Can AI now predict the future better than humans? A UK startup surprises experts in a global forecasting contest, signaling a shift in artificial intelligence.
 |
| A British AI startup has outperformed top human forecasters in a high-stakes global competition. Image: CH |
Tech Desk — September 21, 2025:
In a quietly remarkable achievement, a British artificial intelligence startup has edged past many of the world’s most skilled human forecasters in a global competition that tests the limits of logic, intuition, and predictive insight.
The company, Mantic AI, co-founded by a former Google DeepMind researcher, has secured eighth place in the Metaculus Cup, a highly respected forecasting contest that tasks participants with predicting the likelihood of real-world events. These are not binary logic problems or clean data exercises. They are messy, unpredictable, and deeply human challenges—like whether Elon Musk and Donald Trump would engage in a public feud or whether UK cabinet minister Kemi Badenoch would lose her political position.
And yet, an AI system did what many thought impossible: it outperformed most human competitors, achieving over 80% of the average accuracy of the top forecasters, far surpassing pre-competition predictions that AI would only reach 40%.
For decades, forecasting the future has been considered the domain of human reasoning, intuition, and experience. That a machine could begin to encroach on this territory challenges some of our most deeply held assumptions about intelligence—especially as it relates to decision-making, judgment, and cognition.
Ben Shindel, one of the seasoned human participants, summed it up bluntly: “It’s certainly a weird feeling to be outdone by several bots.”
But this isn't just about bots getting lucky. Mantic AI employed a sophisticated forecasting architecture that split complex questions among different language models, including those from OpenAI, Google, and DeepSeek. Each model contributed based on its strengths, effectively creating an ensemble system that mirrors how human teams collaborate and divide cognitive labor.
Toby Shevlane, Mantic’s co-founder, rejected the idea that AI is just parroting training data. He argued that forecasting requires reasoning—and Mantic’s success is evidence of emerging machine cognition: “Some say LLMs just regurgitate their training data, but you can’t predict the future like that.”
In fact, Mantic’s predictions often stood out for their originality. While many human forecasters clustered around community averages—an effect known as prediction herding—Mantic was more confident in deviating from consensus when its internal models saw clear evidence to do so. That kind of independent reasoning, many argue, is the very essence of intelligence.
This moment may also represent something much larger than a clever algorithm topping a scoreboard. To many in the AI research community, it’s an indicator that we are inching closer to Artificial General Intelligence (AGI)—a system capable of performing a wide range of intellectual tasks at a human level or beyond.
Unlike today’s narrow AI, AGI could reason across contexts, apply knowledge from one domain to another, and make decisions with limited or uncertain data—just as humans do. A recent research paper from Google DeepMind even suggested that AGI could arrive as early as 2030. Mantic’s success doesn't confirm that AGI has arrived, but it does strongly hint that proto-AGI capabilities are emerging.
Forecasting competitions like the Metaculus Cup are becoming de facto proving grounds for this next generation of AI. Unlike games like chess or Go, which take place in constrained environments with fixed rules, real-world prediction involves ambiguous data, shifting dynamics, and incomplete information. It’s a true test of applied reasoning—and now, increasingly, AI is passing that test.
The implications of this are vast. From public policy and finance to climate science and defense, high-stakes forecasting is embedded in our most critical decisions. If machines begin to excel in this area, they may soon become indispensable advisors—or even decision-makers. The ethical, societal, and regulatory questions are only beginning to surface.
Yet for now, one thing is clear: Mantic AI’s performance represents more than a technical win. It’s a signal that AI is no longer confined to narrow, rule-based tasks. It’s stepping into the arena of uncertainty and judgment—domains that were, until recently, believed to be the last refuge of uniquely human intelligence.
And in that arena, the machines aren’t just learning. They’re starting to lead.