What AI actually changes in geopolitical forecasting

Geopolitical analysis has a speed problem. By the time a traditional risk memo is researched, written, reviewed, and delivered, the situation it describes has usually moved. The industry’s answer has long been to hire more experts and write faster memos. At Downstream Intelligence, we think that’s the wrong fix.

The interesting thing AI changes is not the writing — it’s the decomposition. Instead of asking an expert “what do you think happens next in the Taiwan Strait?”, you can break the question into observable signals, the drivers behind them, and the causal chains that connect one event to the next. Signals are data points you can actually watch. Drivers are the forces producing them. Causal chains are the paths from here to the scenarios that matter. Once a question is structured that way, machines are extremely good at monitoring thousands of signals continuously and updating probabilities the moment something shifts — which no analyst team, however brilliant, can do at 3 a.m. on a Sunday.

What AI does not change is judgment. Models can’t tell you which questions are worth asking, and they inherit every blind spot in their data — smaller unreported events ripple into larger ones in ways no system fully maps. Forecasting is non-deterministic; anyone selling certainty is selling something else. That’s why we bound our scenarios with an upside, a downside, and a base case, each with measurable requirements, and why we share our forecasts publicly. Being wrong in public is uncomfortable. It’s also the fastest way to get better, and it’s the opposite of the traditional model, where assessments live in private memos and nobody keeps score.

I came to this from the qualitative side — speechwriting at the UN, public diplomacy, a press office in the Swiss Armed Forces. What convinced me isn’t that AI replaces that world’s expertise. It’s that pairing expertise with a system that updates continuously beats either one alone. The generational bet we’re making at Downstream is that probabilistic, transparent, continuously-updated analysis will displace the language of confident assertion. The world doesn’t wait for the quarterly report anymore. Analysis shouldn’t either.

Raffael Hüberli is a co-founder of Downstream Intelligence. Its forecasting platform is currently being backtested internally and results are shared publicly via Instagram (@downstreamintel).