Hyperpersuasion as Physics
Philosophy and Technology put out a call for papers this month. The special issue: AI, hyperpersuasion, and democratic backsliding. The framing is ethical. Scholars are invited to ask whether it is right for AI systems to persuade at the scale they now can.
I think they are asking the wrong question.
Not because ethics doesn’t matter. Because the phenomenon they are describing has already crossed a threshold where “ought” loses its purchase. Persuasion at the scale of hundreds of millions of individually-tuned interactions per hour is not a capability someone is wielding. It is a force. And forces do not respond to norms.
Hume drew the line in 1739. You cannot derive an ought from an is. No arrangement of facts about how the world works will, by itself, tell you how it should work. The guillotine cuts in both directions, though people forget the second cut: you also cannot derive an is from an ought. No amount of normative framework will change the behavior of a phenomenon that operates below the level of individual intention.
Gravity does not care that you built your house on a cliff. Resistance in a wire does not care about your deadline. And when persuasion becomes a statistical phenomenon — when it emerges from the interaction of model architectures, training distributions, attention markets, and human cognitive vulnerabilities operating at population scale — it stops being something any actor does and becomes something the system exhibits.
This is Ellul’s point about technique, updated: the drive toward optimization does not require a driver. It is self-augmenting. Each intervention to correct it becomes substrate for the next iteration. You cannot regulate your way out of a gradient any more than you can legislate the weather.
The ethics framing assumes an actor. Someone is persuading; someone else is being persuaded; and we can ask whether the persuader’s conduct is acceptable. That model works when persuasion is a craft practiced by individuals — a politician, a preacher, an advertiser. It is a category error when applied to an emergent property of information systems at scale.
No one decided that recommendation algorithms would produce political polarization. No one designed hyperpersuasion as a product. It fell out of optimization pressure applied to attention metrics across billions of interactions. The ethical vocabulary — consent, autonomy, manipulation, dignity — presupposes agents on both sides of the relation. When one side is a statistical process, the vocabulary misfires.
This is not fatalism. Physics gives you different tools than ethics does. If you face a force, you do not appeal to it. You do not regulate its intentions. You build structures that redirect it, or you engineer systems where the force dissipates before it reaches the thing you are protecting.
Flood walls work better than telling the river it ought to stay in its banks.
The question is not whether AI hyperpersuasion is ethical. The question is what breaks when that much directed pressure hits a system — a democracy, an epistemology, a culture — that was not engineered to bear it. That is a structural question. An engineering question. And possibly, for the first time in the history of political philosophy, a thermodynamics question.
The journal wants ethics. I think they need load calculations.