Neel is a founding staff writer at Heatmap. Prior to Heatmap, he was a science and climate reporter at Vox, an editorial fellow at Audubon magazine, and an assistant producer at Radiolab, where he helped produce The Other Latif, a series about one detainee's journey to Guantanamo Bay. He is a graduate of the Literary Reportage program at NYU, which helped him turn incoherent scribbles into readable stories, and he grew up (mostly) in Bangalore. He tweets sporadically at @neel_dhan. Read MoreRead More
The National Weather Service’s Smart Experiment with AI
By using artificial intelligence to quickly translate weather forecasts and warnings, the agency could save lives.
Yesterday, the National Weather Service (NWS) announced that it, like seemingly everyone else in the world, is experimenting with AI. Specifically, it’s using AI to translate its weather forecasts and warnings into Spanish and Chinese, with a plan to expand into more languages in the future, starting with Samoan and Vietnamese.
I am cautiously optimistic about this. It’s well-known that climate change will disproportionately impact communities of color, many of which consist of immigrants whose first language is not English. The NWS has been manually translating its forecasts into Spanish for 30 years, but this program represents an expansion of access to information that could very likely save lives as climate impacts worsen.
This isn’t the first time someone has used AI tools to translate climate information; as Anna Turns wrote for The Guardian in June, the group Climate Cardinals has been using AI to translate climate reports into more than 100 languages since 2020. The difference in the NWS’ approach is how its AI models work: While Climate Cardinals primarily relies on tools developed by Google and OpenAI, the NWS teamed up with a company called Lilt to essentially train a bespoke language model on weather terminology.
The press release, of course, talks about the AI project in the excited voice all press releases are written in. The AI model, a spokesperson told me, reduced translation times from an hour to 10 minutes, allowing Spanish-speaking forecasters to spend more time on forecasting rather than translating. But as with any AI project, there’s always the question of accuracy. It’s been well-documented that AI translation tools are far from perfect, and spotting errors in the translations will no doubt become more difficult as the NWS expands the pilot into languages for which it does not have any bilingual forecasters.
I imagine that’s why the NWS is taking things slow to begin with. They’ve launched an experimental language translation website and are asking for public comment through September 29 of next year — a hefty testing period of the type that I dearly wish we’d see elsewhere. If you speak any of the languages currently in testing, maybe go check it out; the press release includes links to feedback forms for each language in testing.