Researchers studying New Zealand's most extreme rainfall used generative artificial intelligence to simulate hundreds of ways this century might unfold, and found the rarest downpours intensify faster as the planet warms.

In a high-emissions future, rain falling in a one-in-100-year daily event increased by about 9% per degree of warming, averaged across New Zealand and the futures simulated, while more frequent heavy rain rose about 5.6% per degree.

The heavier the daily downpour, the greater the potential for flooding and damage.

The rarest storms intensify more than everyday heavy rain, reversing what many readers might expect from a simple warming atmosphere.

The study also found that warming widens the gap between milder and more severe possible futures: in the harsher simulations, extreme rainfall increased substantially more than the averages above.

That means variability itself grows with warming, widening the range of possible rainfall extremes faster than their average intensity rises — a widening spread of outcomes, not just a shift in the mean.

Conventional climate models use the laws of physics, and detailed regional models demand such powerful computers that researchers can explore only a handful of futures.

The team trained its AI on 140 years of regional climate simulations covering historical and warmer future conditions, teaching it how global weather patterns translate into local rainfall.

After about a day of training, the model generated more than 15,000 simulated years in six hours on four specialised computer chips — work the authors say conventional regional models would have taken years to produce. Each simulation represented New Zealand in grid cells about 12 kilometres wide.

The authors compare the problem to the All Blacks playing Australia at Eden Park: everyone expects the stronger team to win, but every match unfolds differently, and replays can end in upsets. Six replays give only a rough picture of rare outcomes, while hundreds reveal surprising possibilities.

Climate works the same way — rising greenhouse gases push temperatures up, but natural variability, including the El Niño Southern Oscillation, shapes when and where extremes strike. Small differences in starting conditions lead to different outcomes even under identical emissions.

The researchers are candid about limits: AI can produce poor outputs under unfamiliar conditions, a problem that plagued some earlier studies.

Training on both historical and warmer climates, alongside a training strategy their prior work showed produces realistic extreme events, helped address this.

More simulations cannot say exactly where or when the next record-breaking downpour will hit, but they reveal extremes beyond anything communities have experienced in living memory.

With more groups now using AI this way, work is underway to coordinate methods globally and develop shared standards, and an earlier review found similar approaches could work in other regions.

The New Zealand result lands alongside broader warnings about intensifying extremes.

The annual State of the Climate report, published October 14 in the journal BioScience, found seven of 16 monitored climate indicators at record highs and estimated the warming rate rose from about 0.05°C per decade between 1880 and 1970 to roughly 0.36°C per decade between 2010 and 2024 — putting the 2°C threshold in reach before 2050, per Time's coverage of the report.

Faster local tools like the New Zealand AI study give regions a clearer picture of the rare, damaging events they must prepare for.