A language model will now write you a better speech than most people you know. Better structure, cleaner argument, no dead lines. Ask for a wedding toast, a fundraising appeal, a rallying memo to a demoralized team, and what comes back is competent, often more than competent.
What I keep noticing is that this matters less than it sounds like it should. The polished text exists, and rooms full of people remain unmoved by polished text, much as they were unmoved by the polished text humans wrote before. I think the explanation is old. Aristotle, in the Rhetoric, listed three means of persuasion and wrote that the speaker's character, meaning what the audience believes about the person talking, may almost be the most effective of them. The words themselves were only ever one ingredient, and by his account not the strongest. He was working from orators in Athens, but nothing I've watched in the past two years contradicts him.
The speech is not the speaker
The clearest case I know involves the most famous American speech of the twentieth century. The 'I have a dream' passage wasn't in the prepared text of King's remarks at the March on Washington. He had used the dream material at earlier events, and some of his advisers thought it was worn out. Partway through the written speech, as people on the platform later told it, Mahalia Jackson called out to him to tell the crowd about the dream, and he pushed the pages aside and did. Whatever moved a quarter of a million people that afternoon, it wasn't fresh composition. It was a man they had watched go to jail for the thing he was asking of them, departing from his script in front of everyone.
So when I say the ability to move people, I mean something specific: getting others to trust you, commit, and act against their own immediate comfort, because of who you are and what you're asking. Not persuasion in the abstract. The version with a face on it.
It works on one person at a time
There's a quieter form of the same thing, the ability to make a single person feel met rather than managed, and it can be worth more than any address to a crowd. My favorite documented case: in May 1969, Fred Rogers testified before a Senate subcommittee that was weighing cuts to public broadcasting. The chairman, John Pastore, opened the exchange gruff and openly impatient. Rogers didn't argue budget numbers. He talked for about six minutes about what he tried to do for children, recited the words of a song about what a child can do with anger, and Pastore, who had never seen his program, told him he had goose bumps and that he'd just earned the twenty million dollars. You can watch the footage, and nothing in the transcript explains it. What did the work was the fact, obvious to everyone in the room, that Rogers meant every word.
What it isn't
Worth separating this from its lookalikes, though two of them can be dismissed quickly. Volume isn't it; the loudest person in the room moves no one, they just postpone disagreement. Manipulation is closer, the near enemy: a manipulator produces real movement, but you can usually feel the difference, sometimes only afterward, and once it's seen it poisons everything that person built with you. The confusion I'd actually flag is reach. Plenty of accounts with millions of followers can't get anyone to do one costly thing, and there are people unknown outside a small circle who can ask others for enormous things and be answered. One is a distribution number. The other is something people grant you slowly and can revoke.
What unlimited persuasion does
The economics here aren't complicated. Machines can produce persuasive language at close to zero cost, and things in unlimited supply lose value; there's no exemption for words. Email made correspondence free, and within a couple of decades we had built folders to hide it from ourselves and learned to delete anything that smelled automated. I expect the same reflex to form around optimized persuasion, and honestly it's already forming in me: I've started skimming anything that reads too smoothly, which is an awkward thing for a writer to admit.
Here is where I should concede the weak point. In blind conditions, people often can't tell machine writing from human writing, and some people are genuinely comforted, even moved, by conversations with chatbots. So the strong claim, that a machine cannot move anyone, is one I've quietly retired. The claim I still believe is narrower, and it's about following: going somewhere costly because of who asked you to. A model has no exposure. It won't be there if the thing fails, it forfeits nothing, and at some level we price that in. We'll take its drafts and its advice. Taking a risk for it is a different matter, and I've seen no sign of that.
Almost everyone has a teacher who redirected their life. Ask what that person actually had, and it was rarely the best command of the material; it was that they saw something in you, believed it, and got you to believe it, and you would have worked twice as hard rather than disappoint them. That's the skill I mean, and nothing about it requires a stage or a following. It gets built in ordinary rooms, by being present when it would be easier not to be, and by asking people for things you are visibly prepared to carry your share of. I don't know of anyone even attempting to automate that.