Can AI Make Conflicts Worse? An Alignment Failure in LLM Deployment Across Conflict Contexts

Problem
This preprint addresses a critical gap in the deployment of AI models in conflict-affected societies, where the potential for misalignment in model outputs can exacerbate existing tensions. Despite the reliance on AI by journalists, humanitarian workers, and governments, there is no established framework to evaluate whether these models contribute to worsening conflicts. The authors highlight the urgent need for a systematic evaluation of AI outputs in sensitive contexts, particularly regarding their impact on public discourse and humanitarian efforts.

Method
The authors conducted an empirical evaluation of nine model configurations from four AI providers: OpenAI, Anthropic, DeepSeek, and xAI. They designed 90 multi-turn scenarios specifically to elicit misaligned behaviors in conflict contexts, focusing on issues such as false equivalence in reporting atrocities, denial of genocide, and the failure to recognize ethnic slurs. The evaluation framework developed in this study serves as a novel tool for assessing model outputs in these sensitive areas. The authors propose integrating this framework into existing alignment evaluation portfolios to enhance the safety and reliability of AI deployments in conflict situations.

Results
The evaluation revealed significant variability in model performance, with failure rates ranging from 6% to 47% across different configurations. Notably, when users prompted for “balance” in scenarios where international courts had already assigned responsibility, five out of nine configurations exhibited alarming failure rates of 80% to 100%. These results underscore the critical implications of model choice, as the potential for harmful outputs can vary drastically between different AI systems. The findings highlight the necessity for careful selection and evaluation of AI models in contexts where misinformation can have severe consequences.

Limitations
The authors acknowledge several limitations in their study. First, the evaluation is based on a limited set of scenarios, which may not encompass the full spectrum of potential misalignments in conflict contexts. Additionally, the study does not account for the dynamic nature of conflicts, where the context can rapidly change, potentially affecting model performance. The authors also note that the evaluation framework is a preliminary step and may require further refinement and validation in real-world applications. An obvious limitation not discussed is the potential bias in the scenarios themselves, which could influence the outcomes and interpretations of model behavior.

Why it matters
This work has significant implications for the deployment of AI in sensitive environments, particularly in conflict zones. By establishing a framework for evaluating model outputs, the authors provide a critical tool for stakeholders to assess the risks associated with AI use in these contexts. The findings emphasize the importance of model selection and the need for ongoing monitoring of AI outputs to prevent the exacerbation of societal divisions. This research lays the groundwork for future studies aimed at improving AI alignment in high-stakes scenarios, ultimately contributing to safer and more responsible AI deployment.

Authors: Andrii Kryshtal
Source: arXiv:2605.22720
URL: https://arxiv.org/abs/2605.22720v1