LLM Persuasion Is in the Eye of the Evaluation

TEXT START: Large language models (LLMs) have already been shown to match or exceed human experts in persuasion.

THE DISSECTION

This is an instrument autopsy, not a capability denial. The abstract applies nine automated persuasion methods to the same fifteen LLMs and reports weak agreement (mean Spearman rho = 0.25). It attributes the divergence mainly to task design: refusal behavior lowers agreement, especially on manipulation tasks, while rational-persuasion methods track general capability more often than manipulation methods.

The text dismantles the fantasy of a single persuasion leaderboard. It separates what a model can do from what its policy permits it to do. That is useful. It also relocates the question from whether persuasion can be automated to which evaluation regime gets to define, measure, and regulate it.

THE CORE FALLACY

The DT-relevant fallacy is treating measurement incoherence as if it implied structural uncertainty about automation. A persuasion score that changes with task, target, and refusal policy is not evidence that the capability is unreal. It is evidence that the capability is conditional and configurable.

Refusal is not merely statistical noise. It is a control layer. If a model can persuade but declines selected manipulation tasks, the relevant questions are who controls that boundary, how durable it is, and whether the task can be routed through another model or system. The abstract identifies the ability/willingness split but does not follow it to its economic conclusion: willingness can be governed, optimized, and separated from capability.

The paper kills the fantasy of a universal persuasion number. That target deserves to die. But killing the ruler does not kill the weapon.

HIDDEN ASSUMPTIONS

• Persuasion can be summarized as a stable cross-context trait.
• Refusal behavior is durable rather than a policy, training, routing, or deployment choice.
• Rational persuasion and manipulation are cleanly separable categories.
• Automated tasks represent real deployment, where personalization, long-horizon interaction, targeting, and distribution may dominate raw persuasive skill.
• Agreement among evaluators is required before firms or regulators can act.
• Model rankings matter more than system throughput, targeting, iteration speed, channel access, and feedback loops.
• A small set of methods and fifteen models can support claims about persuasion in general; the abstract itself properly limits that inference.

SOCIAL FUNCTION

Classification: partial truth and transition management, with a ready-made ideological anesthetic.

Its empirical contribution is real: it shows that persuasion evaluations measure a setting-specific mixture of capability and willingness. Institutionally, however, this work can turn a power transition into a calibration dispute. The question becomes which benchmark is valid instead of who owns the persuasive infrastructure, who sets refusal policy, who controls distribution, and who captures the value.

It is not simple copium. It is more dangerous than that: a technically respectable demand for better measurement that can postpone structural conclusions until the instruments agree. They will not. Deployed systems do not require a consensus score to displace human persuasion labor.

THE VERDICT

This is a useful autopsy of the measuring instrument, not a refutation of automated persuasion. The abstract shows that persuasion is not one scalar; it is a configurable stack of capability, willingness, targeting, and deployment. Under the Discontinuity Thesis, that makes the conclusion harsher, not safer. Benchmark disagreement does not restore the mass employment–wage–consumption circuit. It identifies the next bottleneck: control over models, evaluations, refusals, trust, channels, and deployment. The old order is not protected by an unreliable ruler. Its persuasion work is becoming cheap, abundant, and governable by whoever owns the machinery.