AI Forecasting in 2026: What 11 Analyses Say

According to a synthesis of analyses from LessWrong, human forecasters on the Metaculus platform are still outperforming AI bots at predicting future events—though the gap is narrowing. The review examined eleven studies spanning late two thousand twenty-four through mid-two thousand twenty-six, comparing top human forecasters against dozens of bot systems. The key finding for AI builders: picking a frontier reasoning model matters more than anything else. A simple AI bot using a modern frontier model placed in the top five, while sophisticated approaches with older models lagged behind. But here's what's striking: good scaffolding—the prompts and techniques that guide the AI's thinking—appears to be worth roughly nine months of base model improvement. The research also found that the best forecasting systems combine multiple approaches: using different sources to gather information, searching iteratively rather than one-shot, and then aggregating predictions from multiple models before applying calibration adjustments. Eighty-six percent of the competition winners used this ensemble approach. The takeaway: humans aren't losing their forecasting edge yet, but the gap is tightening, and thoughtful system design matters as much as raw model power.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton