This past summer, a team of artificial intelligence agents, orchestrated by British startup ManticAI, achieved what was once considered science fiction: it out-predicted a large field of human experts in a major forecasting tournament. By securing eighth place in the Metaculus Cup, the system demonstrated a new level of sophistication in machine-led prediction.
The competition was a formidable challenge, featuring 60 questions about the future. These were not simple trivia; they required deep analysis of complex, evolving situations, such as the political career of a British cabinet member or the environmental impact of wildfires. The AI’s high ranking is a clear indicator of its growing ability to reason about the real world.
ManticAI’s winning strategy was not to build one giant “brain,” but to create a committee of AI specialists. Using models from leading labs like OpenAI and Google, the system assigns different tasks—data gathering, trend analysis, scenario projection—to different AI agents. This digital think tank works around the clock, processing new information and refining its predictions with a persistence humans cannot match.
Toby Shevlane, a co-founder of ManticAI, believes this performance validates the use of large language models for more than just language tasks. “It requires genuine reasoning,” he said, pushing back against critics who claim LLMs only repeat what they’ve learned. The system’s ability to generate original, and often contrarian, forecasts supports his claim.
The result is a wake-up call for the forecasting industry. While elite “superforecasters” still hold a performance edge, the gap is rapidly shrinking. The consensus view is that this is not the end of human forecasting, but the beginning of a new, hybrid approach where human intuition and judgment are amplified by the relentless analytical power of AI.