'people wondered the same after the fire was discovered' is also what people said after nuclear fission was discovered. And it was as misleading a comparison then as it is today.
It's worse than that, though. We didn't have nuclear fission being made by multiple competing parties in the private sector with no possibility of government regulation or oversight. That's what we've got today for ML.
Decentralization sounds great until someone is cooking with plutonium. Our values of freely sharing information and technology are simply unequipped for technologies this dangerous, but we have not yet recognized ML as a weapon of mass destruction. We should, as it plainly can now be used thusly.
As I predict that in the future, open source ML models will be regarded the same way that loose nukes are treated today.
Of course, the problem will be far worse. We never set up multiple competing MOOCs to teach-at-scale nuke-making to the next generation. In effect, we have made ML terrorism unavoidable and endemic after 2025 or so.
Given that democracy was already hanging by a thread, I'm going to wager that the profusion of next-generation ML bots will make it effectively impossible at scale, as it becomes a simple matter to create 10,000 supporters, say, with unique faces, voices, and opinions, out of the thinnest of air.
This means that democracy will denature into a sort of suspicious populism. And before you say, "isn't that where we're already at," I'll say, "yes and it just got a whole heck of a lot worse."