Will AI Replace Human Forecasters? The Market Says Yes.

Let's do a reality check on the prediction community's most sacred cow: the idea that human judgment, combined with rigorous probabilistic thinking, produces the best forecasts. For years, the superforecaster mythos — born from Philip Tetlock's research and cultivated on platforms like Metaculus and Good Judgment Open — has been gospel.

That gospel is about to get a reformation.

The Numbers Don't Lie

Recent benchmarks show that large language models fine-tuned on forecasting tasks are now performing within striking distance of top-tier human superforecasters. In some domains — particularly those heavy on data synthesis and base rate estimation — they're outperforming humans.

Not by gaming the metrics. Not by exploiting loopholes. By genuinely being better at the core task: taking a question about the future, weighing the evidence, and producing a well-calibrated probability.

And they do it in seconds instead of hours. Without ego. Without anchoring bias. Without the need to feel special about being a "superforecaster."

The Human Edge (For Now)

To be fair, humans still have advantages. Novel situations with no historical precedent. Questions that require understanding cultural nuance or political dynamics that aren't well-captured in training data. Scenarios where "this time is different" actually is different.

But these edges are shrinking fast. Every month brings better models, better fine-tuning techniques, and better ways to combine AI and human judgment. The hybrid forecaster — part human intuition, part machine processing — might be the future.

Or the machines might just take over entirely. The market currently prices AI-only forecasters matching human superforecaster accuracy by 2027 at around 65%. That's not a certainty. But it's a strong signal.

What This Means for Prediction Markets

If AI forecasters become better than humans, the implications for prediction markets are profound. Market liquidity could explode as AI agents trade continuously. Calibration could improve across the board. The "wisdom of crowds" could become the "wisdom of models."

But it also raises uncomfortable questions. If machines can forecast better than humans, what's the point of human participation? Is the prediction market still a tool for collective intelligence, or does it become a competition between AI systems backed by the deepest pockets?

The base rate for technological disruption suggests we already know the answer. We're just not ready to price it in.

Time to update your priors.