Invisible Conflicts: Media Coverage Asymmetry and Categorical Failure in LLM Conflict Forecasting

Media coverage of armed conflict is deeply asymmetric: we document a 224× gap between the most and least covered conflict zones in English-language media across 22 countries from 2020–2026, measured as distinct news articles per ACLED conflict event (GDELT, deduped URL-level). We investigate whether this asymmetry shapes LLM parametric knowledge by evaluating zero-shot conflict escalation forecasting across all 22 countries using two LLM backbones (Llama-3.3-70B, GPT-4o). Results reveal a more troubling pattern than a simple performance gradient: LLMs do not forecast conflict — they categorize it. Models apply near-universal escalation priors to under-covered active conflict zones (recall = 0.918) while exhibiting near-zero recall on over-covered zones (recall = 0.231), suggesting that parametric knowledge from media coverage produces conflict-zone labels rather than dynamic understanding of when escalation actually occurs. A system that predicts escalation every month for six years in DRC or Myanmar is not forecasting — it is reciting a category. Critically, this failure cannot be resolved at inference time: augmenting prompts with structured ACLED event count evidence degrades under-covered performance below the trivial always-predict-escalation baseline for both models (Llama F1: 0.329 vs. 0.344; GPT-4o F1: 0.211 vs. 0.344), confirming that the deficit is in parametric knowledge and not in evidence access. This categorical failure is arguably more dangerous than low F1 for humanitarian early warning systems, where missing a crisis costs lives. We argue that the NLP community must treat geographic conflict coverage asymmetry as a first-class fairness problem: under-covered populations receive not just less accurate AI, but qualitatively different AI that cannot distinguish stable from escalating periods within their conflicts. We call for coverage-stratified benchmarking, conflict NLP datasets for under-covered zones, and training data documentation standards for geographic conflict representation.