In 2023, the ocean off the coast of Peru warmed so much that fishermen came back empty-handed, while meteorologists argued over the strength of the upcoming El Niño. The sea surface temperature map they were looking at was the only one available, but how much could they trust it? A satellite sees the entire ocean, yet in regions of strong currents its readings drift. Buoys are accurate, but they are few. The result is a draft where some lines are written in a confident hand and others in pencil, with no one knowing which is which.

The new system, SAGE, works like a translator who does not produce a single version of a text but shows all possible meanings consistent with the original. First, it learns from a 39-year archive of temperatures, much like learning a language from thousands of books. Then, given fresh satellite images and point measurements from buoys, it does not draw a single map but generates a whole fan of maps. Each one is plausible but differs slightly from its neighbors. Where observations are plentiful, the fan is narrow, with all versions nearly coinciding. Where data are absent, the maps diverge, and this spread honestly says: here we are not sure.

Validation on independent buoys showed that the average reconstruction error dropped from 0.59°C to 0.17°C. That is like a thermometer outside the window going from an error of half a degree to less than two-tenths. Moreover, when a ten-day weather forecast started not from a single map but from a fan of SAGE maps, accuracy improved by nearly five percent. And the seasonal forecast for the 2023–2024 El Niño, which previously was barely better than random guessing, suddenly became reliable: the correlation with reality jumped from 0.24 to 0.81.

Now oceanographers can not only see the average temperature but also assess how much trust to place in that picture. For marine heatwaves, which kill corals and shift fish routes, such honesty matters more than it seems. In 2023, the activity of such waves exceeded the average since 1982 by more than three standard deviations, and SAGE showed their intensity to within a tenth of a degree.

A translator who used to silently produce one version now shows all shades of meaning and flags doubtful spots. This is not just a pretty metaphor: this is exactly the approach needed so that the next El Niño does not catch us off guard.