Mantic Beat Human Forecasters. It Still Isn't an Oracle
Matthew Mbaka · September 18, 2026 · AI
Mantic has earned a useful result and a dangerous nickname at the same time.
The London AI startup beat every human entrant in the summer 2026 Metaculus Cup and finished behind only one other bot, according to Reuters' September 18 report. It also raised US$25 million in a seed round led by Radical Ventures, with backing from Microsoft M12, Thinking Machines Lab and Balderton Capital.
The tempting summary is that an AI can now predict the future better than people. That is too broad. Mantic showed that its system can assign probabilities unusually well across a defined set of political, economic and cultural questions. That is impressive. It is not the same as knowing what will happen next.
What the contest actually measured
The Metaculus Cup asked participants to forecast questions that could later be resolved. A good forecaster does not simply call an outcome. It states a probability and updates that number as evidence changes.
That distinction matters. A forecast of 40% can be reasonable even when the event does not happen. You judge a forecasting system across many questions and over time, not by celebrating one surprise or mocking one miss.
Mantic says it specializes frontier models from other companies, tests the system against historical outcomes and uses the results to improve it. In the tournament, it sometimes broke away from the human consensus. Reuters reported that it gave Colombia's eventual presidential winner about a 40% chance early in the contest, compared with a roughly 30% community forecast.
That is evidence of useful independent judgement. It is not proof that disagreement is automatically smarter. A contrarian forecast can be right for the wrong reason, and a consensus can still contain information the model has missed.
The next test is outside the tournament
A public forecasting competition has clean questions, known deadlines and agreed ways to resolve outcomes. Real decisions are messier. A government may want to know whether a supply chain will remain usable, not merely whether a named port closes. A company may care about the cost of being wrong more than the raw forecast score.
Before using a system like Mantic for planning, an organization should ask:
- Was the forecast made before the relevant information became public? A timestamped audit trail matters.
- Does it stay well calibrated in this subject? Good election forecasts do not establish good drug, security or revenue forecasts.
- How often does it update? A stale 70% estimate can be worse than a current human judgement.
- Can reviewers see the evidence? A probability without sources is difficult to challenge.
- What happens when the question is poorly written? Ambiguous wording can produce a precise answer to the wrong problem.
Mapletechie made a similar point in its look at a headline AI benchmark: the testing setup tells you what a score means. The same discipline belongs here.
Where it could still be genuinely useful
The best use is not to hand a major decision to a machine. It is to give a decision team another forecast that can be logged, compared and challenged.
For a Canadian manufacturer, that might mean tracking the probability of a component shortage alongside supplier reports. A public agency could compare the model with internal analysts and a simple baseline. An investor could require that every forecast include its time horizon, sources and revision history before money moves.
Radical Ventures' involvement gives the story a Canadian connection, but the more important question is whether customers can reproduce the tournament advantage in their own work. Mantic has not named the organizations already using its forecasts, and independent evidence from those deployments is not yet public.
The Metaculus result deserves attention because it beats people on a real forecasting task. It should raise the standard for evaluating AI forecasts, not lower the standard for acting on them.
Tags: Mantic, AI forecasting, Metaculus, Radical Ventures