Large language models increasingly translate and explain politically contested events, yet their outputs can introduce narrative bias while preserving the underlying facts.
We create two datasets covering the Russia–Ukraine war and the Israel–Palestine conflict, pairing neutral event descriptions with competing narratives curated from news outlets and Telegram channels.
We introduce a framework for measuring perspectival divergence: systematic alignment with competing narratives about the same events. We characterize model-specific tendencies through narrative framing, emotional manipulation, source credibility, social identity, and contested naming.
Across both datasets and four languages, we show that apparent neutrality varies with the conflict, language, and prompting context. To mitigate these effects, we adapt self-distancing, a psychological intervention that prompts models to adopt a distant observer’s perspective.
Across eight models, self-distancing reduces measured divergence by 45% on average while preserving event content in 93% of outputs. We further introduce Self-Distancing Activation Steering (SDAS), which extracts the internal representational shift induced by self-distancing and applies it during generation without fine-tuning or modifying the original prompt.
Across five open-weight models, SDAS reduces divergence by 58% while preserving event content in 91% of outputs. Steering directions also transfer between conflicts without additional tuning, offering a practical approach to measuring and reducing narrative bias without prescribing a preferred political viewpoint.
