You can find out whether it will rain on Saturday. Roughly whether it will be fine in ten days. But if you ask what the weather will be like in a month, no serious meteorologist will give you a confident answer.
Between two and six weeks, the weather is almost uncharted territory. Specialists sometimes call it the "desert of predictability". A team of researchers has just planted a first flag there.
What they achieved
In a study published on Wednesday in the journal Nature, researchers from MIT and Microsoft present a method that markedly improves forecasts between two and six weeks ahead, including for extreme events such as heatwaves.
According to them, a model now outperforms climate averages over that period for the first time. That is the key point: until now, to know what the weather would be like in a month, relying on the averages of past years was just as effective as any forecast. One of the researchers sums it up: nobody had yet managed to show that a model did better.
Their method also won an international forecasting competition.
How it works
The trick is elegant. The AI does not replace the existing large weather models, which simulate the atmosphere using the laws of physics. It learns their mistakes.
Imagine a sailor who knows his compass always deviates three degrees to the east. He does not throw away his compass: he mentally corrects every reading. The AI does the same thing with weather models: it has learned in which situations they go wrong, and in which direction, and it corrects their forecasts accordingly.
If long-range weather is so difficult, it is because of a discovery made by the American meteorologist Edward Lorenz in the 1960s. By rerunning a simulation with figures rounded only slightly differently, he obtained completely different weather a few weeks later. In 1972, he popularised the idea in a lecture with a title that has remained famous: can the flap of a butterfly's wings in Brazil trigger a tornado in Texas? Tiny differences amplify over time. That is why no tool will ever be able to predict the weather on a specific day six weeks out, only trends.
Why it matters
A reliable trend a few weeks ahead changes a great deal. A farmer can decide when to sow or harvest. A hospital can prepare for a heatwave. An electricity supplier can anticipate a peak in demand. Emergency services can pre-position equipment before a period of heavy rain.
And the tool has another asset: it is free, published in open access, and it does not need a supercomputer to run. Countries that cannot afford to run large weather models, Kenya for example, can use it to improve their own forecasts.
The limits
This is not about knowing whether it will rain on Saturday 21 November. Forecasts several weeks ahead remain probabilities and trends: hotter or colder than normal, drier or wetter. The improvement is real and measured, but it does not turn the uncertain into the certain.
It is the same logic we explained regarding the importance of being able to say "I don't know": a good forecaster is not the one who claims to know everything, but the one whose probabilities are right.
What to take away
There is a lot of talk about AI that writes, draws or programs. This one does something more discreet, and perhaps more valuable: it gives us a few weeks' head start on the weather.
In a world where extreme events are multiplying, those few weeks can save harvests, and lives.