I disagree with your assessment. Modern machine learning methods will result in a far more complex algorithm than any human could device on their own. It will be "cheap" to execute, but only because of the vast amount of data it's been trained on. That's the strength. The raw number of calculations it can do per second means it can sift through more data in a day than a human could in its lifetime.
The biggest danger is hidden local minimums that can potentially distort new results and findings. But that's not unique to machine learning. And you can let the AI grade its result with how certain it is. Automatically flag any result that falls below a certain threshold. And again, this is where the danger of local minimum come in. Because it can raise the certainty of results when maybe it shouldn't be so certain. Though you will find out quickly when results are false with high certainty. Then you go back and you keep working at it.
As of this particular topic of earthquakes. There are no traditional tools of accurately predicting them. We've never been able to do it before. Now they think this might be able to. So we will just have to wait and see if its predictions are accurate or not.
What is it you think we're talking about? This isn't some LLM. It's a very specific purposebuilt piece of software to analyse data from earthquakes and recordable shifts in the tectonic plates.