes” share pays 1 and currently costs 0.36, the market is saying there is roughly a 36% chance the event happens.
If you think the real chance is higher, you buy and push the price up. If you think it is lower, you sell or bet on “no” and push it down. The price settles where buyers and sellers balance out. That is why the price can be read as the crowd’s probability.
The main platforms
- Manifold
- Uses play money called mana. Anyone can create a question, which is why it has thousands of markets on niche topics, including superintelligence. SafeASI uses Manifold data.
- Polymarket
- A real-money market built on cryptocurrency, known for large markets on politics and world events.
- Kalshi
- A regulated real-money exchange in the United States for event contracts.
- Metaculus
- Not a market but a forecasting platform. People submit probabilities and are scored on accuracy, and the site combines them into a community forecast.
Which platforms you can legally use depends on where you live, and the rules have been changing.
Why they often work
Markets reward people for being right and punish them for being wrong. Over time, traders who are well informed or careful gain influence, and those who guess badly lose money or play money. Studies going back to the Iowa Electronic Markets in the late 1980s found that market prices often beat polls and individual experts. Several platforms publish calibration charts showing that events priced at around 30% do in fact happen around 30% of the time.
Where they go wrong
- Few traders. A market with twenty people can be moved by one confident bettor. SafeASI shows the number of traders next to every question for this reason.
- Play money. When nothing real is at stake, some people bet for fun or to make a point.
- Long time horizons. A market that resolves in 2050 locks up your money for decades, so careful traders have little reason to fix a wrong price. Long-term prices are therefore less reliable than short-term ones.
- Unclear questions. If “superintelligence” is defined vaguely, traders are partly betting on how the creator will interpret it.
- Who collects the winnings. If superintelligence led to a catastrophe, nobody would be around to collect on a “yes” bet. That can push prices on some questions in strange directions.
How to use them
Treat market prices as a useful summary of what informed people currently believe, not as facts about the future. They are strongest when many traders are involved, the question is clear and the deadline is close. For superintelligence, at least two of those three are usually missing, which is why the methodology page lists the limits of the SafeASI numbers.