A growing cohort of startups is wagering that artificial intelligence can forecast the future better than the most skilled human analysts, across subjects ranging from stock market swings to the reopening of the Strait of Hormuz and the chart performance of a pop single. The claim is no longer purely speculative. According to reporting by Agence France-Presse, AI models have reached the level of the best human "superforecasters," and in early September two AI systems placed at the top of the Metaculus Cup, a competition open to both human and machine predictors.
From Superforecasters to Machines
Superforecasters are individuals whose probability estimates for future events consistently beat those of ordinary people and, in many cases, of trained specialists. Their advantage rests on disciplined habits: breaking a question into smaller parts, updating beliefs as new evidence arrives, and avoiding overconfidence. Those same habits are what modern AI systems can imitate at scale, because they can ingest vast quantities of public and private data and check many interlinked assumptions at once.
Dan Schwarz, cofounder of the forecasting company FutureSearch, argues that the shift will be sweeping. He predicts that by 2030 organizations will look back on their current decision-making as deeply outdated. His company employs a human superforecaster, Tom Liptay, yet Schwarz acknowledges that Liptay's position is not secure over the medium term, since the systems are expected to surpass him.
Markets, Prediction Platforms and Automation
Finance is the most immediate testing ground. Algorithms have shaped Wall Street for more than a decade, but AI forecasting extends that automation from executing trades to estimating the likelihood of events that move prices. Preseen, cofounded by Theo Summer and Veniamin Veselovsky, reports using AI tools to grow a $35 stake into $1.9 million in seven months on the Kalshi prediction platform. Such results are anecdotal and should be read as the claims of their backers rather than as independently verified performance.
The broader concern is who benefits. Veselovsky notes that hedge funds, which can act quickly on a perceived edge, are typically the first buyers of such tools. He warns that retail traders without access to expensive AI systems may find consumer-oriented platforms such as Kalshi and Polymarket increasingly dominated by automated forecasters, narrowing the room for ordinary participants to profit.
Beyond Finance: Policy, Science and Risk
The most advanced systems no longer simply answer forecasting questions. They can generate their own questions about the future and assign probabilities to the range of possible outcomes, a capability Toby Shevlane, cofounder of the startup Mantic, describes as a significant step. His company supplies a large American firm with scenarios for regulatory change in its sector and has also worked with the British government.
The potential applications are wide. Companies want to understand and monitor risk, insurers may use probabilistic assessments in underwriting, and researchers could use forecasts to decide which medical treatments merit funding. Each use case depends on the same underlying premise: that a calibrated probability is more useful to decision-makers than a confident-sounding prediction.
Open Questions and Strategic Risks
Widespread access to these tools carries costs that are not yet well understood. Veselovsky says he sees real benefits but is uneasy about geopolitical competition, asking whether rival states such as Russia or China may already be building more capable systems. Forecasting models trained on sensitive data also raise familiar questions about transparency, accountability and who controls the outputs. As these systems move from trading desks into government and corporate planning, the standards for testing their accuracy and explaining their reasoning will matter as much as the forecasts themselves.
