It is easy to talk about argumentative fallacies: label a personal attack as a logical error and file it away. But in real political debates, ethos is not background noise; it is a strategic resource. A new study tested LLMs on exactly this ground, comparing model-generated dialogues with the ElecDeb60to16-fallacy corpus of U.S. presidential debates. The result is a substantial difference: human agents alternate logical defences and ethotic counterattacks, while LLMs almost always retreat to logical defence.

The study describes a reduced strategic action space. This is not a matter of compute capacity or tokens: it is safety fine-tuning that narrows the acceptable moves. When a model is aligned to avoid personal attacks, it ends up treating ad hominem argumentation as an error not to replicate. In a norm-governed debate, this rigidity produces correct but unrealistic responses, especially in domains where character contestation is expected by interlocutors.

The control issue for self-hosted deployments

For AI-RADAR, the point is not whether LLMs should make personal attacks, but that alignment is a product choice that shapes strategic behaviour. API-based models apply filters that users cannot modify; open-weight models allow deeper intervention, but they shift the cost elsewhere: you need GPUs with adequate VRAM, a fine-tuning pipeline, and the skills to evaluate results. TCO is not just the price of the service; it is the cost of control.

This changes incentives. Anyone using an LLM in a negotiation or political analysis context probably does not want an assistant that dismisses ad hominem as a fallacy; they want an agent that recognises the move and responds consistently with the context. For an organisation running on-premise for data sovereignty, the question becomes: how much behaviour can I redefine without depending on the vendor? The answer depends on weight availability and the possibility of targeted fine-tuning.

No invented benchmarks are needed to see the direction: if alignment ethics are applied uniformly, they can erase an entire class of communicative moves. The trade-off is not only between safety and realism, but between ready-to-use models and systems that require local intervention. For those evaluating on-premise deployment, AI-RADAR gathers analytical frameworks at /llm-onpremise to map these equilibria.