A new study celebrates artificial intelligence for translating National Weather Service forecasts into multiple languages, claiming accuracy above 95 percent and dramatically faster turnaround times. Yet the history of weather communication shows that translation failures often stem from culture, context, and human interpretation—areas where AI still faces significant challenges.
CHAMPAIGN, Ill. — Nearly 69 million people in the United States speak a language other than English at home, yet weather warnings have long been issued almost exclusively in English. A new study documents how the National Weather Service is using artificial intelligence to change that, developing a comprehensive translation program that delivers life-saving forecasts and alerts in Spanish, Chinese, Vietnamese, Samoan, French and other languages.
The new multidisciplinary study was led by University of Illinois Urbana-Champaign climate, meteorology and atmospheric sciences professor Joseph Trujillo-Falcón and NWS scientist Monica Bozeman and is published in the journal Artificial Intelligence for the Earth Systems.
“Translating weather forecasts has always been a critical, time-consuming task, often added to the plates of bilingual forecasters managing full operational responsibilities, but it provides information for the 68.8 million people who do not speak English at home,” said Trujillo-Falcón, who also leads the ALERTAS lab in partnership with the department of communication at Illinois.
The urgency for accurate and culturally sensitive translation of weather forecasts in the U.S. intensified in the 1970s and ’80s. In 1987, an F4 tornado struck the town of Saragosa, Texas, and 151 of the town’s 183 residents suffered injuries or casualties. The town, which had many Spanish-only speakers, had only one Spanish-language radio station, which did provide the NWS warning; however, the literal translation of the English message failed because the word “warning” has no Spanish equivalent.
The new program partners the NWS with the AI translation platform LILT, whose patented training process enables large language models to adapt neural machine translation tools for weather terminology and messaging.
“Before this program, bilingual NWS forecasters spent up to an hour manually translating a single product. The AI system now completes the same translation in 5 to 7 minutes for some NWS products, with accuracy scores above 95%,” Trujillo-Falcón said. “More than 30 NWS offices from across the country are using it and, starting with the 2025 hurricane season, the National Hurricane Center officially began issuing AI-translated Spanish advisories for all hurricanes in the Pacific and Atlantic coasts, providing lifesaving information to the U.S. and even Latin America.”
Beyond the operational aspect of the new program, the team is also very interested in how it will benefit the American people and economy.
“Tourists from across the world come back time and time again, benefiting local businesses, and we can provide them with life-saving information,” Trujillo-Falcón said. “And with the World Cup approaching, the NWS has made an agency-wide effort to provide decision-support services in the event of dangerous storms approaching the stadiums.”
Trujillo-Falcón emphasized the importance of the multidisciplinary focus the U. of I. brings to the program. For example, Illinois geography and geographic information science graduate student and study co-author Liam Llewellyn has overlaid geographic information systems, weather and census data to identify which language translations are needed for all 122 NWS offices across the country.
“We plan to expand these collaborations in the future here at Illinois,” Trujillo-Falcón said. “For example, I have been working with Spanish and Portuguese professor Salvatore Callesano, who is helping with the social science evaluation of how the public responds to AI tools.”
Researchers from the University of North Dakota, Pace University, Colorado State University and the University of Oklahoma also contributed to this study.
Translating garbage gets worse garbage not better garbage. Named storms and constant amber alerts are like Peter crying Wolf, translating it does not improve the outcome.
Valid points. However as litigation crazy as we have become, not giving the higher risk warning leads to lawsuits.
+1 Somehow we, society, has determined that weather forecasting is an exact science and those that forecast should be held accountable for variations from those forecasts. Add that to an increasingly more litigious society and you have CYA increasingly more important than the forecast.
re: “Translation Tools & Paths to Public”
Here in Dallas (DFW Metroplex) we have CH 23 KUVN (kay uu vay enay) a full-service Spanish language station that does a fine job on “translation” …
O/T Blighty madness update, no translation required.
Sainsbury’s swaps brown eggs for white in net zero push
The supermarket said white eggs have a 12.7 per cent lower carbon footprint than brown alternatives, as the hens that lay them are typically smaller and eat less feed. – Retail Gazette
BP is preparing to quit the North Sea after more than 60 years as Labour ramps up taxes on oil and gas. – DS
Mr Miliband scoffs that his critics are ‘climate deniers’. He lifts his chin to the Moon and howls his pride that we are setting an example to the rest of the world, even while the rest of the world thinks we are nuts. – DM
AI is the least of our worries, Miliband is out to get revenge for that bacon sandwich…
oddball Ed, like the deluded geek Malvolio in Shakespeare’s Twelfth Night, was exacting revenge on an unkind world.
Trouble is, all these prognostications are based on… RCP8.5
The Justification for Net Zero Has Vanished with the Demise of RCP8.5
There was no justification for Net Zero. Rationalization at best.
Justification, which is rooted in justice, requires evidence of which there has been none to date.
Surely smaller hens produce smaller eggs? Sainsbury’s may come to regret this decision.
Most commercial eggs in Australia are laid by ISA Brown hens. The ISA Brown is a hybrid which lays prolifically for a couple of years, but effectively ages very rapidly.
They have replaced other egg-laying breeds because they are so prolific.
White eggs tend to be laid by White Leghorns, which lay far less prolifically, so eat more per egg, but perhaps the laying breeds are quite different in the UK.
It’s possible that the hens which lay white eggs eat less per egg while they are on the lay, but ISA Browns are more economical overall, largely due to their lack of downtime. A hen which lays an egg every day beats one which stops laying for 3 months of the year.
Always love to see “patented modelling” … that is rather amusing and must be like climate models .. something so far from any science or mathematical norm you can patent it.
Historically I knew you couldn’t patent models so it click baited me into a search
Answer: No. Pure math, formulas, scientific theories, business models and abstract ideas are considered building blocks of knowledge and cannot be patented.
However: A recent addition which stinks of lots of AI companies getting laws changed.
AI models can be patented if they demonstrate how the model drives a tangible, concrete improvement.
I can’t wait for someone to fall foul of an AI patent and have that played out in court because the AI ruling makes no sense relative to the general case. If the general case is no the AI case should be upheld to be no.
a tangible, concrete improvement.
Oh for one of those.
One would need a bilingual meteorologist to proofread any AI product to deal with hallucinations and other flaws with the product.
And that bilingual meteorologist would take 1 HOUR to translate the weather report?
Call me skeptical.
I wonder how well the Google translation program does with these weather alerts.
On the other hand, English is a global language. Example: Air Traffic Controllers and pilots.
Perhaps a bit more English literacy is needed.
Do not misconstrue. I disfavor cancel culture. I disfavor putting lives at risk when a valid solution can be developed.
This is a watch and see. AI, for all of its down sides and benefits, might be a good tool if the data used is not the Internet.
The way to measure “how well” would be to let AI to translate the alert to French and back.
Reminds me of an old joke when a computer was tasked to translate “Out of sight, out of mind” to Russian and back: “Invisible idiot”.
There were some other humorousness involved with an earlier MSWord thesaurus.
Not just French, of course. That test would need to be applied to all languages addressed. I submit a refinement would be to use 2 AIs. First would do the translation. The other would do the translate back to English.
Am I the only one who believes that everyone (in US and UK) should use the English language for everything outside the home?
There are 96 others that so believe. 🙂
There is no consensus, so it is not science.
Depends how many uber drivers your country needs.
While that is a point to be discussed, the article addresses people inside their homes getting the alerts. That does not diminish the value of your thought.
Had an aunt and grandmother that were bilingual, Czech and English. In their homes they would speak Czech to each other. My mom understood Czech fairly well but didn’t speak it too well. But out in public I never heard them speak anything but English.
How long would it take for the English warning to come back around as the various translations cycle through?
The event might be over before the English warning comes back around!
Those weekly test of The Emergency System that interrupt TV shows, would the shows be over before it ends?
It’s not a bad idea to provide the warnings in other languages but I see a potential practical problem in actually delivering them.
Hmmm… How about people download apps to their cell phones or computers capable of translation to their preferred language.
This translation effort will end up like Chinese whispers:
Yes, sometimes I do have too much time on my hands
I changed your weather warning to ““Severe thunderstorm warning in effect until 6:15 pm for your area.” Portuguese: “Aviso de tempestade severa em vigor até às 18:15 para a sua área.”
Part of me is getting tired of having to pay extra to “help” people who come here voluntarily, and then once here, refuse to learn the language.
I don’t mind helping them learn, but they have to at least make an effort.
I have a philosophy.
If you want to talk to me, you talk in my language and do not require I learn yours.
The converse also. I want to talk to you, I do so in your language.
I pity the police officers who must become proficient in 130 different languages because someone running a stop sign might not speak English.
Hopefully AI models will progress to the point of dispensing with biased, agenda-driven climate con humans like Anthropic says generally about AI. We need to be saved from ourselves.
Sounds like a whole bunch of people living in an English speaking country need to learn English. My German and Czech ancestors did it back around the year 1900. Surely people today can do the same thing.
Let me know when AI (artificial intelligence) can give me an accurate 10-day forecast. Until then…
They can’t. Try a forecast for January 1st, 2100. That one the climate science can do.
LOL! Of course they can. By then most of us will be dead, and the rest won’t be able to remember it.
I’m a bit skeptical AI is the golden future it’s cracked up to be due to malicious actors not mention the likelihood it could all disappear up its fundamental orifice regurgitating itself-
Anthropic urges global freeze on AI as it warns of losing control
The sublime irony could well be we reach a stage whereby you can’t trust anything digitally presented so we go back to over the counter analogue dealing with anything financial not to mention the challenge that everything can be fake news and imagery. That’s just beginning to dawn on some it seems.