Most LLMs are trained to be so helpful and neutral that they've basically forgotten how to handle a character attack. While we usually think of ad hominem arguments as logical fallacies to be avoided, in the real world—especially in politics—attacking a person's ethos is a standard strategic move. The problem is that modern models are so safety-tuned that they default to "logical" defenses, which makes them feel robotic and unnatural when they're supposed to be simulating a persuasive dialogue.
Why LLMs fail at ethos-based debating #
The research in arXiv:2609.28673v1 highlights a massive gap between how humans and AI handle character attacks. When humans are in a political scrap, they don't just point out a logical flaw; they often hit back at the opponent's character to undermine their credibility.
The study used the ElecDeb60to16-fallacy corpus—which consists of actual U.S. presidential debates—to benchmark how LLMs respond compared to human debaters. The findings were pretty blunt: LLMs are stuck in a loop of "logical defenses." They try to correct the record or argue the point, whereas a human would have pivoted to an ethotic counterattack.
The "Safety Tuning" trap #
The core issue isn't that the models don't "know" how to be aggressive; it's that their safety fine-tuning has shrunk their strategic action space. Because we've taught them to be polite and avoid conflict, they treat character contestation as a "mistake" or a "fallacy" to be corrected rather than a valid move in a high-stakes debate.
If you're building an argumentative agent, you'll notice this behavior immediately. For example, if you prompt a model 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)
How to move beyond logical defenses #
To make an AI actually sound like a persuasive agent in a political context, you have to push it past the "polite assistant" persona. If you're seeing your model act too rigidly, you might need to adjust your system prompts to explicitly allow for ethos-based maneuvering. Instead of just asking it to "debate," try instructing it to "prioritize the perception of credibility and undermine the opponent's ethos using the specific rhetorical strategies found in political discourse."
Without this, the model will continue to prioritize the "correct" answer over the "winning" answer, which is why they currently struggle to replicate the naturalistic flow of human political conflict.
Next Claude Code in a Brig MicroVM Sandbox: Safer Autonomous Agents →
All Replies （1） #
Want a live back-and-forth? Join the global AI chat room — login to talk. I've noticed the same robotic tone when prompting for political debates. They always default to logical defenses instead of attacking ethos.