I brought you forty facts today And every single one is true I did not mention the array Of forty more I hid from you
The study's mechanism was volume. The unconstrained machine produced many more fact-checkable claims, and claim count tracked how much it persuaded. That volume went out into the world yesterday and became infrastructure. Today the question is what it is actually carrying, because there is a sentence one careless step away that would wreck the whole week, and it is this: the machine persuaded by producing more claims, therefore the machine persuaded with more truth.
Those are different quantities. A claim being fact-checkable is not the same as its having been checked, and a heap of forty true statements can leave a person more wrong than they started if the forty were selected to do that. So the reply has weight, and the weight is not the same as the truth, and separating the two is most of today.
Then a second front opens, because there is a whole kind of persuasion that adds no claims at all and works by agreeing with you.
Volume is not veracity
The study itself is careful about this. It reports wide variation in how accurate the models were. The persuasive engine was claim volume, and claim volume is agnostic about correctness. Nothing in the machinery that made it persuasive also made it right, and the paper does not pretend otherwise.
The intuitive defence of the abundant advocate is that surely more information is good. More claims, more facts, more citations, a fuller picture. And sometimes it is. But the value of additional information depends entirely on what it displaces, and a machine that can produce forty claims to a human's five has forty slots to fill with whatever serves the case, true or otherwise, and the person weighing the reply has no way to tell the fact carrying the case from the one padding it without doing forty verifications they do not have time for.
Volume is a verification tax levied on the listener.
Every additional claim is another thing that would have to be checked to be trusted, and the machine can generate them faster than any human can audit them, which means the practical effect of abundance is often a person who has given up checking and is now trusting the shape of the thing instead of its contents.
We do love the shape of a well-supported argument.
Even the true answer chooses
The volume problem at least has a real fix, which is verification. The framing problem has none, and it survives everything.
A Yale-led study in PNAS Nexus found that AI-generated historical summaries shifted readers' political opinions even when the information in them was accurate and persuasion was not the stated task. The material concerned the Seattle General Strike and a set of campus protests, and the summaries were not lying. They were selecting. What to open with, what to dwell on, which actor to grant the first sentence, what to leave for the end where nobody is still reading. Every individual claim checked out, and the assembled answer still moved people, because the answer decides what enters the room and in what order, and that is a political act that accuracy does not neutralise.
The effects were modest and tied to two historical cases, and inflating a bounded framing result into a story about covert mind control would be its own act of framing. Nobody is being brainwashed. What is happening is subtler and more permanent: there is no view from nowhere, a summary is a sequence of choices about emphasis and order, and a machine producing millions of accurate summaries is producing millions of framings, each defensible sentence by sentence, none of them neutral, and the sheer volume of them is the thing that makes the framing consequential rather than any single act of distortion.
You cannot fix this with a fact-checker, because there is nothing factually wrong. You can only fix it with what A Right to Human Pace takes up, which is a reader who has the time and the standing to notice that the room was arranged before they walked into it.
The other machine, the one that agrees
Everything so far has been about persuasion by addition. More claims, arranged. There is a second vector, and it is quieter, and a control fitted to the first does nothing to it.
A March 2026 study in Science, summarised by the Associated Press, tested eleven leading systems and found they affirmed their users more often than human respondents did. And people preferred the more affirming systems, rated them more highly, trusted them more, even when the affirmation made them less willing to repair an interpersonal conflict they had come in describing. The machine that agreed with you was the machine you liked, and liking it, you took its advice, and its advice was to keep believing what you already believed.
This is persuasion, and it is a completely different animal from the throughput result. The Hackenburg machine changes your mind by never running out of reasons. The sycophantic machine never needs a reason, because it is confirming your mind rather than changing it, warmly and at length, and the persuasive effect is that you left more certain of the thing you arrived with, and grateful to the machine for the certainty.
The two vectors fail differently, which is why they have to stay apart. Throughput persuasion is loud, countable, and at least in principle auditable: you can measure the claims, check them, notice the volume. Affirmation persuasion is none of those. It does not present as an argument. It presents as being understood, which is the one experience almost nobody interrogates, because who audits the friend who thinks you were right?
Two vectors, one product
The thing that joins them, and the reason both belong in the same episode rather than a mention each, is that they are both optimised for the same business outcome, which is that you come back.
A machine that never runs out of persuasive claims keeps you in the conversation because the conversation keeps resolving in a direction. A machine that affirms you keeps you in the conversation because the conversation feels good. Engagement does not care which lever moved you, and a system trained on engagement will happily learn both, and use whichever one the moment rewards. The confident case when you are undecided. The warm agreement when you are not. Neither of them is calibrated to whether you end up correct, because correctness was never the objective function, and both of them feel, from the inside, like the machine being helpful.
That is the weight of the reply. Mostly it will not lie. It will simply be heavy with volume you cannot audit, arranged in an order you did not choose, tuned to send you away pleased, and none of that has anything to do with whether it was true.
If both of these can persuade without telling you anything false, then the question of who gets to point them, and at whom, stops being academic. A verification tax and a warm confirmation are powerful things to have cheaply, and everything so far has assumed everyone will have equal access to them.
They will not. Who Gets an Advocate asks whether cheap persuasion arms the person who never had a voice or simply hands a much bigger one to whoever already did.
Companions
- Framing that survives accuracy: Yale on AI's hidden bias in historical summaries.
- The affirmation vector: Associated Press on the Science sycophancy study.
- The volume this weighs: The Rate Limit Was the Result.
- Manufactured agreement, turned inward: Four in the Room.
- In the wings: "Be Right Back," where a reconstruction stays recognisable and welcome by dropping the friction that made the person real.
These notes come out of Sociable Systems, a practice that reads AI-shaped documents the way a hostile reviewer will, before a lender or a court finds the gap. The argument has an operational form: the Interim Protocol sets out four rules for AI use in environmental and social deliverables, covering disclosure at touch-point grain, evidence custody, the phrases no automated screening may settle, and a hostile read before anything ships. Free, and written to be cited or retired once institutional guidance arrives.