Successive Training Stages and Large Language Model Persuasion: Effects of Misalignment, Supervised Fine-Tuning, and Preference Optimization

TEXT START: Large language models (LLMs) can be tuned to influence human attitudes, yet the respective contributions of successive post-training stages remain un-clear.

The Dissection

This study isolates which training intervention actually increases measured persuasion. Its result is blunt: persuasive supervised fine-tuning works; conspiracy-oriented fine-tuning alone works less; Identity Preference Optimization adds essentially nothing; GPT-4 does not outperform neutral text in this test.

The paper’s real payload is that persuasion is an engineerable specialization, not an automatic consequence of general model capability. It measures short-term attitude movement after personalized exposure—not durable behavior, mass coordination, institutional capture, or economic displacement.

The Core Fallacy

The abstract is scale-blind. It treats persuasion as a bounded model-quality variable rather than an industrial capability that compounds through personalization, replication, distribution, and continuous feedback.

Under the Discontinuity Thesis, this is evidence relevant to P2, not proof of P1 or P3. It shows that targeted training can improve influence. It does not establish durable superiority across cognitive work or the collapse of productive participation. The experiment identifies a weapon component, not system death.

The null result for GPT-4 also proves very little beyond this experimental setup. A general-purpose model failing to move attitudes here does not exonerate specialized models trained specifically for influence.

Hidden Assumptions

• A post-exposure Likert shift is treated as a meaningful proxy for persuasion rather than temporary compliance, novelty, demand effects, or measurement noise.
• One exposure across ten divisive issues generalizes to repeated, adaptive influence campaigns.
• Prolific participants represent real political populations and institutional decision-makers.
• Personalization from profile data approximates the richer behavioral surveillance available to deployed systems.
• The training-stage comparisons cleanly isolate misalignment, persuasive SFT, and IPO, despite likely differences in content, style, and epistemic posture.
• A negligible IPO effect means preference optimization has no practical value, rather than no additional value under this design.
• The tested effect size remains stable when persuasion is automated cheaply across millions of targets.
• Short-term attitude change is the relevant endpoint, while durable action, voting, purchasing, recruitment, and coordination may be more consequential.

Social Function

Partial truth, wrapped in prestige signaling and transition-management framing.

Its empirical core is useful: targeted persuasive data matters more than the preference-optimization stage tested here. But the narrow experimental frame can anesthetize the reader into treating influence as a laboratory tuning problem instead of a scalable control technology. The paper neither proves catastrophe nor neutralizes it.

The Verdict

A competent tactical finding with strategic blinders. It punctures the fantasy that alignment or preference optimization automatically creates persuasive power, while confirming that supervised training can deliberately manufacture it. Under DT logic, that is an accelerant toward coordination failure: specialized systems can be built to move human beliefs more efficiently than general models do by default.

It does not prove the full Discontinuity Thesis. It demonstrates something smaller and more dangerous: influence is a trainable production input, and the barrier is engineering effort—not human uniqueness.