AI Forecasters Are Catching Up With the Humans Who Predict the Future

For years, the art of predicting the future belonged to a small and unusual elite. So-called superforecasters — hobbyists, retired analysts, statisticians and curious amateurs who consistently outperform their peers in prediction tournaments — became famous for beating intelligence analysts at their own game. Now, according to a report in The Economist, artificial intelligence is beginning to match and in some cases beat even the best of them.

The claim marks a milestone in a field that has quietly become one of the more revealing tests of machine intelligence. Unlike exams or coding puzzles, forecasting cannot be solved by memorisation. A question about whether a ceasefire will hold, a central bank will cut rates, or a disease outbreak will cross a border has no answer sitting in a textbook. It demands that a forecaster gather scattered evidence, weigh competing explanations, assign a probability, and then update that probability as the world changes. That combination of research, reasoning and calibrated humility is precisely what large language models have historically been bad at.

Why forecasting is a good test

Prediction tournaments offer something that much of AI evaluation lacks: a clean scoreboard. Forecasts are expressed as probabilities, and probabilities can be scored against what actually happened. A model that says an event is 90% likely and is repeatedly wrong will be punished by the maths. There is little room for the vague impressiveness that makes chatbot demonstrations so easy to over-interpret.

That also makes the results harder to dismiss. If AI systems are now producing forecasts whose accuracy rivals that of elite human groups, it suggests the models are doing more than parroting plausible-sounding text. Good calibration — knowing how confident to be — is a skill that even well-informed humans struggle to acquire.

Speed and scale are the real advantage

Even if machines merely equal the best humans, the economics are striking. Superforecasters are scarce, slow and expensive. Assembling a panel and aggregating their judgements takes time, and each question consumes hours of human attention. A model can be pointed at thousands of questions at once, at negligible cost, and can refresh its estimates whenever new information appears.

That opens possibilities that were previously impractical: continuously updated risk dashboards for insurers, early-warning systems for governments, or probability estimates attached to every line of a corporate strategy document. Human forecasting talent could be reserved for the handful of questions where judgement matters most, with machines handling the long tail.

Reasons for caution

Several caveats deserve emphasis. Benchmarks can flatter the systems designed to pass them, and results depend heavily on which questions are asked and over what time horizon. Models trained on historical text may also enjoy subtle advantages when tested on events that are, in effect, already partly known. Short-horizon questions with abundant news coverage are far easier than genuinely novel, long-range ones.

There is also the question of what happens when machine forecasts start influencing the world they describe. If markets, ministries and newsrooms all lean on similar models, they may converge on the same blind spots — and be surprised together.

Still, the direction of travel is clear. Forecasting has long been treated as a distinctively human craft, requiring intuition honed by experience. The emerging evidence suggests it is instead a skill that can be measured, learned and, increasingly, automated. The interesting question is no longer whether machines can forecast, but how much we should let them decide what to expect. Read More