MIT researchers use AI to learn and correct forecast-model errors, doubling 2–6 week forecast skill and beating climatology for the first time (reports)

Verified (fact)

MIT researchers developed a method that uses AI to learn and correct the errors of existing weather forecast models. It targets subseasonal forecasting, which looks two to six weeks ahead. Applying the method to a top forecast model doubled its forecast skill for that range. The researchers said they also showed for the first time that it outperforms forecasts based on long-term climate averages.

Needs fact-check (claim)

The doubling of forecast skill and the advantage over climate averages are results reported by the researchers.

Why it matters (interpretation)

Two-to-six-week forecasts sit between short-range and seasonal forecasts and have long had low accuracy. Improvements in this range could be put to direct use in agriculture and energy supply planning and in preparing for extreme weather such as cold waves and floods.

Confidence medium · official source pending

Sources

Press "AI로 2~6주 장기 기상예보 정확도 2배 높였다"…MIT 연구진, 기후 평균 넘어선 예측 실현 AI News Korea

More about MIT

2026-10-12 P2 Research Paper Safety & alignment Language models Carnegie Mellon University UC Berkeley MIT University of Oxford UCLA OpenAI
Study: just 10 minutes of AI use makes people give up more easily once help is removed; 3 RCTs, 1,222 participants (report)
Researchers led by first author Grace Liu of Carnegie Mellon University, with UC Berkeley's Center for Human-Compatible AI (CHAI), MIT,...
2026-10-08 P3 Research Paper Robotics & physical AI Safety & alignment MIT
MIT develops 'SANDO' to guarantee collision-free drone paths in dynamic environments without maps (report)
MIT researchers have developed SANDO, an autonomous flight path-planning technique that mathematically guarantees collision-free drone...