As the tech industry seeks more uses for artificial intelligence, some people have turned their attention to weather forecasting.

Newer AI models like Google’s GraphCast can generate a 10-day weather forecast in minutes, GBH meteorologist Dave Epstein said.

But while those forecasts can be accurate, they also have limitations — especially in a changing climate.

To understand how they work, Epstein said, it’s important to first take a look at the models he and other meteorologists use now.

“Everybody has heard pretty much about the Euro and the GFS and all those different models,” he said. “If I showed you an AI prediction and showed you the Euro prediction, you would say, I don't really see the difference. … But again, it's not based on the atmosphere on March 28. It's based on all the historical stuff that's occurred, and based on that, here's what probably is going to happen. It's a much narrower forecast.”

The current models feed real-time data collected from weather balloons into algorithms to try and predict what conditions might look like in the near future — the next hour, the next day, the next 10 days, he said. It’s a process that can take a while.

“It's a very computer-intensive process. It takes hours, so two to three hours from the time you start ingesting the data until you get the forecast,” Epstein said.

AI models, in contrast, compare current conditions to historical data and produce an output of what’s happened in the past when conditions were similar, he said.

“It’s basically not looking at the real data in the same way,” Epstein said. “It's just saying, okay, a storm off the coast of the Carolinas with high pressure here and low pressure there and the upper levels looking like this — did this in the past. And that's what it predicts.”

The upside: A forecast that takes minutes instead of hours to generate. The downside is that the AI models can only use previously-recorded data that has been fed to them, so they may not be able to forecast unprecedented weather events, like Hurricane Harvey, which devastated parts of Texas in 2017.

“As we go into more extreme weather and we see things like 60 inches down in Houston during that hurricane a few years ago, … there's nothing that's occurred like that in the past, so it's going to have trouble predicting those more extreme events because it doesn't have an extreme event in it's sort of historical knowledge,” he said. “That's where the operational models and an actual human being are really going to come into play.”

Epstein said he looks at forecasts not just for the most likely scenarios, but for mid-range and low-range probability conditions.

“You can't do that yet with any of the AI stuff,” at least at this point, he said.

Though the AI models are not yet in widespread use, he said he is keeping an eye on new developments.

“I definitely think that it's going to be part of the suite of data that we all use to think about how forecasts are going to be made,” Epstein said. “And I think that over time, we will gain more knowledge as to just how accurate the AI models are.”

And he doesn’t expect meteorologists will be out of a job.

“I think you're still going to have to interpret it,” he said.