Modern LLMs are too polite to fight dirty in political debates.

Most LLMs are trained to be helpful and neutral, leaving them ill-equipped for character attacks. Although ad hominem arguments are generally treated as logical fallacies to avoid, political conflict in the real world often treats an attack on someone’s ethos as a standard strategic move. Safety-tuned models tend to default to “logical” defenses, making simulated persuasive exchanges feel robotic and unnatural.

Why LLMs falter at ethos-based debate

Research in arXiv:2609.28673v1 identifies a massive gap between how humans and AI handle character attacks. During a political scrap, humans often do more than expose a logical flaw; they attack the opponent’s character to undermine their credibility.

The study used the ElecDeb60to16-fallacy corpus—which contains actual U.S. presidential debates—to benchmark how LLMs respond compared with human debaters. Its findings were blunt: LLMs remain trapped in “logical defenses.” They correct the record or argue the point, while a human would instead launch an ethotic counterattack.

The “Safety Tuning” trap

The problem is not that models cannot “know” how to be aggressive. Their safety fine-tuning constrains the range of strategic moves available to them. After being trained to remain polite and avoid conflict, they treat character contestation as a “mistake” or “fallacy” to correct rather than a valid move in high-stakes debate.

This behavior becomes obvious when building an argumentative agent. For example, when prompted to simulate a political debate:

• Human strategy: “My opponent claims I'm inexperienced, but let's look at his record of failure over the last decade.” (Ethotic counterattack)
• LLM strategy: “It is inaccurate to say I am inexperienced because I have served in X position for Y years.” (Logical defense/correction)

The LLM strategy may be technically correct, but it is strategically weak. It focuses on “fact-checking” while the opponent is managing “perception.”

Moving beyond logical defenses

To make AI sound like a persuasive agent in a political context, it must move beyond the “polite assistant” persona. If a model appears too rigid, adjust its system prompts to explicitly permit ethos-based maneuvering. Instead of merely asking it to “debate,” instruct it to “prioritize the perception of credibility and undermine the opponent's ethos using the specific rhetorical strategies found in political discourse.”

Otherwise, the model will continue prioritizing the “correct” answer over the “winning” answer, which explains its current difficulty reproducing the naturalistic flow of human political conflict.