On 16 September 2025, Émile Servan-Schreiber announced that The Forecasting Machine was open to everyone. The platform had spent a few months in closed beta. Servan-Schreiber described it as the product of 18 months of research, experimentation and teamwork, incubated at Hypermind, the forecasting company he has run since its NewsFutures days. New users got their first 15 forecasts free.
The problem it sets out to solve
The launch post framed the problem as one of supply. There are far too few trained forecasters for the number of important questions that deserve a forecast, and the shortfall leads to costly mistakes. Crowd forecasting and prediction markets, Hypermind’s background, produce good probabilities. But they need many engaged people per question, so they cannot cover everything a decision maker cares about.
Servan-Schreiber’s argument was that AI now makes quality forecasting possible at scale. A machine can be asked about almost any topic, at any hour, and keep its answer current as the news changes.
How the machine works
As described at launch, the platform combines two ingredients:
- Real-time global news. It continuously scrapes news so that each forecast rests on current information rather than a model’s frozen training data.
- An ensemble of AI models. Several large language models forecast the same question independently, and their answers are combined. Servan-Schreiber calls this the “supercollective intelligence” of the ensemble, an echo of the crowd principle Hypermind applies to people.
- Dashboards and automatic updates. Each question becomes a tracked probability that is revised as events unfold, with a written rationale explaining the reasoning.
The output is not a single prediction. It is a probability for each possible outcome, which is easier to act on and easier to check later. Our explainer on why probabilities beat predictions covers the reasoning.
A changing line-up of models
The ensemble is not fixed. On 27 September 2025, eleven days after launch, The Forecasting Machine’s company page announced that Grok 4 had joined GPT-5, o3 and Gemini 2.5 Pro. The announcement was illustrated with a screenshot of a social-media post claiming Grok 4 ranked first on FutureX. FutureX is an academic live benchmark that scores AI agents on predictions about real future events, and it had appeared on arXiv a few weeks earlier. We have not checked that ranking ourselves.
The forecasts Servan-Schreiber shared over the following weeks show the ensemble in practice. Individual model samples from o3, GPT-5, Claude Sonnet 4, Gemini 2.5 Pro and Grok 4 appear alongside the aggregate. As of October 2026, the platform’s homepage lists Claude Opus, Claude Sonnet, Gemini Pro and GPT models. Swapping models in and out as they improve is part of the design.
First endorsements and early users
The launch post quoted Rafał Kierzenkowski, who leads strategic foresight at the OECD, on what the tool had done for his team during the beta:
The Forecasting Machine helps conduct horizon scanning at scale and speed which hasn’t been possible until now.
Rafał Kierzenkowski, OECD, quoted by Émile Servan-Schreiber on LinkedIn
“Horizon scanning” is the foresight practice of systematically watching for early signs of change. It usually means analysts reading widely and flagging what might matter. A machine that turns those signals into tracked probabilities speeds up that work rather than replacing it.
Servan-Schreiber thanked several beta testers by name, including Kevin Ryan and Charles-Albert Lehalle. In the comments, Lehalle wrote that he had been testing the machine for months. He used it to map the news onto a set of competing narratives whose weights add up to one, so that he could see which story the evidence currently favoured. His testimonial on the homepage puts it more briefly:
a great tool to convert an avalanche of news into a collection of time series
Charles-Albert Lehalle, testimonial on theforecastingmachine.com
The team named at launch was Maurice, Theodore and Arthur Balick, Florian Jacques, Anzhelika Nastashchuk and Younes J. Since then the homepage has listed the OECD, Institut Montaigne, CMAP at École Polytechnique (IP Paris), Hypermind, Agenda Pública, beBartlet, Renault Group and EDF among the organisations using the platform.
What to watch
A forecasting tool earns trust through its track record, not its launch. The questions put to the machine in its first weeks, including the length of the October 2025 US government shutdown, now have known answers. We look at how those early forecasts turned out in a separate review, including the one the machine got wrong.
