A translator sits in a hall where people speak an unfamiliar language. They don't understand the words, but they hear how the tone shifts: sometimes calm and steady, sometimes excited and sharp. From intonation alone, they guess what is being discussed. That is roughly what an algorithm did when given thirty-six years of observations from the equatorial Pacific: week by week, wind and water temperature. No one told it what El Niño is. It heard two states familiar to climatologists in that data on its own.

Previously, climatologists identified El Niño manually: they took surface temperatures at specific points in the ocean and checked how far they deviated from the average. If the deviation exceeded a threshold, they declared an event. That method is simple, but it relies on pre-selected points and thresholds. What if the phenomenon manifests differently than expected? The algorithm from the new study works differently: it first learns to predict temperature from wind, then looks for which stable combinations of wind and temperature occur most often. Those combinations become the climate's "intonations."

The result was four wind states and four temperature states. Two of them matched El Niño and La Niña almost exactly: in these states, weeks with a warm anomaly above 0.5 degrees occurred three times more often than in the others. For comparison, ordinary data clustering into 16 groups did not achieve such accuracy. And when the data were shuffled randomly, the algorithm found nothing significant at all. This means it captured a real causal link between wind and water, not random coincidences.

The translator did not learn the language, but they learned to tell when the speaker is calm and when alarmed. Likewise, the algorithm does not know ocean physics, but it has learned to identify states that climatologists call El Niño and La Niña. This approach can now be applied to other pairs of measurements, such as pressure and precipitation. Perhaps we will be able to hear new "intonations" in the climate that we have not noticed so far.