On 2 March 2026, Delphine Sabattier opened her Smart Tech programme on B SMART 4Change by describing an experiment. She had spent time testing The Forecasting Machine, which she called less a crystal ball than an engine for anticipation. Her guest was Émile Servan-Schreiber, co-founder and Chief Scientific Advisor of The Forecasting Machine and CEO of Hypermind. The six-minute segment, titled “L’IA et l’anticipation stratégique” (“AI and strategic anticipation”), set out how the tool works and who it is for.
A “small yes” on Europe’s tech sovereignty
Sabattier had asked the machine whether Europe stood a real chance of carrying out its strategy to reduce its digital dependencies. She reported the answer as a “small yes” at 53%, in other words barely better than even odds rather than a confident call. She said the number itself interested her less than what sat behind it: the sources used, the criteria applied, the explanation of the result, and the fact that the AI models involved did not all agree. Grok, she noted, was more sceptical than the others about Europe’s prospects. Servan-Schreiber cautioned that this depends on the question, and Sabattier agreed not to read too much into it.
From human crowds to crowds of models
Servan-Schreiber placed the product in a longer story. For 25 years, he said, his teams have produced forecasts for companies and governments using collective intelligence, which means pooling the judgements of many people. Prediction markets such as Polymarket now do something similar for the public. Nobody fully masters the future, he said, but together people can make very good forecasts.
What changed, he said, came with large language models (LLMs), the systems behind ChatGPT. Over about three years, their forecasts went from roughly acceptable to steadily better. Prompted correctly and used together, because none is reliable enough alone, they now produce forecasts he described as broadly on a par with groups of human experts. When Sabattier asked whether this meant AI had overtaken collective intelligence, he answered yes, then reframed the question: AI models, he said, are themselves collective intelligences.
In other words, a model trained on vast amounts of human writing already aggregates many viewpoints. Combining several such models adds another layer of aggregation on top.
Not just asking a chatbot
Servan-Schreiber stressed that the product does not simply ask ChatGPT a question. A chatbot gives one answer and has to be asked again for an update, which becomes unmanageable for hundreds of questions. He said some power users of The Forecasting Machine follow hundreds of questions that are refreshed every day. He described the pipeline as follows:
- the question is reformulated with the user, and possible answers are proposed, so that it says exactly what was meant;
- the system identifies the causal factors that could affect the outcome;
- it searches for relevant documents online in every language and weighs the most pertinent ones;
- it passes everything to several LLMs as one large, well-informed prompt, then synthesises their forecasts with a fully sourced explanation;
- forecasts update at a chosen frequency, and users get notifications when something changes.
Sabattier said she no longer needed to master the “art of the prompt”, since the tool checks and rewrites the question at the start. Servan-Schreiber added that the user stays in charge and can edit everything. The aim, he said, is not a top-down machine that tells people what to think, but one that does the laborious work.
Who it is for
The target users, he said, are strategic analysts in companies and governments. In his view, people alone can no longer keep pace with how fast and how interconnected the world has become. A buyer of sunflower oil at a wholesaler like Metro has to follow the war in Ukraine, and a carmaker has to follow the battery market and regulation involving China. Business decisions increasingly rest on geopolitical judgements several years out. He argued that this calls for machines that multiply forecasters’ capacity by at least 100. That figure is his own description of what the tool does, not an independently measured result. Sabattier closed by noting that a few free questions are available to try it.
