Most local TV stations air just a few weather reports a day: a forecast in the morning, perhaps an update in the afternoon, followed by an outlook for the next day at night. But when a winter storm was blanketing much of the U.S. earlier this month, Cleveland residents were able to get hourly updates on the situation on Haystack News—a TV news aggregator that recently began to use AI to produce local weather reports.
Haystack built an automated workflow on top of Amazon’s AWS that uses LLMs, text-to-speech, and an image generator to churn out new forecast videos every hour. It has been releasing these AI-generated weather updates for 30 cities, including Cleveland, Houston, Cupertino, Kansas City, and St. Louis. The company has plans to expand the program to more than 100 locations over the next two weeks, and aims to be in every U.S. city within a matter of months. “We will try to get it done 100% this quarter,” says Haystack News cofounder and CEO Daniel Barreto.
The use of artificial intelligence by news organizations has been a divisive subject (particularly when the industry finds itself at a moment of turmoil). AI has become a shiny new toy for some companies looking to demonstrate their future-readiness, and TV networks around the world have begun to experiment with AI-generated news anchors. At the same time, AI has been used by some publishers to generate heaps of cheap, SEO-optimized web content, which has often been riddled with errors.
Haystack has no plans to replace traditional news coverage with AI. The company has licensing deals with big media companies such as ABC, Fox, Nexstar, Hearst, and Scripps. Thanks to these partnerships, Haystack is able to aggregate local news for 97% of U.S. TV markets in its mobile and smart TV apps. However, Barreto believes that there is a role for AI to augment some news coverage—especially in noncontroversial areas. “Weather is probably the best [use case], because it’s data,” he argues.
It’s also a subject whose coverage is woefully lacking, at least on local TV news. Take the San Francisco Bay Area, for example, where Haystack News is headquartered. Local broadcasters try to cover the entire region in a single weather report, despite the fact that San Francisco itself is known for widely varying micro-climates. Add the surrounding cities, and you frequently have temperature differences of 20 degrees or more. “When you live in San Jose, the weather is very different [than in San Francisco)],” Barreto says.
These discrepancies can be even more pronounced in poorer rural areas, which are often disproportionately affected by extreme weather. Data from the Pew Research Center shows that 71% of rural residents believe that weather is important to their daily lives. Yet, 57% of the same respondents told researchers that the local news doesn’t cover the areas where they live—a sentiment that is shared by only 35% of urban residents.
Haystack, which was founded in 2013 and received financial backing from European TV-maker Vestel and others, began to develop its AI-generated weather forecasts six months ago. The process involved “a lot of trial and error, a lot of internal testing,” according to Barreto. Over time, the company zeroed in on a simple formula: Local weather data is being fed into a large language model, which generates a script for a roughly one-minute forecast. The script is then run through a text-to-speech engine, whose output is accompanied by automatically generated infographics.
All of Haystack’s weather data comes from the National Weather Service and major commercial and professional vendors, reducing the risk of what’s known in the AI field as hallucinations—i.e., AI models making up facts. However, early on, the company did have a few instances of AI misinterpreting the data, admits Barreto. “We had examples where it was cold, and it told you [that] it’s going to be nice outside.” The AI simply associated clear and sunny skies with T-shirt weather, even when it was actually freezing cold.
Making sure that those flubs are far and few between will be important for Haystack, says Erickson Strategy & Insights analyst Paul Erickson. Getting the weather wrong one time too many could lead to people abandoning the service altogether as they conflate its AI weather reports with the rest of its programming. “Haystack doesn’t produce its own news, but the consumer may not see it that way,” Erickson cautions. “Over time, people may lose trust in it.”
Just as important as teaching the AI model about temperature was to make the process itself scale by keeping computational costs under control. “This type of content expires within hours,” says Barreto. Unlike other videos, weather reports can’t get repurposed and monetized on social media days after the fact.
The short lifespan of news video also raises questions about the viability of some other AI ventures in this space. In December, a company called Channel 1 announced plans to launch an entire news network with AI anchors, synthetic voice-overs, and some AI-generated imagery. Other news networks, including South Korea’s SBS and Taiwan’s FTV, have begun to experiment with AI-generated anchors.
However, these experiments generally still require a lot of manual intervention, which increases production costs. SBS’ AI anchor Zae-In, for instance, has been puppeteered by a live actor, with AI simply swapping out the actor’s face and voice for that of the synthetic host—an approach that’s not easily scalable.
Haystack, meanwhile, is already exploring the use of AI for other data-heavy topics, including local traffic reports. And Barreto fully expects that we will see the use of AI grow, even on-air, as part of the program of traditional TV broadcasters. “In the long term, AI is going to be everywhere,” he says.
{Categories} _Category: Applications,*ALL*{/Categories}
{URL}https://www.fastcompany.com/91017834/ai-is-coming-for-your-local-weather-report{/URL}
{Author}Janko Roettgers{/Author}
{Image}https://images.fastcompany.net/image/upload/w_1280,q_auto,f_auto,fl_lossy/wp-cms/uploads/2024/01/p-1-91017834-ai-is-coming-for-your-local-weather-report.jpg{/Image}
{Keywords}Tech{/Keywords}
{Source}All{/Source}
{Thumb}{/Thumb}