In May 2025, Émile Servan-Schreiber spent a week in Prague talking about how humans and machines should share the work of forecasting. On 20 May, he posted the last slide of his presentation at the Europe as a Task conference, titled “Hybrid Forecasting With AI”. It sets out a simple division of labour, and an argument about where each kind of forecaster is strong.
Two talks in one week
The first was at EAGxPrague, the effective-altruism conference held on 9–11 May. His talk, “State of the art in forecasting by crowds and machines”, started from a familiar problem: forecasting is hard, even for experts, but crowd-forecasting algorithms can turn many flawed judgements into reliable collective forecasts. His abstract ended on an open question:
Yet, in the last couple of years, artificial intelligence has started to catch up to the human gold standard.
Émile Servan-Schreiber, EAGxPrague 2025 abstract
The second was Europe as a Task, a conference on Europe’s future organised by the Institute of International Relations, EUROPEUM and the Association for International Affairs, with support from the Czech foreign ministry. President Petr Pavel opened it at Prague Castle on 12 May, and the working day followed at Czernin Palace, home of the foreign ministry.
There, the Czech think tank České priority organised a roundtable, “Guided by Foresight: Strategies for an Uncertain World”. Pavlína Žáková, the Czech deputy minister for European affairs, moderated it. The speakers were Jaana Tapanainen-Thiess of Finland’s Government Foresight Group in the Prime Minister’s Office, Marek Havrda of the European Commission’s Regulatory Scrutiny Board, and Servan-Schreiber. Jan Kleňha of České priority had invited him.
Foresight and forecasting
The roundtable started from a distinction. Foresight explores the range of possible futures, usually as scenarios. Forecasting estimates how probable specific developments are. According to České priority’s summary, Tapanainen-Thiess described how foresight is built into all twelve Finnish ministries. Havrda suggested that scenarios should come with key indicators that are monitored and, ideally, forecast, so they work as an early-warning system. That is where forecasting fits in. Servan-Schreiber presented Hypermind’s approach and Glimt, the Swedish government’s crowd-forecasting platform for Ukraine, which the summary said had about 19,000 volunteer forecasters at the time.
The slide: two tasks, two winners
His slide splits forecasting into two tasks done in sequence: information gathering, then guessing. The guess is unavoidable, he wrote, “because you’ll never gather all the information necessary to attain certainty.” For each task, the slide compares three kinds of forecaster:
- Information gathering. An individual is held back by limited time and resources. A crowd is faster and more thorough, as Wikipedia shows. AI is “faster, better, cheaper”, and the slide marks it as best.
- Guessing. An individual’s judgement is distorted by subjective noise, and AI is marked risk averse. A crowd is marked best, because individual noise cancels out while information adds up, which is the “wisdom of crowds”.
In his post, Servan-Schreiber put it plainly. For gathering information, “individuals are worst, crowds are much better (e.g., Wikipedia), but AI now rules.” For guessing, AI “seems generally penalized by risk aversion”, while humans, and crowds in particular, remain better at taking well-calculated risks.
Risk aversion matters because a forecaster who stays near 50%, or near the base rate, rarely makes a dramatic mistake but also rarely adds information. Accuracy scores such as the Brier score reward forecasters who move confidently towards the right answer. The Brier score is the average squared gap between the stated probability and what actually happened, so lower is better. Hedging every forecast keeps it mediocre.
The design it implies
His conclusion was a design rule. A system aiming for the highest accuracy should use AI to deliver as much relevant information as a crowd can digest, “perhaps even including a baseline conservative forecast”. Then the crowd, large or small, makes the actual guess. The machine does the reading. The people decide how far to move from a cautious starting point.
The model is a snapshot of mid-2025. The same two-step split appeared in the Glimt presentation in Paris that October. By March 2026, Servan-Schreiber was saying that AI could forecast as well as a diverse crowd of smart humans, which is the premise of The Forecasting Machine. The question his EAGxPrague abstract left open has kept moving.
